{"id": "2606.14784", "title": "LLM-Based Synthetic Ground Truth Generation for Audio-Based Emotion Classification via In-Context Learning", "abstract": "Understanding human states and interaction dynamics is a core goal of human-computer interaction (HCI). As interaction paradigms become more immersive, virtual reality (VR) has emerged as a powerful platform for studying collaborative work. In such settings, evaluating team collaboration states, including team performance and team resilience, requires continuous and reliable inference of latent team-level cognitive and affective states from multi-modal sensor data, such as speech signals. However, generating ground truth labels for these latent states remains challenging due to sensor-induced noise, contextual variability, and sparse expert annotations. Traditional self-reporting approaches provide only static and delayed measurements and are therefore insufficient for capturing dynamic team processes reflected in continuous speech data. In this work, we propose a large language model (LLM)-driven, agentic inference workflow for automated emotion-related synthetic ground truth generation from streaming speech data in multi-user VR environments. Leveraging the generalization capabilities of LLMs, we use In-Context Learning (ICL) with few-shot demonstrations of paired audio-based samples and their corresponding transcriptions. ICL tends to achieve task adaptation comparable to model fine-tuning while circumventing the computational overhead of parameter updates. To construct informative and robust in-context prompts, we adopt a retrieval-based selection strategy that dynamically identifies relevant audio demonstrations based on similarity in the acoustic feature space.", "date": "2026-06-10", "categories": []}
{"id": "2606.19568", "title": "Exploring Feature Extraction Technique Parameters for Acoustic Gunshot Classification", "abstract": "Acoustic gunshot detection is a problem with applications across civilian public safety, military operations, and wildlife conservation, yet the field lacks a rigorous exploration of feature extraction techniques with a focus on generalization to realistic data. The mixed effectiveness of commercial gunshot detection and classification systems indicates an open problem that is not adequately addressed by the current literature. In this paper, we present a systematic investigation of common feature extraction techniques using a dataset of 23,000 gunshot recordings across 85 firearms and 21 calibers. We benchmark three feature extraction techniques with 12 total unique parameter sets using ResNet-18. Our results demonstrate that using the correct feature extraction technique can improve top-1 accuracy by up to 20%, and utilizing the correct parameters for a given feature extraction technique can improve that value by up to 4.7%.", "date": "2026-06-17", "categories": []}
{"id": "2606.19579", "title": "FlowFake: Liquid Networks for Audio Deepfake Detection", "abstract": "Audio deepfakes generated by neural text-to-speech and voice-cloning systems threaten speaker verification and public discourse at scale. The core challenge is cross-dataset generalization: detectors trained on one synthesis pipeline collapse on unseen forgeries. We argue that this failure is primarily because of structural synthetic speech artifacts which are multi-timescale trajectory anomalies. Though every existing detector aggregates a fixed-window frame statistics, this misaligns the architecture with the signal. We propose FlowFake, a Liquid Time-Constant (LTC) architecture whose hidden state evolves via a learned ODE, with per-neuron adaptive time constants simultaneously resolving spectral (10ms) and prosodic (2s) cues. At only 34K parameters FlowFake achieves formal BIBO stability and O(dt^4) integration error. On a four-dataset cross domain benchmark (ASVspoof2019-LA, FakeOrReal, InTheWild, MLAAD), FlowFake reaches 75.29% on ASVspoof2019 trained only on FakeOrReal and 79.97% trained only on MLAAD. It outperforms RawGAT-ST and Whisper-DF on every evaluated pair and matching SSL Wav2vec2 (300x larger) at 0.01% of its parameter count. The source code is available on : https://github.com/GhostRider2023/FlowFake", "date": "2026-06-17", "categories": []}
{"id": "2606.19597", "title": "PrefSQA: Pairwise Preference Prediction for Speech Quality Assessment and the Critical Role of High Quality Datasets", "abstract": "Mean opinion scores (MOS) are widely used for speech quality assessment, yet scalar labels are sensitive to rater variability and listening test differences. This introduces labeling noise, which limits the reliability of MOS prediction. Preference prediction reduces this variability as listeners compare signals directly, producing cleaner labels. We study MOS-free preference prediction and propose PrefSQA, which incorporates uncertainty-aware logits, an impairment attention head, and a module based on non-matching-reference comparisons. We use and refine five datasets, including MOS-derived and low-noise simulated sets with matching and non-matching content, experiment with human preference sets, and test on unseen data. Experiments show small improvements on MOS-derived data, while other sets reveal clear improvement over the baselines, highlighting the value of high-quality preference data and demonstrating the effectiveness of the proposed method.", "date": "2026-06-17", "categories": []}
{"id": "2606.19629", "title": "RIVET: Robust Idempotent Voice Attribute Editing", "abstract": "Voice attribute editing models modify characteristics such as age and gender while preserving speaker identity. In large-scale speech datasets, however, attribute annotations are often noisy or inconsistent, which can cause conditional generative models to produce unstable edits. In this work, we show that idempotency provides an effective mechanism for improving robustness to noisy labels. An idempotent operator is one for which repeated application does not change the result, i.e., f(f(x)) = f(x). Enforcing this property acts as an implicit regularizer that reduces sensitivity to mislabeled examples. We introduce RIVET, a training framework that incorporates an idempotency objective to improve robustness to label noise. We evaluate RIVET under controlled label noise and on the GLOBE dataset with naturally noisy annotations. RIVET improves editing success and better preserves speaker identity than standard training, showing that idempotency improves robustness in voice editing models.", "date": "2026-06-17", "categories": []}
{"id": "2606.20101", "title": "Hybrid Diffusion Transformer for Instruction-Guided Audio Editing via Rectified Flow", "abstract": "Audio editing aims to modify specific content in an existing audio clip according to a natural language instruction while preserving the remaining acoustic content. Despite the remarkable progress of diffusion models, existing training-based editing methods mainly rely on the local inductive biases and cross-attention interaction in convolutional U-Net backbones, which often hinder long-range semantic alignment and precise understanding and localization of instructions. In contrast, diffusion transformers provide stronger global modeling and multimodal fusion, but existing editing architectures usually adopt a simple stack of MMDiT and DiT blocks. Applying joint attention over concatenated audio and text tokens in all blocks results in quadratic complexity with respect to token length. To balance editing performance and efficiency, we propose a hybrid two-stage diffusion transformer architecture for instruction-guided audio editing based on rectified flow matching. It performs joint attention over audio and text tokens to establish coarse semantic alignment at low-resolution stage, then switches to alternating joint-attention and cross-attention blocks to refine editing details at high-resolution stage. This coarse-to-fine strategy enables efficient and accurate instruction-guided audio editing. Experiments show that the proposed framework achieves notable performance gains on challenging editing tasks involving overlapping audio events and complex instructions, while substantially improving editing efficiency with a compact model.", "date": "2026-06-18", "categories": []}
{"id": "2603.10827", "title": "Speaker Verification with Speech-Aware LLMs: Evaluation and Augmentation", "abstract": "Speech-aware large language models (LLMs) can accept speech inputs, yet their training objectives largely emphasize linguistic content or specific fields such as emotions or the speaker's gender, leaving it unclear whether they encode speaker identity. First, we propose a model-agnostic scoring protocol that produces continuous verification scores for both API-only and open-weight models, using confidence scores or log-likelihood ratios from the Yes/No token probabilities. Using this protocol, we benchmark recent speech-aware LLMs and observe weak speaker discrimination (EERs above 20% on VoxCeleb1). Second, we introduce a lightweight augmentation that equips an LLM with ASV capability by injecting frozen ECAPA-TDNN speaker embeddings through a learned projection and training only LoRA adapters. On TinyLLaMA-1.1B, the resulting ECAPA-LLM achieves 1.03% EER on VoxCeleb1-E, approaching a dedicated speaker verification system while preserving a natural-language interface.", "date": "2026-03-11", "categories": []}
{"id": "2606.18273", "title": "Continuous Audio Thinking for Large Audio Language Models", "abstract": "Large audio language models (LALMs) have shown impressive capabilities on diverse audio understanding tasks, ranging from speech transcription to music analysis. However, because LALMs are typically trained to produce text-aligned responses, their hidden states are progressively shaped for text generation rather than for preserving acoustic information. As a result, the diverse acoustic content that audio carries, such as phonetic detail, prosody, sound events, affect, and pitch, is lost along the way and difficult to leverage in the response. We introduce Continuous Audio Thinking (CoAT), a framework that equips audio language models with a continuous latent workspace for organizing acoustic information prior to response generation, grounded by distillation from audio experts. Within the thinking space, the model can utilize the rich acoustic information provided by expert distillation when generating its response. Furthermore, the proposed continuous thinking block can be processed in a single prefill, so CoAT does not require additional autoregressive decoding cost over the baseline. Across three LALMs, Qwen2-Audio, Qwen2.5-Omni-7B, and Audio Flamingo~3, performance gains on a broad benchmark suite spanning audio reasoning, audio understanding, music classification, speech emotion, and speech transcription demonstrate the effectiveness of CoAT. Further analysis confirms that the auxiliary supervision propagates from the thinking positions to the model's textual responses.", "date": "2026-06-05", "categories": []}
{"id": "2606.18584", "title": "Speech-Driven End-to-End Language Discrimination towards Chinese Dialects", "abstract": "Language discrimination among similar languages, varieties, and dialects is a challenging natural language processing task. The traditional text-driven focus leads to poor results. In this paper, we explore the effectiveness of speech-driven features towards language discrimination among Chinese dialects. First, we systematically explore the appropriateness of speech-driven MFCC features towards CNN-based language discrimination. Then, we design an end-to-end speech recognition model based on HMM-DNN to predict Chinese dialect words. We adopt attention to extract the discriminative words related to different Chinese dialects. Finally, through a CNN, we combine the word-level embedding and the MFCC-based features. Evaluation of two benchmark Chinese dialect corpora shows the appropriateness and effectiveness of the proposed speech-driven approach to fine-grained Chinese dialect discrimination compared to the state-of-the-art methods.", "date": "2026-06-17", "categories": []}
{"id": "2606.17301", "title": "Turning music identification into a neural forward pass", "abstract": "Search, a foundational operation in computer science, maps a query to a matching item in a collection. It is typically implemented as a System-2 like, rule-based pipeline in which a key is computed, an index is probed, and candidates are verified. By contrast, human recognition resembles a System-1 like, associative model of identity recovery, in which even partial cues can trigger a recall without explicitly enumerating, ranking, or even accessing discrete candidates. Here, we show that music sound identification, a difficult search problem, can be performed in a single neural feed-forward pass by a generative transformer. Trained on an audio dataset, the model predicts the corresponding track identifier from a short audio excerpt. This approach surpasses state-of-the-art acoustic fingerprinting, with the largest gains for short audio segments (1 second), demonstrating the method is not only viable but advantageous. Moreover, it reduces external storage to 0.33% of the baseline footprint and improves inference latency by 2.3x (p95). Furthermore, the model can reject queries for unseen tracks, supporting open-set operation while reducing misattribution risk. Using music track identification as an example, this work reframes search, bringing it closer in spirit to human associative recognition and away from algorithmic database lookup.", "date": "2026-06-15", "categories": []}
{"id": "2606.18135", "title": "Descriptor: Certus Caliber Classification Gunshot Dataset (C3GD)", "abstract": "In this work, we introduce the Certus Caliber Classification Gunshot Dataset (C3GD), a publicly accessible data set developed for the analysis of firearm muzzle blast sounds. The dataset aims to provide a wide variety of firearms, calibers, cartridges, microphones, and microphone locations with metadata detailed beyond what is currently otherwise available. It comprises more than 8000 field-collected data points from 28 firearms across 16 calibers. Because data collection in the field is costly, much of the existing research has been done using gunshot audio collected from the internet, which increases the risk of low-quality data and label noise. This dataset is primarily focused on caliber classification, but can also be used for gunshot detection, audio separation, and audio signal processing, providing a diversified and real-world reference. The dataset aims to provide enough diversity to be able to generalize to more real-world applications while also providing enough metadata for detailed academic analysis.", "date": "2026-06-16", "categories": []}
{"id": "2509.14959", "title": "Discrete optimal transport is a strong audio adversarial attack", "abstract": "In this paper, we investigate discrete optimal transport (DOT) as a black-box attack against modern automatic speaker verification (ASV) and anti-spoofing countermeasure (CM) systems.\n  Our attack operates as a post-processing distribution-alignment step. Frame-level WavLM embeddings of generated speech (or another person speech) are aligned to an unpaired bona fide speech pool using entropic optimal transport and a top-k barycentric projection, followed by neural vocoding. Unlike gradient-based attacks, the proposed method requires no access to model parameters, gradients, or training data.\n  Experiments on ASVspoof2019 and ASVspoof5 demonstrate that DOT attack substantially increases CM EER and substantially degrades ASV performance across multiple spoofing attacks. The attack transfers across datasets and remains effective after CM fine-tuning. Analysis using speaker similarity, Fr\\'echet Audio Distance, and visualization of embedding distributions suggests that DOT succeeds by shifting source speech toward bona fide regions of the representation space rather than by maximizing speaker similarity.\n  These results indicate that optimal-transport-based distribution alignment represents a previously underexplored attack vector for contemporary ASV and anti-spoofing systems.", "date": "2025-09-18", "categories": []}
{"id": "2512.15313", "title": "Time-Varying Audio Effect Modeling by End-to-End Adversarial Training", "abstract": "Deep learning has become a standard approach for the modeling of audio effects, yet strictly black-box modeling remains problematic for time-varying systems. Unlike time-invariant effects, training models on devices with internal modulation typically requires the recording or extraction of control signals to ensure the time-alignment required by standard loss functions. This paper introduces a Generative Adversarial Network (GAN) framework to model such effects using only input-output audio recordings, without requiring a modulation signal extraction. We propose a convolutional-recurrent architecture trained via a two-stage strategy: an initial adversarial phase allows the model to learn the distribution of the modulation behavior without strict phase constraints, followed by a supervised fine-tuning phase where a State Prediction Network (SPN) estimates the initial internal states required to synchronize the model with the target. Additionally, a new metric based on chirp-train signals is developed to quantify modulation accuracy. Experiments modeling a vintage hardware phaser demonstrate the method's ability to capture time-varying dynamics in a fully black-box context.", "date": "2025-12-17", "categories": []}
{"id": "2602.01394", "title": "SSNAPS: Audio-Visual Separation of Speech and Background Noise with Diffusion Inverse Sampling", "abstract": "This paper addresses the challenge of audio-visual single-microphone speech separation and enhancement in the presence of real-world environmental noise. Our approach is based on generative inverse sampling, where we model clean speech and ambient noise with dedicated diffusion priors and jointly leverage them to recover all underlying sources. To achieve this, reformulate a recent inverse sampler to match our setting. We evaluate on mixtures of 1, 2, and 3 speakers with noise and show that, despite being entirely unsupervised, our method consistently outperforms leading supervised baselines in WER across all conditions. We further extend our framework to handle off-screen speaker separation. Moreover, the high fidelity of the separated noise component makes it suitable for downstream detection of the acoustic scene. Code and pretrained models will become available upon acceptance. Demo page: https://ssnaps2026.github.io/ssnaps2026/", "date": "2026-02-01", "categories": []}
{"id": "2606.08898", "title": "Few-shot Class-variable Incremental Audio Classification via Prototype Adaptation and Pseudo Class-variable Training", "abstract": "In the task of few-shot class-incremental audio classification, the number of classes is assumed to always increase without considering the possibility of decrease. However, the number of classes generally increases or decreases in practice. In this paper, we investigate a problem of Few-shot Class-variable Incremental Audio Classification (FCIAC), in which the number of classes increases or decreases. We propose a FCIAC method using prototype adaptation and pseudo class-variable training. The model in our method consists of an encoder and a classifier. The classifier is initialized by a class-variable prototype adaptation network, whose structure dynamically changes with the change of classes. In addition, we design a pseudo class-variable training strategy to enhance the model's adaptability to changing classes. Experiments on three public datasets show that our method exceeds previous methods in average accuracy. The code is at: https://github.com/cgq2971-afk/FCIAC.", "date": "2026-06-08", "categories": []}
{"id": "2606.14791", "title": "From Physics to Representation: Audio Learning with Synthetic Pre-training via Procedural Generation", "abstract": "Self-supervised learning advances audio representation for multimedia analysis. However, prevailing data-centric approaches rely on massive real-world corpora, increasing training costs, curation burdens, and privacy barriers. To address this, we present AudioPG, a procedural synthesis framework eliminating real audio recordings during pre-training. AudioPG trains a Transformer-based masked autoencoder on waveforms generated on-the-fly from basic acoustic primitives and composition rules. The encoder transfers effectively to real audio benchmarks, achieving 90.60% accuracy on ESC-50, 0.546 mAP on FSD50K, 88.17% on UrbanSound8K, and 97.03% on Speech Commands V2. Notably, pre-training completes in under 20 minutes on a single GPU. Latent space analysis reveals physical factors, including fundamental frequency and relative intensity, emerge in orthogonal subspaces, making representations linearly decodable. These results establish procedural synthesis as an efficient, interpretable pre-training signal when large-scale corpora are unavailable. Our code is available at: https://github.com/Freyliu0516/audioPG.", "date": "2026-06-11", "categories": []}
{"id": "2606.14820", "title": "Spectro-Temporal Interference Confounds Phase Encoding in Spatial Audio Foundation Models", "abstract": "Recent spatial self supervised audio models achieve high performance on localization tasks, raising questions about their encoding of microsecond interaural phase fine structures. We propose a psychoacoustic benchmark based on the binaural masking level difference to evaluate this. Using an equalization cancellation baseline and a GCC PHAT positive control we evaluate nine frozen audio models spanning binaural SSL, monaural SSL, and neural audio codecs. Four monaural negative controls yield zero BMLD confirming binaural specificity. Two general purpose binaural SSL models exhibit minimal phase sensitivity while dedicated binaural spatial SSL models achieve BMLD comparable to the analytical baseline. Progressive physical ablations show that general purpose binaural SSL models rely on spectro temporal interference textures rather than cross channel phase computation. High detection rates in speech reflect a confounding reliance on broadband envelopes rather than genuine phase encoding.", "date": "2026-06-12", "categories": []}
{"id": "2606.15436", "title": "Beyond Classification: A Cough Regression Benchmark for Respiratory Acoustic Foundation Models", "abstract": "Respiratory acoustic foundation models (FMs) excel at cough classification, yet their ability to predict continuous health quantities from cough audio remains largely unexplored, despite the clinical value of passive age, BMI, and disease probability estimation in settings where physical measurements are unavailable. We introduce the multi-model, multi-target cough regression benchmark evaluating five FMs (OPERA-CT, OPERA-CE, OPERA-GT, HeAR, M2D+Resp) across six targets on three datasets under subject-disjoint protocols, comparing linear, MLP-small, and full MLP regression heads. MLP-small beats the mean-predictor baseline on all tasks and linear probing in 23 of 30 model x task cases, with full MLP overfitting on small clinical data but recovering on larger sets, revealing a dataset size x head-capacity trade-off. HeAR leads within-dataset age regression on Coswara (9.12 yr MAE); its CIDRZ result is excluded from headline claims owing to possible HeAR-CIDRZ pretraining overlap. OPERA-GT is favored over OPERA-CT on age in all three datasets, with the CIDRZ margin within seed variance, extending a generative-pretraining advantage from breath to cough. HeAR and M2D+Resp reach near-full performance at N = 50 samples while OPERA models require N = 400. Cross-dataset transfer is strongly asymmetric as large diverse data generalises to small clinical populations (CoughVID to CIDRZ: -0.17 yr) but not vice versa (CIDRZ to Coswara: +2.43 yr, +26.6%).", "date": "2026-06-13", "categories": []}
{"id": "2606.15751", "title": "Acoustic Prompting via Stage-wise Modulation for Few-Shot Learning in Audio Language Models", "abstract": "Audio-Language Models (ALMs) have shown remarkable success in zero-shot audio classification by aligning audio waveforms with text. Recent efforts to improve downstream performance focus on learning optimal text prompts. However, previous approaches focus on the text encoder, leaving the potential of learnable prompts within the audio encoder unexplored. In this paper, we propose a novel framework that introduces trainable prompts into the audio encoder to capture task-specific acoustic features. We demonstrate that integrating audio-side prompt learning with existing text-side approaches enhances few-shot adaptation. Through extensive experiments across 11 datasets show that integrating our method as a plug-and-play module alongside existing text prompt tuning generally leads to performance improvements. These findings suggest that explicitly modulating the audio representation space effectively complements text-only prompting approaches. The code is available at https://github.com/hyebin-c/aspl.", "date": "2026-06-14", "categories": []}
{"id": "2606.15888", "title": "NVMOS: Non-Verbal Vocalization Quality Assessment in Speech", "abstract": "Non-verbal vocalizations (NVs), such as laughter, sighs, and coughs, are important acoustic cues for emotion and intent. Existing speech quality assessment methods typically focus on overall naturalness, while non-verbal TTS evaluations mainly examine whether a target NV appears with the correct type and position. However, the perceptual quality of NV events themselves remains underexplored. To address this gap, we construct an NV-MOS dataset containing outputs from multiple NV-TTS systems and naturally occurring NV samples, with ratings collected from three acoustic experts on a perceptual quality scale. We further analyze audio-capable multimodal large language models such as Gemini and find clear inconsistencies between their scores and expert ratings. These results suggest that general-purpose multimodal models cannot reliably replace human judgments for NV quality assessment. We then propose NVMOS, to our knowledge the first model that can reliably predict the perceptual quality of NV events in speech. Experimental results show that, with a local NV-event focusing module, NVMOS reaches expert-level or stronger agreement with human MOS.", "date": "2026-06-14", "categories": []}
{"id": "2606.16472", "title": "From Awareness to Adherence: Bridging the Context Gap in Spoken Dialogue Systems via Context-Aware Decoding", "abstract": "Despite the success of end-to-end (E2E) spoken dialogue systems, maintaining strict context adherence in multi-round conversations remains a challenge. While prior works attribute these failures to models forgetting dialogue history, we highlight an equally critical but overlooked bottleneck: a gap between latent context awareness and active adherence. Although models internally recognize relevant past utterances, strong parametric priors often overshadow these signals during decoding. To bridge this gap, we propose an audio-adapted Context-Aware Decoding (CAD) approach. By leveraging internal attention mechanisms to isolate key historical rounds, our approach contrasts output distributions with and without this key context during inference, directly amplifying multimodal contextual signals. Evaluations on the Audio MultiChallenge benchmark demonstrate significant improvements in Semantic Memory and Self Coherence subtasks, successfully enforcing strict, context-faithful adherence.", "date": "2026-06-15", "categories": []}
{"id": "2606.16532", "title": "Dual-Granularity Orthogonal Disentanglement for Generalizable Audio Deepfake Detection", "abstract": "Audio deepfake detectors often fail to generalize across speakers, as they learn speaker-identity features rather than synthesis artifacts, known as implicit identity leakage. Existing methods address this but incur architectural complexity or training instability. This paper proposes a dual-granularity orthogonal disentanglement framework enforcing feature independence at two levels: sample-level cosine orthogonality captures directional decorrelation, while batch-level cross-covariance regularization eliminates linear correlations across embedding dimensions. A curriculum disentanglement schedule progressively strengthens the orthogonality constraint without auxiliary networks or adversarial dynamics. Experiments on ASVspoof 2019 LA, ASVspoof 2021 DF, and In-the-Wild datasets demonstrate that the proposed method achieves 1.35%, 7.88%, and 21.58% equal error rates (EER), respectively, surpassing gradient reversal disentanglement by 2.60% absolute on cross-dataset transfer.", "date": "2026-06-15", "categories": []}
{"id": "2606.16612", "title": "Beyond Artifacts: Towards Generalizable Synthetic Song Detection via Music-Intrinsic Features", "abstract": "The rapid advancement of AI music generators highlights the urgent need for reliable Synthetic Song Detection (SSD). Existing SSD methods often rely on low-level artifacts or fixed feature assumptions, struggling to capture generator-agnostic cues. To address this, we propose Sofia (Synthetic-song detection framework via music features), a flexible framework that models music-intrinsic attributes via feature-specific experts and an adaptive Mixture-of-Experts (MoE) module. By configuring Sofia with representative Vocal, Audio-effect, Global structure features, and their combinations, we present their individual and complementary contributions. To comprehensively evaluate our framework, we further construct MUSIC8K, a challenging benchmark featuring lastest emerging generators and realistic audio perturbations. Experiments show that Sofia learns generator-agnostic representations from music-intrinsic features, improving the F1 score by 18.5 points over the strongest baseline on MUSIC8K-O while maintaining strong robustness.", "date": "2026-06-15", "categories": []}
{"id": "2606.16837", "title": "Robust Spoofed Speech Detection via Temporal Pyramid Modeling", "abstract": "Spoofed speech detection is increasingly challenged by realistic synthesis, voice conversion, and replay attacks, with cross-dataset generalization remaining a major limitation. This work we propose a Temporal Pyramid Adapter that utilize parallel temporal convolutions with varying receptive fields to capture multi-scale spoofing cues, ranging from local artifacts to global prosodic irregularities. We also integrated self-supervised XLS-R representations combined with front-end adapters, including Mel, Sinc, and a Temporal Pyramid design for multi-scale temporal modeling. The proposed model is evaluated cross multiple benchmark including ASVspoof 2017, ASVspoof 2021 (DF/LA), PartialSpoof, DiffSSD, and multilingual HQ-MPSD datasets. Experimental results demonstrate that Temporal Pyramid model obtained AUC of 99.24% and a EER of 3.87% on the PartialSpoof database, which is significantly outperforming the base model and several SOTA baseline such as LCNN-BLSTM (9.87% EER) and TRACE (8.08% EER). Additionally, multilingual evaluations confirm that while spoofing artifact are independent from language. While self-supervised representations improve robustness, performance degrades under domain and language shifts, highlighting the need for better adaptation and calibration strategies.", "date": "2026-06-15", "categories": []}
{"id": "2606.16969", "title": "Probing Low Frame Rate Degradation in Neural Audio Codecs", "abstract": "Low frame rates in neural audio codecs are attractive for autoregressive speech synthesis, where the generation cost scales linearly with the sequence length. Recent work has demonstrated that codecs can operate at 12.5 Hz and below, but the mechanisms underlying low frame rate degradation remain insufficiently understood. We investigate these mechanisms through a controlled frame rate ablation. We reproduce a quality cliff at 6.25 Hz reported in previous works and evaluate candidate explanations: phonemic collisions and codebook saturation, neither of which shows evidence of a fundamental barrier. The cliff is instead caused by suboptimal training configuration: fixed clip duration during training yields too few tokens at low frame rates, starving the decoder of inter-token context. Once corrected, WER degrades smoothly with phonemic load down to 3.1 Hz and 1.6 Hz, suggesting the inference-time efficiency gains of low frame rate codecs are more accessible than previously assumed.", "date": "2026-06-15", "categories": []}
{"id": "2602.05670", "title": "HyperPotter: Spell the Charm of High-Order Interactions in Audio Deepfake Detection", "abstract": "Advances in AIGC technologies have enabled the synthesis of highly realistic audio deepfakes capable of deceiving human auditory perception. Although numerous audio deepfake detection (ADD) methods have been developed, most rely on local temporal/spectral features or pairwise relations, overlooking high-order interactions (HOIs). HOIs capture discriminative patterns that emerge from multiple feature components beyond their individual contributions. We propose HyperPotter, a hypergraph-based framework designed to capture high-order relations associated with synergistic patterns through clustering-based hyperedges with class-aware prototype initialization. Extensive experiments on 13 test sets show that HyperPotter improves over the baseline on 11 sets, yielding an average relative EER reduction of 12.68\\% across all test sets and 22.15\\% on the improved sets. These results demonstrate strong cross-scenario generalization, while also revealing robustness limits under severe codec or channel distortion.", "date": "2026-02-05", "categories": []}
{"id": "2606.14049", "title": "FoleyGenEx: Unified Video-to-Audio Generation with Multi-Modal Control, Temporal Alignment, and Semantic Precision", "abstract": "We present FoleyGenEx, a unified video-to-audio (VTA) framework integrating multi-modal control, frame-level temporal alignment, and fine-grained semantics, enabling synchronized, versatile audio synthesis for diverse tasks. Existing VTA methods either have multi-modal control but weak temporal alignment or strong alignment but lack reference audio conditioning and semantic precision. FoleyGenEx fills this gap via three core innovations: a conditional injection mechanism for audio-controlled VTA and Foley extension, a multi-modal dynamic masking strategy preserving training synchronization, and an adverb-based data augmentation algorithm leveraging signal processing and large language models to enhance textual supervision with nuanced semantics. Experiments on AudioCaps, VGGSound, and Greatest Hits demonstrate its competitive controllable VTA performance against existing methods. Demo samples are available at https://foleygenex.github.io/FoleyGenEx.", "date": "2026-06-12", "categories": []}
{"id": "2606.14141", "title": "Spatio-Temporal Audio Language Modeling for Dynamic Sound Sources", "abstract": "Sound events are entities with semantic identities, locations, and trajectories, but current audio-language models usually reason about clips as global event content. Conversely, sound event localization models track source directions over time but offer limited semantic coverage for language reasoning. To address this gap, we introduce ST-AudioQA, a spatio-temporal audio QA dataset and benchmark built from first-order ambisonic (FOA) renderings of static and moving sound sources. Each scene provides source identity, activity, direction, distance, and motion metadata, enabling dense trajectory supervision and questions about what is sounding, where it is, how it moves, and how sources relate. We further propose ST-Audio Encoder, a time-resolved FOA audio encoder that learns event semantics together with source trajectories, and ST-AudioLM, which connects the audio tokens from the encoder to an LLM for spatio-temporal audio QA. Experiments show that this representation improves the semantic-localization tradeoff and yields stronger reasoning performance than static spatial and localization-oriented baselines.", "date": "2026-06-12", "categories": []}
{"id": "2606.14591", "title": "AudioDER: A Deduplication-Enhanced Reasoning Dataset for Post-Training Large Audio-Language Models", "abstract": "Large Audio-Language Models (LALMs) have shown strong performance on a wide range of audio understanding tasks, yet they still struggle with complex audio reasoning. A practical way to improve such capabilities is post-training, whose effectiveness critically depends on the quality and diversity of training data. However, existing audio-language datasets often contain substantial redundancy, where many samples are highly similar in acoustic content and thus provide overlapping supervisory signals. Such redundancy not only increases annotation cost, but also limits corpus diversity and reduces the effectiveness of post-training. To address this issue, we propose a redundancy-aware data construction pipeline for building reasoning-oriented supervision for LALMs. Specifically, we first perform acoustic similarity-based deduplication across raw audio datasets to improve corpus diversity. We then integrate existing audio captions and question-answer pairs into a unified multiple-choice format. Based on these unified annotations, we leverage Qwen3-30B to generate chain-of-thought (CoT) rationales for reasoning-oriented supervision. Based on this pipeline, we construct AudioDER, a reasoning-oriented post-training dataset containing approximately 191k samples spanning sound, speech, and music. Each sample consists of an audio clip, a multiple-choice question, four answer candidates, an audio caption, and a CoT rationale. Extensive experiments show that post-training on AudioDER consistently improves the performance of Qwen2-Audio-7B-Instruct on multiple audio reasoning benchmarks, including MMAU-mini, MMSU, and MMAR. We hope AudioDER can serve as a valuable resource for advancing audio reasoning research and the development of more capable LALMs.", "date": "2026-06-12", "categories": []}
{"id": "2509.04682", "title": "GetNetUPAM: Ecologically Informed Nested Cross-Validation and Noise-Robust Attention for Marine Bioacoustic Monitoring", "abstract": "Deploying reliable bioacoustic monitoring systems requires models that generalize under high-noise, low-SNR conditions and evaluation protocols that expose deployment-relevant failure modes, gaps largely unaddressed in current UPAM practice. Intrinsic noise, variable propagation, and mixed biological and anthropogenic sources induce distribution shifts that conventional models and single-split evaluations obscure, inflating performance and masking instability. We introduce GetNetUPAM, a hierarchical nested cross-validation framework that uses the nested stage to quantify model stability rather than tune for inflated hold-out scores. By partitioning data into site-year blocks, GetNetUPAM preserves ecological heterogeneity and forces each outer fold to represent a distinct environmental regime, preventing overfitting to localized noise or sensor artifacts. Inner stratified folds measure generalization across the full UPAM signal distribution, enforcing strict separation between model development and the outer held-out deployment condition. Using GetNetUPAM, we evaluate the Adaptive Resolution Pooling and Attention Network (ARPA-N), a CNN architecture for irregular spectrogram dimensions. ARPA-N integrates CBAM spatial attention as a learned noise suppressor, producing attention maps that localize true call structure and avoid the global, non-biological cues exploited by standard CNNs on long-window data. Under GetNetUPAM, ARPA-N generalizes robustly across diverse environmental regimes. In the zero-training support Balleny Islands region, it reduces false positives per hour by over an order of magnitude (approximately 10x) at fixed 90 percent recall, yielding consistently improved metrics across folds. These advances provide a reproducible benchmark and move UPAM toward scalable, deployment-reliable ecological monitoring.", "date": "2025-09-04", "categories": []}
{"id": "2606.12503", "title": "Dolph2Vec: Self-Supervised Representations of Dolphin Vocalizations", "abstract": "Self-supervised learning (SSL) has opened new opportunities in bioacoustics by enabling scalable modeling of animal vocalizations without the need for expensive manual annotation. However, current SSL models in this domain prioritize broad generalization across species and are not optimized for uncovering the fine-grained structure of individual communication systems. In this work, we collect and release a novel dataset of over five years of longitudinal recordings, from five known dolphins in a semi-naturalistic marine environment, an unprecedented resource for studying dolphin communication. We adapt the Wav2Vec2.0 Baevski et al. (2020) architecture to this domain and introduce Dolph2Vec, the first large-scale, species-specific SSL model trained exclusively on this data. We benchmark our model on two biologically relevant tasks: signature whistle classification and whistle detection. Dolph2Vec significantly outperforms general-purpose baselines in both tasks. Beyond performance, we show that learned embeddings and codebook structure capture interpretable acoustic units aligned with dolphin whistle categories and possibly sub-whistle structure, enabling fine-grained analysis of communication patterns. Our findings demonstrate how SSL can serve as both a model and a scientific tool to explore hypotheses in animal communication research.", "date": "2026-06-10", "categories": []}
{"id": "2606.13236", "title": "Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier", "abstract": "Passive acoustic monitoring holds great promise for ecological inference, yet existing automated tools are typically narrowly trained and non-transferable. We address these limitations with PULSE, a semi-supervised, multi-task framework for Orthoptera bioacoustics, combining weakly-supervised species classification, self-supervised learning on unlabelled field audio, and knowledge distillation from a general-purpose bioacoustic model. Our domain-adapted specialist model outperforms a state-of-the-art general model across all metrics (macro F1: 0.21 vs. 0.07; AUC: 0.74 vs. 0.45; AP: 0.32 vs. 0.19), with active learning further raising F1 to 0.34 and AUC to 0.84. Beyond classification, the learned embeddings encode ecologically meaningful structure, exposed through an interactive visualisation tool for ecological discovery.", "date": "2026-06-11", "categories": []}
{"id": "2606.10046", "title": "Inside the Latent Flow: Causal Deciphering of Attention Dynamics in Audio Separation Foundation Models", "abstract": "Flow-matching transformers achieve strong audio separation, yet their attention dynamics are opaque. We adapt established causal-intervention principles into a deterministic, inference-time probing protocol for SAM Audio. Orthogonal probing uncovers a dual-pathway text-conditioning mechanism: additive injections control semantic identity, while cross-attention refines acoustic structure. We observe an asynchronous layerwise convergence: stable layers build temporal scaffolds early, whereas fast layers continue resolving artifacts during sampling. The model also attenuates temporal segmentation cues to maintain continuous-flow stability. Using these insights, we propose Layer-Selective Attention Caching (LSAC), a training-free acceleration method that caches attention in stable layers. Across acoustic complexities, LSAC cuts self-attention computation by about ~25% with negligible quality loss and yields up to 6.7x higher quality retention than naive step reduction.", "date": "2026-06-08", "categories": []}
{"id": "2606.11542", "title": "Pretrained self-supervised speech models can recognize unseen consonants", "abstract": "Modern pretrained self-supervised automatic speech recognition models are trained on large-scale audio data to encode speech into contextualized representations. However, their training data are heavily skewed toward high-resource languages with little data from low-resource languages, raising concerns about the potential underrepresentation of typologically uncommon speech sounds such as click consonants primarily found in Khoisan languages. This leads to our central research question: Can these models recognize click consonants as accurately as other speech sounds? To address this question, we fine-tune and compare pretrained self-supervised speech models (Wav2Vec2 and HuBERT) on data from two click-rich Khoisan languages (G|ui and West !Xoon). Our results reveal that the fine-tuned models consistently recognize clicks more accurately than non-clicks, suggesting that self-supervision enables generalization across human speech sounds including rare phonemes.", "date": "2026-06-10", "categories": []}
{"id": "2606.11639", "title": "Evaluating Bias in Phoneme-Based Automatic Speech Recognition Systems: An Analysis of IPA Transcription Models", "abstract": "The popularization of automatic speech recognition (ASR) systems has increased exploration of the demographic biases related to race, age, gender, and accent, often formed from imbalanced training data. Most of these studies focused on standard grapheme-based ASR systems with comparatively little emphasis on phoneme-based systems, such as models that produce International Phonetic Alphabet (IPA) representations. As ASR systems shift toward multilingual support and low-resource language modeling, IPA-based layers serve as a critical, language-agnostic foundation. In this study, we evaluate the performance of two state-of-the-art open-source ASR systems, WhisperIPA and ZIPA, that generate IPA transcriptions across diverse accents and language sources. Our evaluation includes existing multilingual speech corpora and demographically annotated English-language corpora. We measure model performance by comparing model-generated IPA transcriptions against grapheme-to-phoneme (G2P) systems using both standard phoneme error rate (PER) and a proposed Soft PER metric that tolerates linguistically similar phoneme substitutions. Our analysis examines how performance varies across languages and demographic groups such as gender, accent, ethnicity, and age, revealing persistent disparities even after accounting for acceptable phonemic variation. These findings provide insight into potential sources of bias and inform the development of more inclusive and linguistically robust phoneme-based ASR systems. Our code and data will be made publicly available to the community.", "date": "2026-06-10", "categories": []}
{"id": "2606.11915", "title": "Quality Adaptive Angular Margin Learning for Respiratory Sound Classification", "abstract": "We present a quality-adaptive angular-margin learning framework that improves feature generalization by enforcing intra-class compactness and inter-class separability. Our framework, titled QLung, introduces a no-reference audio quality margin derived from spectral entropy and root-mean-square energy, which adaptively scales angular margins based on recording quality. To this end, we propose a log-scaled angular margin that stabilizes training under severe class imbalance. We also use an angular classifier that normalizes features and class weights, ensuring margin penalties are applied consistently on the unit hypersphere. Our approach improves in-distribution performance on the ICBHI dataset by 2.46\\% over the cross-entropy baseline, and most significantly, achieves the strongest out-of-distribution performance on the SPRSound dataset compared to prior state-of-the-art methods. Code is available at https://github.com/RSC-Toolkit/QLung.", "date": "2026-06-10", "categories": []}
{"id": "2606.11922", "title": "Lung-SRAD: Spectral-Aware Regularized Audio DASS with Dual-Axis Patch-Mix Contrastive Learning for Respiratory Sound Classification", "abstract": "Recent respiratory sound classification (RSC) studies largely rely on CLS-token driven self-attention architectures such as the Audio Spectrogram Transformer (AST). While effective at modeling global context, recent analyses suggest a low-pass filtering behavior that may reduce sensitivity to localized abnormal patterns. In this work, we investigate State Space Models (SSMs) as an alternative backbone for RSC. Using the Distilled Audio State Space model, we analyze intermediate representations through spectral response curves and observe stronger preservation of mid-to-high spatial-frequency components. Based on these observations, we introduce spectral-aware layer regularization using Gaussian convolution applied to selected layers. We further propose Dual-Axis Patch-Mix contrastive learning tailored to SSM-based audio models for robust representation learning. Experiments on the ICBHI benchmark show that our approach achieves 64.48% score, outperforming the AST baseline by 5%. Code is available at https://github.com/RSC-Toolkit/Lung-SRAD.", "date": "2026-06-10", "categories": []}
{"id": "2606.12282", "title": "PianoKontext: Expressive Performance Rendering from Deadpan Context", "abstract": "Expressive performance rendering (EPR) aims to generate realistic performances constrained on sequences of notes. However, flow matching audio editing models manipulate only synchronized music samples of the same duration, limiting their understanding of expressive timing. We introduce PianoKontext, a flow matching rendering model for classical piano music that generates variable-length performances in the latent space of a pretrained Music2Latent model. We synthesize MIDI scores into deadpan audio and employ Dynamic Time Warping (DTW) in the latent space to construct paired data for training. The aligned embeddings are concatenated in DiT blocks, allowing for a simple and effective learning of the dependencies between the score and performances. Audio samples are available at our demo page: https://realfolkcode.github.io/pianokontext_demo/.", "date": "2026-06-10", "categories": []}
{"id": "2412.11449", "title": "Whisper-GPT -- Continuous Discrete Hybrid Representation Language Models For Speech And Music", "abstract": "We propose WHISPER-GPT: A generative large language model (LLM) for speech and music that allows us to work with continuous audio representations and discrete tokens simultaneously as part of a single architecture. There has been a huge surge in generative audio, speech, and music models that utilize discrete audio tokens derived from neural compression algorithms, e.g. ENCODEC. However, one of the major drawbacks of this approach is handling the context length. It blows up for high-fidelity generative architecture if one has to account for all the audio contents at various frequencies for the next token prediction. By combining continuous audio representation like the spectrogram and discrete acoustic tokens, we retain the best of both worlds: Have all the information needed from the audio at a specific time instance in a single token, yet allow LLM to predict the future token to allow for sampling and other benefits discrete space provides. We show how our architecture improves the perplexity and negative log-likelihood scores for the next token prediction compared to a token-based LLM for speech and music.", "date": "2024-12-16", "categories": []}
{"id": "2606.10233", "title": "ANCHOR: Autoregressive Non-intrusive Chunk-Ordered Refinement for Joint Multi-Resolution Speech Quality Modeling", "abstract": "While speech quality is typically assessed on complete utterances, streaming and generative systems require incremental estimation from partial audio. Existing predictors assume full context, degrading on prefix-constrained inputs. Extending ARECHO, we propose ANCHOR, reformulating incremental assessment as a multi-resolution autoregressive task. It models chunk- and utterance-level quality within a single decoder using dual-resolution tokens and a resolution-aware hierarchy for coarse-to-fine refinement. Experiments show substantial robustness under partial input, including a 48% PLCMOS error reduction on 2-second prefixes. Convergence analysis reveals a 4-6 s effective perceptual context horizon. A stress test further isolates structured extrapolation biases under localized corruption. Results demonstrate that hierarchical supervision improves incremental prediction and elucidates how perceptual quality accumulates over time.", "date": "2026-06-08", "categories": []}
{"id": "2606.10246", "title": "Linguistically Augmented Audio Speech Data (LinguAS)", "abstract": "Maliciously-created fake speech, including deepfaked and spoofed audio, is proliferating at an alarming rate, and detection models are racing to stay ahead of the curve. Yet, most detection models are trained to make inference on frame-level audio features alone without leveraging valuable linguistic cues at larger timescales. To address this gap, we present Linguistically Augmented Audio Speech Data (LinguAS), a dataset of genuine and deepfaked audio samples annotated with five strategically-chosen, Expert-Defined Linguistic Features (EDLFs) that occur frequently in spoken English and are characteristic of natural human speech. LinguAS contains over 800 audio samples, each of which are annotated with EDLFs. The dataset has a balanced number of four spoofed audio attack types and a proportionate number of genuine speech samples. We also include metadata on speaker gender and the generator/source for each spoofed audio sample, offering more granularity for model training. We found that models trained on data augmented with EDLFs had improved model performance significantly beyond the ASVspoof 2021 deep learning baselines and SSL models like HuBert and XLSR. LinguAS's augmented linguistic, gender, and generator metadata provide audio deepfake researchers with a dataset that emphasizes real human language traits to improve model inference of faked speech. Data and code are publicly available.", "date": "2026-06-08", "categories": []}
{"id": "2606.10407", "title": "Time-frequency localization of bird calls in dense soundscapes", "abstract": "Passive acoustic monitoring enables large-scale observation of wildlife, but most bioacoustic classifiers only predict species presence in a time window without localizing vocalizations precisely in time or frequency, limiting downstream analyses. We formulate bird vocalization detection as an object detection task on spectrograms and train YOLO11 models to localize bird calls in dense tropical soundscapes from Singapore. We additionally introduce an open-source browser-based annotation tool and propose Intersection over Minimum (IoMin), an evaluation metric that better handles ambiguous acoustic boundaries than standard IoU and is better suited to the problem at hand. The best YOLO model nearly doubles baseline performance on in-distribution soundscapes from Singapore (81.8% vs. 42.1% IoMin@50 F1-score) while still outperforming the baseline on unseen out-of-distribution recordings from Hawaii (58.6% vs. 48.6%). These results suggest that object detection frameworks are a promising approach to time-frequency localization of animal vocalizations in complex soundscapes.", "date": "2026-06-09", "categories": []}
{"id": "2606.10738", "title": "Spatial-Omni: Spatial Audio Understanding Integration in Multimodal LLMs via FOA Encoding", "abstract": "Recent multimodal large language models mainly process audio as monaural signals, thereby discarding the spatial cues contained in spatial audio for sound localization, spatial relation reasoning, and spatial scene understanding. We propose Spatial-Omni, a lightweight method that implements SO-Encoder to inject First-Order Ambisonics (FOA) spatial audio into existing Omni LLMs as an independent modality, without modifying their original audio encoders. SO-Encoder provides spatial tokens with limited additional context cost and improves spatial audio understanding through efficient staged training. To support training and evaluation, we construct SO-Dataset, SO-QA, and SO-Bench from open-source data, real recordings, and simulations, containing 400K FOA spatial audio clips and 2.1M spatial question answering pairs. SO-Bench covers 16 spatial audio understanding subtasks, including basic detection and location estimation, spatial relation understanding, and complex spatial reasoning. Experiments show that Spatial-Omni outperforms existing open-source Large Audio-Language Models (LALMs) and Omni LLM models on spatial audio understanding tasks while retaining a reasonable level of general audio understanding. Code and data are available at https://github.com/dieKarotte/Spatial-Omni.", "date": "2026-06-09", "categories": []}
{"id": "2606.10811", "title": "Deep learning for echo sounder data", "abstract": "There is no doubt that over the last decade, techniques from the field of machine learning have revolutionized how we process and interpret data, especially images and text. For underwater observations acoustics is a primary source of information, and naturally, deep learning methods have been applied to echograms and other acoustics data, but so far with rather modest results. Here, we argue that due to intrinsic properties of acoustic data, substantial advances will likely require research into deep learning methods beyond mere recycling of models and techniques from image processing. Currently, the potential for breakthroughs in method development is hindered by the lack of standard data formats and organization, and even more by the lack of readily available, high quality data sets with established performance goals. To advance the field, these shortcomings should be remedied", "date": "2026-06-09", "categories": []}
{"id": "2606.10908", "title": "RAT: Reference-Augmented Training for ASV Anti-Spoofing", "abstract": "We introduce a spoofing countermeasure architecture conditioned on speaker-reference recordings, but observe that it converges to a solution that effectively ignores the reference during inference. Surprisingly, training with a reference channel induces invariance that improves deepfake detection, even when the reference is absent or mismatched during inference. Based on this observation, we propose a Reference-Augmented Training (RAT) strategy. RAT yields improved detection performance compared to single-utterance baselines, even when the reference recording is replaced with a zero vector at inference. Through rigorous analysis, we demonstrate that the optimization process rapidly diminishes the reference contributions, leading to inference largely independent of the reference channel. Using RAT, we achieve state-of-the-art 2.57% EER and 0.074 minDCF on the ASVspoof 5 benchmark with a single detector, surpassing even large ensemble systems.", "date": "2026-06-09", "categories": []}
{"id": "2606.10972", "title": "Optimizing 2D Input Representations and Sub-phase Fusion Strategies for Differential Diagnosis of Asthma and COPD Using CNN- and GRU-Based Networks", "abstract": "This study aims to explore the performance of the VAR model in comparison with mel-frequency cepstral coefficient (MFCC) matrices and log-mel spectrograms using deep learning. In pulmonary sound classification, spectrogram-based representations suffer from inconsistent temporal dimensions due to varying respiratory cycle durations. Along with traditional trimming/zero-padding, adaptive-length windowing was presented to fix their temporal dimensions. Their spectral and temporal dimensions were optimized by testing a range of parameters. Different convolutional neural network (CNN) architectures were employed to extract features from the two-dimensional representations obtained over the sub-phases. The extracted sub-phase features were then fused using various strategies including direct concatenation, gated recurrent unit (GRU) network and GRU with attention mechanism. Model performances were assessed through respiratory cycle-based evaluation and subject-based evaluation comprising multiple respiratory cycles. Several data augmentation techniques were also studied to cope with limitations in data size. The best cycle-based F1-score (0.877) was obtained using the MFCC matrices with thirteen coefficients and 64-point time resolution per sub-phase representation followed by direct feature concatenation, and the best subject-based F1-score (0.855) was obtained using the MFCC matrices with thirteen coefficients and 256-point time resolution per full-cycle representation, both obtained by adaptive-length windowing. Augmentation degraded the performance of models overall, yet mixup augmentation was the best among the methods tested. MFCC outperformed log-mel spectrogram and VAR model in differentiation of asthma and COPD. Sophisticated fusion strategies did not improve the diagnosis. Augmentation did not contribute, demonstrating the significance of authentic data in pulmonary sound studies.", "date": "2026-06-09", "categories": []}
{"id": "2502.16584", "title": "Audio-FLAN: An Instruction-Following Dataset for Unified Audio Understanding and Generation of Speech, Music, and Sound", "abstract": "Recent advancements in audio tokenization have significantly enhanced the integration of audio capabilities into large language models (LLMs). However, audio understanding and generation are often treated as distinct tasks, hindering the development of truly unified audio-language models. While instruction tuning has demonstrated remarkable success in improving generalization and zero-shot learning across text and vision, its application to audio remains largely unexplored. A major obstacle is the lack of comprehensive datasets that unify audio understanding and generation. To address this, we introduce Audio-FLAN, a large-scale instruction-tuning dataset covering 80 diverse tasks across speech, music, and sound domains, with over 100 million instances. Audio-FLAN lays the foundation for unified audio-language models that can seamlessly handle both understanding (e.g., transcription, comprehension) and generation (e.g., speech, music, sound) tasks across a wide range of audio domains in a zero-shot manner. The Audio-FLAN dataset is available on HuggingFace and GitHub.", "date": "2025-02-23", "categories": []}
{"id": "2511.18421", "title": "DHAuDS: A Dynamic and Heterogeneous Audio Benchmark for Test-Time Adaptation", "abstract": "Existing Test-time Adaptation (TTA) studies rely heavily on static and homogeneous corruption protocols, such as ImageNet-C and CIFAR-10-C/100-C, leading to inconsistent evaluation settings and potentially inflated robustness estimates that are compared with real-world situations. TTA lacks a standardized evaluation infrastructure capable of modeling realistic heterogeneous acoustic degradation. We introduce DHAuDS, a standardized benchmark suite for evaluating audio classification TTA robustness under dynamic corruption severity and heterogeneous noise mixtures. Rather than proposing a new TTA algorithm, DHAuDS focuses on exposing robustness limitations that remain hidden under conventional fixed-noise evaluation protocols.", "date": "2025-11-23", "categories": []}
{"id": "2512.20978", "title": "GenTSE: Enhancing Target Speaker Extraction via a Coarse-to-Fine Generative Language Model", "abstract": "Language Model (LM)-based generative modeling has emerged as a promising direction for TSE, offering potential for improved generalization and high-fidelity speech. We propose GenTSE, a two-stage decoder-only generative LM for TSE: Stage-1 predicts coarse semantic tokens, and Stage-2 generates fine acoustic tokens. Separating semantics and acoustics stabilizes decoding and yields more accurate target speech. Both stages use continuous SSL or codec embeddings, offering richer context than discretized-prompt methods. To reduce exposure bias, we employ a Frozen-LM Conditioning training strategy that conditions the LMs on predicted tokens from earlier checkpoints to reduce the gap between teacher-forcing training and autoregressive inference. We further apply DPO to better align outputs with perceptual preferences. Experiments on Libri2Mix show that GenTSE surpasses previous LM-based systems in speech quality, intelligibility, and speaker consistency.", "date": "2025-12-24", "categories": []}
{"id": "2602.20967", "title": "Training-Free Intelligibility-Guided Observation Addition for Noisy ASR", "abstract": "Automatic speech recognition (ASR) degrades severely in noisy environments. Although speech enhancement (SE) front-ends effectively suppress background noise, they often introduce artifacts that harm recognition. Observation addition (OA) addressed this issue by fusing noisy and SE enhanced speech, improving recognition without modifying the parameters of the SE or ASR models. This paper proposes an intelligibility-guided OA method, where fusion weights are derived from intelligibility estimates obtained directly from the backend ASR. Unlike prior OA methods based on trained neural predictors, the proposed method is training-free, reducing complexity and enhances generalization. Extensive experiments across diverse SE-ASR combinations and datasets demonstrate strong robustness and improvements over existing OA baselines. Additional analyses of intelligibility-guided switching-based alternatives and frame versus utterance-level OA further validate the proposed design.", "date": "2026-02-24", "categories": []}
{"id": "2605.06582", "title": "PairAlign: A Framework for Sequence Tokenization via Self-Alignment with Applications to Audio Tokenization", "abstract": "Many operations on sensory data -- comparison, memory, retrieval, and reasoning -- are naturally expressed over discrete symbolic structures. In language this interface is given by tokens; in audio, it must be learned. Existing audio tokenizers rely on quantization, clustering, or codec reconstruction, assigning tokens locally, so sequence consistency, compactness, length control, termination, and edit similarity are rarely optimized directly.\n  We introduce PairAlign, a framework for compact audio tokenization through sequence-level self-alignment. PairAlign treats tokenization as conditional sequence generation: an encoder maps speech to a continuous condition, and an autoregressive decoder generates tokens from BOS, learning token identity, order, length, and EOS placement. Given two content-preserving views, each view's sequence is trained to be likely under the other's representation, while unrelated examples provide competing sequences. This gives a scalable surrogate for edit-distance preservation while discouraging many-to-one collapse.\n  PairAlign starts from VQ-style tokenization and refines it with EMA-teacher targets, cross-paired teacher forcing, prefix corruption, likelihood contrast, and length control.\n  On 3-second speech, PairAlign learns compact, non-degenerate sequences with broad vocabulary usage and strong cross-view consistency. On retrieval tests, it preserves edit-distance search while reducing archive token count by 55%. A continuous-sweep probe shows lower local overlap than a dense geometric tokenizer, but stronger length control and bounded edit trajectories under 100 ms shifts. PairAlign is a sequence-symbolic predictive learner: like JEPA-style objectives, it predicts an abstract target from another view as a learned variable-length symbolic sequence, not a continuous latent.", "date": "2026-05-07", "categories": []}
{"id": "2606.02341", "title": "Parameter-efficient Dual-encoder Architecture with Differentiable Choquet Integral Fusion for Underwater Acoustic Classification", "abstract": "Underwater acoustic classification has a wide array of oceanic applications, but faces challenges due to an increasingly complex acoustic environment. Waveform and spectrogram representations have been primarily used as acoustic data features for classification tasks in this domain. Spectrograms model harmonic dependencies, but these reduced representations can filter out acoustic features relevant for discrimination. While phase information from the waveform allows full characterization of the signal, the original waveform can be noisy and complex, rendering this representation difficult for models to process directly. This paper proposes a dual-encoder neural architecture to simultaneously process acoustic waveforms and spectrograms, leveraging pre-trained backbones and parameter-efficient fine-tuning modules, enabling a domain adaptation. To combine these adapted branches, a novel differentiable fuzzy aggregation mechanism based on the Choquet integral is introduced to balance the temporal and spectral representations. This fusion strategy not only yields higher classification accuracy but also provides interpretability. Specifically, by analyzing the learned fuzzy measures, insights are revealed about class-specific shifts in the network's representation reliance. By dynamically shifting attention to the representation least corrupted by potential asymmetric channel distortions, the proposed gating mechanism mitigates the non-stationary challenges of the underwater environment. Evaluations on the DeepShip and ShipsEar datasets demonstrate that the proposed architecture achieves classification improvements over independent single-encoder baselines, while simultaneously restricting the trainable parameter space. This mitigates the risk of overfitting on limited acoustic datasets while alleviating the computational costs associated with fully fine-tuning foundation models.", "date": "2026-06-01", "categories": []}
{"id": "2606.08194", "title": "GlobeAudio: A Multilingual Multicultural Benchmark for Naturalistic Evaluation of Large Audio-Language Models", "abstract": "Large Audio-Language Models (LALMs) integrate audio perception and language understanding within a unified framework, enabling a wide range of real-world applications. Despite recent advances, evaluation for LALMs remains heavily underspecified relative to real-world requirements: most lack true linguistic and cultural authenticity, while others fail to capture acoustic realism. To bridge this gap, we propose GlobeAudio, a multilingual and multicultural benchmark designed to evaluate naturalistic audio understanding. GlobeAudio consists of 5,637 multiple-choice questions across six typologically diverse languages, expertly crafted by native speakers grounded on naturally occurring audio. In order to do well, models must possess higher-level auditory reasoning skills and culturally grounded interpretation. We systematically evaluate representative closed-source and open-source LALMs, as well as cascaded ASR-LLM pipelines. Our experiments reveal substantial performance gaps under natural acoustic conditions, particularly for open-source models and low-resource languages. These findings highlight critical limitations of current LALMs and underscore the importance of naturalistic audio evaluation for future audio-language systems. GlobeAudio can be found at https://huggingface.co/datasets/iNLP-Lab/GlobeAudio .", "date": "2026-06-06", "categories": []}
{"id": "2606.08210", "title": "Paediatric-HGNN: A Hybrid Heterogeneous Graph Neural Network for Detecting Disfluency in Children's Speech via Multiscale Acoustic Fusion", "abstract": "Automated stuttering detection (ASD) systems struggle with paediatric speech due to high acoustic variability in developing voices and the subtle distinction between pathological stuttering and typical developmental disfluencies. We introduce Paediatric-HGNN, a framework using a Context-aware Part-whole Interaction Network (CaPIN) tailored for paediatric data. Instead of conventional 1D signal modelling, our approach builds a heterogeneous graph capturing hierarchical relationships between lexical units (word nodes) and fine-grained acoustic segments (frame nodes). Trained on curated paediatric corpora (UCLASS and FluencyBank), Paediatric-HGNN achieves 82.4% weighted accuracy and a Typical Disfluency F1-score of 0.386. Modelling hierarchical lexical-acoustic interactions captures developmental \"searching\" behaviour, offering a more robust and interpretable tool for early clinical intervention.", "date": "2026-06-06", "categories": []}
{"id": "2606.08425", "title": "TinyGiantALM: A Compact Audio-Language Model for Intent-Aware Reasoning under Resource Constraints", "abstract": "Current advancements in Audio Reasoning rely on massive Large Audio-Language Models (LALMs), hindering deployment in resource-constrained environments. We introduce TinyGiantALM, a compact 1.5B efficiency-oriented alternative. Instead of brute-force scaling, we propose an Instruction-Aware Feature Refinement framework using a Query-guided Projector and Semantic Gating to filter acoustic signals based on user intent. On the MMAR benchmark, TinyGiantALM achieves 46.4% zero-shot accuracy, significantly outperforming 7B-13B baselines. While a reasoning gap in logical narrative remains versus 30B+ models and certain trade-offs exist in overly dense or spatial scenes, our approach notably surpasses models up to 8x larger in disentangling mixed-modality environments. These findings demonstrate that architectural precision offers a tangible pathway to secure robust perception capabilities on edge-friendly scales.", "date": "2026-06-07", "categories": []}
{"id": "2606.08669", "title": "A Comparison of SSL-Based Feature Extractors and Back-End Classifiers for Spoofing Detection: A Multi-Corpus Training and Cross-Linguistic Analysis", "abstract": "Voice biometric systems face growing threats from spoofing attacks, yet the evaluation of detection models remains inconsistent across datasets. To investigate these unpredictable fluctuations, we conduct a comprehensive benchmark of four self-supervised learning feature extractors paired with four back-end classifiers. We compare the hierarchical local feature extraction of ResNet with the global sequence and relational modeling of attention and graph-based back-ends. Through multi-corpus training across three scenarios and six evaluation datasets, our empirical analysis yields two critical findings. First, we expose a domain bias within the ASVspoof 5 dataset, showing that naive data scaling actively degrades performance. Second, our cross-linguistic analysis reveals that fine-tuning with just 8 hours of target-language data enhances detection robustness. Together, these findings emphasize the critical need for domain-aware and language-specific adaptation in spoofing detection.", "date": "2026-06-07", "categories": []}
{"id": "2603.08683", "title": "Benchmarking Language Modeling for Lossless Compression of Full-Fidelity Audio", "abstract": "Autoregressive \"language\" models (LMs) trained on raw waveforms can be repurposed for lossless audio compression, but prior work is limited to 8-bit audio, leaving open whether such approaches work for practical settings (16/24-bit) and can compete with existing codecs. We benchmark LM-based compression on full-fidelity audio across diverse domains (music, speech, bioacoustics), sampling rates (16kHz-48kHz), and bit depths (8, 16, 24-bit). Standard sample-level tokenization becomes intractable at higher bit depths due to vocabulary size (65K for 16-bit; 16.7M for 24-bit). We propose Trilobyte, a byte-level tokenization schema for full resolution audio, improving vocabulary scaling from $O(2^{b})$ to $O(1)$ and enabling the first tractable 24-bit LM-based lossless compression. While LMs consistently outperform FLAC and yield state-of-the-art compression at 8-bit and 16-bit, we observe that compression gains become more modest as bit depth increases beyond 8-bit.", "date": "2026-03-09", "categories": []}
{"id": "2606.01802", "title": "MOSS-Audio Technical Report", "abstract": "MOSS-Audio is a unified audio-language model for speech, environmental sound, and music understanding, supporting audio captioning, time-aware question answering, timestamped transcription, and audio-grounded reasoning. MOSS-Audio couples a dedicated audio encoder with a modality adapter and a large language model: the encoder produces 12.5 Hz temporal representations, the adapter projects them into the decoder space, and the decoder generates autoregressive text outputs. Two design choices are central to the system: DeepStack cross-layer feature injection, which exposes the decoder to acoustic information from multiple encoder depths, and time markers, which provide explicit temporal cues by inserting timestamp markers into the audio-token stream. At the data level, we design an event-preserving audio annotation pipeline that segments raw audio at coherent event boundaries, applies branch-specific annotation to speech, music, and general audio, and merges the results into unified captions for pretraining. The intermediate branch-specific captions are further retained to support the construction of task-oriented SFT data. The model is pretrained on large-scale audio-language data, with time-aware objectives incorporated to support temporal grounding, and then undergoes multi-stage post-training to enhance instruction following and audio-grounded reasoning. We release 4B and 8B variants in both Instruct and Thinking configurations. MOSS-Audio achieves strong performance across general audio understanding, speech captioning, ASR, and timestamped ASR, positioning it as a promising understanding foundation for future voice agents.", "date": "2026-06-01", "categories": []}
{"id": "2606.06615", "title": "FIGMA: Towards FIne-Grained Music retrievAl", "abstract": "Retrieving music using natural language descriptions has improved with contrastive audio-text models such as CLAP, but current systems remain limited to coarse semantic queries. When descriptions specify fine-grained musical attributes such as tempo, key, chord progression, or rhythmic structure, existing models often fail to retrieve the correct audio. We show that this limitation stems from the contrastive learning objective itself: despite being trained on long captions, CLAP-based models effectively utilize only the first few tokens, discarding much of the information encoded in detailed prompts. Then, we propose FIGMA (FIne-Grained Music RetrievAl), a multi-view contrastive architecture that addresses this limitation by jointly optimizing global audio-text alignment and frame-level, token-wise alignment. This design enables FIGMA to capture both high-level semantic context and fine-grained musical attributes within a unified representation space. Moreover, we formalize the task of Fine-Grained Music Retrieval and construct Fine-Grained Music Caption dataset (FGMCaps), a large-scale dataset of 380K music-caption pairs for training along with a 10K test set, both annotated with tempo, key, chord progression, beat count, as well as genre and mood. Extensive experiments demonstrate that FIGMA consistently outperforms existing CLAP-based music retrieval models across multiple music retrieval benchmarks, including out-of-domain evaluations, with relative improvements of up to 73.3%.", "date": "2026-06-04", "categories": []}
{"id": "2606.06743", "title": "HybridCodec: Fast Dual-Stream, Semantically Enhanced Neural Audio Codec", "abstract": "The popularity of neural audio codecs as speech tokenizers has surged with the advent of Multimodal Large Language Models. New codec architectures with semantic and acoustic disentanglement have emerged. There are two main approaches to introduce semantic information into codec models: one distills semantic information from SSL representations into the first RVQ layer, while the other maintains separate streams for semantic and acoustic features. We propose HybridCodec, a unified architecture that combines both paradigms. It employs separate semantic and acoustic branches while distilling SSL representations into the semantic stream. This design ensures strong disentanglement without requiring an SSL model during inference. HybridCodec shows superior semantic specialization (RVQ-1) on in-domain test set and competitive reconstruction (RVQ-all). We demonstrate its robustness in out-of-domain and zero-shot cross-lingual settings, achieving a 3x speedup over existing dual-stream models.", "date": "2026-06-04", "categories": []}
{"id": "2606.06907", "title": "SpectCount: Spectrotemporal Counting via Synthetic Signals Improves Large Audio Language Models", "abstract": "Large audio language models (LALMs) extend large language models with an audio encoder and large-scale audio data. However, the scarcity of high-quality annotated audio data remains a fundamental bottleneck for scaling. Through probing signal detectability analysis, we identify fine-grained spectrotemporal perceptual weaknesses in a foundation LALM. To address these challenges, we propose Spectrotemporal Counting (SpectCount), a data-efficient fine-tuning approach based on fully synthetic audio signals generated on-the-fly, without relying on real-world audio, annotations, or pretrained generative models. SpectCount not only resolves the observed weaknesses but also improves performance on diverse auditory benchmarks spanning sound, music, and speech, unseen during fine-tuning. These results suggest that weakness-targeted synthetic signals provide a data-efficient path toward enhanced auditory understanding capabilities in LALMs.", "date": "2026-06-05", "categories": []}
{"id": "2606.07033", "title": "Hierarchical Semantic-Constrained Heterogeneous Graph for Audio-Visual Event Localization", "abstract": "Open-vocabulary audio-visual event localization (OV-AVEL) jointly models audio-visual cues to recognize and temporally localize events, including categories unseen during training. Existing methods primarily learn joint audio-visual representations in Euclidean space, but still face two significant challenges. First, the lack of supervision signals for unseen categories makes it difficult to maintain audio-visual consistency across multiple temporal scales. Second, the lack of hierarchical constraints between segment- and video-level semantics prevents the model from establishing semantic consistency across different levels. To address these challenges, we propose a hierarchical semantic constrained heterogeneous graph (HSCHG) for audio-visual event localization framework. We first construct a heterogeneous hierarchical graph in Euclidean space, which includes audio and visual segment nodes and their corresponding video-level nodes. We use multi-directional temporal edges to capture complete temporal information within each modality. Simultaneously, we employ a dual-threshold filtering gated fusion strategy, introducing cross-modal information only when the alignment confidence is high. Furthermore, we introduce bidirectional semantic constraints between segment- and video-level representations to achieve semantic consistency across different levels. Based on this, we map the multi-level audio-visual representations and text prototypes uniformly into hyperbolic space. We use a hierarchical entailment regularization loss to characterize the hierarchical relationships between videos and segments. Extensive experimental results show that our method outperforms existing methods on the OV-AVEL benchmark. Ablation studies further validate the effectiveness of our method.", "date": "2026-06-05", "categories": []}
{"id": "2606.07229", "title": "MMAE: A Massive Multitask Audio Editing Benchmark", "abstract": "We introduce MMAE, a Massive Multitask Audio Editing benchmark, serving as the first comprehensive evaluation testbed designed for general-purpose instruction-based audio editing. Spurred by the shift toward intelligent creation, interactive editing has rapidly expanded from visual domains, pioneered by models like Nano-banana 2 for images and Gemini-Omni for video, into audio. However, the current evaluation infrastructure lags severely, remaining highly fragmented and restricted to specific subdomains or basic operations. Unlike existing benchmarks that are limited in scope, MMAE extends to a broad spectrum of real-world scenarios, encompassing 7 distinct audio modalities, including sound, speech, music, and their mixtures. Furthermore, we establish a comprehensive taxonomy spanning 6 levels of task complexity, from basic modifications to multi-hop reasoning and multi-round editing, 2 levels of granularity, and 8 distinct operation types. Meticulously curated through human-agent collaboration, MMAE comprises 2,000 high-fidelity samples paired with a pioneering rubric-based evaluation framework. By decomposing free-form tasks into 17,741 verifiable criteria, this robust rubric-based paradigm enables a precise, multi-dimensional assessment of both instruction following and context consistency. Our extensive evaluation of leading models reveals that current systems remain far from achieving reliable edits. Strikingly, the Exact Match Rate (EMR) consistently falls below 5% and plummets to an absolute 0% in complex, mixed-modality tasks, exposing critical bottlenecks in precise execution and structural robustness. We hope MMAE will serve as a catalyst for future advances in the intelligent creation community, providing a clear diagnostic roadmap and establishing a standardized, long-lasting evaluation paradigm for next-generation audio editing systems.", "date": "2026-06-05", "categories": []}
{"id": "2606.05857", "title": "Forgive or forget: Understanding the context of hate in audio retrieval systems", "abstract": "Handling toxic retrieval in text-to-audio systems is challenging due to contextual dependencies. Existing strategies (e.g., rephrasing, summarization) risk altering intent or omitting details. We propose a post hoc causal debiasing framework with a sentiment-controlled mediator to preserve semantic relevance while suppressing harmful speech. Our approach is model-agnostic and integrates seamlessly with existing retrieval pipelines. We introduce two variants: Forgive, which re-ranks and filters toxic audio via logit adjustment, and Forget, which generates counterfactual toxic prompts to mitigate harmful retrievals. Experiments show consistent toxicity reduction with minimal loss in retrieval accuracy, improving both safety and reliability.", "date": "2026-06-04", "categories": []}
{"id": "2606.06168", "title": "ProSarc: Prosody-Aware Sarcasm Recognition Framework via Temporal Prosodic Incongruity", "abstract": "We present ProSarc, an audio-only framework that detects sarcasm by modelling temporal prosodic incongruity, that is, the mismatch between local prosodic dynamics and the utterance-level emotional baseline. Dual encoding paths, a Global Emotion Encoder and a Temporal Prosody Encoder (BiLSTM + multi-head attention), feed a Prosodic Incongruity Analyzer that produces a scalar incongruity score for classification. Monte Carlo dropout provides uncertainty estimates, and an attention-based mechanism localises sarcastic onset without frame-level labels. ProSarc outperforms prior audio-only methods on MUStARD++ (F1=75.3) and generalises to spontaneous (PodSarc, F1=62.9) and cross-lingual speech (MuSaG, F1=65.6). Ten-run validation confirms the contribution of incongruity modelling (Wilcoxon p=0.002, Cohen's d=1.51). Human evaluation shows that model uncertainty tracks perceptual ambiguity and predicted onsets align with human-annotated temporal windows.", "date": "2026-06-04", "categories": []}
{"id": "2606.06211", "title": "FiLM-Based Speaker Conditioning of a SpeechLLM for Pathological Speech Recognition", "abstract": "Automatic speech recognition (ASR) has advanced remarkably for standard speech; however, pathological speech from neurological conditions remains a significant challenge. We investigate speaker conditioning via Feature-wise Linear Modulation (FiLM), injecting x-vector-derived information into each transformer layer of a frozen ASR encoder to adapt internal representations to individual pathological speakers without modifying base model weights. We benchmark this for the ASR task against standard and parameter-efficient fine-tuning baselines, complemented by post-processing, on Spanish and English pathological speech. Additionally, we evaluate if the adapted model preserves the ability to answer speech-related questions. Results show that speaker-conditioned ASR is competitive with established adaptation strategies while retaining performance on non-conditioned speech.", "date": "2026-06-04", "categories": []}
{"id": "2606.06357", "title": "F3-Tokenizer: Taming Audio Autoencoder Latents for Understanding and Generation", "abstract": "Continuous audio autoencoders reconstruct waveforms well but often produce latents with weak structure for understanding, while self-supervised audio encoders capture semantics but are not directly decodable. This mismatch complicates a single audio tokenizer that must support both understanding and generation. We adapt continuous autoencoder latents to this setting with two components: a noise-regularized autoencoder bottleneck and a latent-side representation encoder. The bottleneck uses channel normalization and stochastic perturbation instead of KL-based variational training, yielding scale-controlled continuous latents for reconstruction and autoregressive generation. The representation encoder is trained on frozen autoencoder latents with RQ-MTP and frozen-LLM supervision. The resulting tokenizer provides high-dimensional representations for understanding while preserving normalized continuous latents as generation targets", "date": "2026-06-04", "categories": []}
{"id": "2606.06444", "title": "USAD 2.0: Scaling Representation Distillation for Universal Audio Understanding", "abstract": "Audio encoders are critical to modern audio applications as large language models (LLMs) increasingly rely on a single encoder for diverse inputs. While self-supervised learning (SSL) has yielded strong domain-specific encoders like speech or music experts, multi-domain approaches like USAD and SPEAR remain limited in coverage and evaluation. Recent studies also suggest supervised encoders align better with audio LLMs. We present USAD 2.0, a universal encoder integrating knowledge from both SSL and supervised foundation models. USAD 2.0 introduces domain-aware distillation to address teacher mismatch, extends coverage to the music domain, and adds second-stage supervised distillation for downstream use. We further scale the model to one billion parameters via depth scaling. Experiments show USAD 2.0 achieves strong or state-of-the-art performance across probing and LLM-based evaluations.", "date": "2026-06-04", "categories": []}
{"id": "2509.21597", "title": "AUDDT: A Unified Benchmark Toolkit for Audio and Speech Deepfake Detectors", "abstract": "With the prevalence of artificial intelligence (AI)-generated content, such as audio deepfakes, a large body of recent work has focused on developing deepfake detection techniques. However, existing benchmarks employ a narrow set of datasets, leaving detector generalization to real-world conditions uncertain. In this paper, we systematically review 31 existing audio deepfake datasets and present an open-source benchmarking toolkit called AUDDT (https://github.com/MuSAELab/AUDDT). The goal of this toolkit is to automate the evaluation of pretrained detectors across a wide range of speech and non-speech audio datasets, giving users direct feedback on the advantages and shortcomings of their deepfake detectors under diverse manipulation types and recording conditions. We start by showcasing the usage of the developed toolkit, the composition of our benchmark, and the breakdown of different deepfake subgroups. Next, we highlight how AUDDT differs from existing benchmarking efforts by enabling large-scale, diverse evaluation across modern spoofing methods and richer attribute-level analysis through comprehensive metadata annotation. Using a widely adopted pretrained deepfake detector, we present in- and out-of-domain detection results, revealing notable performance variability across different conditions and audio manipulation types. Lastly, we also analyze the limitations of these existing datasets and their gaps relative to practical deployment scenarios.", "date": "2025-09-25", "categories": []}
{"id": "2603.09391", "title": "Physics-Informed Neural Engine Sound Modeling with Differentiable Pulse-Train Synthesis", "abstract": "Engine sounds originate from sequential exhaust pressure pulses rather than sustained harmonic oscillations. While neural synthesis methods typically aim to approximate the resulting spectral characteristics, we propose directly modeling the underlying pulse shapes and temporal structure. We present the Pulse-Train-Resonator (PTR) model, a differentiable synthesis architecture that generates engine audio as parameterized pulse trains aligned to engine firing patterns and propagates them through recursive Karplus-Strong resonators simulating exhaust acoustics. The architecture integrates physics-informed inductive biases including harmonic decay, thermodynamic pitch modulation, valve-dynamics envelopes, exhaust system resonances and derived engine operating modes such as throttle operation and Deceleration Fuel Cutoff (DFCO). Validated on three diverse engine types totaling 7.5 hours of audio, PTR achieves a 21% improvement in harmonic reconstruction and a 5.7% reduction in total loss over a harmonic-plus-noise baseline model, while providing interpretable parameters corresponding to physical phenomena. Complete code, model weights, and audio examples are openly available.", "date": "2026-03-10", "categories": []}
{"id": "2605.13672", "title": "SpurAudio: A Benchmark for Studying Shortcut Learning in Few-Shot Audio Classification", "abstract": "Few-shot classification (FSC) is widely used for learning from limited labeled data, yet most evaluations implicitly assume that target concepts are independent of contextual cues. In real-world settings, however, examples often appear within rich contexts, allowing models to exploit spurious correlations between foreground content and background signals. While such effects have been studied in few-shot image classification, their role in few-shot audio classification remains largely unexplored, and existing audio benchmarks offer limited control over contextual structure. We introduce SpurAudio, a benchmark that leverages the natural separability of foreground events and background environments in audio to enable controlled, multi-level evaluation of contextual shifts across support and query sets. Using this benchmark, we show that many state-of-the-art few-shot methods suffer severe performance degradation when background correlations are disrupted, despite achieving similar accuracy under standard evaluation protocols. Crucially, this vulnerability persists even in large pretrained audio foundation models, ruling out limited backbone capacity as an explanation. Moreover, methods that appear comparable under conventional benchmarks can exhibit markedly different sensitivity to spurious correlations, revealing systematic algorithmic strengths and vulnerabilities tied to how feature representations interact with classifier heads at inference time. These findings provide new insight into the behavior of few-shot methods in audio and highlight the need for benchmarks that explicitly probe context dependence when evaluating FSC models.", "date": "2026-05-13", "categories": []}
{"id": "2605.30457", "title": "Extracting accent features in spoken Brazilian Portuguese without sociolinguistic labels", "abstract": "Regional accent classification in Brazilian Portuguese (pt-BR) suffers from the need for reliable labeling. While large self-supervised learning (SSL) speech models are powerful, their training pipelines dilute sociophonetic information, since accent labels are generally not reliable or are not used in training objectives. This work introduces a novel workflow for feature extraction using only acoustic labels. By isolating explicit regional accent landmarks and using a phoneme-based forced aligner (ZIPA), our targeted feature set captures dialectal variance more effectively than utterance embeddings, demonstrating that localized features can outperform general-purpose architectures on accent-related tasks using minimal and objective data labels.", "date": "2026-05-28", "categories": []}
{"id": "2606.04103", "title": "The Differentiable Auditory Loop (DAL): An ML Framework for Hyper-Personalized Hearing Aids", "abstract": "Conventional hearing aids rely on fixed, frequency-dependent amplification and compression to manage reduced sensitivity, which often fails to provide sufficient listening support in complex environments, such as situations with multiple speakers (the ``cocktail party'' problem). To more comprehensively address the underlying encoding dysfunctions of hearing loss, we introduce the Differentiable Auditory Loop (DAL), a new open-source framework for personalized hearing aid design and fitting. Our first implementation of DAL incorporates CARFAC, a differentiable model of human cochlear function, which we ported to JAX, to optimize a deep neural network to match impaired auditory neural activity patterns with a normal-hearing reference. To build a hearing aid with the fine-grained spectro-temporal signal processing required, we adopt SEANet, a waveform-to-waveform fully convolutional UNet generator. We fine-tune the network by comparing the outputs of a CARFAC model fitted to normal hearing with that of a CARFAC model fitted to match each subject's individual hearing impairment. The comparison is done using loss functions derived from the respective CARFAC neural activity pattern (NAP) outputs and stabilized auditory images (SAIs), the latter providing a 2D representation that captures phase-insensitive temporal structure in the auditory nerve output. Through gradient descent, the SEANet model learns to both denoise the input and compensate for the hearing loss modelled by the impaired CARFAC model. Across neural-representation and signal-fidelity metrics, the DAL-optimized SEANet model outperformed the tested master hearing aid (MHA) baselines. The DAL framework provides a practical path toward model-based, machine-learning-driven personalization of hearing aid signal processing. Next steps include hardware deployment to enable real-world clinical testing.", "date": "2026-06-02", "categories": []}
{"id": "2606.04210", "title": "Representation Matters in Randomized Smoothing for Audio Classification", "abstract": "Randomized smoothing (RS) certifies robustness in the vector space where Gaussian noise is added. In audio classification, this space is often not uniquely defined as standard pipelines normalize, range-control, and transform waveforms into log-mel or other spectral features. We show that direct RS is therefore under-specified unless the certified object and preprocessing policy are explicit. On two audio benchmarks, keyword spotting and environmental-sound classification, we study waveform, feature-space, and post-processed smoothing. Our diagnostics show why representation-aware reporting is necessary: at the same smoothing level $\\sigma=0.0025$, the two datasets share the same median raw radius $.007996$, but different waveform energies yield different SNR-equivalent scales ($83.98$ vs. $90.97$ dB); log-mel smoothing gives higher positive-radius certified accuracy on environmental sounds ($68.42\\%$ vs. $65.53\\%$), certifying more examples with nonzero radius but over features rather than waveforms; and clipping or peak normalization changes the effective perturbation norm by roughly $230$--$351\\times$. We therefore recommend that audio RS studies choose and report the task-specific certified object and perturbation model, including the perturbation location, gain policy, raw radius, and any post-noise geometry changes.", "date": "2026-06-02", "categories": []}
{"id": "2606.04418", "title": "CleanCodec: Efficient and Robust Speech Tokenization via Perceptually Guided Encoding", "abstract": "Neural audio codecs are a key component of speech processing pipelines, compressing audio into discrete tokens for downstream modeling. However, existing codecs struggle to balance reconstruction quality with token efficiency, often encoding perceptually irrelevant information such as background noise and recording artifacts at the expense of linguistically and acoustically meaningful content. We reframe audio tokenization as a selective information bottleneck problem and propose CleanCodec, a denoising audio codec which learns to encode only perceptually important features and discard imperceptible information. At just 12.5 tokens per second, CleanCodec achieves state-of-the-art tokenization efficiency, substantially outperforming existing codecs in speaker similarity and speech intelligibility. Evaluations on downstream text-to-speech and voice conversion tasks further demonstrate improved performance and up to 17x faster inference, highlighting significant efficiency gains.", "date": "2026-06-03", "categories": []}
{"id": "2606.04844", "title": "Drift-Augmented Scoring: Text-Derived Noise Robustness for Zero-Shot Audio-Language Classification", "abstract": "Contrastive audio-language models such as CLAP enable zero-shot audio classification: a sound is labelled by matching its embedding to text prompt embeddings, with no labelled audio. This matching breaks down under acoustic noise, where accuracy and mAP fall by 12-30 percentage points at 0 dB SNR on standard benchmarks. We propose Drift Augmented Scoring (DAS), a small per-class bonus added to the cosine score. The bonus rewards a class when the noisy audio embedding drifts in the direction that the class's noise-conditioned text prompts predict. It is derived from text alone, computed once and cached, and adds a single inner product per class at inference, with no gradients and no test-time batch. On a LAION CLAP backbone, we compare DAS against the four variants of Acevedo et al.'s concurrent method on UrbanSound8K and the full FSD50K eval set, mixing each clip with urban acoustic scene noise across a range of SNRs. DAS improves the metric on every test condition: by +2.60 to +5.75 accuracy points on UrbanSound8K and +1.50 to +1.74 mAP points on FSD50K.", "date": "2026-06-03", "categories": []}
{"id": "2606.05101", "title": "FoeGlass: Simple In-Context Learning Is Enough for Red Teaming Audio Deepfake Detectors", "abstract": "Audio deepfake detection (ADD) models are critical for countering the malicious use of text-to-speech (TTS) models. Evaluating and strengthening ADD models requires developing datasets that span the space of generated audio and highlight high-error regions. Existing dataset development strategies face two challenges: (i) manual collection, and (ii) inefficient discovery of blind spots in the ADD models. To address these challenges, we propose FoeGlass, the first black-box automated red-teaming method for ADDs, which effectively discovers ADD failure modes in the space of generated audio underexplored by state-of-the-art deepfake benchmarks. FoeGlass uses the in-context learning capabilities of an LLM to explore the input space of a TTS model, generating audio samples that fool the target ADD using only black-box access to all components. By using a carefully designed context based on diversity measurements, FoeGlass mitigates the common problem of mode collapse in automated red-teaming systems. Empirical evaluations on several open-source ADD and TTS models demonstrate that data generated from FoeGlass substantially improves the false negative rates over unconditional sampling baselines and recent spoofing datasets by up to 94%, while requiring no manual supervision. Furthermore, we show that the attacks generated by FoeGlass are transferable across different target ADDs, demonstrating its broad applicability and ease of use for the automated red teaming of ADD systems. Finally, fine-tuning ADD models on FoeGlass-generated samples notably enhances the robustness of the detectors (up 41%).", "date": "2026-06-03", "categories": []}
{"id": "2606.05121", "title": "Audio Interaction Model", "abstract": "Audio is an inherently interactive modality, yet today's Large Audio Language Models (LALMs) are offline, and streaming audio models each handle only a single task such as streaming ASR or voice chatting. It is time to unify them into one online LALM: a model that, through an always-on perceive-decide-respond loop, listens to sound, environment, and instructions in real time and reacts on the fly. We formalize this regime as the Audio Interaction Model, and realize it with Audio-Interaction, a unified streaming model that retains offline task execution while adding online general audio instruction following, from dialogue to full voice chatting, deciding when to respond from the semantics of the stream. To enable this, we propose SoundFlow, a framework that instantiates the perceive-decide-respond loop end to end, from data to training to deployment, through streaming-native data construction, comprehension-aware training, and asynchronous low-latency inference for stable real-time interaction. We further construct StreamAudio-2M, a 2.6M-item streaming corpus spanning 7 fundamental abilities and 28 sub-tasks, and Proactive-Sound-Bench for evaluating proactive audio intervention. Across 8 benchmarks, Audio-Interaction preserves competitive performance on mainstream audio tasks while unlocking capabilities inaccessible to offline LALMs, including real-time ASR, streaming audio instruction following, and proactive help.", "date": "2026-06-03", "categories": []}
{"id": "2606.02739", "title": "EntangleCodec: A Unified Discrete Audio Tokenizer via Semantic-Acoustic Entanglement", "abstract": "Audio tokenizers serve as the discrete interface between continuous audio and Audio Language Models (ALMs), but existing tokenizers often struggle to support both understanding and generation. Reconstruction-oriented codecs preserve acoustic fidelity but lack rich semantics, while semantic-aware tokenizers typically rely on separate semantic and acoustic streams, introducing redundancy or misalignment.\n  We propose \\textbf{EntangleCodec}, a unified discrete audio tokenizer that learns caption-aligned semantic-acoustic representations before quantization. By aligning audio with rich captions rather than ASR transcripts, EntangleCodec captures linguistic content, speaker identity, emotion, prosody, and acoustic scenes within a compact token stream. A flow-matching diffusion decoder further enables high-quality reconstruction across speech, music, and general audio.\n  EntangleCodec achieves reconstruction quality competitive with specialized codecs, outperforms all codec-based baselines on audio understanding by up to \\textbf{+7.4\\%} on MMAR, and supports both TTS and TTA generation in a unified framework. Furthermore, EntangleCodec-based audio language models demonstrate strong scaling behavior: even at \\textit{0.6B} parameters, the model surpasses specialized continuous-representation LLMs with over \\textit{13B} parameters across three benchmarks using \\textbf{22$\\times$} fewer parameters; scaling to \\textit{8B} further establishes new state-of-the-art results on MMAR, highlighting that representation quality is as critical as model scale in audio language modeling. Code and model weights are available at https://github.com/luckyerr/EntangleCodec.", "date": "2026-06-01", "categories": []}
{"id": "2606.02998", "title": "CoughSense: Five-Class Respiratory Disease Classification via Whisper Encoder Fine-Tuning and Dual-Encoder Cross-Attention Fusion with Balanced Contrastive Learning", "abstract": "Automated cough analysis offers a path to low-cost respiratory screening, but most existing work stops at binary COVID-19 detection. A practical tool needs to tell apart several respiratory conditions from one cough recording on a consumer smartphone. We present CoughSense, a system that sorts cough recordings into five classes. These are healthy, COVID-19, asthma or respiratory condition, bronchitis, and pneumonia. We aggregated 18,301 recordings from four public datasets (Coswara, CoughVID, Virufy, and the West China Hospital Pediatric Cough Dataset) and used the OpenAI Whisper encoder as a pretrained backbone for cough disease classification. The main contribution is active-frame QKV attention pooling, which restricts attention to the first 200 of 1500 encoder tokens. This avoids the silence-dilution problem that arises because a 3-second cough fills only 150 tokens of Whisper's 30-second input window. Other training parts handle the 19 to 1 class imbalance and the four-dataset domain shift. These include WeightedRandomSampler, SpecAugment, Balanced Mixup with forced minority pairing, a supervised contrastive auxiliary loss, FiLM symptom conditioning, and gradient-reversal domain adaptation. A dual-encoder model fuses Whisper with the OPERA-CT respiratory foundation model through cross-attention. CoughSense (Whisper-tiny, 8.6M parameters) reached 82.3 percent balanced accuracy on five-fold cross-validation (macro-F1 of 0.817, AUC of 0.941). It beat an ImageNet-pretrained EfficientNet-B2 by 11.1 points and a ViT trained from scratch by 29.6 points. All five classes passed 74 percent recall and four of five passed 80 percent. The dual-encoder model reached 85.4 percent balanced accuracy. Active-frame pooling is the largest single contributor across all ablation components at 5.1 points, which should help any short-audio task using Whisper as a backbone.", "date": "2026-06-02", "categories": []}
{"id": "2606.03116", "title": "AnyAudio-Judge: A Dynamic Rubric-Based Benchmark and Evaluator for Audio Instruction Following", "abstract": "The rapid advancement of instruction-guided audio generation has highlighted the critical need for robust alignment evaluation. Current automated evaluation methods heavily rely on holistic scoring from general-purpose large language models, which struggle to decouple complex instructions, lack interpretability, and fail to capture fine-grained attribute mismatches. To address this, we introduce a novel dynamic rubric-based evaluation paradigm that adaptively decomposes complex audio captions into a variable number of independent, verifiable binary rubric items. To rigorously benchmark this capability, we propose the AnyAudio-Judge Bench, a comprehensive, bilingual benchmark comprising 7,920 meticulously curated samples across four diverse audio domains (speech, sound, music, and mixed), featuring deliberately constructed hard negatives. Furthermore, we construct a large-scale corpus of 105K samples with explicit Chain-of-Thought (CoT) rationales to train our dedicated evaluator, the AnyAudio-Judge model. By employing a training pipeline that combines Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO), our model successfully aligns its reasoning paths with the rubric-based scoring mechanism. Extensive experiments demonstrate that AnyAudio-Judge not only significantly enhances zero-shot alignment detection compared to state-of-the-art baselines, but also provides precise and interpretable reward signals that substantially improve instruction alignment in downstream reinforcement learning for audio generation.", "date": "2026-06-02", "categories": []}
{"id": "2412.03771", "title": "Embedding-Space Diffusion for Zero-Shot Environmental Sound Classification", "abstract": "Zero-shot learning enables models to generalise to unseen classes by leveraging semantic information, bridging the gap between training and testing sets with non-overlapping classes. While much research has focused on zero-shot learning in computer vision, the application of these methods to environmental audio remains underexplored, with poor performance in existing studies. Generative methods, which have demonstrated success in computer vision, are notably absent from zero-shot environmental sound classification studies.\n  To address this gap, this work investigates generative methods for zero-shot learning in environmental audio. Two successful generative models from computer vision are adapted: a cross-aligned and distribution-aligned variational autoencoder (CADA-VAE) and a leveraging invariant side generative adversarial network (LisGAN). Additionally, we introduced a novel diffusion model conditioned on class auxiliary data. Synthetic embeddings generated by the diffusion model are combined with seen class embeddings to train a classifier.\n  Experiments are conducted on five environmental audio datasets, ESC-50, ARCA23K-FSD, FSC22, UrbanSound8k and TAU Urban Acoustics 2019, and one music classification dataset, GTZAN. Results show that the diffusion model outperforms all baseline methods on average across six audio datasets.\n  This work establishes the diffusion model as a promising approach for zero-shot learning and introduces the first benchmark of generative methods for zero-shot environmental sound classification, providing a foundation for future research.", "date": "2024-12-04", "categories": []}
{"id": "2510.01891", "title": "HRTFformer: A Spatially-Aware Transformer for Individual HRTF Upsampling in Immersive Audio Rendering", "abstract": "Individual Head-Related Transfer Functions (HRTFs) are starting to be introduced in many commercial immersive audio applications and are crucial for realistic spatial audio rendering. However, one of the main hesitations regarding their introduction is that creating individual HRTFs is impractical at scale due to the complexities of the HRTF measurement process. To mitigate this drawback, HRTF spatial upsampling has been proposed with the aim of reducing the measurements required. While prior work has seen success with different machine learning (ML) approaches, these models often struggle with long-range preservation of local spatial variation patterns across neighbouring source directions and generalization at high upsampling factors. In this paper, we propose a novel transformer-based architecture for HRTF upsampling, leveraging the attention mechanism to better capture spatial correlations across the HRTF sphere. Working in the spherical harmonic (SH) domain, our model learns to reconstruct high-resolution HRTFs from sparse input measurements with significantly improved accuracy. To enhance spatial coherence, we introduce a neighbour dissimilarity loss that promotes magnitude smoothness, yielding more realistic upsampling. We evaluate our method using both perceptual localization models and objective spectral distortion metrics. Experiments show that our model outperforms existing methods across several evaluation metrics in generating realistic, high-fidelity HRTFs.", "date": "2025-10-02", "categories": []}
{"id": "2511.13487", "title": "Systematic Evaluation of Time-Frequency Features for Binaural Sound Source Localization", "abstract": "This study presents a systematic evaluation of time-frequency feature design for binaural sound source localization (SSL), focusing on how feature selection influences model performance across diverse conditions. We investigate the performance of a convolutional neural network (CNN) model using various combinations of amplitude-based features (magnitude spectrogram, interaural level difference - ILD) and phase-based features (phase spectrogram, interaural phase difference - IPD). Evaluations on in-domain and out-of-domain data with mismatched head-related transfer functions (HRTFs) reveal that carefully chosen feature combinations often outperform increases in model complexity. While two-feature sets such as ILD + IPD are sufficient for in-domain SSL, generalization to diverse content requires richer inputs combining channel spectrograms with both ILD and IPD. Using the optimal feature sets, our low-complexity CNN model achieves competitive performance. Our findings underscore the importance of feature design in binaural SSL and provide practical guidance for both domain-specific and general-purpose localization.", "date": "2025-11-17", "categories": []}
{"id": "2512.10120", "title": "VocSim: A Training-free Benchmark for Zero-shot Content Identity in Single-source Audio", "abstract": "General-purpose audio representations aim to map acoustically variable instances of the same event to nearby points, resolving content identity in a zero-shot setting. Unlike supervised classification benchmarks that measure adaptability via parameter updates, we introduce VocSim, a training-free benchmark probing the intrinsic geometric alignment of frozen embeddings, with no parameters updated and no labels used (a label-free PCA whitening is fit per subset to correct anisotropy). VocSim aggregates 125k single-source clips from 19 corpora spanning human speech, animal vocalizations, and environmental sounds, isolating content representation from source separation (polyphonic mixtures are out of scope). We evaluate embeddings with Precision@k for local purity and the Global Separation Rate (GSR) for point-wise class separation, calibrated by lift over an empirical permutation baseline. A simple pipeline of frozen Whisper features, time-frequency pooling, and label-free PCA yields strong zero-shot performance with stable GSR rankings across domains (Kendall's tau = 0.60). However, on blind low-resource speech (Shipibo-Conibo, Chintang), local retrieval collapses while remaining above chance, exposing a cross-lingual speech generalization gap. As external validation, our top embeddings predict avian perceptual similarity, improve bioacoustic classification, and achieve state-of-the-art on the HEAR benchmark. We release data, code, and a public leaderboard.", "date": "2025-12-10", "categories": []}
{"id": "2601.03615", "title": "SARA: Stress Test Reasoning in Audio Deepfake Detection", "abstract": "Audio Language Models (ALMs) offer a promising shift towards explainable audio deepfake detections (ADD), moving beyond \\textit{black-box} classifiers by providing transparency to their predictions via reasoning traces. However, such reasoning may not support the model predictions, reflecting poor coherence, or, worse, may rationalize incorrect predictions with plausible but misleading explanation. Moreover, the behavior of ALM reasoning under adversarial attacks remains under-explored, raising questions about the practical reliability of such explanation capabilities. To address this gap, this study introduces \\textbf{SARA} (\\textbf{S}hift \\textbf{A}nalysis of \\textbf{R}easoning in \\textbf{A}udio), a diagnostic framework that evaluates ALM reasoning across three dimensions: acoustic perception, reasoning-verdict coherence and dissonance. We test five open-source ALMs against both acoustic and linguistic adversarial attacks. We show that acoustic attacks significantly degrade reasoning-verdict coherence (average decrease of 14.20\\%), frequently inducing internal logical conflicts. Conversely, linguistic attacks achieve higher attack success rates while maintaining reasoning coherence. We further demonstrate that the textual coherence of generated reasoning traces also serves as a latent indicator of adversarial inputs, enabling effective detection of perturbed audio (0.78 in F1) \\textit{without accessing the raw acoustic signal}. These findings suggest that reasoning traces provide diagnostic utility that persists even when final classification outputs are compromised.", "date": "2026-01-07", "categories": []}
{"id": "2604.01562", "title": "Acoustic and perceptual differences between standard and accented speech and their voice clones", "abstract": "Voice cloning is often evaluated in terms of overall quality, but less is known about accent preservation and its perceptual consequences. We compare standard and heavily accented Mandarin speech and their voice clones using a combined computational and perceptual design. Embedding-based analyses showed larger original-clone distances for accented speakers in several speaker-discriminative embedding spaces, but this difference disappeared after normalizing against each speaker's within-original baseline variability. In the perception study, clones are rated as more similar to their originals for standard than for accented speakers, and intelligibility increases from original to clone, with a larger gain for accented speech. These results show that accent variation can shape perceived identity match and intelligibility in voice cloning even when it is not reflected in baseline-normalized speaker-embedding distance, and they motivate treating accent preservation as an explicit component of speaker identity preservation, rather than assuming that it is fully captured by off-the-shelf speaker-discriminative embeddings.", "date": "2026-04-02", "categories": []}
{"id": "2604.18360", "title": "Omni-Embed-Audio: Leveraging Multimodal LLMs for Robust Audio-Text Retrieval", "abstract": "Audio-text retrieval systems based on Contrastive Language-Audio Pretraining (CLAP) achieve strong performance on traditional benchmarks; however, these benchmarks rely on caption-style queries that differ substantially from real-world search behavior, limiting their assessment of practical retrieval robustness. We present Omni-Embed-Audio (OEA), a retrieval-oriented encoder leveraging multimodal LLMs with native audio understanding. To systematically evaluate robustness beyond caption-style queries, we introduce User-Intent Queries (UIQs) - five formulations reflecting natural search behaviors: questions, commands, keyword tags, paraphrases, and exclusion-based negative queries. For negative queries, we develop a hard negative mining pipeline and propose discrimination metrics (HNSR, TFR) assessing models' ability to suppress acoustically similar distractors. Experiments on AudioCaps, Clotho, and MECAT show that OEA achieves comparable text-to-audio retrieval performance to state-of-the-art M2D-CLAP, while demonstrating clear advantages in two critical areas: (1) dominant text-to-text retrieval (+22% relative improvement), and (2) substantially superior hard negative discrimination (+4.3%p HNSR@10, +34.7% relative TFR@10), revealing that LLM backbones provide superior semantic understanding of complex queries.", "date": "2026-04-20", "categories": []}
{"id": "2605.26310", "title": "Classification and detection of multiple UAVs using rational Gaussian wavelet neural networks", "abstract": "The detection of unmanned aerial vehicles (UAVs) is important for the protection of civilian and military infrastructure. In this paper we propose a cost effective UAV detection system using sound signals obtained from microphones. The recorded signals are passed through a signal processing pipeline which employs interpretable adaptive feature extractors using so-called rational Gaussian wavelets. These adaptive wavelet transformations are embedded into and trained together with an underlying small neural network which detects and classifies UAVs based on the obtained features. This leads to a physically interpretable machine learning algorithm that in addition to classifying UAVs is also capable of detecting UAV swarms. We demonstrate our results using data collected in indoor studio and noisy outdoor environments. We conclude that the proposed method outperforms traditional machine learning approaches for detecting and classifying single UAVs as well as drone swarms, while retaining a high degree of interpretability. Our implementation of the proposed methods is made publicly available for reproducibility.", "date": "2026-05-25", "categories": []}
{"id": "2605.29948", "title": "HoliTok:A Coutinuous Holistic Tokenization with Robust Dual Capabilities of Speech Generation and Understanding", "abstract": "Unified speech foundation models require a holistic tokenization space that is both learnable by language models and decodable into high-quality waveforms. Existing speech tokenizers, however, often fail to satisfy these requirements simultaneously, leading to increased architectural complexity and more involved training designs. We propose HoliTok, a continuous Holistic speech Tokenization model designed for unified generation-understanding modeling. HoliTok encodes 48~kHz speech into a compact 25~Hz sequence of 128-dimensional latents. It is trained with a progressive strategy that jointly preserves signal-level fidelity, incorporates semantic information, and maintains strong latent learnability. Based on this tokenization, we build a unified AR+DiT model for speech synthesis and recognition, where the same latent sequence supports both generation-specific and unified generation-understanding tasks. Experiments show that HoliTok achieves competitive reconstruction fidelity, improves generative learnability for high-quality and controllable synthesis, and, among the evaluated representations, is the only one that operates robustly in our unified generation-understanding architecture without additional optimization tricks. These results suggest that HoliTok serves as an effective speech tokenizer and a foundational representation interface for unified spoken language modeling. The code is available at: https://github.com/bovod-sjtu/HoliTok.", "date": "2026-05-28", "categories": []}
{"id": "2606.01134", "title": "Context-aware child-directed speech detection from long-form recordings", "abstract": "Automatically distinguishing child-directed speech from adult-directed speech in long-form recordings is key to scalable analyses of children's language environments. Existing approaches process utterances in isolation and have been evaluated primarily on English. We address these gaps along three dimensions. First, we fine-tune and evaluate six-self supervised models on a multilingual dataset of 182 children, showing that in-domain pre-training on child-centered recordings substantially outperforms models trained on adult speech. Second, we demonstrate that incorporating surrounding context substantially improves classification, with an absolute gain of 13.8% in average F1-score. Third, we evaluate our model in a realistic end-to-end pipeline, from adult speech detection to addressee classification, showing that performance drops under automatic segmentation but still consistently outperforms a rule-based baseline.", "date": "2026-05-31", "categories": []}
{"id": "2606.01686", "title": "HAIM: Human-AI Music Datasets for AI Music Production Tracking Benchmark", "abstract": "As generative platforms such as Suno and Udio reach human-grade audio quality, the scope of AI's utility has expanded across the entire music production workflow. Beyond simple track generation, these advancements have catalyzed the adoption of AI-driven methodologies in diverse forms. These include vocal synthesis, arrangement, and professional mastering. However, current detection research remains largely confined to a binary `AI-or-human' paradigm. It fails to reflect the realities of contemporary music production workflows. In real-world production, AI tools are increasingly used to refine or master human-produced tracks, and human engineers likewise post-process AI-generated material to ensure professional quality. Moreover, users often employ adversarial tactics to bypass AI detectors, such as applying human mastering to AI-generated tracks. This creates a grey area that a simple binary classification fails to capture. In this paper, we define and investigate ``AI Music Tracking'': the challenge of identifying specific AI integration across the multifaceted spectrum of music production. To this end, we introduce HAIM, a dataset with diverse labels for stages of music production. It is designed to isolate stages of AI intervention, including hybrid production and agent-level tracking. Our evaluation of state-of-the-art detectors reveals systemic flaws. By releasing HAIM, we propose a new benchmark that shifts the field beyond binary classification toward a granular, structured evaluation of AI music.", "date": "2026-06-01", "categories": []}
{"id": "2509.24901", "title": "Unmute the Patch Tokens: Rethinking Probing in Multi-Label Audio Classification", "abstract": "Although probing frozen models has become a standard evaluation paradigm, self-supervised learning in audio defaults to fine-tuning when pursuing state-of-the-art on AudioSet. A key reason is that global pooling creates an information bottleneck causing linear probes to misrepresent the embedding quality: The $\\texttt{cls}$-token discards crucial token information about dispersed, localized events in audio. This weakness is rooted in the mismatch between the pretraining objective (globally) and the downstream task (localized). Across a comprehensive benchmark of 13 datasets and 6 spectrogram-based encoders, we investigate the global pooling bottleneck. We introduce binarized prototypical probes: a lightweight and simple pooling method that learns prototypes to perform class-wise information aggregation. Despite its simplicity, our method notably outperforms linear and attentive probing. Our work establishes probing as a competitive and efficient paradigm for evaluating audio SSL models, challenging the reliance on costly fine-tuning.", "date": "2025-09-29", "categories": []}
{"id": "2602.16305", "title": "BAT: Better Audio Transformer Guided by Convex Gated Probing", "abstract": "Probing is widely adopted in computer vision to faithfully evaluate self-supervised learning (SSL) embeddings, as finetuning may misrepresent their inherent quality. In contrast, audio SSL models still rely on finetuning because simple probing fails to unlock their full potential and alters their rankings when competing on AudioSet. Hence, a robust and efficient probing mechanism is required to guide the trajectory of audio SSL towards reliable and reproducible methods. We introduce Convex Gated Probing (CGP), a prototype-based method that significantly closes the gap between finetuning and probing in audio. CGP efficiently utilizes all frozen layers via a gating mechanism and exposes the location of latent task-relevant information. Guided by CGP as a reliable post-hoc evaluation probe, we rework the entire SSL pipeline of current best performing audio models that use legacy implementations of prior SSL methods. By refining data preprocessing, model architecture, and pretraining recipe, we introduce Better Audio Transformer (BAT), and establish new SOTA on audio benchmarks.", "date": "2026-02-18", "categories": []}
{"id": "2605.31521", "title": "UniAudio-Token: Empowering Semantic Speech Tokenizers with General Audio Perception", "abstract": "Semantic speech tokenizers have become a widely used interface for Audio-LLMs, owing to their compact single-codebook design and strong linguistic alignment. However, their focus on linguistic abstraction induces acoustic blindness, limiting their applicability beyond speech-centric tasks. We propose UniAudio-Token, a framework that empowers semantic tokenizers with general audio perception without compromising speech ability. Instead of altering the semantic paradigm, UniAudio-Token mitigates its information loss through two key innovations: (1) Semantic-Acoustic Primitives (SAP) provide structured supervision by decomposing audio into linguistic content, vocal attributes, and auditory-scene primitives; and (2) Semantic-Acoustic Equilibrium (SAE) introduces a content-aware gating mechanism that adaptively restores fine-grained acoustic details from shallow layers. Extensive evaluations show that UniAudio-Token learns comprehensive universal representations while preserving high-fidelity speech generation. When integrated with downstream LLMs, it outperforms all single-codebook baseline tokenizers on both understanding and generation tasks, effectively serving as a unified audio interface. We publicly release all our code, including training and inference scripts, together with the model checkpoints at https://github.com/Tencent/Universal_Audio_Tokenizer.", "date": "2026-05-29", "categories": []}
{"id": "2502.20838", "title": "Weakly Supervised Detection and Temporal Localization of Whale Calls in Long-Duration Bioacoustic Data", "abstract": "Passive acoustic monitoring (PAM) systems generate continuous recordings spanning months, yet automated bioacoustic analysis of whale calls requires two separate annotation efforts: binary presence labels for classification and precise temporal boundaries for localization. A binary label for a multi-minute recording can be assigned in seconds, but timestamping every call within it requires hours of expert effort. Providing both is infeasible at operational scale. We present DSMIL-LocNet, a weakly supervised multiple instance learning (MIL) framework that performs both classification and temporal localization using only recording-level presence/absence labels. Our dual-stream architecture integrates spectral and temporal features to process recordings of 2--30 minutes without the temporal compression that degrades existing CNN methods on long inputs. On the AcousticTrends BlueFinLibrary, DSMIL-LocNet achieves F1 scores of 0.88--0.91 on recordings of 300--1800s, where fully supervised CNN baselines degrade to 0.19--0.64. It also provides temporal localization that these baselines cannot produce without frame-level annotation. Code: https://github.com/Ragib-Amin-Nihal/DSMIL-Loc", "date": "2025-02-28", "categories": []}
{"id": "2602.08979", "title": "Beyond Transcripts: A Renewed Perspective on Audio Chaptering", "abstract": "Audio chaptering, the task of segmenting long-form audio into coherent sections, is increasingly important for navigating podcasts, lectures, and videos. Despite its relevance, research remains limited and text-based, leaving key questions unresolved about leveraging audio information, handling ASR errors, and transcript-free evaluation. We address these gaps through three contributions: (1) a systematic comparison between text-based models with acoustic features, a novel audio-only architecture (AudioSeg) operating on learned audio representations, and multimodal LLMs; (2) empirical analysis of factors affecting performance, including transcript quality, acoustic features, duration, and speaker composition; and (3) formalized evaluation protocols contrasting transcript-dependent text-space protocols with transcript-invariant time-space protocols. Our experiments on YTSeg reveal that AudioSeg substantially outperforms text-based approaches, pauses provide the largest acoustic gains, and MLLMs remain limited by context length and weak instruction following, yet MLLMs are promising on shorter audio.", "date": "2026-02-09", "categories": []}
{"id": "2603.01006", "title": "AG-REPA: Causal Layer Selection for Representation Alignment in Audio Flow Matching", "abstract": "REPresentation Alignment (REPA) improves the training of generative flow models by aligning intermediate hidden states with pretrained teacher features, but its effectiveness in token-conditioned audio Flow Matching critically depends on the choice of supervised layers, which is typically made heuristically based on the depth. In this work, we introduce Attribution-Guided REPresentation Alignment (AG-REPA), a novel causal layer selection strategy for representation alignment in audio Flow Matching. Firstly, we find that layers that best store semantic/acoustic information (high teacher-space similarity) are not necessarily the layers that contribute most to the velocity field that drives generation, and we call it Store-Contribute Dissociation (SCD). To turn this insight into an actionable training guidance, we propose a forward-only gate ablation (FoG-A) that quantifies each layer's causal contribution via the induced change in the predicted velocity field, enabling sparse layer selection and adaptive weighting for alignment. Across unified speech and general-audio training (LibriSpeech + AudioSet) under different token-conditioning topologies, AG-REPA consistently outperforms REPA baselines. Overall, our results show that alignment is most effective when applied to the causally dominant layers that drive the velocity field, rather than to layers that are representationally rich but functionally passive.", "date": "2026-03-01", "categories": []}
{"id": "2603.27667", "title": "EvA: An Evidence-First Audio Understanding Paradigm for LALMs", "abstract": "Large Audio Language Models (LALMs) still struggle in complex acoustic scenes because they often fail to preserve task-relevant acoustic evidence before reasoning begins. We identify this error pattern as the evidence bottleneck: state-of-the-art systems show larger deficits in acoustic evidence extraction than in downstream reasoning, suggesting that upstream perception is often the limiting factor. To address this problem, we propose EvA (Evidence-First Audio), a dual-path architecture that enhances acoustic evidence preservation through hierarchical aggregation and non-compressive, time-aligned fusion. We also build EvA-Perception, a large-scale training set with about 54K event-ordered captions and 500K evidence-grounded QA pairs. Under a unified zero-shot protocol, EvA achieves the best open-source \\emph{Perception} results on MMAU, MMAR, and MMSU, with the largest gains on perception-heavy splits. Human evaluation on open-ended captioning further shows improved fine-grained acoustic coverage and caption quality. These results support the evidence-first hypothesis: stronger audio understanding depends on preserving acoustic evidence before reasoning. Project can be found at https://satsuki2486441738.github.io/EvA/.", "date": "2026-03-29", "categories": []}
{"id": "2605.28833", "title": "Transcribing Children's Speech: ASR Performance and Obtaining Reliable Orthographic Transcriptions", "abstract": "Automatic speech recognition (ASR) has the potential to substantially reduce manual annotation effort in child speech research by generating automatic transcriptions. However, obtaining reliably high-quality ASR transcriptions for child speech remains challenging in low-resource languages due to limited child-specific pre-trained models and highly diverse noise conditions. This study investigates the effectiveness of state-of-the-art ASR models on child speech through two research questions, by evaluating nine ASR models from three model families (Whisper, Parakeet, and Wav2Vec2) on two Dutch child speech datasets, JASMIN and DART. Research question 1 examines the performance of ASR-models applied to child speech. The fine-tuned Whisper-medium model achieves the best overall performance, with a WER of 5.54% on JASMIN and 70.37% on DART, showing that the noisy DART data are clearly more challenging. Research question 2 examines to what extent it is possible to select a subset for which reliable orthographic transcriptions can be obtained automatically, without the need for manual verification. We use an utterance-level selection method that compares ASR output with the original read prompt to identify correctly pronounced recordings. Using the proposed selection method, 42.0% [for JASMIN] and 18.1% [for DART] of the utterances can be automatically identified as correctly pronounced with high confidence, resulting in very low error rates on an utterance level (precisions of 98.3% and higher) and reducing the need for manual verification.", "date": "2026-04-10", "categories": []}
{"id": "2605.29300", "title": "MusTBENCH: Benchmarking and Advancing Temporal Grounding in Music LLMs", "abstract": "Recent Large Audio-Language Models (LALMs) have demonstrated promising abilities in understanding musical content. However, whether their responses are grounded in the correct temporal regions of the audio remains underexplored. This limitation is particularly critical for music understanding, where key information often occurs as temporally localized events, such as instrument entries and rhythmic transitions. To address this gap, we introduce MusTBENCH, a music-expert-validated benchmark designed to evaluate temporal grounding in LALMs through five temporally grounded question-answering tasks. To further improve temporal grounding in existing models, we propose MusT, a novel four-stage temporal optimization recipe spanning music encoder adaptation, LLM adaptation, LLM supervised fine-tuning, and RL-based optimization. Experiments on MusTBENCH show that existing LALMs struggle with precise temporal grounding, while MusT brings significant improvements over strong baselines. These results establish temporal grounding as a key missing capability in current LALMs and position MusTBENCH as a challenging benchmark for future research in temporally grounded music understanding.", "date": "2026-05-28", "categories": []}
{"id": "2605.29531", "title": "Audio Deepfake Detection with Half-Truth Localisation Using Cross-Attentive Feature Fusion", "abstract": "Audio deepfake detection is well-studied as a binary problem, but partially manipulated speech, where a short synthesised segment is spliced into an otherwise genuine utterance, poses a harder and more realistic threat. Detecting such half-truth audio requires not only distinguishing it from real and fully fake speech, but also localising where the manipulation occurs. We present CAFNet, a 576k-parameter architecture that addresses both tasks jointly: it performs ternary classification (real, fully-fake, or half-truth) and regresses the temporal boundaries of the synthesised region in a single forward pass. CAFNet fuses Mel-Frequency Cepstral Coefficient (MFCC), Linear-Frequency Cepstral Coefficient (LFCC), and Chroma Short-Time Fourier Transform (Chroma-STFT) features through parallel depthwise-separable convolution branches with cross-attention, followed by a Bidirectional Long Short-Term Memory (BiLSTM) regression head for boundary prediction. On the combined Multi-Lingual Audio Deepfake Detection Corpus (MLADDC) T2+T3 test set, CAFNet achieves 92.71% accuracy and macro Area Under the Curve (AUC) of 0.9910, with boundary localisation Mean Absolute Error (MAE) of 0.075s and a median error of 0.052s. On binary detection, it achieves 96.76% accuracy and 3.20% Equal Error Rate (EER), outperforming fine-tuned XLS-R 300M (78.31%) and AST 87M (93.03%) at over 500 times fewer parameters. A cross-dataset study further shows that standard fine-tuning collapses cross-domain representations even under reduced backbone learning rates.", "date": "2026-05-28", "categories": []}
{"id": "2605.29862", "title": "Mitigating Stethoscope-Induced Shortcuts in Respiratory Sound Classification under Federated Domain Generalization with Causality-Inspired Interventions", "abstract": "AI-driven respiratory sound classification (RSC) is promising for automated pulmonary disease detection, yet multi-site deployment is hindered by inter-stethoscope variability. We introduce a federated domain generalization (FedDG) formulation for RSC under stethoscope-induced device shifts, where clients use heterogeneous devices and the model is evaluated on unseen devices. Our empirical analysis shows that stethoscope-induced style and disease-specific content are tightly entangled, making deterministic style removal unreliable. In response, we propose a causality-inspired multimodal FedDG framework that combines: (i) a causality-inspired device style intervention network that performs content-preserving style perturbations, (ii) counterfactual text augmentation that neutralizes metadata shortcuts, and (iii) gradient alignment that facilitates device-invariant representations across clients. Built on a multimodal language-audio pretraining model, it outperforms conventional data augmentation and federated learning baselines in leave-one-device-out validation on ICBHI and SPRSound datasets. Code will be released upon publication.", "date": "2026-05-28", "categories": []}
{"id": "2605.27840", "title": "LoSATok: Low-dimensional Semantic-Acoustic Tokenizer for Cross-Domain Audio Understanding and Generation", "abstract": "Audio tokenizers are fundamental to unifying audio understanding and generation. Understanding requires high-level semantics, while generation demands semantic and acoustic details. Existing unified tokenizers jointly encode both in high-dimensional continuous latents, which increases the modeling burden of Diffusion Transformers (DiTs) for generation. We propose LoSATok, a low-dimensional audio tokenizer for cross-domain audio understanding and generation. Motivated by the observation that 1280-dimensional semantic encoder features are compressible, we introduce a Semantic Bottleneck that compresses them into 128 dimensions, regularized by the proposed time-relation loss for temporal feature consistency. We further design a dual-level semantic supervision method that leverages both high- and low-dimensional semantic signals, enabling the tokenizer to jointly capture semantics and acoustic details within a compact latent space. Experiments on speech, music, and general audio show that SemBo preserves strong low-dimensional semantic capacity and LoSATok retains competitive understanding performance compared with several semantic representations, while consistently improving DiT modeling performance on speech, music, and audio generation. These results demonstrate that LoSATok's low-dimensional representations can effectively support audio understanding and generation. Our code is provided at https://github.com/wxzyd123/LoSATok.", "date": "2026-05-27", "categories": []}
{"id": "2605.28139", "title": "Data-Efficient On-Policy Distillation for Automatic Speech Recognition", "abstract": "Building competitive automatic speech recognition (ASR) models usually requires large-scale au- dio supervision, which makes reproduction and specialization expensive. We study Ark-ASR, a 0.6B- parameter audio-conditioned language model trained with 100k hours of speech, and examine whether a strong Qwen-ASR teacher can transfer additional recognition capability through on-policy distillation. Across Mandarin and English ASR benchmarks, the proposed training recipe consistently improves over supervised fine-tuning alone and outperforms the same-scale Qwen3-ASR-0.6B baseline on four of five evaluation sets. This is achieved with only 100k hours of speech, compared with the 20M hours of super- vised audio reported for the Qwen3-Omni AuT encoder. The larger Qwen3-ASR-1.7B remains stronger, but the results show that teacher-guided on-policy training can substantially close the gap for compact ASR models under a much smaller audio budget. A support-overlap diagnostic further suggests that the teacher-data stage improves local student-teacher compatibility, matching recent analyses of when on-policy distillation is effective.", "date": "2026-05-27", "categories": []}
{"id": "2605.03929", "title": "PHALAR: Phasors for Learned Musical Audio Representations", "abstract": "Stem retrieval, the task of matching missing stems to a given audio submix, is a key challenge currently limited by models that discard temporal information. We introduce PHALAR, a contrastive framework achieving a relative accuracy increase of up to $\\approx 70\\%$ over the state-of-the-art while requiring $<50\\%$ of the parameters and a 7$\\times$ training speedup. By utilizing a Learned Spectral Pooling layer and a complex-valued head, PHALAR enforces pitch-equivariant and phase-equivariant biases. PHALAR establishes new retrieval state-of-the-art across MoisesDB, Slakh, and ChocoChorales, correlating significantly higher with human coherence judgment than semantic baselines. Finally, zero-shot beat tracking and linear chord probing confirm that PHALAR captures robust musical structures beyond the retrieval task.", "date": "2026-05-05", "categories": []}
{"id": "2605.26176", "title": "PitchBench: Measuring Pitch Hearing in Audio-Language Models", "abstract": "Audio-language models (ALMs) are increasingly used in real-world applications that require understanding music, from music tutoring and transcription to captioning, recommendation systems, and music production. More broadly, they are becoming an important component of multimodal AI systems that must reason from sensory input rather than text alone. This makes reliable musical perception a critical prerequisite: if a model cannot accurately hear the structure of sound, it cannot be trusted to reason about, teach, transcribe, or act on audio in the real world. Yet existing benchmarks rarely assess one of the most fundamental musical abilities underlying such perception: pitch hearing. Current evaluations tend to probe pitch hearing only indirectly, through higher-level tasks and often in multiple-choice formats, leaving open how reliably ALMs identify fine-grained pitch across instruments, acoustic conditions, and response formats. We introduce PitchBench, an evaluation suite that systematically measures pitch hearing in ALMs. PitchBench comprises 28 experiments spanning absolute and relative pitch perception within sequences and chords, while varying loudness, note duration, sound source, time stretching, background noise, and other acoustic conditions. Tasks range from identifying individual pitches in isolation to tracking a melodic line within a four-part musical texture. Evaluating frontier ALMs, we find that pitch hearing remains highly unreliable: models perform consistently poorly across settings, with accuracy varying sharply by sound source, note duration, and notation format. Current ALMs do not yet possess stable pitch perception, even for controlled synthetic and instrumental stimuli. Alongside the benchmark, we release PitchBench as a Python package containing the evaluation data and data generation tools to support future work on pitch-aware audio-language modeling.", "date": "2026-05-25", "categories": []}
{"id": "2605.26523", "title": "StreamSplit: Continuous Audio Representation Learning via Uncertainty-Guided Adaptive Splitting", "abstract": "Large-batch Contrastive Learning (CL), the foundation of modern representation learning, is fundamentally incompatible with the volatile resource constraints of edge devices. This conflict creates a dilemma: small on-device batches degrade model fidelity, while offloading to the cloud incurs unacceptable latency and bandwidth costs. Existing solutions often resort to static model compression, which fails to adapt to the runtime volatility of edge environments. To bridge this gap, we present StreamSplit, a novel framework that makes streaming CL practical across heterogeneous ARM client platforms. StreamSplit resolves the conflict between the continuous nature of ambient audio and the discrete batch requirements of models like CLAP and COLA. We introduce: (1) A distribution-based streaming framework that decouples representation quality from local batch size, using a tractable Hybrid Loss to maintain fidelity despite sparse updates; and (2) An Uncertainty-Guided Adaptive Splitter that uses a lightweight Reinforcement Learning (RL) policy to dynamically partition computation. Uniquely, this policy integrates real-time resource monitoring with embedding ambiguity to optimize the accuracy-latency trade-off on the fly. We evaluate StreamSplit on diverse hardware, from the resource-constrained Raspberry Pi 4 to the high-performance Apple M2. Results demonstrate that StreamSplit reduces per-sample latency by up to 4.7x and cuts bandwidth by 77.1% and energy by 52.3% compared to server-centric baselines. Crucially, it maintains accuracy within 2.2% of server-centric models, proving that adaptive, distributed learning is a viable path for the modern edge ecosystem.", "date": "2026-05-26", "categories": []}
{"id": "2601.21463", "title": "Unifying Speech Editing Detection and Content Localization via Prior-Enhanced Audio LLMs", "abstract": "Existing speech editing detection (SED) datasets are predominantly constructed using manual splicing or limited editing operations, resulting in restricted diversity and poor coverage of realistic editing scenarios. Meanwhile, current SED methods rely heavily on frame-level supervision to detect observable acoustic anomalies, which fundamentally limits their ability to handle deletion-type edits, where the manipulated content is entirely absent from the signal. To address these challenges, we present a unified framework that bridges speech editing detection and content localization through a generative formulation based on Audio Large Language Models (Audio LLMs). We first introduce AiEdit, https://huggingface.co/datasets/JunXueTech/AiEdit, a large-scale bilingual dataset (approximately 140 hours) that covers addition, deletion, and modification operations using state-of-the-art end-to-end speech editing systems, providing a more realistic benchmark for modern threats. Building upon this, we reformulate SED as a structured text generation task, enabling joint reasoning over edit type identification, and content localization. To enhance the grounding of generative models in acoustic evidence, we propose a prior-enhanced prompting strategy that injects word-level probabilistic cues derived from a frame-level detector. Furthermore, we introduce an acoustic consistency-aware loss that explicitly enforces the separation between normal and anomalous acoustic representations in the latent space. Experimental results demonstrate that the proposed approach consistently outperforms existing methods across both detection and localization tasks.", "date": "2026-01-29", "categories": []}
{"id": "2605.04700", "title": "Sparse Tokens Suffice: Jailbreaking Audio Language Models via Token-Aware Gradient Optimization", "abstract": "Jailbreak attacks on audio language models (ALMs) optimize audio perturbations to elicit unsafe generations, and they typically update the entire waveform densely throughout optimization. In this work, we investigate the necessity of such dense optimization by analyzing the structure of token-aligned gradients in ALMs. We find that gradient energy is highly non-uniform across audio tokens, indicating that only a small subset of token-aligned audio regions dominates the optimization signal. Motivated by this observation, we propose Token-Aware Gradient Optimization (TAGO), which enables sparse jailbreak optimization by retaining only waveform gradients aligned with audio tokens that have high gradient energy, while masking the remaining gradients at each iteration. Across three ALMs, TAGO outperforms baselines, and substantial sparsification preserves strong attack success rates (e.g. on Qwen3-Omni, $\\mathrm{ASR}_{l}$ remains at 86% with a token retention ratio of 0.25, compared to 87% with full token retention). These results demonstrate that dense waveform updates are largely redundant, and we advocate that future audio jailbreak and safety alignment research should further leverage this heterogeneous token-level gradient structure.", "date": "2026-05-06", "categories": []}
{"id": "2605.24193", "title": "Music Transcription with (Almost) No Supervision", "abstract": "Competitive music transcription models require large amounts of paired audio-score data, which is scarce due to collection costs, alignment difficulty, and copyright restrictions. Meanwhile, vast quantities of unpaired audio recordings and symbolic scores are freely available but have gone unused. We adopt a cycle-consistent translation framework in which a small amount of paired data acts as a minimal anchor, unlocking the full potential of the unpaired pool. We find that: unpaired data yields surprisingly large gains, especially under limited supervision; unpaired audio contributes more than unpaired scores; incorporating unlabeled audio from a new instrument during training improves transcription for that instrument without any paired supervision. Together, these results suggest that scaling unpaired data offers a practical path toward high-quality transcription for instruments where labeled data remains scarce.", "date": "2026-05-22", "categories": []}
{"id": "2605.24291", "title": "Rubato: Transcribing Piano Music with Timestamps", "abstract": "We consider the conversion of musical recordings into human-readable sheet music annotated with timestamps. Such output lets a listener clearly visualize rubato (temporally expressive playing), a learner diagnose ensemble precision and timing choices against the written music, and a musicology scholar compare performance styles across recordings of the same work. We introduce (1) a prompt-conditioned encoder-decoder model, named Rubato, trained to output (2) a new textual representation for polyphonic music, named InterMo, which we designed for compatibility with sequence-to-sequence training. Our experiments demonstrate that Rubato produces timestamped piano sheet music from audio with higher notational accuracy than the best existing approaches, which are based on cascades. We find that even if the cascade is given ground-truth MIDI instead of audio, Rubato performs better, suggesting that the ceiling of existing approaches is primarily representational, not acoustic. Further, because Rubato is trained on several related tasks (with prompts), it competes with or outperforms the best single-task systems on related but simpler tasks like MIDI note grounding and beat/downbeat detection. A demo is available at https://nctamer.github.io/rubato-transcription .", "date": "2026-05-22", "categories": []}
{"id": "2605.25179", "title": "Locality Matters for Training-Free Audio Token Compression in Audio-Language Models", "abstract": "Audio-language models (ALMs) are increasingly used for audio captioning, question answering, and open-ended audio understanding, but their inference cost remains high when audio inputs are represented as long prefix-token sequences. These audio prefixes consume context budget, increase memory usage, and make deployment harder in resource-constrained or latency-sensitive settings. Existing training-free audio-token reduction methods mainly rely on fixed pooling or score-based pruning. Fixed pooling is content-agnostic, while score-based pruning can preserve isolated salient tokens but discard nearby acoustic context. We propose Local Temporal Bipartite Merging (LTBM), a training-free encoder-space compression method that merges similar nearby audio tokens under an explicit temporal window constraint. Beyond introducing LTBM, we use a controlled Global Merge variant to isolate whether temporal locality itself is a useful inductive bias for audio-token compression. Experiments on AudioCaps, Clotho, and MMAU with Qwen2-Audio show evidence of a task-dependent locality effect: locality-aware merging is more favorable for captioning at several compression settings, especially under stronger compression, while global matching is more competitive for multiple-choice audio understanding. A cross-backbone validation on Audio Flamingo 3 further supports the captioning-side advantage of locality-aware merging under moderate and aggressive compression.", "date": "2026-05-24", "categories": []}
{"id": "2502.04230", "title": "XAttnMark: Learning Robust Audio Watermarking with Cross-Attention", "abstract": "The rapid proliferation of generative audio synthesis and editing technologies has raised serious concerns about copyright infringement, data provenance, and the spread of misinformation via deepfake audio. Watermarking offers a proactive solution by embedding imperceptible yet identifiable and traceable signals into audio content. While recent neural network-based watermarking methods like WavMark and AudioSeal have improved robustness and quality, they struggle to jointly optimize both robust detection and accurate attribution. This paper introduces Cross-Attention Robust Audio Watermark (XATTNMARK), which bridges this gap by leveraging partial parameter sharing between the generator and the detector, a cross-attention mechanism for efficient message retrieval, and a temporal conditioning module for improved message distribution. Additionally, we propose a psychoacoustic-aligned time-frequency (TF) masking loss that captures fine-grained auditory masking effects, improving watermark imperceptibility. XATTNMARK achieves state-of-the-art performance in both detection and attribution, demonstrating superior robustness against a wide range of audio transformations, including challenging generative editing at varying strengths. This work advances audio watermarking for protecting intellectual property and ensuring authenticity in the era of generative AI.", "date": "2025-02-06", "categories": []}
{"id": "2605.20519", "title": "Codec-Robust Attacks on Audio LLMs", "abstract": "Prior attacks on Audio Large Language Models (Audio LLMs) demonstrated that carefully crafted waveform-domain perturbations can force targeted adversarial outputs. As a defense mechanism against these attacks, real-world codec compression preprocessing has been studied to both detect and remove the perturbations. Yet no existing attack has demonstrated robustness against these compressions. We introduce CodecAttack, which optimizes a perturbation in a neural audio codec's continuous latent space rather than directly perturbing the audio waveform. We show that the codec's compression channel, which discards waveform perturbations, transmits perturbations crafted in its own latent space. To further harden the attack across real-world compression channels, we apply multi-bitrate straight-through Expectation-over-Transformation (EoT), all without modifying the target model. Across three realistic Audio LLM deployment scenarios and three target models, CodecAttack achieves an average 85.5% target-substring attack success rate (ASR) on Opus at moderate bitrates, while the waveform baseline trained with identical EoT hardening does not exceed 26% at any bitrate. The attack transfers to held-out codecs, reaching up to 100% ASR on MP3 and 84% on AAC-LC without retraining. A per-band energy analysis shows that the latent perturbation concentrates below 4kHz, exactly where codecs allocate the most bits, while the waveform baseline spreads into higher frequencies that codecs discard. These results demonstrate that lossy compression is not a reliable defense against adversarial audio and that codec-aware attacks pose a practical threat to deployed Audio LLM systems.", "date": "2026-05-19", "categories": []}
{"id": "2509.15151", "title": "Exploring How Audio Effects Alter Emotion with Foundation Models", "abstract": "Audio effects (FX) such as reverberation, distortion, modulation, and dynamic range processing play a pivotal role in shaping emotional responses during music listening. While prior studies have examined links between low-level audio features and affective perception, the systematic impact of audio FX on emotion remains underexplored. This work investigates how foundation models - large-scale neural architectures pretrained on multimodal data - can be leveraged to analyze these effects. Such models encode rich associations between musical structure, timbre, and affective meaning, offering a powerful framework for probing the emotional consequences of sound design techniques. By applying various probing methods to embeddings from deep learning models, we examine the complex, nonlinear relationships between audio FX and estimated emotion, uncovering patterns tied to specific effects and evaluating the robustness of foundation audio models. Our findings aim to advance understanding of the perceptual impact of audio production practices, with implications for music cognition, performance, and affective computing.", "date": "2025-09-18", "categories": []}
{"id": "2605.03934", "title": "Towards Open World Sound Event Detection", "abstract": "Sound Event Detection (SED) plays a vital role in audio understanding, with applications in surveillance, smart cities, healthcare, and multimedia indexing. However, conventional SED systems operate under a closed-world assumption, limiting their effectiveness in real-world environments where novel acoustic events frequently emerge. Inspired by the success of open-world learning in computer vision, we introduce the Open-World Sound Event Detection (OW-SED) paradigm, where models must detect known events, identify unseen ones, and incrementally learn from them. To tackle the unique challenges of OW-SED, such as overlapping and ambiguous events, we propose a 1D Deformable architecture that leverages deformable attention to adaptively focus on salient temporal regions. Furthermore, we design a novel Open-World Deformable Sound Event Detection Transformer (WOOT) framework incorporating feature disentanglement to separate class-specific and class-agnostic representations, together with a one-to-many matching strategy and a diversity loss to enhance representation diversity. Experimental results demonstrate that our method achieves marginally superior performance compared to existing leading techniques in closed-world settings and significantly improves over existing baselines in open-world scenarios.", "date": "2026-05-05", "categories": []}
{"id": "2605.21143", "title": "CoarseSoundNet: Building a reliable model for ecological soundscape analysis", "abstract": "A soundscape is composed of three types of sound: biophony (sounds made by animals), geophony (natural abiotic sounds) and anthropophony (sounds made by humans). A key research question in the field of soundscape ecology is how these components interact with each other, specifically how biophony responds to geophony and anthropophony. Nevertheless, as of today, there are not many analytical instruments that enable the distinct quantification of these elements. Recent machine learning (ML) approaches aim to support automated analysis but often rely on task-specific or clean data, limiting generalisation to noisy passive acoustic monitoring (PAM) recordings. This study presents a clear and reproducible structure to build ML models for coarse soundscape classification and introduces CoarseSoundNet, a deep learning model trained to distinguish biophony, geophony, and anthropophony under realistic PAM conditions. We systematically investigate model architectures, the influence of an additional training class, data composition, and evaluation strategies. Our findings suggest that model performance improves with additional PAM data, especially when similar to the target domain, and by introducing an explicit silence class during training. Class-specific decision thresholds and duration-based constraints further enhance performance, particularly for anthropophony and geophony. Error analyses exhibit challenges for anthropophony due to masking effects and confusions for silence and insect sounds for geophony and biophony. Finally, we conduct an ecological case study which shows that pre-filtering recordings with CoarseSoundNet yields acoustic index trends comparable to ground-truth filtering, supporting its use as an effective preprocessing tool for ecoacoustic analyses.", "date": "2026-05-20", "categories": []}
{"id": "2605.22262", "title": "Automatic Contextual Audio Denoising", "abstract": "Audio context determines which sound components and sources are relevant and which can be perceived as irrelevant (noise) by listeners. For example, traffic noise is informative in urban surveillance but noise for a phone call at the same location. Most current audio denoising systems apply fixed target-noise definitions, often removing useful components in one context while failing to suppress irrelevant components. To address this, we introduce the concept automatic contextual audio denoising (ACAD) which defines target and noise based on the inferred context. In this work, we restrict context to be associated with an acoustic scene class. We label sound events outside the event distribution of a scene class (noise) as out-of-context (OC) and events typical for that scene as in-context (IC). We implement a deep learning method that automatically infers the context of the audio signal and removes OC components, and benchmark it against variants: without context inference, with oracle context, and with separately provided uninformative context. On paired clean/noisy data across diverse contexts, where OC components in one context may be IC in another, our proposed method outperforms other approaches across standard objective metrics, indicating that the model can infer context and context-dependent processing can enhance denoising.", "date": "2026-05-21", "categories": []}
{"id": "2506.14148", "title": "Acoustic scattering AI for non-invasive object classifications: A case study on hair assessment", "abstract": "This paper presents a novel non-invasive object classification approach using acoustic scattering, demonstrated through a case study on hair assessment. When an incident wave interacts with an object, it generates a scattered acoustic field encoding structural and material properties. By emitting acoustic stimuli and capturing the scattered signals from head-with-hair-sample objects, we classify hair type and moisture using AI-driven, deep-learning-based sound classification. We benchmark comprehensive methods, including (i) fully supervised deep learning, (ii) embedding-based classification, (iii) supervised foundation model fine-tuning, and (iv) self-supervised model fine-tuning. Our best strategy achieves nearly 90% classification accuracy by fine-tuning all parameters of a self-supervised model. These results highlight acoustic scattering as a privacy-preserving, non-contact alternative to visual classification, opening huge potential for applications in various industries.", "date": "2025-06-17", "categories": []}
{"id": "2511.21577", "title": "HarmonicAttack: An Adaptive Cross-Domain Audio Watermark Removal", "abstract": "The availability of high-quality, AI-generated audio raises security challenges such as misinformation campaigns and voice-cloning fraud. A key defense against the misuse of AI-generated audio is by watermarking it, so that it can be easily distinguished from genuine audio. Those seeking to misuse AI-generated audio may attempt to remove audio watermarks, so studying effective watermark removal techniques is critical to objectively evaluate the robustness of audio watermarks. Previous watermark removal schemes typically assume access to the target watermark detector during the removal process. This assumption is often impractical, which may lead to a false sense of confidence in current watermark schemes. We introduce HarmonicAttack, a novel audio watermark removal method that requires no access to the target watermark algorithm. It only needs a number of original and watermarked samples to train a general model capable of removing watermarks from audio samples. We also find that training samples do not need to share the same distribution as target samples, as our attack generalizes to out-of-distribution samples with minimal degradation. Compared with existing watermark removal attacks, HarmonicAttack is more effective at removing watermarks from state-of-the-art schemes, including AudioSeal, WavMark, SilentCipher, and AudioMarkNet, while maintaining high perceptual quality. Although HarmonicAttack is trained on the LibriSpeech dataset against AudioSeal, it generalizes across unseen datasets and watermarking schemes. For instance, on VCTK, HarmonicAttack achieves a 92% ASR against AudioMarkNet, substantially outperforming the best baseline at 38%. On FMA, HarmonicAttack reaches 100% ASR against all watermarks, whereas the best baseline achieves only 2% against AudioSeal and 44% against WavMark.", "date": "2025-11-26", "categories": []}
{"id": "2602.11910", "title": "TADA! Tuning Audio Diffusion Models through Activation Steering", "abstract": "Audio diffusion models can synthesize high-fidelity music from text, yet achieving fine-grained control over specific musical attributes remains challenging, as their internal mechanisms for representing high-level concepts are poorly understood. In this work, we use activation patching to demonstrate that recent audio diffusion architectures exhibit a semantic bottleneck, where a small, shared subset of consecutive attention layers controls distinct musical concepts, such as the presence of specific instruments, vocals, or genres. Building on this, we systematically evaluate a broad spectrum of steering paradigms, comparing activation steering against prompt-level, score-space, and weight-space interventions, analyzing the interaction between the steering mechanism and the intervention site. Our new benchmark, supported by an extensive user study, demonstrates that localized activation steering establishes a new state-of-the-art in audio concept modulation.", "date": "2026-02-12", "categories": []}
{"id": "2603.25476", "title": "How Class Ontology and Data Scale Affect Audio Transfer Learning", "abstract": "Transfer learning is a crucial concept within deep learning that allows artificial neural networks to benefit from a large pre-training data basis when confronted with a task of limited data. Despite its ubiquitous use and clear benefits, there are still many open questions regarding the inner workings of transfer learning and, in particular, regarding the understanding of when and how well it works. To that extent, we perform a rigorous study focusing on audio-to-audio transfer learning, in which we pre-train various model states on (ontology-based) subsets of AudioSet and fine-tune them on three computer audition tasks, namely acoustic scene recognition, bird activity recognition, and speech command recognition. We report that increasing the number of samples and classes in the pre-training data both have a positive impact on transfer learning. This is, however, generally surpassed by similarity between pre-training and the downstream task, which can lead the model to learn comparable features.", "date": "2026-03-26", "categories": []}
{"id": "2605.17370", "title": "CBT-Audio: Evaluating Audio Language Models for Patient-Side Distress Intensity Estimation in CBT Session Recordings", "abstract": "Cognitive behavioural therapy is widely used to help patients understand and manage psychological distress. It is often delivered through spoken conversation, where therapists attend not only to what patients say, but also to how they say it, because these cues can help therapists decide how to respond and adapt treatment. Progress in building AI systems for CBT remains largely limited to text, partly because most available datasets are text based and shareable spoken CBT data are scarce under ethical and privacy constraints. This creates a blind spot because text based models and evaluations cannot capture the mismatch between the transcript and the patient's voice, even though therapists often rely on this mismatch to understand patient distress. We introduce CBT-Audio, a dataset for evaluating patient distress estimation from spoken CBT sessions with audio language models. CBT-Audio contains 1,802 patient turns from 96 publicly available CBT recordings, with turn-level distress labels validated on an experts-annotated subset. We evaluate 10 open source audio language models under three input conditions, where models receive only patient audio, only the transcript, or both audio and transcript. Our results show that audio can provide useful information beyond text, especially when combined with transcripts. Adding audio to transcript input improves distress estimation over using the transcript alone in 8 of 10 model families, with significant gains in 4, and case studies show the clearest benefit when verbal content and vocal delivery diverge. CBT-Audio makes spoken patient behaviour measurable for AI evaluation in CBT-related tasks and supports future work on audio language models for mental health interaction.", "date": "2026-05-17", "categories": []}
{"id": "2605.19833", "title": "Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation", "abstract": "Despite rapid advances in automatic speech recognition (ASR) and large audio-language models, robust recognition in real-world environments remains limited by an \"acoustic robustness bottleneck\": models often lose acoustic grounding and produce omissions or hallucinations under severe, compositional distortions. We propose Mega-ASR, a unified ASR-in-the-wild framework that combines scalable compound-data construction with progressive acoustic-to-semantic optimization. We introduce Voices-in-the-Wild-2M, covering 7 classic acoustic phenomena and 54 physically plausible compound scenarios, and train Mega-ASR with Acoustic-to-Semantic Progressive Supervised Fine-Tuning and Dual-Granularity WER-Gated Policy Optimization. Extensive experiments demonstrate that Mega-ASR achieves significant advantages over prior state-of-the-art systems on adverse-condition ASR benchmarks (45.69% vs. 54.01% on VOiCES R4-B-F, and 21.49% vs. 29.34% on NOIZEUS Sta-0). On complex compositional acoustic scenarios, Mega-ASR further delivers over 30% relative WER reduction against strong open- and closed-source baselines, establishing a scalable paradigm for robust ASR in-the-wild.", "date": "2026-05-19", "categories": []}
{"id": "2404.00470", "title": "Classification of Short Segment Pediatric Heart Sounds Based on a Transformer-Based Convolutional Neural Network", "abstract": "Congenital anomalies arising as a result of a defect in the structure of the heart and great vessels are known as congenital heart diseases or CHDs. A PCG can provide essential details about the mechanical conduction system of the heart and point out specific patterns linked to different kinds of CHD. This study aims to investigate the minimum signal duration required for the automatic classification of heart sounds. This study also investigated the optimum signal quality assessment indicator (Root Mean Square of Successive Differences) RMSSD and (Zero Crossings Rate) ZCR value. Mel-frequency cepstral coefficients (MFCCs) based feature is used as an input to build a Transformer-Based residual one-dimensional convolutional neural network, which is then used for classifying the heart sound. The study showed that 0.4 is the ideal threshold for getting suitable signals for the RMSSD and ZCR indicators. Moreover, a minimum signal length of 5s is required for effective heart sound classification. It also shows that a shorter signal (3 s heart sound) does not have enough information to categorize heart sounds accurately, and the longer signal (15 s heart sound) may contain more noise. The best accuracy, 93.69%, is obtained for the 5s signal to distinguish the heart sound.", "date": "2024-03-30", "categories": []}
{"id": "2408.17352", "title": "AASIST3: KAN-Enhanced AASIST Speech Deepfake Detection using SSL Features and Additional Regularization for the ASVspoof 2024 Challenge", "abstract": "Automatic Speaker Verification (ASV) systems, which identify speakers based on their voice characteristics, have numerous applications, such as user authentication in financial transactions, exclusive access control in smart devices, and forensic fraud detection. However, the advancement of deep learning algorithms has enabled the generation of synthetic audio through Text-to-Speech (TTS) and Voice Conversion (VC) systems, exposing ASV systems to potential vulnerabilities. To counteract this, we propose a novel architecture named AASIST3. By enhancing the existing AASIST framework with Kolmogorov-Arnold networks, additional layers, encoders, and pre-emphasis techniques, AASIST3 achieves a more than twofold improvement in performance. It demonstrates minDCF results of 0.5357 in the closed condition and 0.1414 in the open condition, significantly enhancing the detection of synthetic voices and improving ASV security. \\textbf{The new version of the model is publicly available at \\href{https://huggingface.co/lab260/Spectra-AASIST3}{\\underline{HuggingFace (2026)}}}", "date": "2024-08-30", "categories": []}
{"id": "2605.17085", "title": "Taming Audio VAEs via Target-KL Regularization", "abstract": "Latent diffusion models have emerged as the dominant paradigm for many generation tasks including audio generation such as text-to-audio, text-to-music and text-to-speech. A key component of latent diffusion is an autoencoder (VAE) that compresses high-dimensional signals into a low frame rate continuous representation that is conducive for downstream prediction. Regularizing these VAEs is challenging, as there is a trade-off between over-regularized (poor output quality) and under-regularized (difficult to predict) latent representations. We propose a framework for studying this trade-off through compression and train Audio VAEs at specific bitrates via target-KL regularization. This allows direct comparison to well-studied discrete neural audio codec models, and the construction of rate-distortion curves for audio VAEs. We evaluate the impact of target-KL regularization on text-to-sound generation and find that sweeping compression rates is helpful in identifying the optimal generation setting.", "date": "2026-05-16", "categories": []}
{"id": "2605.17991", "title": "Stable Audio 3", "abstract": "Stable Audio 3 is a family of fast latent diffusion models (small, medium, large) for variable-length audio generation and editing. Since our models can generate several minutes of audio, variable-length generations are key to avoid the cost of producing full-length generations for short sounds. We also support inpainting, enabling targeted audio editing and the continuation of short recordings. Our latent diffusion models operate on top of a novel semantic-acoustic autoencoder that projects audio into a compact latent space, enabling efficient diffusion-based generation while preserving audio fidelity and encouraging semantic structure in the latent. Finally, we run adversarial post-training to both accelerate inference and improve generation quality, reducing the number of inference steps while improving fidelity and prompt adherence. Stable Audio 3 models are trained on licensed and Creative Commons data to generate music and sounds in less than a 2s on an H200 GPU and less than a few seconds on a MacBook Pro M4. We release the weights of small and medium, that can run on consumer-grade hardware, together with their training and inference pipeline.", "date": "2026-05-18", "categories": []}
{"id": "2605.18613", "title": "SAME: A Semantically-Aligned Music Autoencoder", "abstract": "Latent representations are at the heart of the majority of modern generative models. In the audio domain they are typically produced by a neural-audio-codec autoencoder. In this work we introduce SAME (Semantically-Aligned Music autoEncoder), an autoencoder for stereo music and general audio that reaches a 4096$\\times$ temporal compression ratio while maintaining reconstruction quality and downstream generative performance. We achieve this by combining a tranformer-based backbone with set of semantic regularisation approaches, phase-aware reconstruction losses and improved discriminator designs. The architecture delivers substantial computational cost benefits, through both its high compression ratio and its reliance on well-optimised transformer primitives. Two variants (a large SAME-L and a CPU-deployable SAME-S) are released in open-weights form.", "date": "2026-05-18", "categories": []}
{"id": "2605.18749", "title": "WavFlow: Audio Generation in Waveform Space", "abstract": "Modern audio generation predominantly relies on latent-space compression, introducing additional complexity and potential information loss. In this work, we challenge this paradigm with WavFlow, a framework that generates high-fidelity audio directly in raw waveform space without intermediate representations. To overcome the inherent difficulties of modeling high-dimensional and low-energy signals, we reshape audio into 2D token grids through waveform patchify and introduce amplitude lifting to align signal scales, enabling stable optimization via direct x-prediction in flow matching. To capture complex semantic alignment and temporal synchronization, we leverage an automated data pipeline to curate 5 million high-quality video-text-audio triplets, allowing the model to learn fine-grained acoustic patterns from scratch. Experimental results show that WavFlow achieves competitive performance on the video-to-audio benchmark VGGSound (FD_PaSST: 59.98, IS_PANNs: 17.40, DeSync: 0.44) and the text-to-audio benchmark AudioCaps (FD_PANNs: 10.63, IS_PANNs: 12.62), matching or exceeding the performance of established latent-based methods. Our work demonstrates that intermediate compression is not a prerequisite for high-quality synthesis, offering a simpler and more scalable alternative for multimodal audio generation.", "date": "2026-05-18", "categories": []}
{"id": "2507.15970", "title": "CIS-BWE: Chaos-Informed Speech Bandwidth Extension", "abstract": "Recovering high-frequency components lost to bandwidth constraints is crucial for applications ranging from telecommunications to high-fidelity audio on limited resources. We introduce NDSI-BWE, a new adversarial Band Width Extension (BWE) framework that leverage four new discriminators inspired by nonlinear dynamical system to capture diverse temporal behaviors: a Multi-Resolution Lyapunov Discriminator (MRLD) for determining sensitivity to initial conditions by capturing deterministic chaos, a Multi-Scale Recurrence Discriminator (MS-RD) for self-similar recurrence dynamics, a Multi-Scale Detrended Fractal Analysis Discriminator (MSDFA) for long range slow variant scale invariant relationship, a Multi-Resolution Poincar\\'e Plot Discriminator (MR-PPD) for capturing hidden latent space relationship, a Multi-Period Discriminator (MPD) for cyclical patterns, a Multi-Resolution Amplitude Discriminator (MRAD) and Multi-Resolution Phase Discriminator (MRPD) for capturing intricate amplitude-phase transition statistics. By using depth-wise convolution at the core of the convolutional block with in each discriminators, NDSI-BWE attains an eight-times parameter reduction. These seven discriminators guide a complex-valued ConformerNeXt based genetor with a dual stream Lattice-Net based architecture for simultaneous refinement of magnitude and phase. The genertor leverage the transformer based conformer's global dependency modeling and ConvNeXt block's local temporal modeling capability. Across six objective evaluation metrics and subjective based texts comprises of five human judges, NDSI-BWE establishes a new SoTA in BWE.", "date": "2025-07-21", "categories": []}
{"id": "2605.14736", "title": "IsoNet: Spatially-aware audio-visual target speech extraction in complex acoustic environments", "abstract": "Target speech extraction remains difficult for compact devices because monaural neural models lack spatial evidence and classical beamformers lose resolving power when the microphone aperture is only a few centimetres. We present IsoNet, a user-selectable audio-visual target speech extraction system for a compact 4-microphone array. IsoNet combines complex multi-channel STFT features, GCC-PHAT spatial cues, face-conditioned visual embeddings, and auxiliary direction-of-arrival supervision inside a U-Net mask estimation network. Three curriculum variants were trained on 25,000 simulated VoxCeleb mixtures with progressively difficult SNR regimes. On a hard test set spanning -1 to 10 dB SNR, IsoNet-CL1 achieves 9.31 dB SI-SDR, a 4.85 dB improvement over the mixture, with PESQ 2.13 and STOI 0.84. Oracle delay-and-sum and MVDR beamformers degrade the same mixtures by 4.82 dB and 6.08 dB SI-SDRi, respectively, showing that the proposed learned multimodal conditioning solves a regime where conventional spatial filtering is ineffective. Ablation studies show consistent gains from visual conditioning, GCC-PHAT features, and extended delay-bin encoding. The results establish a compact-array, face-selectable speech extraction baseline under controlled simulation and identify the remaining barriers to real deployment, especially phase reconstruction, multi-interferer mixtures, and simulation-to-real transfer.", "date": "2026-05-14", "categories": []}
{"id": "2605.15831", "title": "Modeling Music as a Time-Frequency Image: A 2D Tokenizer for Music Generation", "abstract": "Autoregressive music generation depends strongly on the audio tokenizer. Existing high-fidelity codecs often use residual multi-codebook quantization, which preserves reconstruction quality but complicates language modeling after sequence flattening, as the residual hierarchy imposes strong sequential dependencies and can amplify error accumulation. We propose BandTok, a generation-oriented 2D Mel-spectrogram tokenizer that represents each frame with Mel-frequency band tokens from a single shared codebook. This design yields a physically interpretable time-frequency token grid with a more independent token structure, making it better suited for autoregressive modeling. BandTok improves reconstruction with a multi-scale PatchGAN objective and EMA codebook updates. We further introduce an autoregressive language model with 2D Rotary Position Embedding (2D RoPE) to preserve temporal and frequency-band structure during generation. Experiments show that BandTok improves over residual-codebook tokenizers and achieves strong results in a data-limited setting. The source code and generation demos for this work are publicly available.", "date": "2026-05-15", "categories": []}
{"id": "2605.15984", "title": "Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues", "abstract": "Toxic speech detection has become a crucial challenge in maintaining safe online communication environments. However, existing approaches to toxic speech detection often neglect the contribution of paralinguistic cues, such as emotion, intonation, and speech rate, which are key to detecting speech toxicity. Moreover, current toxic speech datasets are predominantly text-based, limiting the development of models that can capture paralinguistic cues.To address these challenges, we present ToxiAlert-Bench, a large-scale audio dataset comprising over 30,000 audio clips annotated with seven major toxic categories and twenty fine-grained toxic labels. Uniquely, our dataset annotates toxicity sources -- distinguishing between textual content and paralinguistic origins -- for comprehensive toxic speech analysis.Furthermore, we propose a dual-head neural network with a multi-stage training strategy tailored for toxic speech detection. This architecture features two task-specific classification headers: one for identifying the source of sensitivity (textual or paralinguistic), and the other for categorizing the specific toxic type. The training process involves independent head training followed by joint fine-tuning to reduce task interference. To mitigate data class imbalance, we incorporate class-balanced sampling and weighted loss functions.Our experimental results show that leveraging paralinguistic features significantly improves detection performance. Our method consistently outperforms existing baselines across multiple evaluation metrics, with a 21.1% relative improvement in Macro-F1 score and a 13.0% relative gain in accuracy over the strongest baseline, highlighting its enhanced effectiveness and practical applicability.", "date": "2026-05-15", "categories": []}
{"id": "2508.11845", "title": "AVEX: What Matters for Animal Vocalization Encoding", "abstract": "Bioacoustics, the study of sounds produced by living organisms, plays a vital role in conservation, biodiversity monitoring, and behavioral studies. Many tasks in this field, such as species, individual, and behavior classification and detection, are well-suited to machine learning. However, they often suffer from limited annotated data, highlighting the need for a general-purpose bioacoustic encoder capable of extracting useful representations for diverse downstream tasks. Such encoders have been proposed before, but are often limited in scope due to a focus on a narrow range of species (typically birds), and a reliance on a single model architecture or training paradigm. Moreover, they are usually evaluated on a small set of tasks and datasets. In this work, we present a large-scale empirical study that covers aspects of bioacoustics that are relevant to research but have previously been scarcely considered: training data diversity and scale, model architectures and training recipes, and the breadth of evaluation tasks and datasets. We obtain encoders that are state-of-the-art on the existing and proposed benchmarks. We also identify what matters for training these encoders, such that this work can be extended when more data are available or better architectures are proposed. Specifically, across 26 datasets with tasks including species classification, detection, individual ID, and vocal repertoire discovery, we find self-supervised pre-training followed by supervised post-training on a mixed bioacoustics + general-audio corpus yields the strongest in- and out-of-distribution performance. We show the importance of data diversity in both stages. To support ongoing research and application, we will release the model checkpoints.", "date": "2025-08-15", "categories": []}
{"id": "2511.21247", "title": "The Spheres Dataset: Multitrack Orchestral Recordings for Music Source Separation and Information Retrieval", "abstract": "This paper introduces The Spheres dataset, multitrack orchestral recordings designed to advance machine learning research in music source separation and related MIR tasks within the classical music domain. The dataset is composed of over one hour recordings of musical pieces performed by the Colibr\\`i Ensemble at The Spheres recording studio, capturing two canonical works - Tchaikovsky's Romeo and Juliet and Mozart's Symphony No. 40 - along with chromatic scales and solo excerpts for each instrument. The recording setup employed 23 microphones, including close spot, main, and ambient microphones, enabling the creation of realistic stereo mixes with controlled bleeding and providing isolated stems for supervised training of source separation models. In addition, room impulse responses were estimated for each instrument position, offering valuable acoustic characterization of the recording space. We present the dataset structure, acoustic analysis, and baseline evaluations using X-UMX based models for orchestral family separation and microphone debleeding. Results highlight both the potential and the challenges of source separation in complex orchestral scenarios, underscoring the dataset's value for benchmarking and for exploring new approaches to separation, localization, dereverberation, and immersive rendering of classical music.", "date": "2025-11-26", "categories": []}
{"id": "2512.03637", "title": "AaSP: Aliasing-aware Self-Supervised Pre-Training for Audio Spectrogram Transformers", "abstract": "Transformer-based audio self-supervised learning (SSL) models commonly use spectrograms, vision-style Transformers, and masked modeling objectives. However, convolutional patchification with temporal downsampling lowers the effective Nyquist frequency and introduces aliasing, while na\\\"ive low-pass filtering may remove task-relevant high-frequency cues. We present AaSP, an aliasing-aware self-supervised pre-training framework for audio spectrogram transformers. AaSP combines an aliasing-aware patch representation, teacher-student masked modeling, a cross-attention predictor, and multi-mask contrastive regularization to learn representations that integrate features from alias-prone modulation bands while remaining stable across masked views. Its patch-embedding module, Aliasing-aware Patch Embedding (AaPE), augments standard patch tokens with features from alias-prone modulation bands using a band-limited complex sinusoidal kernel with a two-sided exponential window. The kernel's frequency and decay parameters are estimated from the input, enabling adaptive subband analysis whose outputs are fused with standard patch tokens. We pre-train on AudioSet and evaluate the learned representations by fine-tuning and linear evaluation on acoustic/environmental, speech, and music recognition benchmarks. Under fine-tuning, the full AaSP framework achieves state-of-the-art results on AS-20K, ESC-50, and NSynth among compared self-supervised baselines, while remaining competitive elsewhere. Linear evaluation shows a similar trend, including gains on US8K and NSynth. Overall, AaSP learns representations that are more stable under aliasing-sensitive temporal perturbations and competitive for downstream transfer.", "date": "2025-12-03", "categories": []}
{"id": "2605.12534", "title": "BioSEN: A Bio-acoustic Signal Enhancement Network for Animal Vocalizations", "abstract": "Most work in audio enhancement targets human speech, while bioacoustics is less studied due to noisy recordings and the distinct traits of animal sounds. To fill this gap, we adapt speech enhancement methods and build BioSEN, a model made for bioacoustic signals. BioSEN has three modules: a multi-scale dual-axis attention unit for time-frequency feature extraction, a bio-harmonic multi-scale enhancement unit for capturing harmonic structures, and an\n  energy-adaptive gating connection unit that uses frequency weights to keep vocalizations from being removed as noise. Tests on three bioacoustic datasets show that BioSEN matches or exceeds state-of-the-art speech enhancement models while using far less computation. These results show BioSEN's strength for bioacoustic audio enhancement and its promise for biodiversity monitoring and conservation.", "date": "2026-05-02", "categories": []}
{"id": "2605.14031", "title": "Masked Autoencoders with Limited Data: Does It Work? A Fine-Grained Bioacoustics Case Study", "abstract": "Bioacoustic recognition requires fine-grained acoustic understanding to distinguish similar-sounding species. However, many large-scale data repositories such as iNaturalist are weakly annotated, often with only a single positive species label per recording, making supervised learning particularly challenging. Inspired by advances in computer vision, recent approaches have shifted toward self-supervised learning to capture the underlying structure of audio without relying on exhaustive annotations. In particular, masked autoencoders (MAE) have shown strong transferability on massive audio corpora, yet their effectiveness in more modest bioacoustic settings remains underexplored. In this work, we conduct a systematic study of MAE pretraining for species classification on iNatSounds, analyzing the impacts of pretraining data scale, domain specificity, data curation, and transfer strategies. Consistent with prior work, we find that models pretrained on diverse general audio data achieve the best transfer performance on iNatSounds. Contrary to observations from large-scale audio benchmarks, we find that (1) additional masked reconstruction pretraining on domain-specific data provides limited benefits and may even degrade performance relative to off-the-shelf models, and (2) selective data filtering offers a negligible advantage when the overall data scale is limited. Our results indicate that, in moderate-sized fine-grained bioacoustic settings, pretraining scale dominates objective design. These findings further clarify when MAE-based pretraining is effective and provide practical guidance for model selection under limited supervision.", "date": "2026-05-13", "categories": []}
{"id": "2605.14231", "title": "AudioMosaic: Contrastive Masked Audio Representation Learning", "abstract": "Audio self-supervised learning (SSL) aims to learn general-purpose representations from large-scale unlabeled audio data. While recent advances have been driven mainly by generative reconstruction objectives, contrastive approaches remain less explored, partly due to the difficulty of designing effective audio augmentations and the large batch sizes required for contrastive pre-training. We introduce \\textbf{AudioMosaic}, a contrastive learning-based audio encoder for general audio understanding. During pre-training, AudioMosaic constructs positive pairs by applying structured time-frequency masking to spectrogram patches, which reduces memory usage and enables efficient large-batch training. Compared with generative approaches, the AudioMosaic encoder learns more discriminative utterance-level representations that demonstrate strong transferability across datasets, domains, and acoustic conditions. Extensive experiments show that AudioMosaic achieves state-of-the-art performance on several standard audio benchmarks under both linear probing and fine-tuning. We further show that integrating the pretrained AudioMosaic encoder into audio-language models improves performance on audio-language tasks. The code is publicly available in our \\href{https://github.com/HanxunH/AudioMosaic}{GitHub repository}.", "date": "2026-05-14", "categories": []}
{"id": "2605.14555", "title": "Break-the-Beat! Controllable MIDI-to-Drum Audio Synthesis", "abstract": "Current methods for creating drum loop audio in digital music production, such as using one-shot samples or resampling, often demand non-trivial efforts of creators. While recent generative models achieve high fidelity and adhere to text, they lack the specific control needed for such a task. Existing symbolic-to-audio research often focuses on single, tonal instruments, leaving the challenge of polyphonic, percussive drum synthesis unaddressed. We address this gap by introducing ``Break-the-Beat!,'' a model capable of rendering a drum MIDI with the timbre of a reference audio. It is built by fine-tuning a pre-trained text-to-audio model with our proposed content encoder and a effective hybrid conditioning mechanism. To enable this, we construct a new dataset of paired target-reference drum audio from existing drum audio datasets. Experiments demonstrate that our model generates high-quality drum audio that follows high-resolution drum MIDI, achieving strong performance across metrics of audio quality, rhythmic alignment, and beat continuity. This offer producers a new, controllable tool for creative production. Demo page: https://ik4sumii.github.io/break-the-beat/", "date": "2026-05-14", "categories": []}
{"id": "2605.14896", "title": "Text-Dependent Speaker Verification (TdSV) Challenge 2024: Team Naive System Report", "abstract": "This paper presents a system for the 2024 Text-Dependent Speaker Verification (TdSV) Challenge. The system achieved a Minimum Detection Cost Function (MinDCF) of 0.0461 and an Equal Error Rate (EER) of 1.3\\%. Our approach focused on adapting existing state-of-the-art neural networks, ResNet-TDNN and NeXt-TDNN, originally trained on the VoxCeleb dataset. This strategy was chosen because of the limited challenge duration and the available resources at the time. In addition, we designed a lightweight and resource-efficient model, EfficientNet-A0, trained specifically on the challenge dataset to improve adaptation and strengthen the ensemble approach. Our system combines advanced neural architectures, extensive data augmentation, and optimised hyperparameters. These components helped achieve strong performance in text-dependent speaker verification. The results also demonstrate the effectiveness of multi-model ensemble learning for both speaker and phrase verification.", "date": "2026-05-14", "categories": []}
{"id": "2605.15044", "title": "SpeakerLLM: A Speaker-Specialized Audio-LLM for Speaker Understanding and Verification Reasoning", "abstract": "As audio-first agents become increasingly common in physical AI, conversational robots, and screenless wearables, audio large language models (audio-LLMs) must integrate speaker-specific understanding to support user authorization, personalization, and context-aware interaction. This requires modeling who is speaking, how the voice sounds, and how recording conditions affect speaker cues. Conventional speaker verification systems provide strong scalar scores but little linguistic evidence, while current audio-LLMs and speaker-aware language models have limited ability to organize speaker information beyond binary labels or descriptive profiles. We present SpeakerLLM, a speaker-specialized audio-LLM framework that unifies single-utterance speaker profiling, recording-condition understanding, utterance-pair speaker comparison, and evidence-organized verification reasoning within a natural-language interface. We construct verification-reasoning targets and a decision-composition policy that separate profile-level evidence from the final same-or-different decision and organize recording condition, profile evidence, and the decision into a structured trace. At its core, SpeakerLLM uses a hierarchical speaker tokenizer designed to capture multiple granularities of speaker evidence. Utterance-level speaker embeddings summarize identity and profile-level cues, whereas frame-level speaker features preserve fine-grained acoustic descriptors. Experiments show that SpeakerLLM-Base improves speaker-profile and recording-condition understanding over general audio-LLMs, while SpeakerLLM-VR preserves strong generated-verdict accuracy and produces decision traces grounded in the supervised verification reasoning schema. We will release the metadata-enriched supervision dataset and target-construction code for reproducibility.", "date": "2026-05-14", "categories": []}
{"id": "2411.15913", "title": "Repurposing Image Diffusion Models for Training-Free Music Style Transfer on Mel-spectrograms", "abstract": "Music style transfer blends source structure with reference style to enable personalized music creation. However, existing zero-shot methods often struggle to capture fine-grained audio nuances, relying on coarse text descriptions or requiring expensive task-specific training. We propose Stylus, a training-free framework that repurposes pretrained image diffusion models for music style transfer in the Mel-spectrogram domain. By treating audio as structured time-frequency images, Stylus manipulates self-attention by injecting style keys and values while preserving source structural queries. To ensure high fidelity, we introduce a phase-preserving reconstruction strategy to mitigate spectrogram inversion artifacts, alongside a classifier-free-guidance-inspired control for adjustable stylization. Extensive evaluations including 2,925 human ratings demonstrate that Stylus outperforms state-of-the-art baselines, achieving 34.1% higher content preservation and 25.7% better perceptual quality. Our work validates that generic image priors can be effectively leveraged for the training-free transformation of structured Mel-spectrograms. Code and materials are available at https://github.com/Sooyyoungg/Stylus.git.", "date": "2024-11-24", "categories": []}
{"id": "2502.20427", "title": "DeePen: Penetration Testing for Audio Deepfake Detection", "abstract": "Deepfakes - manipulated or forged audio and video media - pose significant security risks to individuals, organizations, and society at large. To address these challenges, machine learning-based classifiers are commonly employed to detect deepfake content. In this paper, we assess the robustness of such classifiers through a systematic penetration testing methodology, which we introduce as DeePen. Our approach operates without prior knowledge of or access to the target deepfake detection models. Instead, it leverages a set of carefully selected signal processing modifications - referred to as attacks - to evaluate model vulnerabilities. Using DeePen, we analyze both real-world production systems and publicly available academic model checkpoints, demonstrating that all tested systems exhibit weaknesses and can be reliably deceived by simple manipulations such as time-stretching or echo addition. Furthermore, our findings reveal that while some attacks can be mitigated by retraining detection systems with knowledge of the specific attack, others remain persistently effective.", "date": "2025-02-27", "categories": []}
{"id": "2603.02245", "title": "LMU-Based Sequential Learning and Posterior Ensemble Fusion for Cross-Domain Infant Cry Classification", "abstract": "Decoding infant cry causes remains challenging for healthcare monitoring due to short nonstationary signals, limited annotations, and strong domain shifts across infants and datasets. We propose a compact acoustic framework that fuses mel-frequency cepstral coefficients (MFCCs), short-time Fourier transform (STFT) features, and fundamental-frequency (F0) contours within a multi-branch convolutional neural network (CNN) encoder, and models temporal dynamics using an enhanced Legendre Memory Unit (LMU). Compared to LSTMs, the LMU backbone provides stable sequence modeling with substantially fewer recurrent parameters, supporting efficient deployment. To improve cross-dataset generalization, we introduce calibrated posterior ensemble fusion with entropy-gated weighting to preserve domain-specific expertise while mitigating dataset bias. Experiments on Baby2020 and Baby Crying demonstrate improved macro-F1 under cross-domain evaluation, along with leakage aware splits and real-time feasibility for on-device monitoring.", "date": "2026-02-24", "categories": []}
{"id": "2605.06309", "title": "MultiLinguahah : A New Unsupervised Multilingual Acoustic Laughter Segmentation Method", "abstract": "Laughter is a social non-vocalization that is universal across cultures and languages, and is crucial for human communication, including social bonding and communication signaling. However, detecting laughter in audio is a challenging task, and segmenting is even more difficult. Currently, Machine Learning methods generally rely on costly manual annotation, and their datasets are mostly based on English contexts. Thus, we propose an unsupervised multilingual method that sets up the laughter segmentation task as an anomaly detection of energy-based segmented audio sequences. Our method applies an Isolation Forest on audio representations learned from BYOL-A encoder. We compare our method with several state-of-the-art laughter detection algorithms on four datasets, including stand-up comedy, sitcoms, and general short audio from AudioSet. Our results show that state-of-the-art methods are not optimized for multilingual contexts, while our method outperforms them in non-English settings.", "date": "2026-05-07", "categories": []}
{"id": "2605.13651", "title": "NAACA: Training-Free NeuroAuditory Attentive Cognitive Architecture with Oscillatory Working Memory for Salience-Driven Attention Gating", "abstract": "Audio provides critical situational cues, yet current Audio Language Models (ALMs) face an attention bottleneck in long-form recordings where dominant background patterns can dilute rare, salient events. We introduce NAACA, a training-free NeuroAuditory Attentive Cognitive Architecture that reframes attention allocation as an auditory salience filtering problem. At its core is OWM, a neuro-inspired Oscillatory Working Memory that maintains stable attractor-like states and triggers higher-cognition ALM processing only when adaptive energy fluctuations signal perceptual salience, triggering higher-level reasoning. On XD-Violence, NAACA improves AudioQwen's average precision (AP) from 53.50% to 70.60% while reducing unnecessary ALM invocations. Furthermore, qualitative case studies on the Urban Soundscapes of the World (USoW) dataset show that OWM captures novel events and subcategory shifts while remaining robust to transient pauses and ambient urban noise.", "date": "2026-05-13", "categories": []}
{"id": "2605.11192", "title": "Exploring Token-Space Manipulation in Latent Audio Tokenizers", "abstract": "Neural audio codecs provide compact discrete representations for speech generation and manipulation. However, most codecs organize tokens as frame-level sequences, making it difficult to study or intervene on global factors of variation. In this work, we propose the Latent Audio Tokenizer for Token-space Editing (LATTE) that appends a fixed set of learnable latent tokens to the audio feature sequence and retains only these tokens for quantization and decoding. This design produces a compact, non-temporally aligned bottleneck in which each token can aggregate global information across the full utterance. We show that the resulting tokenizer preserves competitive reconstruction quality in low-bitrate speech coding settings while enabling simple token-space interventions. In particular, we find that swapping selected latent token positions between utterances can modify global attributes, such as speaker identity and background noise, and we evaluate these interventions on voice conversion and denoising tasks. Our results suggest that compact latent audio tokenizers can support controllable audio manipulation without supervision in task-specific editing models.", "date": "2026-05-11", "categories": []}
{"id": "2605.12135", "title": "STRUM: A Spectral Transcription and Rhythm Understanding Model for End-to-End Generation of Playable Rhythm-Game Charts", "abstract": "We present STRUM (Spectral Transcription and Rhythm Understanding Model), an audio-to-chart pipeline that converts raw recordings into playable Clone Hero / YARG charts for drums, guitar, bass, vocals, and keys without any oracle metadata. STRUM is a multi-stage hybrid: a two-stage CRNN onset detector and a six-model ensemble classifier for drums; neural onset detectors with monophonic pitch tracking for guitar and bass; word-aligned ASR for vocals; and spectral keyboard detection for keys. We evaluate on a 30-song in-envelope benchmark constructed by screening candidate songs on a single audio-quality criterion -- the median 1-second drum-stem RMS after htdemucs_6s source separation. On this benchmark STRUM achieves drums onset F1 = 0.838, bass F1 = 0.694, guitar F1 = 0.651, and vocals F1 = 0.539 at a +/- 100 ms tolerance with per-song global offset search. We report a complete ablation of seven drum-pipeline components with paired per-song Wilcoxon tests, an analysis of ground-truth-to-audio timing distributions in community Clone Hero charts, and a per-class confusion matrix for the drum classifier. Code, model weights, and the full benchmark manifest are released.", "date": "2026-05-12", "categories": []}
{"id": "2505.18184", "title": "AI- Enhanced Stethoscope in Remote Diagnostics for Cardiopulmonary Diseases", "abstract": "The increase in cardiac and pulmonary diseases presents an alarming and pervasive health challenge on a global scale responsible for unexpected and premature mortalities. In spite of how serious these conditions are, existing methods of detection and treatment encounter challenges, particularly in achieving timely diagnosis for effective medical intervention. Manual screening processes commonly used for primary detection of cardiac and respiratory problems face inherent limitations, increased by a scarcity of skilled medical practitioners in remote or under-resourced areas. To address this, our study introduces an innovative yet efficient model which integrates AI for diagnosing lung and heart conditions concurrently using the auscultation sounds. Unlike the already high-priced digital stethoscope, our proposed model has been particularly designed to deploy on low-cost embedded devices and thus ensure applicability in under-developed regions that actually face an issue of accessing medical care. Our proposed model incorporates MFCC feature extraction and engineering techniques to ensure that the signal is well analyzed for accurate diagnostics through the hybrid model combining Gated Recurrent Unit with CNN in processing audio signals recorded from the low-cost stethoscope. Beyond its diagnostic capabilities, the model generates digital audio records that facilitate in classifying six pulmonary and five cardiovascular diseases. Hence, the integration of a cost effective stethoscope with an efficient AI empowered model deployed on a web app providing real-time analysis, represents a transformative step towards standardized healthcare", "date": "2025-05-18", "categories": []}
{"id": "2507.23511", "title": "MECAT: A Multi-Experts Constructed Benchmark for Fine-Grained Audio Understanding Tasks", "abstract": "While large audio-language models have advanced open-ended audio understanding, they still fall short of nuanced human-level comprehension. This gap persists largely because current benchmarks, limited by data annotations and evaluation metrics, fail to reliably distinguish between generic and highly detailed model outputs. To this end, this work introduces MECAT, a Multi-Expert Constructed Benchmark for Fine-Grained Audio Understanding Tasks. Generated via a pipeline that integrates analysis from specialized expert models with Chain-of-Thought large language model reasoning, MECAT provides multi-perspective, fine-grained captions and open-set question-answering pairs. The benchmark is complemented by a novel metric: DATE (Discriminative-Enhanced Audio Text Evaluation). This metric penalizes generic terms and rewards detailed descriptions by combining single-sample semantic similarity with cross-sample discriminability. A comprehensive evaluation of state-of-the-art audio models is also presented, providing new insights into their current capabilities and limitations. The data and code are available at https://github.com/xiaomi-research/mecat", "date": "2025-07-31", "categories": []}
{"id": "2510.19414", "title": "EchoFake: A Replay-Aware Dataset for Practical Speech Deepfake Detection", "abstract": "The growing prevalence of speech deepfakes has raised serious concerns, particularly in real-world scenarios such as telephone fraud and identity theft. While many anti-spoofing systems have demonstrated promising performance on lab-generated synthetic speech, they often fail when confronted with physical replay attacks-a common and low-cost form of attack used in practical settings. Our experiments show that models trained on existing datasets exhibit severe performance degradation, with average accuracy dropping to 59.6% when evaluated on replayed audio. To bridge this gap, we present EchoFake, a comprehensive dataset comprising more than 120 hours of audio from over 13,000 speakers, featuring both cutting-edge zero-shot text-to-speech (TTS) speech and physical replay recordings collected under varied devices and real-world environmental settings. Additionally, we evaluate three baseline detection models and show that models trained on EchoFake achieve lower average EERs across datasets, indicating better generalization. By introducing more practical challenges relevant to real-world deployment, EchoFake offers a more realistic foundation for advancing spoofing detection methods.", "date": "2025-10-22", "categories": []}
{"id": "2601.02954", "title": "The World is Not Mono: Enabling Spatial Understanding in Large Audio-Language Models", "abstract": "Large audio-language models have made rapid progress in recognizing what is present in an audio clip, but spatial audio-language understanding still lacks a clear task interface. A model must also decide where sound events occur, which semantic and spatial attributes belong to the same auditory object, how multiple objects are arranged, and whether a scene-level answer is physically plausible. We formalize this capability as audio scene analysis (ASA), a three-level problem spanning atomic perception, relational integration, and cognitive reasoning. We propose The World is Not Mono (TWNM), a framework that equips audio-language models with explicit spatial evidence. TWNM uses physically grounded First-Order Ambisonics (FOA) simulation for controllable supervision, learns slot-regularized spatial representations from multichannel audio, fuses them with semantic audio features, and trains with a progressive curriculum ending in preference optimization over metadata-derived answers and auxiliary format/evidence rewards. To operationalize ASA, we build a controlled benchmark from scene metadata, covering localization, attribute binding, spatial comparison, scene abduction, and counterfactual reasoning. On this benchmark, TWNM achieves 70.8% overall accuracy, 66.4% on spatial-family tasks, and 79.76% on mixed L3 scene-level multiple-choice QA. We also audit monaural and binaural reference systems as diagnostic references with explicit audit labels, since they differ in spatial input, training interface, and output format. The supported claim is that a clearly defined ASA hierarchy, FOA-conditioned spatial representations, and metadata-grounded training enable controlled, auditable spatial audio-language reasoning, with STARSS23 providing a limited real-recording diagnostic.", "date": "2026-01-06", "categories": []}
{"id": "2601.12248", "title": "AQUA-Bench: Beyond Finding Answers to Knowing When There Are None in Audio Question Answering", "abstract": "Recent advances in audio-aware large language models have shown strong performance on audio question answering. However, existing benchmarks mainly cover answerable questions and overlook the challenge of unanswerable ones, where no reliable answer can be inferred from the audio. Such cases are common in real-world settings, where questions may be misleading, ill-posed, or incompatible with the information. To address this gap, we present AQUA-Bench, a benchmark for Audio Question Unanswerability Assessment. It systematically evaluates three scenarios: Absent Answer Detection (the correct option is missing), Incompatible Answer Set Detection (choices are categorically mismatched with the question), and Incompatible Audio Question Detection (the question is irrelevant or lacks sufficient grounding in the audio). By assessing these cases, AQUA-Bench offers a rigorous measure of model reliability and promotes the development of audio-language systems that are more robust and trustworthy. Our experiments suggest that while models excel on standard answerable tasks, they often face notable challenges with unanswerable ones, pointing to a blind spot in current audio-language understanding.", "date": "2026-01-18", "categories": []}
{"id": "2603.09007", "title": "Gender Fairness in Audio Deepfake Detection: Performance and Disparity Analysis", "abstract": "Audio deepfake detection aims to detect real human voices from those generated by Artificial Intelligence (AI) and has emerged as a significant problem in the field of voice biometrics systems. With the ever-improving quality of synthetic voice, the probability of such a voice being exploited for illicit practices like identity thest and impersonation increases. Although significant progress has been made in the field of Audio Deepfake Detection in recent times, the issue of gender bias remains underexplored and in its nascent stage In this paper, we have attempted a thorough analysis of gender dependent performance and fairness in audio deepfake detection models. We have used the ASVspoof 5 dataset and train a ResNet-18 classifier and evaluate detection performance across four different audio features, and compared the performance with baseline AASIST model. Beyond conventional metrics such as Equal Error Rate (EER %), we incorporated five established fairness metrics to quantify gender disparities in the model. Our results show that even when the overall EER difference between genders appears low, fairness-aware evaluation reveals disparities in error distribution that are obscured by aggregate performance measures. These findings demonstrate that reliance on standard metrics is unreliable, whereas fairness metrics provide critical insights into demographic-specific failure modes. This work highlights the importance of fairness-aware evaluation for developing a more equitable, robust, and trustworthy audio deepfake detection system.", "date": "2026-03-09", "categories": []}
{"id": "2605.08750", "title": "Communicating Sound Through Natural Language", "abstract": "Natural language is widely used to describe, prompt, and control audio systems, but rarely serves as the representation carrying audio itself. We introduce lexical acoustic coding (LAC), a framework in which pre-trained LLM sender and receiver agents transmit sound through natural language. Under fixed system prompts, the agents write their own analysis and synthesis code, communicating only through a lexical sentence, shared vocabulary, and optional symbolic music structure. The sender analyzes an input waveform into interpretable, non-learned acoustic descriptors, quantizes each with a feature-specific interval vocabulary, and verbalizes the lexical code as English. The receiver parses the sentence back into lexical-acoustic constraints and renders a waveform through closed-loop refinement. The transmitted text serves as both a rich caption and as the transport representation itself. We frame LAC as a finite-rate lossy quantizer, exposing trade-offs between vocabulary size, rate, and fidelity. Experiments on short sounds and symbolic music transfer show that plain text preserves measurable acoustic structure while remaining interpretable, editable, and native to LLM-mediated communication.", "date": "2026-05-09", "categories": []}
{"id": "2605.09087", "title": "Towards Trustworthy Audio Deepfake Detection: A Systematic Framework for Diagnosing and Mitigating Gender Bias", "abstract": "Audio deepfake detection systems are increasingly deployed in high-stakes security applications, yet their fairness across demographic groups remains critically underexamined. Prior work measures gender disparity but does not investigate where it comes from or how to fix it systematically. We present the first diagnosis-first framework that identifies bias source before applying targeted mitigation, evaluated on two models, AASIST and Wav2Vec2+ResNet18, on ASVSpoof5. Our diagnosis shows that bias does not stem from imbalanced training data but from acoustic representation differences, gender leakage in learned features, and structural evaluation asymmetry. We test mitigation strategies across in-processing, post-processing and combined families, including novel methods introduced in this work. Adjusting the decision threshold separately per gender reduces unfairness by 54% to 75% at no cost to detection accuracy, and our new epoch-level fairness regularisation method outperforms existing per-batch approaches. Adversarial debiasing succeeds only when gender leakage is localised, and fails when it is diffuse, an outcome correctly predicted by our diagnosis before training. No single method fully closes the fairness gap, confirming that bias sources must be identified before fixes are applied and that fairer benchmark design is equally important", "date": "2026-05-09", "categories": []}
{"id": "2605.09570", "title": "End-to-End Keyword Spotting on FPGA Using Graph Neural Networks with a Neuromorphic Auditory Sensor", "abstract": "With the rapid growth of mobile robotics and embedded intelligence, there is an increasing demand for efficient on-device data processing on edge platforms. A promising research direction is the use of neuromorphic sensors inspired by human sensory systems, which generate sparse, event-based data encoding changes in the environment. In this work, we present the first end-to-end FPGA implementation of a keyword spotting system that integrates a Neuromorphic Auditory Sensor (NAS) and a graph neural network (GNN) on a single FPGA device, enabling real-time processing of raw audio data. The proposed architecture eliminates conventional signal preprocessing and operates directly on event-based audio streams. Leveraging a compute-near-memory network architecture, the system achieves efficient inference with low latency and low power consumption. Experimental results demonstrate an accuracy of 87.43% after quantization on the Google Speech Commands v2 dataset processed through the neuromorphic sensor, with end-to-end latency below 35 us and average power consumption of 1.12 W. The processed datasets, software models, and hardware modules are available at https://github.com/vision-agh/NAS-GNN-KWS.", "date": "2026-05-10", "categories": []}
{"id": "2605.09908", "title": "Voice Biomarkers for Depression and Anxiety", "abstract": "Current approaches to detecting depression and anxiety from speech primarily rely on machine learning techniques that utilize hand-engineered paralinguistic features and related acoustic descriptors derived from time- and frequency-domain representations of speech signals. Applying deep learning methods directly to raw speech signals has the potential to produce biomarker representations with substantially greater predictive power. However, these approaches typically require large volumes of carefully annotated data to learn robust and clinically meaningful representations of the underlying biomarkers. In this paper, we describe our efforts toward developing a deep learning model trained on a large-scale proprietary dataset comprising ~65,000 utterances collected from more than 23,000 subjects representative of relevant United States demographics. We present the techniques employed and analyze their impact on model performance. Our results demonstrate that the proposed models can extract content-agnostic biomarker information, which, when combined with lexical features extracted from audio, yields improved predictive performance in production settings. Our models are evaluated on ~5000 unique subjects and achieve performance of 71% in terms of sensitivity and specificity. To foster further research in mental health assessment from speech, we release the best-performing model described in this paper on HuggingFace.", "date": "2026-05-11", "categories": []}
{"id": "2605.10084", "title": "PoDAR: Power-Disentangled Audio Representation for Generative Modeling", "abstract": "The performance of audio latent diffusion models is primarily governed by generator expressivity and the modelability of the underlying latent space. While recent research has focused primarily on the former, as well as improving the reconstruction fidelity of audio codecs, we demonstrate that latent modelability can be significantly improved through explicit factor disentanglement. We present PoDAR (Power-Disentangled Audio Representation), a framework that utilizes a randomized power augmentation and latent consistency objective to decouple signal power from invariant semantic content. This factorization makes the latent space easier to model, which both accelerates the convergence of downstream generative models and improves final overall performance. When applied to a Stable Audio 1.0 VAE with an F5-TTS generator, PoDAR achieves about a $2\\times$ acceleration in convergence to match baseline performance, while increasing final speaker similarity by 0.055 and UTMOS by 0.22 on the LibriSpeech-PC dataset. Furthermore, isolating power into dedicated channels enables the application of CFG exclusively to power-invariant content, effectively extending the stable guidance regime to higher scales.", "date": "2026-05-11", "categories": []}
{"id": "2605.10153", "title": "APEX: Audio Prototype EXplanations for Classification Tasks", "abstract": "Explainable AI (XAI) has achieved remarkable success in image classification, yet the audio domain lacks equally mature solutions. Current methods apply vision-based attribution techniques to spectrograms, overlooking fundamental differences between visual and acoustic signals. While prototype reasoning is promising, acoustic similarity remains multidimensional. We introduce APEX (Audio Prototype EXplanations), a post-hoc framework for interpreting pre-trained audio classifiers. Crucially, APEX requires no fine-tuning of the original backbone and strictly preserves output invariance. APEX disentangles explanations into four perspectives: Square-based prototypes to localize transient events, Time-based for temporal patterns, Frequency-based highlighting spectral bands, and Time-Frequency-based integrating both. This yields intuitive, example-based explanations that respect acoustic properties, providing greater semantic clarity than standard gradient-based methods.", "date": "2026-05-11", "categories": []}
{"id": "2605.10256", "title": "A Cold Diffusion Approach for Percussive Dereverberation", "abstract": "Most recent advances in audio dereverberation focus almost exclusively on speech, leaving percussive and drum signals largely unexplored despite their importance in music production. Percussive dereverberation poses distinct challenges due to sharp transients and dense temporal structure. In this work, we propose a cold diffusion framework for dereverberating stereo drum stems (downmixes), modeling reverberation as a deterministic degradation process that progressively transforms anechoic signals into reverberant ones. We investigate two reverse-process parameterizations, Direct (next-state) and a Delta-normalized residual (velocity-style) prediction, and implement the framework using both a UNet and a diffusion Transformer backbone. The models are trained and evaluated on curated datasets comprising both acoustic and electronic drum recordings, with reverberation generated using a combination of synthetic and real room impulse responses. Extensive experiments on in-domain and fully out-of-domain test sets demonstrate that the proposed method consistently outperforms strong score-based and conditional diffusion baselines, evaluated using signal-based and perceptual metrics tailored to percussive audio.", "date": "2026-05-11", "categories": []}
{"id": "2605.10281", "title": "Drum Synthesis from Expressive Drum Grids via Neural Audio Codecs", "abstract": "Generating realistic drum audio directly from symbolic representations is a challenging task at the intersection of music perception and machine learning. We propose a system that transforms an expressive drum grid, a time-aligned MIDI representation with microtiming and velocity information, into drum audio by predicting discrete codes of a neural audio codec. Our approach uses a Transformer-based model to map the drum grid input to a sequence of codec tokens, which are then converted to waveform audio via a pre-trained codec decoder. We experiment with multiple state-of-the-art neural codecs, namely EnCodec, DAC, and X-Codec, to assess how the choice of audio representation impacts the quality of the generated drums. The system is trained and evaluated on the Expanded Groove MIDI Dataset, E-GMD, a large collection of human drum performances with paired MIDI and audio. We evaluate the fidelity and musical alignment of the generated audio using objective metrics. Overall, our results establish codec-token prediction as an effective route for drum grid-to-audio generation and provide practical insights into selecting audio tokenizers for percussive synthesis.", "date": "2026-05-11", "categories": []}
{"id": "2605.10494", "title": "Multi-layer attentive probing improves transfer of audio representations for bioacoustics", "abstract": "Probing heads map the representations learned from audio by a machine learning model to downstream task labels and are a key component in evaluating representation learning. Most bioacoustic benchmarks use a fixed, low-capacity probe, such as a linear layer on the final encoder layer. While this standardization enables model comparisons, it may bias results by overlooking the interaction between encoder features and probe design. In this work, we systematically study different probing strategies across two bioacoustic benchmarks, BEANs and BirdSet. We evaluate last- and multi-layer probing, across linear and attention probes. We show that larger probe heads that leverage time information have superior performance. Our results suggest that current benchmarks may misrepresent encoder quality when relying on a last-layer probing setup. Multi-layer probing improves downstream task performance across all tested models, while attention probing has superior performance to linear probing for transformer models.", "date": "2026-05-11", "categories": []}
{"id": "2605.07903", "title": "BeeVe: Unsupervised Acoustic State Discovery in Honey Bee Buzzing", "abstract": "Discovering structure in biological signals without supervision is a fundamental problem in computational intelligence, yet existing bioacoustic methods assume vocal production models or predefined semantic units, leaving non-vocal species poorly served. This work introduces BeeVe, an unsupervised framework for acoustic state discovery in collective honey bee buzzing. BeeVe uses the self-supervised Patchout Spectrogram Transformer (PaSST) as a frozen feature extractor, then trains a Vector-Quantized Variational Autoencoder (VQ-VAE) without labels on those embeddings, learning a finite discrete codebook of acoustic tokens directly from unlabelled hive audio. No labels, pretext tasks, or contrastive objectives are used at any stage. Post-hoc evaluation against known queen status reveals that the learned tokens separate queenright and queenless conditions with Jensen-Shannon Divergence values between 0.609 and 0.688, and that the queenless condition further decomposes into three internally coherent sub-states stable across experiments with different codebook sizes and random seeds. Token transition analysis confirms non-random sequential structure (p << 0.001) across all experiments. Generalisation to unseen recordings preserves both token overlap (Jaccard = 0.947) and global manifold topology. These results demonstrate that unsupervised discrete codebook learning can recover repeatable acoustic structure from a non-vocal biological signal without annotation, opening a path toward non-invasive acoustic hive health monitoring.", "date": "2026-05-08", "categories": []}
{"id": "2601.20362", "title": "Switchcodec: Adaptive residual-expert sparse quantization for high-fidelity neural audio coding", "abstract": "Recent neural audio compression models often rely on residual vector quantization for high-fidelity coding, but using a fixed number of per-frame codebooks is suboptimal for the wide variability of audio content-especially for signals that are either very simple or highly complex. To address this limitation, we propose SwitchCodec, a neural audio codec based on Residual Experts Vector Quantization (REVQ). REVQ combines a shared quantizer with dynamically routed expert quantizers that are activated according to the input audio, decoupling bitrate from codebook capacity and improving compression efficiency. This design ensures full training and utilization of each quantizer. In addition, a variable-bitrate mechanism adjusts the number of active expert quantizers at inference, enabling multi-bitrate operation without retraining. Experiments demonstrate that SwitchCodec surpasses existing baselines on both objective metrics and subjective listening tests.", "date": "2026-01-28", "categories": []}
{"id": "2604.20229", "title": "Enhancing Speaker Verification with Whispered Speech via Post-Processing", "abstract": "Speaker verification is a task of confirming an individual's identity through the analysis of their voice. Whispered speech differs from phonated speech in acoustic characteristics, which degrades the performance of speaker verification systems in real-life scenarios, including avoiding fully phonated speech to protect privacy, disrupt others, or when the lack of full vocalization is dictated by a disease. In this paper we propose a model with a training recipe to obtain more robust representations against whispered speech hindrances. The proposed system employs an encoder--decoder structure built atop a fine-tuned speaker verification backbone, optimized jointly using cosine similarity--based classification and triplet loss. We gain relative improvement of 22.26\\% compared to the baseline (baseline 6.77\\% vs ours 5.27\\%) in normal vs whispered speech trials, achieving AUC of 98.16\\%. In tests comparing whispered to whispered, our model attains an EER of 1.88\\% with AUC equal to 99.73\\%, which represents a 15\\% relative enhancement over the prior leading ReDimNet-B2. We also offer a summary of the most popular and state-of-the-art speaker verification models in terms of their performance with whispered speech. Additionally, we evaluate how these models perform under noisy audios, obtaining that generally the same relative level of noise degrades the performance of speaker verification more significantly on whispered speech than on normal speech.", "date": "2026-04-22", "categories": []}
{"id": "2605.04505", "title": "JASTIN: Aligning LLMs for Zero-Shot Audio and Speech Evaluation via Natural Language Instructions", "abstract": "The rapid advancement of generative audio models has outpaced the development of robust evaluation methodologies. Existing objective metrics and general multimodal large language models (MLLMs) often struggle with domain generalization, zero-shot capabilities, and instructional flexibility. To address these bottlenecks, we propose JASTIN, a generalizable, instruction-driven audio evaluation framework that formulates audio assessment as a self-instructed reasoning task. JASTIN bridges a frozen high-performance audio encoder with a fine-tuned LLM backbone via a trainable audio adapter. To ensure robust zero-shot generalization, we introduce a comprehensive instruction following data preparation pipeline, incorporating Multi-Source, Multi-Task, Multi-Calibration, and Multi-Description data. Experimental results demonstrate that JASTIN achieves state-of-the-art Pearson and Spearman correlations with human subjective ratings. It consistently outperforms general MLLMs across speech, sound, music, and out-of-domain evaluation tasks without the need for task-specific retraining.", "date": "2026-05-06", "categories": []}
{"id": "2605.04547", "title": "Stage-adaptive audio diffusion modeling", "abstract": "Recent progress in diffusion-based audio generation and restoration has substantially improved performance across heterogeneous conditioning regimes, including text-conditioned audio generation and audio-conditioned super-resolution. However, training audio diffusion models remains computationally expensive, and most existing pipelines still rely on static optimization recipes that treat the relative importance of training signals as fixed throughout learning. In this work, we argue that a major source of inefficiency lies in the evolving balance between semantic acquisition and generation-oriented refinement. Early training places stronger emphasis on acquiring condition-aligned semantic structure and coarse global organization, whereas later training increasingly emphasizes temporal consistency, perceptual fidelity, and fine-detail refinement. To characterize this evolving balance, we introduce a progress-based regime variable derived from the training-time slope of an SSL-space discrepancy, which measures semantic progress during training. Based on this signal, we develop three complementary stage-aware mechanisms: decayed SSL guidance for early semantic bootstrapping, self-adaptive timestep sampling driven by the regime variable, and structure-aware regularization activated from convergent grouped organization in parameter space. We evaluate these mechanisms on text-conditioned audio generation and audio-conditioned super-resolution. Across both settings, the proposed stage-aware strategies improve convergence behavior and yield gains on the primary generation and spectral reconstruction metrics over standard static baselines. These results support the view that efficient audio diffusion training can benefit from treating external guidance, internal organization, and optimization emphasis as stage-dependent components rather than fixed ingredients.", "date": "2026-05-06", "categories": []}
{"id": "2605.04556", "title": "Benchmarking LLMs on the Massive Sound Embedding Benchmark (MSEB)", "abstract": "The Massive Sound Embedding Benchmark (MSEB) has emerged as a standard for evaluating the functional breadth of audio models. While initial baselines focused on specialized encoders, the shift toward \"audio-native\" Large Language Models (LLMs) suggests a new paradigm where a single multimodal backbone may replace complex, task-specific pipelines. This paper provides a rigorous empirical evaluation of leading LLMs - including members from the Gemini and GPT families - across the eight core MSEB capabilities to assess their efficacy and audio-text parity. Our results indicate that while a significant modality gap persists regarding performance and robustness, the empirical evidence for an \"optimal\" modeling approach remains inconclusive. Ultimately, the choice between audionative and cascaded architectures depends heavily on specific use-case requirements and the underlying assumptions regarding latency, cost, and reasoning depth.", "date": "2026-05-06", "categories": []}
{"id": "2508.03448", "title": "SonicMaster: Towards Controllable All-in-One Music Restoration and Mastering", "abstract": "Music recordings often suffer from audio quality issues such as excessive reverberation, distortion, clipping, tonal imbalances, and a narrowed stereo image, especially when created in non-professional settings without specialized equipment or expertise. These problems are typically corrected using separate specialized tools and manual adjustments. In this paper, we introduce SonicMaster, the first unified generative model for music restoration and mastering that addresses a broad spectrum of audio artifacts with text-based control. SonicMaster is conditioned on natural language instructions to apply targeted enhancements, or can operate in an automatic mode for general restoration. To train this model, we construct the SonicMaster dataset, a large dataset of paired degraded and high-quality tracks by simulating common degradation types with nineteen degradation functions belonging to five enhancements groups: equalization, dynamics, reverb, amplitude, and stereo. Our approach leverages a flow-matching generative training paradigm to learn an audio transformation that maps degraded inputs to their cleaned, mastered versions guided by text prompts. Objective audio quality metrics demonstrate that SonicMaster significantly improves sound quality across all artifact categories. Furthermore, subjective listening tests confirm that listeners prefer SonicMaster's enhanced outputs over other baselines.", "date": "2025-08-05", "categories": []}
{"id": "2602.00594", "title": "Kanade: A Simple Disentangled Tokenizer for Spoken Language Modeling", "abstract": "A good language model starts with a good tokenizer. Tokenization is especially important for speech modeling, which must handle continuous signals that mix linguistic and non-linguistic information. A speech tokenizer should extract phonetics and prosody, suppress linguistically irrelevant information like speaker identity, and enable high-quality synthesis. We present Kanade, a single-layer disentangled speech tokenizer that realizes this ideal. Kanade separates out acoustic constants to create a single stream of tokens that captures rich phonetics and prosody. It does so without the need for auxiliary methods that existing disentangled codecs often rely on. Experiments show that Kanade achieves state-of-the-art speaker disentanglement and lexical availability, while maintaining excellent reconstruction quality.", "date": "2026-01-31", "categories": []}
{"id": "2603.06542", "title": "RAMoEA-QA: Hierarchical Specialization for Robust Respiratory Audio Question Answering", "abstract": "Conversational generative AI is increasingly explored in healthcare, where models must integrate heterogeneous patient signals and support diverse interaction styles while producing clinically meaningful outputs. In respiratory care, non-invasive audio recordings captured with sensing devices offer a scalable route to screening and longitudinal monitoring, but heterogeneity is particularly acute: recordings vary across devices, environments, and acquisition protocols, and queries may vary in intent, answer format, and prediction objective. Existing biomedical audio-language question answering systems for respiratory assessment are starting to emerge, but they are typically built as single-path models, processing all inputs through the same acoustic and language pathway despite variation in recording conditions and query types. They are also usually evaluated in relatively limited settings, leaving open their robustness under realistic distribution shifts, including changes in acquisition domains, modality, and clinical task. To address this gap, we introduce RAMoEA-QA, the first RA QA model designed to support input-dependent specialization across heterogeneous recordings and query types within a unified hierarchical two-stage framework. We study this design in a unified RA QA setting spanning clinical and self-recorded, multi-device acquisition settings, question formats, and both discrete and continuous targets. Across in-domain and controlled-shift evaluations, RAMoEA-QA improves over matched monolithic baselines and routing controls, reaching 0.72 in in-domain test accuracy (vs. 0.61 and 0.67 for single-path baselines) on discriminative tasks, while also achieving the best regression performance and stronger average transfer under dataset, modality, and task shifts, including gains of up to 23 percentage points in accuracy on the COPD modality-shift setting.", "date": "2026-03-06", "categories": []}
{"id": "2605.02928", "title": "Keyword spotting using convolutional neural network for speech recognition in Hindi", "abstract": "In this study, we investigate the application of keyword spotting (KWS) in the domain of Hindi speech recognition, utilizing a dataset comprising 40,000 audio samples. With a sampling rate of 44 kHz and an average duration of 1.9 seconds per sample, we focus on developing an efficient on-device KWS system tailored for user-specific queries. Leveraging Convolutional Neural Networks (CNNs) for classification, we employ feature engineering techniques to convert raw audio recordings into Mel Frequency Cepstral Coefficients (MFCCs) as an input for our network. Our experiments encompass various CNN architectures, exploring their efficacy in identifying predefined keywords within the continuous speech stream. Our CNN-based approach achieves a commendable accuracy rate of 91.79% through rigorous evaluation, demonstrating promising performance while ensuring computational efficiency and user-specific customization in Hindi speech recognition.", "date": "2026-04-26", "categories": []}
{"id": "2605.03420", "title": "Deepfake Audio Detection Using Self-supervised Fusion Representations", "abstract": "This paper describes a submission to the Environment-Aware Speech and Sound Deepfake Detection Challenge (ESDD2) 2026, which addresses component-level deepfake detection using the CompSpoofV2 dataset, where speech and environmental sounds may be independently manipulated. To address this challenge, a dual-branch deepfake detection framework is proposed to jointly model speech and environmental contextual representations from input audio. Two pretrained models, XLS-R for speech and BEATs for environmental sound, are used to extract complementary contextual representations. A Matching Head is introduced to model representation differences through statistical normalization and representation interaction, enabling estimation of the original class. In parallel, multi-head cross-attention enables effective information exchange between speech and environmental components. The refined representations are processed with residual connections and layer normalization, and passed to an AASIST classifier to predict speech-based and environment-based spoofing probabilities. The model outputs original, speech, and environment predictions. On the test set, the proposed system achieves an F1-score of 70.20% and an environmental EER of 16.54%, outperforming the baseline system.", "date": "2026-05-05", "categories": []}
{"id": "2605.03914", "title": "Ecologically-Constrained Task Arithmetic for Multi-Taxa Bioacoustic Classifiers Without Shared Data", "abstract": "Training data for bioacoustics is scattered across taxa, regions, and institutions. Centralizing it all is often infeasible. We show that independently fine-tuned BEATs encoders can be composed into a unified 661-species classifier via task vector arithmetic without sharing data. We find that bioacoustic task vectors are near-orthogonal (cosine 0.01-0.09). Their separation aligns closely with spectral distribution distance, a gradient consistent with the acoustic niche hypothesis. This geometry makes simple averaging optimal while sign-conflict methods reduce accuracy by one to six percentage points. Composition also creates an asymmetric gap: species-rich groups lose accuracy relative to joint training while underrepresented taxa gain, a redistribution useful for equitable biodiversity monitoring. We verify linear mode connectivity across all taxonomic pairs, demonstrate zero-shot transfer to new regions, and identify domain negation as a boundary condition where composition fails. These results enable a collaborative paradigm for bioacoustics where institutions share only task vectors to assemble multi-taxa classifiers, preserving data privacy.", "date": "2026-05-05", "categories": []}
{"id": "2605.01790", "title": "Khala: Scaling Acoustic Token Language Models Toward High-Fidelity Music Generation", "abstract": "A common design pattern in high-quality music generation is to handle structure and fidelity in different representation spaces: a generator first models high-level structure, followed by diffusion-based or neural decoding stages that reconstruct fine details. In this work, we explore an alternative view: both may be progressively modeled within a single deep acoustic-token hierarchy. To study this, we build a 64-layer residual vector quantization (RVQ) acoustic representation and propose a two-stage coarse-to-fine generation framework. A backbone model first generates coarse acoustic tokens for the full track, and a super-resolution model then completes finer tokens within the same acoustic token space. The super-resolution stage works at full-track scale and refines tokens layer by layer while running in parallel over time, leading to a fixed 62-step inference process. To jointly improve lyric alignment and fine-detail reconstruction, we further introduce hybrid-attention training: the alignment objective uses causal attention, while layer-wise refinement uses full attention. A key finding is that text--vocal alignment can emerge within pure acoustic-token language modeling, without requiring a separate semantic token stage. Moreover, initializing the super-resolution model from the trained backbone significantly improves convergence and final quality. Taken together, our results suggest that high-quality music generation can be effectively pursued without separating structure and fidelity into heterogeneous representation spaces. Instead, both can be progressively modeled within a unified acoustic-token hierarchy, pointing toward a simpler and more unified path to high-quality music generation.", "date": "2026-05-03", "categories": []}
{"id": "2604.19652", "title": "Environmental Sound Deepfake Detection Using Deep-Learning Framework", "abstract": "In this paper, we propose a deep-learning framework for environmental sound deepfake detection (ESDD) -- the task of identifying whether the sound scene and sound event in an input audio recording is fake or not. To this end, we conducted extensive experiments to explore how individual spectrograms, a wide range of network architectures and pre-trained models, ensemble of spectrograms or network architectures affect the ESDD task performance. The experimental results on the benchmark datasets of EnvSDD and ESDD-Challenge-TestSet indicate that detecting deepfake audio of sound scene and detecting deepfake audio of sound event should be considered as individual tasks. We also indicate that the approach of finetuning a pre-trained model is more effective compared with training a model from scratch for the ESDD task. Eventually, our best model, which was finetuned from the pre-trained WavLM model with the proposed three-stage training strategy, achieve the Accuracy of 0.98, F1 Score of 0.95, AuC of 0.99 on EnvSDD Test subset and the Accuracy of 0.88, F1 Score of 0.77, and AuC of 0.92 on ESDD-Challenge-TestSet dataset.", "date": "2026-04-21", "categories": []}
{"id": "2605.00022", "title": "Putting HUMANS first: Efficient LAM Evaluation with Human Preference Alignment", "abstract": "The rapid proliferation of large audio models (LAMs) demands efficient approaches for model comparison, yet comprehensive benchmarks are costly. To fill this gap, we investigate whether minimal subsets can reliably evaluate LAMs while reducing costs and data redundancy. Analyzing 10 subset selection methods with 18 audio models across 40 tasks covering major LAM evaluation dimensions, we show that subsets of just 50 examples (0.3% of data) can achieve over 0.93 Pearson correlation with full benchmark scores. To understand how well these scores align with what practitioners ultimately care about, user satisfaction, we collect 776 human preference ratings from realistic voice assistant conversations, finding that both subsets and full benchmark achieve only 0.85 correlation with human. To better predict preferences, we trained regression models on these selected subsets, achieving 0.98 correlation -- outperforming regression models trained on both random subsets and the full benchmark. This demonstrates that in regression modeling, well-curated subsets outpredict the full benchmark, showing quality over quantity. We open-source these regression-weighted subsets as the HUMANS benchmark, an efficient proxy for LAM evaluation that captures both benchmark performance and user preferences.", "date": "2026-04-20", "categories": []}
{"id": "2605.00225", "title": "From Birdsong to Rumbles: Classifying Elephant Calls with Out-of-Species Embeddings", "abstract": "We show that pretrained acoustic embeddings classify elephant vocalisations at a level approaching that of end-to-end supervised neural networks, without any fine-tuning of the embedding model. This result is of practical importance because annotated bioacoustic data are scarce and costly to obtain, leaving conventional supervised approaches prone to overfitting and to poor generalisation under domain shift. A broad range of embedding models drawn from general audio, speech, and bioacoustic domains is evaluated, all of which are either out-of-domain (containing no bioacoustic data) or out-of-species (containing no elephant call data). The embedding networks themselves remain fixed; only the lightweight downstream classifiers, which include a linear model and several small neural networks, are trained. Among the models considered, Perch 2.0 achieves the best cross-validated classification performance, attaining AUCs of 0.849 on African bush elephant (Loxodonta africana) calls and 0.936 on Asian elephant (Elephas maximus) calls, with Perch 1.0 close behind. The best-performing system is within 2.2 % of an end-to-end supervised elephant call classification system. A layerwise analysis of pretrained transformer encoders, considered as embedding models, shows that intermediate representations outperform final-layer outputs. The second layer of both wav2vec2.0 and HuBERT encodes sufficient information for effective elephant call classification; truncation at this layer therefore preserves classification performance whilst retaining only approximately 10 % of the parameters of the full network. Such compact embedding networks are well suited to on-device processing where computational resources are limited.", "date": "2026-04-30", "categories": []}
{"id": "2605.00371", "title": "GaMMA: Towards Joint Global-Temporal Music Understanding in Large Multimodal Models", "abstract": "In this paper, we propose GaMMA, a state-of-the-art (SoTA) large multimodal model (LMM) designed to achieve comprehensive musical content understanding. GaMMA inherits the streamlined encoder-decoder design of LLaVA, enabling effective cross-modal learning between music and language. By incorporating audio encoders in a mixture-of-experts manner, GaMMA effectively unifies both time-series and non-time-series music understanding tasks within one set of parameters. Our approach combines carefully curated datasets at scale with a progressive training pipeline, effectively pushing the boundaries of music understanding via pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL). To comprehensively assess both temporal and non-temporal capability of music LMMs, we introduce MusicBench, the largest music-oriented benchmark, comprising 3,739 human-curated multiple-choice questions covering diverse aspects of musical understanding. Extensive experiments demonstrate that GaMMA establishes new SoTA in the music domain, achieving 79.1% accuracy on MuchoMusic, 79.3% on MusicBench-Temporal, and 81.3% on MusicBench-Global, consistently outperforming previous methods.", "date": "2026-05-01", "categories": []}
{"id": "2605.00431", "title": "MMAudioReverbs: Video-Guided Acoustic Modeling for Dereverberation and Room Impulse Response Estimation", "abstract": "Although recent video-to-audio (V2A) models excelled at synthesizing semantically plausible sounds from visual inputs, they do not explicitly model room-acoustic effects such as reverberation or room impulse responses (RIRs), and thus offer limited controllability over these effects. However, we hypothesize that such V2A models implicitly have semantic knowledge of the relationship between spatial audio and the corresponding vision cues. In this paper, we revisit a V2A model for the sake of the above, and propose the way to utilize the pretrained model as prior for physically grounded room-acoustic processing. Based on one of the state-of-the-art V2A models, MMAudio, we propose MMAudioReverbs that is a unified framework dealing with i) dereverberation and ii) room impulse response (RIR) estimation without network architectural modification, and fine-tuned on a small dataset. Experimental results showed that audio and visual cues respectively have advantage depending on the type of physical room acoustics. It implies that foundation V2A models can be used for physically grounded room-acoustic analysis.", "date": "2026-05-01", "categories": []}
{"id": "2605.00495", "title": "MMAudio-LABEL: Audio Event Labeling via Audio Generation for Silent Video", "abstract": "Recent advances in multimodal generation have enabled high-quality audio generation from silent videos. Practical applications, such as sound production, demand not only the generated audio but also explicit sound event labels detailing the type and timing of sounds. One straightforward approach involves applying a standard sound event detection to the generated audio. However, this post-hoc pipeline is inherently limited, as it is prone to error accumulation. To address this limitation, we propose MMAudio-LABEL (LAtent-Based Event Labeling), an event-aware audio generation framework built on a foundational audio generation model as its backbone that jointly generates audio and frame-aligned sound event predictions from silent videos. We evaluate our method on the Greatest Hits dataset for onset detection and 17-class material classification. Our approach improves onset-detection accuracy from 46.7% to 75.0% and material-classification accuracy from 40.6% to 61.0% over baselines. These results suggest that jointly learning audio generation and event prediction enables a more interpretable and practical video-to-audio synthesis.", "date": "2026-05-01", "categories": []}
{"id": "2605.00777", "title": "LASE: Language-Adversarial Speaker Encoding for Indic Cross-Script Identity Preservation", "abstract": "A speaker encoder used in multilingual voice cloning should treat the same speaker identically regardless of which script the audio was uttered in. Off-the-shelf encoders do not, and the failure is accent-conditional. On a 1043-pair Western-accented voice corpus across English, Hindi, Telugu, and Tamil, WavLM-base-plus-sv loses 0.082 absolute cosine similarity when the same voice changes script and ECAPA-TDNN loses 0.105. On a 1369-pair Indian-accented voice corpus, the gap shrinks to 0.006 (WavLM-SV) and 0.044 (ECAPA-TDNN). The leak is largest where it matters most for cross-script TTS: when a system projects a non-Indic-trained voice into Indic scripts. We present LASE (Language-Adversarial Speaker Encoder), a small projection head over frozen WavLM-base-plus trained with two losses: a supervised contrastive loss over voice identity, and a gradient-reversal cross-entropy against a 4-language classifier that pushes the embedding to be language-uninformative while remaining speaker-informative. Trained on 1118 quality-gated cross-script pairs synthesised from 8 commercial multilingual voices, LASE's residual gap is consistent with zero on both corpora (Delta = 0.013 Western, Delta = 0.026 Indian; both bootstrap 95% CIs include zero) and amplifies the cross-script-vs-floor margin 2.4-2.7x over both baselines. An ECAPA+GRL ablation shows the GRL objective improves either backbone but the WavLM choice contributes too. In synthetic multi-speaker diarisation, LASE matches ECAPA-TDNN on cross-script speaker recall (0.788 vs 0.789) with ~100x less training data. We release the r1 checkpoint, both corpora, and the bootstrap recipe.", "date": "2026-05-01", "categories": []}
{"id": "2602.02980", "title": "WST-X Series: Wavelet Scattering Transform for Interpretable Speech Deepfake Detection", "abstract": "In this work, we focus on front-end design for speech deepfake detectors, the component that determines the discriminative acoustic cues provided to the classifier. Existing approaches are primarily categorized into two types. Hand-crafted filterbank features are transparent but limited in capturing higher-level information. SSL features, in turn, lack interpretability and may overlook fine-grained spectral anomalies. We propose the WST-X series, a novel family of feature extractors that combines the best of both worlds via the wavelet scattering transform (WST), which cascades wavelet convolutions with modulus nonlinearities to produce deformation-stable, multi-scale features. Experiments on the recent Deepfake-Eval-2024 benchmark, together with cross-dataset evaluations on the SpoofCeleb and In-the-Wild, show that WST-X outperforms existing front-ends by a wide margin. Our analysis reveals that a small averaging scale ($J$), combined with high-frequency and directional resolutions ($Q$, $L$), is critical for capturing subtle artifacts. This underscores the value of stable and translation-invariant features for speech deepfake detection. The code is available at https://github.com/xxuan-acoustics/WST-X-Series.", "date": "2026-02-03", "categories": []}
{"id": "2604.27279", "title": "Predicting Upcoming Stuttering Events from Three-Second Audio: Stratified Evaluation Reveals Severity-Selective Precursors, and the Model Deploys Fully On-Device", "abstract": "Audio-based stuttering systems to date have been trained for detection -- what disfluency is present now -- leaving prediction, the capability needed for closed-loop intervention, unstudied at deployable scale. We train a 616K-parameter CNN on SEP-28k (Apple, 20,131 three-second clips) to predict whether the next contiguous clip contains any disfluency.\n  (1) Severity-selective precursor signal: on the episode-grouped test set, aggregate preblock AUC is modest (0.581 [0.542, 0.619]), but stratifying by upcoming event type reveals concentration on clinically severe events -- blocks 0.601 [0.554, 0.651] and sound repetitions 0.617 [0.567, 0.667] both exclude chance, while fillers (0.45) and word repetitions (0.49) are at chance. The aggregate objective converges to a severity-selective predictor because severe events carry prosodic precursors; fillers do not.\n  (2) Cross-population transfer: without fine-tuning, the same checkpoint applied to 1,024 pediatric Children-Who-Stutter utterances (FluencyBank Teaching) attains AUC 0.674 detection and 0.655 prediction; DisfluencySpeech and LibriStutter reach 0.58-0.60 AUC.\n  (3) Deployable on-device: lossless export to CoreML (1.19 MB), ONNX (40 KB), TFLite. Neural-Engine latency per 3 s window: 0.25 ms (iPhone 17 Pro Max, A19 Pro) to 0.55 ms (iPhone SE 3rd-gen and M1 Max). A 4 Hz streaming simulation uses 0.54% of the real-time budget. Platt-calibrated outputs (test ECE 0.010, from 0.177 raw).\n  Five negative ablations -- output-level Future-Guided Learning, multi-clip GRU, time-axis concatenation, asymmetric focal loss, direct block-targeted training -- none improved over the vanilla baseline.", "date": "2026-04-30", "categories": []}
{"id": "2604.27936", "title": "Beyond the Baseband: Adaptive Multi-Band Encoding for Full-Spectrum Bioacoustics Classification", "abstract": "Animals hear and vocalize across frequency ranges that differ substantially from humans, often extending into the ultrasonic domain. Yet most computational bioacoustics systems rely on audio models pre-trained at 16 kHz, restricting their usable bandwidth to the 0-8 kHz baseband and discarding higher-frequency information present in many bioacoustic recordings. We investigate a multi-band encoding framework that decomposes the full spectrum of animal calls into band features and fuses them into a unified representation. Similarity analyses on models show that certain encoders produce decorrelated band embeddings that improve class separation after fusion. Classification experiments on three bioacoustic datasets using eight pre-trained models and five fusion strategies show that fused representations consistently outperform the baseband and time-expansion baselines on two datasets, showing the potential of multi-band methods for full-spectrum encoding of animal calls.", "date": "2026-04-30", "categories": []}
{"id": "2604.26057", "title": "Similarity Choice and Negative Scaling in Supervised Contrastive Learning for Deepfake Audio Detection", "abstract": "Supervised contrastive learning (SupCon) is widely used to shape representations, but has seen limited targeted study for audio deepfake detection. Existing work typically combines contrastive terms with broader pipelines; however, the focus on SupCon itself is missing. In this work, we run a controlled study on wav2vec2 XLS-R (300M) that varies (i) similarity in SupCon (cosine vs angular similarity derived from the hyperspherical angle) and (ii) negative scaling using a warm-started global cross-batch queue. Stage 1 fine-tunes the encoder and projection head with SupCon; Stage 2 freezes them and trains a linear classifier with BCE. Trained on ASVspoof 2019 LA and evaluated on ASV19 eval plus ITW and ASVspoof 2021 DF/LA, Cosine SupCon with a delayed queue achieves the best ITW EER (8.29%) and pooled EER (4.44), while angular similarity performs strongly without queued negatives (ITW 8.70), indicating reduced reliance on large negative sets.", "date": "2026-04-28", "categories": []}
{"id": "2604.26136", "title": "One Voice, Many Tongues: Cross-Lingual Voice Cloning for Scientific Speech", "abstract": "Preserving a speaker's voice identity while generating speech in a different language remains a fundamental challenge in spoken language technology, particularly in specialized domains such as scientific communication. In this paper, we address this challenge through our system submission to the International Conference on Spoken Language Translation (IWSLT 2026), the Cross-Lingual Voice Cloning shared task. First, we evaluate several state-of-the-art voice cloning models for cross-lingual speech generation of scientific texts in Arabic, Chinese, and French. Then, we build voice cloning systems based on the OmniVoice foundation model. We employ data augmentation via multi-model ensemble distillation from the ACL 60/60 corpus. We investigate the effect of using this synthetic data for fine-tuning, demonstrating consistent improvements in intelligibility (WER and CER) across languages while preserving speaker similarity.", "date": "2026-04-28", "categories": []}
{"id": "2604.26281", "title": "DiffAnon: Diffusion-based Prosody Control for Voice Anonymization", "abstract": "To preserve or not to preserve prosody is a central question in voice anonymization. Prosody conveys meaning and affect, yet is tightly coupled with speaker identity. Existing methods either discard prosody for privacy or lack a principled mechanism to control the utility-privacy trade-off, operating at fixed design points. We propose DiffAnon, a diffusion-based anonymization method with classifier-free guidance (CFG) that provides explicit, continuous inference-time control over prosody preservation. DiffAnon refines acoustic detail over semantic embeddings of an RVQ codec, enabling smooth interpolation between anonymization strength and prosodic fidelity within a single model. To the best of our knowledge, it is the first voice anonymization framework to provide structured, interpolatable inference-time prosody control. Experiments demonstrate structured trade-off behavior, achieving strong utility while maintaining competitive privacy across controllable operating points.", "date": "2026-04-29", "categories": []}
{"id": "2604.26568", "title": "Multimodal LLMs are not all you need for Pediatric Speech Language Pathology", "abstract": "Speech Sound Disorders (SSD) affect roughly five percent of children, yet speech-language pathologists face severe staffing shortages and unmanageable caseloads. We test a hierarchical approach to SSD classification on the granular multi-task SLPHelmUltraSuitePlus benchmark. We propose a cascading approach from binary classification to type, and symptom classification. By fine-tuning Speech Representation Models (SRM), and using targeted data augmentation we mitigate biases found by previous works, and improve upon all clinical tasks in the benchmark. We also treat Automatic Speech Recognition (ASR) with our data augmentation approach. Our results demonstrate that SRM consistently outperform the LLM-based state-of-the-art across all evaluated tasks by a large margin. We publish our models and code to foster future research.", "date": "2026-04-29", "categories": []}
{"id": "2511.20006", "title": "BERT-APC: A Reference-free Framework for Automatic Pitch Correction via Musical Context Inference", "abstract": "Automatic Pitch Correction (APC) enhances vocal recordings by aligning pitch deviations with intended musical notes. However, existing APC systems either rely on reference pitches, which limits practical applicability, or employ simple pitch estimation algorithms that often fail to preserve expressiveness and naturalness. We propose BERT-APC, a reference-free APC framework that corrects pitch errors while maintaining the expressiveness and naturalness of vocal performances. In BERT-APC, a stationary pitch predictor first estimates the stationary pitch of each note from the detuned singing voice, where stationary pitch is the continuous pitch from the stable region of a note and approximates its perceived pitch. A context-aware note pitch predictor then infers the intended pitch sequence using a repurposed music language model that incorporates musical context. Finally, a note-level correction algorithm fixes pitch errors while preserving intentional deviations for emotional expression. We also introduce a learnable data augmentation strategy that improves robustness by simulating realistic detuning patterns. Compared to two recent singing voice transcription models, BERT-APC demonstrated superior target note pitch prediction, outperforming the second-best model, ROSVOT, by 10.49 percentage points on highly detuned samples in raw pitch accuracy. In the MOS test, BERT-APC achieved the highest quality rating of $4.32 \\pm 0.15$, significantly higher than Auto-Tune ($3.22 \\pm 0.18$) and Melodyne ($3.08 \\pm 0.18$), while maintaining a comparable ability to preserve expressive nuances. To the best of our knowledge, this is the first APC model that leverages a music language model to achieve reference-free pitch correction with symbolic musical context. The corrected audio samples are available at https://joshua-1995.github.io/BERT-APC-Demo/.", "date": "2025-11-25", "categories": []}
{"id": "2604.24933", "title": "S-SONDO: Self-Supervised Knowledge Distillation for General Audio Foundation Models", "abstract": "General audio foundation models have recently achieved remarkable progress, enabling strong performance across diverse tasks. However, state-of-the-art models remain extremely large, often with hundreds of millions of parameters, leading to high inference costs and limited deployability on edge devices. Knowledge distillation is a proven strategy for model compression, but prior work in audio has mostly focused on supervised settings, relying on class logits, intermediate features, or architecture-specific techniques. Such assumptions exclude models that output only embeddings, such as self-supervised or metric-learning models. We introduce S-SONDO (Self-Supervised KnOwledge DistillatioN for General AuDio FOundation Models), the first framework to distill general audio models using only their output embeddings. By avoiding the need for logits or layer-level alignment, S-SONDO is architecture-agnostic and broadly applicable to embedding-based teachers. We demonstrate its effectiveness by distilling two audio foundation models into three efficient students that are up to 61 times smaller while retaining up to 96% of teacher performance. We also provide practical insights on loss choice and clustering-based balanced data sampling. Code is available here: https://github.com/MedAliAdlouni/ssondo.", "date": "2026-04-27", "categories": []}
{"id": "2604.25819", "title": "Mutual Forcing: Dual-Mode Self-Evolution for Fast Autoregressive Audio-Video Character Generation", "abstract": "In this work, we propose Mutual Forcing, a framework for fast autoregressive audio-video generation with long-horizon audio-video synchronization. Our approach addresses two key challenges: joint audio-video modeling and fast autoregressive generation. To ease joint audio-video optimization, we adopt a two-stage training strategy: we first train uni-modal generators and then couple them into a unified audio-video model for joint training on paired data. For streaming generation, we ask whether a native fast causal audio-video model can be trained directly, instead of following existing streaming distillation pipelines that typically train a bidirectional model first and then convert it into a causal generator through multiple distillation stages. Our answer is Mutual Forcing, which builds directly on native autoregressive model and integrates few-step and multi-step generation within a single weight-shared model, enabling self-distillation and improved training-inference consistency. The multi-step mode improves the few-step mode via self-distillation, while the few-step mode generates historical context during training to improve training-inference consistency; because the two modes share parameters, these two effects reinforce each other within a single model. Compared with prior approaches such as Self-Forcing, Mutual Forcing removes the need for an additional bidirectional teacher model, supports more flexible training sequence lengths, reduces training overhead, and allows the model to improve directly from real paired data rather than a fixed teacher. Experiments show that Mutual Forcing matches or surpasses strong baselines that require around 50 sampling steps while using only 4 to 8 steps, demonstrating substantial advantages in both efficiency and quality. The project page is available at https://mutualforcing.github.io.", "date": "2026-04-28", "categories": []}
{"id": "2509.11717", "title": "CodecSep: Prompt-Driven Universal Sound Separation on Neural Audio Codec Latents", "abstract": "Text-guided sound separation enables flexible audio editing, assistive listening, and open-domain source extraction, but systems such as AudioSep remain too expensive for low-latency edge or codec-mediated deployment. Existing neural audio codec separators are efficient, yet largely restricted to fixed stems or closed taxonomies. We introduce CodecSep, a prompt-driven universal sound separation framework that extracts sources directly in neural audio codec latent space. CodecSep combines a frozen DAC backbone with a lightweight FiLM-conditioned Transformer masker driven by CLAP text embeddings, enabling open-vocabulary separation while preserving codec-native efficiency.\n  Across dnr-v2 and five open-domain benchmarks, CodecSep consistently improves over AudioSep in SI-SDR, remains competitive in ViSQOL, and achieves clear gains in human MOS-LQS. Controlled analyses show that fine-grained prompts outperform coarse labels, and that explicit latent masking is substantially more effective than decoder-style latent generation in codec space. Qualitative diagnostics show that neural audio codec latents retain source-dependent structure, which CodecSep exploits mainly through channel-wise source-conditioned modulation.\n  CodecSep also provides a practical code-stream deployment path. When audio is transmitted as neural audio codec codes, CodecSep maps codes to embeddings, separates directly in codec space, and outputs waveforms or re-quantized codes, avoiding the decode-separate-re-encode loop. In this regime, CodecSep requires only 1.35 GMACs end-to-end: about 54 times less compute than AudioSep in the same pipeline and 25 times lower separator-only compute, with much lower latency and memory. More broadly, CodecSep offers a blueprint for codec-native downstream audio processing.", "date": "2025-09-15", "categories": []}
{"id": "2604.10708", "title": "Audio-Omni: Extending Multi-modal Understanding to Versatile Audio Generation and Editing", "abstract": "Recent progress in multimodal models has spurred rapid advances in audio understanding, generation, and editing. However, these capabilities are typically addressed by specialized models, leaving the development of a truly unified framework that can seamlessly integrate all three tasks underexplored. While some pioneering works have explored unifying audio understanding and generation, they often remain confined to specific domains. To address this, we introduce Audio-Omni, the first end-to-end framework to unify generation and editing across general sound, music, and speech domains, with integrated multi-modal understanding capabilities. Our architecture synergizes a frozen Multimodal Large Language Model for high-level reasoning with a trainable Diffusion Transformer for high-fidelity synthesis. To overcome the critical data scarcity in audio editing, we construct AudioEdit, a new large-scale dataset comprising over one million meticulously curated editing pairs. Extensive experiments demonstrate that Audio-Omni achieves state-of-the-art performance across a suite of benchmarks, outperforming prior unified approaches while achieving performance on par with or superior to specialized expert models. Beyond its core capabilities, Audio-Omni exhibits remarkable inherited capabilities, including knowledge-augmented reasoning generation, in-context generation, and zero-shot cross-lingual control for audio generation, highlighting a promising direction toward universal generative audio intelligence. The code, model, and dataset will be publicly released on https://zeyuet.github.io/Audio-Omni.", "date": "2026-04-12", "categories": []}
{"id": "2604.23284", "title": "Au-M-ol: A Unified Model for Medical Audio and Language Understanding", "abstract": "In this work, we present Au-M-ol, a novel multimodal architecture that extends Large Language Models (LLMs) with audio processing. It is designed to improve performance on clinically relevant tasks such as Automatic Speech Recognition (ASR). Au-M-ol has three main components: (1) an audio encoder that extracts rich acoustic features from medical speech, (2) an adaptation layer that maps audio features into the LLM input space, and (3) a pretrained LLM that performs transcription and clinical language understanding. This design allows the model to interpret spoken medical content directly, improving both accuracy and robustness. In experiments, Au-M-ol reduces Word Error Rate (WER) by 56\\% compared to state-of-the-art baselines on medical transcription tasks. The model also performs well in challenging conditions, including noisy environments, domain-specific terminology, and speaker variability. These results suggest that Au-M-ol is a strong candidate for real-world clinical applications, where reliable and context-aware audio understanding is essential.", "date": "2026-04-25", "categories": []}
{"id": "2604.23717", "title": "HeadRouter: Dynamic Head-Weight Routing for Task-Adaptive Audio Token Pruning in Large Audio Language Models", "abstract": "Recent large audio language models (LALMs) demonstrate remarkable capabilities in processing extended multi-modal sequences, yet incur high inference costs. Token compression is an effective method that directly reduces redundant tokens in the sequence. Existing compression methods usually assume that all attention heads in LALMs contribute equally to various audio tasks and calculate token importance by averaging scores across all heads. However, our analysis demonstrates that attention heads exhibit distinct behaviors across diverse audio domains. We further reveal that only a sparse subset of attention heads actively responds to audio, with completely different performance when handling semantic and acoustic tasks. In light of this observation, we propose HeadRouter, a head-importance-aware token pruning method that perceives the varying importance of attention heads in different audio tasks to maximize the retention of crucial tokens. HeadRouter is training-free and can be applied to various LALMs. Extensive experiments on the AudioMarathon and MMAU-Pro benchmarks demonstrate that HeadRouter achieves state-of-the-art compression performance, exceeding the baseline model even when retaining 70% of the audio tokens and achieving 101.8% and 103.0% of the vanilla average on Qwen2.5-Omni-3B and Qwen2.5-Omni-7B, respectively.", "date": "2026-04-26", "categories": []}
{"id": "2604.24096", "title": "Meta-Ensemble Learning with Diverse Data Splits for Improved Respiratory Sound Classification", "abstract": "Training reliable respiratory sound classification models remains challenging due to the limited size and subject diversity of datasets. Ensemble methods can improve robustness, but when base models are trained on identical data, models tend to overfit and produce highly correlated predictions, thereby reducing the effectiveness of ensembling. In this work, we investigate a meta-ensemble learning methodology that enhances prediction diversity by training base models on diverse data splits and combining their outputs through a trained meta-model. Specifically, we train base models on the ICBHI dataset using two data split settings: fixed 80-20% split and five-fold cross-validation split, under two data granularity settings: patient- and sample-level. The resulting diversity in base model predictions enables the meta-model to better generalize. Our approach achieves new state-of-the-art performance on the ICBHI benchmark, reaching a Score of 66.49% and showing improved generalization on two out-of-distribution datasets, indicating its potential applicability to real-world clinical data.", "date": "2026-04-27", "categories": []}
{"id": "2604.24401", "title": "All That Glitters Is Not Audio: Rethinking Text Priors and Audio Reliance in Audio-Language Evaluation", "abstract": "Large Audio-Language Models show consistent performance gains across speech and audio benchmarks, yet high scores may not reflect true auditory perception. If a model can answer questions without processing the acoustic signal, the benchmark fails as a measure of auditory understanding. We present a diagnostic framework using two axes: text prior, which measures answerability from text and general knowledge alone, and audio reliance, which assesses actual dependency on the acoustic signal. Evaluating eight LALMs across three benchmarks, we find that models retain 60-72% of their full audio scores even without any audio input. Moreover, among items that require audio, only 3.0-4.2% need the complete audio clip; the majority can be resolved using localized fragments. These findings challenge the assumption that benchmark performance equals robust audio understanding, and we conclude with practical guidelines for improving evaluation reliability and benchmark design.", "date": "2026-04-27", "categories": []}
{"id": "2604.22209", "title": "UniSonate: A Unified Model for Speech, Music, and Sound Effect Generation with Text Instructions", "abstract": "Generative audio modeling has largely been fragmented into specialized tasks, text-to-speech (TTS), text-to-music (TTM), and text-to-audio (TTA), each operating under heterogeneous control paradigms. Unifying these modalities remains a fundamental challenge due to the intrinsic dissonance between structured semantic representations (speech/music) and unstructured acoustic textures (sound effects). In this paper, we introduce UniSonate, a unified flow-matching framework capable of synthesizing speech, music, and sound effects through a standardized, reference-free natural language instruction interface. To reconcile structural disparities, we propose a novel dynamic token injection mechanism that projects unstructured environmental sounds into a structured temporal latent space, enabling precise duration control within a phoneme-driven Multimodal Diffusion Transformer (MM-DiT). Coupled with a multi-stage curriculum learning strategy, this approach effectively mitigates cross-modal optimization conflicts. Extensive experiments demonstrate that UniSonate achieves state-of-the-art performance in instruction-based TTS (WER 1.47%) and TTM (SongEval Coherence 3.18), while maintaining competitive fidelity in TTA. Crucially, we observe positive transfer, where joint training on diverse audio data significantly enhances structural coherence and prosodic expressiveness compared to single-task baselines. Audio samples are available at https://qiangchunyu.github.io/UniSonate/.", "date": "2026-04-24", "categories": []}
{"id": "2604.21766", "title": "AUDITA: A New Dataset to Audit Humans vs. AI Skill at Audio QA", "abstract": "Existing audio question answering benchmarks largely emphasize sound event classification or caption-grounded queries, often enabling models to succeed through shortcut strategies, short-duration cues, lexical priors, dataset-specific biases, or even bypassing audio via metadata and captions rather than genuine reasoning Thus, we present AUDITA (Audio Understanding from Diverse Internet Trivia Authors), a large-scale, real-world benchmark to rigorously evaluate audio reasoning beyond surface-level acoustic recognition. AUDITA comprises carefully curated, human-authored trivia questions grounded in real-world audio, designed to stress robust auditory reasoning through challenging distractors and long-range temporal dependencies, using probing queries that cannot be answered from isolated text or sound cues alone. Human average accuracy of 32.13% shows both the challenge of the task while demonstrating meaningful comprehension of the audio. In stark contrast, state of-the-art audio question answering models perform poorly, with average accuracy below 8.86%. Beyond raw accuracy, we apply Item Response Theory (IRT) to estimate latent proficiency, question difficulty, and expose systematic deficiencies of the models and data.", "date": "2026-04-23", "categories": []}
{"id": "2604.12456", "title": "X-VC: Zero-shot Streaming Voice Conversion in Codec Space", "abstract": "Zero-shot voice conversion (VC) aims to convert a source utterance into the voice of an unseen target speaker while preserving its linguistic content. Although recent systems have improved conversion quality, building zero-shot VC systems for interactive scenarios remains challenging because high-fidelity speaker transfer and low-latency streaming inference are difficult to achieve simultaneously. In this work, we present X-VC, a zero-shot streaming VC system that performs one-step conversion in the latent space of a pretrained neural codec. X-VC uses a dual-conditioning acoustic converter that jointly models source codec latents and frame-level acoustic conditions derived from target reference speech, while injecting utterance-level target speaker information through adaptive normalization. To reduce the mismatch between training and inference, we train the model with generated paired data and a role-assignment strategy that combines standard, reconstruction, and reversed modes. For streaming inference, we further adopt a chunkwise inference scheme with overlap smoothing that is aligned with the segment-based training paradigm of the codec. Experiments on Seed-TTS-Eval show that X-VC achieves the best streaming WER in both English and Chinese, strong speaker similarity in same-language and cross-lingual settings, and substantially lower offline real-time factor than the compared baselines. These results suggest that codec-space one-step conversion is a practical approach for building high-quality low-latency zero-shot VC systems. Our audio samples, code and checkpoints are released at https://github.com/Jerrister/X-VC.", "date": "2026-04-14", "categories": []}
{"id": "2406.14294", "title": "DASB - Discrete Audio and Speech Benchmark", "abstract": "Discrete audio tokens have recently gained considerable attention for their potential to bridge audio and language processing, enabling multimodal language models that can both generate and understand audio. However, preserving key information such as phonetic content, speaker identity, and paralinguistic cues remains a major challenge. Identifying the optimal tokenizer and configuration is further complicated by inconsistent evaluation settings across existing studies. To address this, we introduce the Discrete Audio and Speech Benchmark (DASB), a comprehensive framework for benchmarking discrete audio tokens across speech, general audio, and music domains on a range of discriminative and generative tasks. Our results show that discrete representations are less robust than continuous ones and require careful tuning of factors such as model architecture, data size, learning rate, and capacity. Semantic tokens generally outperform acoustic tokens, but a gap remains between discrete tokens and continuous features, highlighting the need for further research. DASB codes, evaluation setup, and leaderboards are publicly available at https://poonehmousavi.github.io/DASB-website/.", "date": "2024-06-20", "categories": []}
{"id": "2604.19300", "title": "HalluAudio: A Comprehensive Benchmark for Hallucination Detection in Large Audio-Language Models", "abstract": "Large Audio-Language Models (LALMs) have recently achieved strong performance across various audio-centric tasks. However, hallucination, where models generate responses that are semantically incorrect or acoustically unsupported, remains largely underexplored in the audio domain. Existing hallucination benchmarks mainly focus on text or vision, while the few audio-oriented studies are limited in scale, modality coverage, and diagnostic depth. We therefore introduce HalluAudio, the first large-scale benchmark for evaluating hallucinations across speech, environmental sound, and music. HalluAudio comprises over 5K human-verified QA pairs and spans diverse task types, including binary judgments, multi-choice reasoning, attribute verification, and open-ended QA. To systematically induce hallucinations, we design adversarial prompts and mixed-audio conditions. Beyond accuracy, our evaluation protocol measures hallucination rate, yes/no bias, error-type analysis, and refusal rate, enabling a fine-grained analysis of LALM failure modes. We benchmark a broad range of open-source and proprietary models, providing the first large-scale comparison across speech, sound, and music. Our results reveal significant deficiencies in acoustic grounding, temporal reasoning, and music attribute understanding, underscoring the need for reliable and robust LALMs.", "date": "2026-04-21", "categories": []}
{"id": "2604.19635", "title": "Towards Streaming Target Speaker Extraction via Chunk-wise Interleaved Splicing of Autoregressive Language Model", "abstract": "While generative models have set new benchmarks for Target Speaker Extraction (TSE), their inherent reliance on global context precludes deployment in real-time applications. Direct adaptation to streaming scenarios often leads to catastrophic inference performance degradation due to the severe mismatch between training and streaming inference. To bridge this gap, we present the first autoregressive (AR) models tailored for streaming TSE. Our approach introduces a Chunk-wise Interleaved Splicing Paradigm that ensures highly efficient and stable streaming inference. To ensure the coherence between the extracted speech segments, we design a historical context refinement mechanism that mitigates boundary discontinuities by leveraging historical information. Experiments on Libri2Mix show that while AR generative baseline exhibits performance degradation at low latencies, our approach maintains 100% stability and superior intelligibility. Furthermore, our streaming results are comparable to or even surpass offline baselines. Additionally, our model achieves a Real-Time-Factor (RTF) of 0.248 on consumer-level GPUs. This work provides empirical evidence that AR generative backbones are viable for latency-sensitive applications through the Chunk-wise Interleaved Splicing Paradigm.", "date": "2026-04-21", "categories": []}
{"id": "2510.08878", "title": "ControlAudio: Tackling Text-Guided, Timing-Indicated and Intelligible Audio Generation via Progressive Diffusion Modeling", "abstract": "Text-to-audio (TTA) generation with fine-grained control signals, e.g., precise timing control or intelligible speech content, has been explored in recent works. However, constrained by data scarcity, their generation performance at scale is still compromised. In this study, we recast controllable TTA generation as a multi-task learning problem and introduce a progressive diffusion modeling approach, ControlAudio. Our method adeptly fits distributions conditioned on more fine-grained information, including text, timing, and phoneme features, through a step-by-step strategy. First, we propose a data construction method spanning both annotation and simulation, augmenting condition information in the sequence of text, timing, and phoneme. Second, at the model training stage, we pretrain a diffusion transformer (DiT) on large-scale text-audio pairs, achieving scalable TTA generation, and then incrementally integrate the timing and phoneme features with unified semantic representations, expanding controllability. Finally, at the inference stage, we propose progressively guided generation, which sequentially emphasizes more fine-grained information, aligning inherently with the coarse-to-fine sampling nature of DiT. Extensive experiments show that ControlAudio achieves state-of-the-art performance in terms of temporal accuracy and speech clarity, significantly outperforming existing methods on both objective and subjective evaluations. Demo samples are available at: https://control-audio.github.io/Control-Audio.", "date": "2025-10-10", "categories": []}
{"id": "2512.03563", "title": "State Space Models for Bioacoustics: A Comparative Evaluation with Transformers", "abstract": "In this study, we evaluate the efficacy of the Mamba architecture bioacoustics by introducing BioMamba, a Mamba-based audio representation model for wildlife sounds. We pre-train a BioMamba using self-supervised learning on a large audio corpus and evaluate it on the BEANS benchmark across diverse classification and detection tasks. Compared to the state-of-the-art Transformer-based model (AVES), BioMamba achieves comparable performance while significantly reducing VRAM consumption. Our results demonstrate Mamba's potential as a computationally efficient alternative for real-world environmental monitoring.", "date": "2025-12-03", "categories": []}
{"id": "2601.04744", "title": "Semi-Supervised Diseased Detection from Speech Dialogues with Multi-Level Data Modeling", "abstract": "Detecting medical conditions from speech acoustics is fundamentally a weakly-supervised learning problem: a single, often noisy, session-level label must be linked to nuanced patterns within a long, complex audio recording. This task is further hampered by severe data scarcity and the subjective nature of clinical annotations. While semi-supervised learning (SSL) offers a viable path to leverage unlabeled data, existing audio methods often fail to address the core challenge that pathological traits are not uniformly expressed in a patient's speech. We propose a novel, audio-only SSL framework that explicitly models this hierarchy by jointly learning from frame-level, segment-level, and session-level representations within unsegmented clinical dialogues. Our end-to-end approach dynamically aggregates these multi-granularity features and generates high-quality pseudo-labels to efficiently utilize unlabeled data. Extensive experiments show the framework is model-agnostic, robust across languages and conditions, and highly data-efficient-achieving, for instance, 90% of fully-supervised performance using only 11 labeled samples. This work provides a principled approach to learning from weak, far-end supervision in medical speech analysis. The code is available at https://github.com/fispresent/semi_pathological.", "date": "2026-01-08", "categories": []}
{"id": "2601.20867", "title": "Generalizable Prompt Tuning for Audio-Language Models via Semantic Expansion", "abstract": "Prompt tuning has achieved remarkable progress in vision-language models (VLMs) and is recently being adopted for audio-language models (ALMs). However, its generalization ability in ALMs remains largely underexplored. We observe that conventional prompt tuning for ALMs also suffers from the Base-New Tradeoff, and we identify that this issue stems from the disrupted semantic structure of the embedding space. To address this issue, we propose Semantically Expanded Prompt Tuning (SEPT)-a plug-and-play framework that explicitly regularizes the prompt embedding space by incorporating semantic neighbors generated by large language models. SEPT introduces a novel semantic expansion loss with margin constraints that promote intra-class compactness and inter-class separability, thereby enhancing the semantic structure of the prompt embedding space. For comprehensive evaluation, we establish the first benchmark setup for prompt generalization in ALMs, covering both base-to-new generalization and cross-dataset transferability. Extensive experiments demonstrate that SEPT consistently improves generalization performance across multiple prompt tuning baselines, while maintaining computational cost during inference.", "date": "2026-01-06", "categories": []}
{"id": "2603.19857", "title": "FoleyDirector: Fine-Grained Temporal Steering for Video-to-Audio Generation via Structured Scripts", "abstract": "Recent Video-to-Audio (V2A) methods have achieved remarkable progress, enabling the synthesis of realistic, high-quality audio. However, they struggle with fine-grained temporal control in multi-event scenarios or when visual cues are insufficient, such as small regions, off-screen sounds, or occluded or partially visible objects. In this paper, we propose FoleyDirector, a framework that, for the first time, enables precise temporal guidance in DiT-based V2A generation while preserving the base model's audio quality and allowing seamless switching between V2A generation and temporally controlled synthesis. FoleyDirector introduces Structured Temporal Scripts (STS), a set of captions corresponding to short temporal segments, to provide richer temporal information. These features are integrated via the Script-Guided Temporal Fusion Module, which employs Temporal Script Attention to fuse STS features coherently. To handle complex multi-event scenarios, we further propose Bi-Frame Sound Synthesis, enabling parallel in-frame and out-of-frame audio generation and improving controllability. To support training and evaluation, we construct the DirectorSound dataset and introduce VGGSoundDirector and DirectorBench. Experiments demonstrate that FoleyDirector substantially enhances temporal controllability while maintaining high audio fidelity, empowering users to act as Foley directors and advancing V2A toward more expressive and controllable generation.", "date": "2026-03-20", "categories": []}
{"id": "2604.16563", "title": "Classification of systolic murmurs in heart sounds using multiresolution complex Gabor dictionary and vision transformer", "abstract": "Systolic murmurs are extra heart sounds that occur during the contraction phase of the cardiac cycle, often indicating heart abnormalities caused by turbulent blood flow. Their intensity, pitch, and quality vary, requiring precise identification for the accurate diagnosis of cardiac disorders. This study presents an automatic classification system for systolic murmurs using a feature extraction module, followed by a classification model. The feature extraction module employs complex orthogonal matching pursuit to project single or multiple murmur segments onto a redundant dictionary composed of multiresolution complex Gabor basis functions (GBFs). The resulting projection weights are split and reshaped into variable-resolution time--frequency feature matrices. Processing multiple segments of a single recording using a shared dictionary mitigates murmur variability. This is achieved by learning the weights for each segment while enforcing that they correspond to the same set of basis functions in the dictionary, promoting consistent time--frequency feature matrices. The classification model is built based on a vision transformer to process multiple input matrices of different resolutions by passing each through a convolutional neural network for patch tokenization. All embedding tokens are then concatenated to form a matrix and forwarded to an encoder layer that includes multihead attention, residual connections, and a convolutional network with a kernel size of one. This integration of multiresolution feature extraction with transformer-based feature classification enhances the accuracy and reliability of heart murmur identification. An experimental analysis of four types of systolic murmurs from the CirCor DigiScope dataset demonstrates the effectiveness of the system, achieving a classification accuracy of $95.96\\%$.", "date": "2026-04-17", "categories": []}
{"id": "2604.16749", "title": "ICLAD: In-Context Learning with Comparison-Guidance for Audio Deepfake Detection", "abstract": "Audio deepfakes pose a significant security threat, yet current state-of-the-art (SOTA) detection systems do not generalize well to realistic in-the-wild deepfakes. We introduce a novel \\textbf{I}n-\\textbf{C}ontext \\textbf{L}earning paradigm with comparison-guidance for \\textbf{A}udio \\textbf{D}eepfake detection (\\textbf{ICLAD}). The framework enables the use of audio language models (ALMs) for training-free generalization to unseen deepfakes and provides textual rationales on the detection outcome. At the core of ICLAD is a pairwise comparative reasoning strategy that guides the ALM to discover and filter hallucinations and deepfake-irrelevant acoustic attributes. The ALM works alongside a specialized deepfake detector, whereby a routing mechanism feeds out-of-distribution samples to the ALM. On in-the-wild datasets, ICLAD improves macro F1 over the specialized detector, with up to $2\\times$ relative improvement. Further analysis demonstrates the flexibility of ICLAD and its potential for deployment on recent open-source ALMs.", "date": "2026-04-17", "categories": []}
{"id": "2604.17823", "title": "A novel LSTM music generator based on the fractional time-frequency feature extraction", "abstract": "In this paper, we propose a novel approach for generating music based on an artificial intelligence (AI) system. We analyze the features of music and use them to fit and predict the music. The fractional Fourier transform (FrFT) and the long short-term memory (LSTM) network are the foundations of our method. The FrFT method is used to extract the spectral features of a music piece, where the music signal is expressed on the time and frequency domains. The LSTM network is used to generate new music based on the extracted features, where we predict the music according to the hidden layer features and real-time inputs using GiantMIDI-Piano dataset. The results of our experiments show that our proposed system is capable of generating high-quality music that is comparable to human-generated music.", "date": "2026-04-20", "categories": []}
{"id": "2604.17986", "title": "Latent Fourier Transform", "abstract": "We introduce the Latent Fourier Transform (LatentFT), a framework that provides novel frequency-domain controls for generative music models. LatentFT combines a diffusion autoencoder with a latent-space Fourier transform to separate musical patterns by timescale. By masking latents in the frequency domain during training, our method yields representations that can be manipulated coherently at inference. This allows us to generate musical variations and blends from reference examples while preserving characteristics at desired timescales, which are specified as frequencies in the latent space. LatentFT parallels the role of the equalizer in music production: while traditional equalizers operates on audible frequencies to shape timbre, LatentFT operates on latent-space frequencies to shape musical structure. Experiments and listening tests show that LatentFT improves condition adherence and quality compared to baselines. We also present a technique for hearing frequencies in the latent space in isolation, and show different musical attributes reside in different regions of the latent spectrum. Our results show how frequency-domain control in latent space provides an intuitive, continuous frequency axis for conditioning and blending, advancing us toward more interpretable and interactive generative music models.", "date": "2026-04-20", "categories": []}
{"id": "2604.18204", "title": "Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages", "abstract": "We present a phoneme-level analysis of automatic speech recognition (ASR) for two low-resourced and phonologically complex East Caucasian languages, Archi and Rutul, based on curated and standardized speech-transcript resources totaling approximately 50 minutes and 1 hour 20 minutes of audio, respectively. Existing recordings and transcriptions are consolidated and processed into a form suitable for ASR training and evaluation. We evaluate several state-of-the-art audio and audio-language models, including wav2vec2, Whisper, and Qwen2-Audio. For wav2vec2, we introduce a language-specific phoneme vocabulary with heuristic output-layer initialization, which yields consistent improvements and achieves performance comparable to or exceeding Whisper in these extremely low-resource settings. Beyond standard word and character error rates, we conduct a detailed phoneme-level error analysis. We find that phoneme recognition accuracy strongly correlates with training frequency, exhibiting a characteristic sigmoid-shaped learning curve. For Archi, this relationship partially breaks for Whisper, pointing to model-specific generalization effects beyond what is predicted by training frequency. Overall, our results indicate that many errors attributed to phonological complexity are better explained by data scarcity. These findings demonstrate the value of phoneme-level evaluation for understanding ASR behavior in low-resource, typologically complex languages.", "date": "2026-04-20", "categories": []}
{"id": "2604.18270", "title": "Incremental learning for audio classification with Hebbian Deep Neural Networks", "abstract": "The ability of humans for lifelong learning is an inspiration for deep learning methods and in particular for continual learning. In this work, we apply Hebbian learning, a biologically inspired learning process, to sound classification. We propose a kernel plasticity approach that selectively modulates network kernels during incremental learning, acting on selected kernels to learn new information and on others to retain previous knowledge. Using the ESC-50 dataset, the proposed method achieves 76.3% overall accuracy over five incremental steps, outperforming a baseline without kernel plasticity (68.7%) and demonstrating significantly greater stability across tasks.", "date": "2026-04-20", "categories": []}
{"id": "2510.09065", "title": "MMAudioSep: Taming Video-to-Audio Generative Model Towards Video/Text-Queried Sound Separation", "abstract": "We introduce MMAudioSep, a generative model for video/text-queried sound separation that is founded on a pretrained video-to-audio model. By leveraging knowledge about the relationship between video/text and audio learned through a pretrained audio generative model, we can train the model more efficiently, i.e., the model does not need to be trained from scratch. We evaluate the performance of MMAudioSep by comparing it to existing separation models, including models based on both deterministic and generative approaches, and find it is superior to the baseline models. Furthermore, we demonstrate that even after acquiring functionality for sound separation via fine-tuning, the model retains the ability for original video-to-audio generation. This highlights the potential of foundational sound generation models to be adopted for sound-related downstream tasks. Our code is available at https://github.com/sony/mmaudiosep.", "date": "2025-10-10", "categories": []}
{"id": "2512.04847", "title": "Language Models as Semantic Teachers: Post-Training Alignment for Medical Audio Understanding", "abstract": "Pre-trained audio models excel at detecting acoustic patterns in auscultation sounds but often fail to grasp their clinical significance, limiting their use and performance in diagnostic tasks. To bridge this gap, we introduce AcuLa (Audio-Clinical Understanding via Language Alignment), a lightweight post-training framework that instills semantic understanding into any audio encoder by aligning it with a medical language model, which acts as a \"semantic teacher.\" To enable alignment at scale, we construct a large-scale dataset by leveraging off-the-shelf large language models to translate the rich, structured metadata accompanying existing audio recordings into coherent clinical reports. Our alignment strategy combines a representation-level contrastive objective with a self-supervised modeling, ensuring that the model learns clinical semantics while preserving fine-grained temporal cues. AcuLa achieves state-of-the-art results across 18 diverse cardio-respiratory tasks from 10 different datasets, improving the mean AUROC on classification benchmarks from 0.68 to 0.79 and, on the most challenging COVID-19 cough detection task, boosting the AUROC from 0.55 to 0.89. Our work demonstrates that this audio-language alignment transforms purely acoustic models into clinically-aware diagnostic tools, establishing a novel paradigm for enhancing physiological understanding in audio-based health monitoring.", "date": "2025-12-04", "categories": []}
{"id": "2604.15383", "title": "Temporal Contrastive Decoding: A Training-Free Method for Large Audio-Language Models", "abstract": "Large audio-language models (LALMs) generalize across speech, sound, and music, but unified decoders can exhibit a \\emph{temporal smoothing bias}: transient acoustic cues may be underutilized in favor of temporally smooth context that is better supported by language priors, leading to less specific audio-grounded outputs. We propose \\emph{Temporal Contrastive Decoding} (TCD), a training-free decoding method for unified LALMs that mitigates this effect at inference time. TCD constructs a temporally blurred slow-path view by smoothing the input waveform and re-encoding it, then contrasts next-token logits from the original and slow-path views. The contrastive signal is applied as a token-level logit update restricted to a small candidate set. A self-normalized stability score sets the blur window and update scale, and a step-wise gate based on uncertainty and audio reliance activates the update only when needed. Experiments on MMAU and AIR-Bench show consistent improvements on strong unified LALMs. We further conduct ablations and an architectural applicability study to analyze the contributions of key components and how TCD behaves across large audio-language model designs.", "date": "2026-04-16", "categories": []}
{"id": "2604.16056", "title": "AST: Adaptive, Seamless, and Training-Free Precise Speech Editing", "abstract": "Text-based speech editing aims to modify specific segments while preserving speaker identity and acoustic context. Existing methods rely on task-specific training, which incurs high data costs and struggles with temporal fidelity in unedited regions. Meanwhile, adapting Text-to-Speech (TTS) models often faces a trade-off between editing quality and consistency. To address these issues, we propose AST, an Adaptive, Seamless, and Training-free precise speech editing framework. Leveraging a pre-trained autoregressive TTS model, AST introduces Latent Recomposition to selectively stitch preserved source segments with newly synthesized targets. Furthermore, AST extends this latent manipulation to enable precise style editing for specific speech segments. To prevent artifacts at these edit boundaries, the framework incorporates Adaptive Weak Fact Guidance (AWFG). AWFG dynamically modulates a mel-space guidance signal, enforcing structural constraints only where necessary without disrupting the generative manifold. To fill the gap of publicly accessible benchmarks, we introduce LibriSpeech-Edit, a new and larger speech editing dataset. As existing metrics poorly evaluate temporal consistency in unedited regions, we propose Word-level Dynamic Time Warping (WDTW). Extensive experiments demonstrate that AST resolves the controllability-quality trade-off without extra training. Compared to the previous most temporally consistent baseline, AST improves consistency while reducing Word Error Rate by nearly 70%. Moreover, applying AST to a foundation TTS model reduces WDTW by 27%, achieving state-of-the-art speaker preservation and temporal fidelity.", "date": "2026-04-17", "categories": []}
{"id": "2509.26007", "title": "MARS: Sound Generation via Multi-Channel Autoregression on Spectrograms", "abstract": "Research on audio generation has progressively developed along both waveform-based and spectrogram-based directions, giving rise to diverse strategies for representing and generating audio. At the same time, advances in image synthesis have shown that autoregression across scales, rather than tokens, improves coherence and detail. Building on these ideas, we introduce MARS (Multi-channel AutoRegression on Spectrograms), which, to the best of our knowledge, is the first adaptation of next-scale autoregressive modeling to the spectrogram domain. MARS treats spectrograms as multi-channel images and employs channel multiplexing (CMX), a reshaping strategy that reduces spatial resolution without information loss. A shared tokenizer provides consistent discrete representations across scales, enabling a transformer-based autoregressor to refine spectrograms from coarse to fine resolutions efficiently. Experiments on a large-scale dataset demonstrate that MARS performs comparably or better than state-of-the-art baselines across multiple evaluation metrics, establishing an efficient and scalable paradigm for high-fidelity sound generation.", "date": "2025-09-30", "categories": []}
{"id": "2604.15086", "title": "ControlFoley: Unified and Controllable Video-to-Audio Generation with Cross-Modal Conflict Handling", "abstract": "Recent advances in video-to-audio (V2A) generation enable high-quality audio synthesis from visual content, yet achieving robust and fine-grained controllability remains challenging. Existing methods suffer from weak textual controllability under visual-text conflict and imprecise stylistic control due to entangled temporal and timbre information in reference audio. Moreover, the lack of standardized benchmarks limits systematic evaluation.\n  We propose ControlFoley, a unified multimodal V2A framework that enables precise control over video, text, and reference audio. We introduce a joint visual encoding paradigm that integrates CLIP with a spatio-temporal audio-visual encoder to improve alignment and textual controllability. We further propose temporal-timbre decoupling to suppress redundant temporal cues while preserving discriminative timbre features. In addition, we design a modality-robust training scheme with unified multimodal representation alignment (REPA) and random modality dropout. We also present VGGSound-TVC, a benchmark for evaluating textual controllability under varying degrees of visual-text conflict.\n  Extensive experiments demonstrate state-of-the-art performance across multiple V2A tasks, including text-guided, text-controlled, and audio-controlled generation. ControlFoley achieves superior controllability under cross-modal conflict while maintaining strong synchronization and audio quality, and shows competitive or better performance compared to an industrial V2A system.\n  Code, models, datasets, and demos are available at: https://yjx-research.github.io/ControlFoley/.", "date": "2026-04-16", "categories": []}
{"id": "2503.10522", "title": "AudioX: A Unified Framework for Anything-to-Audio Generation", "abstract": "Audio and music generation based on flexible multimodal control signals is a widely applicable topic, with the following key challenges: 1) a unified multimodal modeling framework, and 2) large-scale, high-quality training data. As such, we propose AudioX, a unified framework for anything-to-audio generation that integrates varied multimodal conditions (i.e., text, video, and audio signals) in this work. The core design in this framework is a Multimodal Adaptive Fusion module, which enables the effective fusion of diverse multimodal inputs, enhancing cross-modal alignment and improving overall generation quality. To train this unified model, we construct a large-scale, high-quality dataset, IF-caps, comprising over 7 million samples curated through a structured data annotation pipeline. This dataset provides comprehensive supervision for multimodal-conditioned audio generation. We benchmark AudioX against state-of-the-art methods across a wide range of tasks, finding that our model achieves superior performance, especially in text-to-audio and text-to-music generation. These results demonstrate our method is capable of audio generation under multimodal control signals, showing powerful instruction-following potential. The code and datasets will be available at https://zeyuet.github.io/AudioX/.", "date": "2025-03-13", "categories": []}
{"id": "2604.13715", "title": "Towards Fine-grained Temporal Perception: Post-Training Large Audio-Language Models with Audio-Side Time Prompt", "abstract": "Large Audio-Language Models (LALMs) enable general audio understanding and demonstrate remarkable performance across various audio tasks. However, these models still face challenges in temporal perception (e.g., inferring event onset and offset), leading to limited utility in fine-grained scenarios. To address this issue, we propose Audio-Side Time Prompt and leverage Reinforcement Learning (RL) to develop the TimePro-RL framework for fine-grained temporal perception. Specifically, we encode timestamps as embeddings and interleave them within the audio feature sequence as temporal coordinates to prompt the model. Furthermore, we introduce RL following Supervised Fine-Tuning (SFT) to directly optimize temporal alignment performance. Experiments demonstrate that TimePro-RL achieves significant performance gains across a range of audio temporal tasks, such as audio grounding, sound event detection, and dense audio captioning, validating its robust effectiveness.", "date": "2026-04-15", "categories": []}
{"id": "2406.01253", "title": "animal2vec and MeerKAT: A self-supervised transformer for rare-event raw audio input and a large-scale reference dataset for bioacoustics", "abstract": "Bioacoustic research, vital for understanding animal behavior, conservation, and ecology, faces a monumental challenge: analyzing vast datasets where animal vocalizations are rare. While deep learning techniques are becoming standard, adapting them to bioacoustics remains difficult. We address this with animal2vec, an interpretable large transformer model, and a self-supervised training scheme tailored for sparse and unbalanced bioacoustic data. It learns from unlabeled audio and then refines its understanding with labeled data. Furthermore, we introduce and publicly release MeerKAT: Meerkat Kalahari Audio Transcripts, a dataset of meerkat (Suricata suricatta) vocalizations with millisecond-resolution annotations, the largest labeled dataset on non-human terrestrial mammals currently available. Our model outperforms existing methods on MeerKAT and the publicly available NIPS4Bplus birdsong dataset. Moreover, animal2vec performs well even with limited labeled data (few-shot learning). animal2vec and MeerKAT provide a new reference point for bioacoustic research, enabling scientists to analyze large amounts of data even with scarce ground truth information.", "date": "2024-06-03", "categories": []}
{"id": "2509.22220", "title": "StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs", "abstract": "Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; even at high Signal-to-Noise Ratios (SNRs) where speech is perfectly intelligible, their output token sequences can change drastically, increasing the learning burden for downstream LLMs. This instability stems from two flaws: a brittle single-path quantization architecture and a distant training signal indifferent to intermediate token stability. To address this, we introduce StableToken, a tokenizer that achieves stability through a consensus-driven mechanism. Its multi-branch architecture processes audio in parallel, and these representations are merged via a powerful bit-wise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks. Our code and model are publicly available at https://github.com/Tencent/StableToken.", "date": "2025-09-26", "categories": []}
{"id": "2603.27557", "title": "A General Model for Deepfake Speech Detection: Diverse Bonafide Resources or Diverse AI-Based Generators", "abstract": "In this paper, we analyze two main factors of Bonafide Resource (BR) or AI-based Generator (AG) which affect the performance and the generality of a Deepfake Speech Detection (DSD) model. To this end, we first propose a deep-learning based model, referred to as the baseline. Then, we conducted experiments on the baseline by which we indicate how Bonafide Resource (BR) and AI-based Generator (AG) factors affect the threshold score used to detect fake or bonafide input audio in the inference process. Given the experimental results, a dataset, which re-uses public Deepfake Speech Detection (DSD) datasets and shows a balance between Bonafide Resource (BR) or AI-based Generator (AG), is proposed. We then train various deep-learning based models on the proposed dataset and conduct cross-dataset evaluation on different benchmark datasets. The cross-dataset evaluation results prove that the balance of Bonafide Resources (BR) and AI-based Generators (AG) is the key factor to train and achieve a general Deepfake Speech Detection (DSD) model.", "date": "2026-03-29", "categories": []}
{"id": "2604.12292", "title": "CoSyncDiT: Cognitive Synchronous Diffusion Transformer for Movie Dubbing", "abstract": "Movie dubbing aims to synthesize speech that preserves the vocal identity of a reference audio while synchronizing with the lip movements in a target video. Existing methods fail to achieve precise lip-sync and lack naturalness due to explicit alignment at the duration level. While implicit alignment solutions have emerged, they remain susceptible to interference from the reference audio, triggering timbre and pronunciation degradation in in-the-wild scenarios. In this paper, we propose a novel flow matching-based movie dubbing framework driven by the Cognitive Synchronous Diffusion Transformer (CoSync-DiT), inspired by the cognitive process of professional actors. This architecture progressively guides the noise-to-speech generative trajectory by executing acoustic style adapting, fine-grained visual calibrating, and time-aware context aligning. Furthermore, we design the Joint Semantic and Alignment Regularization (JSAR) mechanism to simultaneously constrain frame-level temporal consistency on the contextual outputs and semantic consistency on the flow hidden states, ensuring robust alignment. Extensive experiments on both standard benchmarks and challenging in-the-wild dubbing benchmarks demonstrate that our method achieves the state-of-the-art performance across multiple metrics.", "date": "2026-04-14", "categories": []}
{"id": "2604.12480", "title": "Audio Source Separation in Reverberant Environments using $\\beta$-divergence based Nonnegative Factorization", "abstract": "In Gaussian model-based multichannel audio source separation, the likelihood of observed mixtures of source signals is parametrized by source spectral variances and by associated spatial covariance matrices. These parameters are estimated by maximizing the likelihood through an Expectation-Maximization algorithm and used to separate the signals by means of multichannel Wiener filtering.\n  We propose to estimate these parameters by applying nonnegative factorization based on prior information on source variances. In the nonnegative factorization, spectral basis matrices can be defined as the prior information. The matrices can be either extracted or indirectly made available through a redundant library that is trained in advance. In a separate step, applying nonnegative tensor factorization, two algorithms are proposed in order to either extract or detect the basis matrices that best represent the power spectra of the source signals in the observed mixtures. The factorization is achieved by minimizing the $\\beta$-divergence through multiplicative update rules. The sparsity of factorization can be controlled by tuning the value of $\\beta$.\n  Experiments show that sparsity, rather than the value assigned to $\\beta$ in the training, is crucial in order to increase the separation performance. The proposed method was evaluated in several mixing conditions. It provides better separation quality with respect to other comparable algorithms.", "date": "2026-04-14", "categories": []}
{"id": "2604.12506", "title": "Beyond Transcription: Unified Audio Schema for Perception-Aware AudioLLMs", "abstract": "Recent Audio Large Language Models (AudioLLMs) exhibit a striking performance inversion: while excelling at complex reasoning tasks, they consistently underperform on fine-grained acoustic perception. We attribute this gap to a fundamental limitation of ASR-centric training, which provides precise linguistic targets but implicitly teaches models to suppress paralinguistic cues and acoustic events as noise. To address this, we propose Unified Audio Schema (UAS), a holistic and structured supervision framework that organizes audio information into three explicit components -- Transcription, Paralinguistics, and Non-linguistic Events -- within a unified JSON format. This design achieves comprehensive acoustic coverage without sacrificing the tight audio-text alignment that enables reasoning. We validate the effectiveness of this supervision strategy by applying it to both discrete and continuous AudioLLM architectures. Extensive experiments on MMSU, MMAR, and MMAU demonstrate that UAS-Audio yields consistent improvements, boosting fine-grained perception by 10.9% on MMSU over the same-size state-of-the-art models while preserving robust reasoning capabilities. Our code and model are publicly available at https://github.com/Tencent/Unified_Audio_Schema.", "date": "2026-04-14", "categories": []}
{"id": "2604.12647", "title": "Adaptive Test-Time Scaling for Zero-Shot Respiratory Audio Classification", "abstract": "Automated respiratory audio analysis promises scalable, non-invasive disease screening, yet progress is limited by scarce labeled data and costly expert annotation. Zero-shot inference eliminates task-specific supervision, but existing methods apply uniform computation to every input regardless of difficulty. We introduce TRIAGE, a tiered zero-shot framework that adaptively scales test-time compute by routing each audio sample through progressively richer reasoning stages: fast label-cosine scoring in a joint audio-text embedding space (Tier-L), structured matching with clinician-style descriptors (Tier-M), and retrieval-augmented large language model reasoning (Tier-H). A confidence-based router finalizes easy predictions early while allocating additional computation to ambiguous inputs, enabling nearly half of all samples to exit at the cheapest tier. Across nine respiratory classification tasks without task-specific training, TRIAGE achieves a mean AUROC of 0.744, outperforming prior zero-shot methods and matching or exceeding supervised baselines on multiple tasks. Our analysis show that test-time scaling concentrates gains where they matter: uncertain cases see up to 19% relative improvement while confident predictions remain unchanged at minimal cost.", "date": "2026-04-14", "categories": []}
{"id": "2604.12733", "title": "Transformer Based Machine Fault Detection From Audio Input", "abstract": "In recent years, Sound AI is being increasingly used to predict machine failures. By attaching a microphone to the machine of interest, one can get real time data on machine behavior from the field. Traditionally, Convolutional Neural Net (CNN) architectures have been used to analyze spectrogram images generated from the sounds captured and predict if the machine is functioning as expected. CNN architectures seem to work well empirically even though they have biases like locality and parameter-sharing which may not be completely relevant for spectrogram analysis. With the successful application of transformer-based models in the field of image processing starting with Vision Transformer (ViT) in 2020, there has been significant interest in leveraging these in the field of Sound AI. Since transformer-based architectures have significantly lower inductive biases, they are expected to perform better than CNNs at spectrogram analysis given enough data. This paper demonstrates the effectiveness of transformer-driven architectures in analyzing Sound data and compares the embeddings they generate with CNNs on the specific task of machine fault detection.", "date": "2026-04-14", "categories": []}
{"id": "2601.04867", "title": "Gradient-based Optimisation of Modulation Effects", "abstract": "Modulation effects such as phasers, flangers and chorus effects are heavily used in conjunction with the electric guitar. Machine learning based emulation of analog modulation units has been investigated in recent years, but most methods have either been limited to one class of effect or suffer from a high computational cost or latency compared to canonical digital implementations. Here, we build on previous work and present a framework for modelling flanger, chorus and phaser effects based on differentiable digital signal processing. The model is trained in the time-frequency domain, but at inference operates in the time-domain, requiring zero latency. We investigate the challenges associated with gradient-based optimisation of such effects, and show that low-frequency weighting of loss functions avoids convergence to local minima when learning delay times. We show that when trained against analog effects units, sound output from the model is in some cases perceptually indistinguishable from the reference, but challenges still remain for effects with long delay times and feedback.", "date": "2026-01-08", "categories": []}
{"id": "2604.09675", "title": "Real-Time Voicemail Detection in Telephony Audio Using Temporal Speech Activity Features", "abstract": "Outbound AI calling systems must distinguish voicemail greetings from live human answers in real time to avoid wasted agent interactions and dropped calls. We present a lightweight approach that extracts 15 temporal features from the speech activity pattern of a pre-trained neural voice activity detector (VAD), then classifies with a shallow tree-based ensemble. Across two evaluation sets totaling 764 telephony recordings, the system achieves a combined 96.1% accuracy (734/764), with 99.3% (139/140) on an expert-labeled test set and 95.4% (595/624) on a held-out production set. In production validation over 77,000 calls, it maintained a 0.3% false positive rate and 1.3% false negative rate. End-to-end inference completes in 46 ms on a commodity dual-core CPU with no GPU, supporting 380+ concurrent WebSocket calls. In our search over 3,780 model, feature, and threshold combinations, feature importance was concentrated in three temporal variables. Adding transcription keywords or beep-based features did not improve the best real-time configuration and increased latency substantially. Our results suggest that temporal speech patterns are a strong signal for distinguishing voicemail greetings from live human answers.", "date": "2026-04-02", "categories": []}
{"id": "2604.10021", "title": "Masked Contrastive Pre-Training Improves Music Audio Key Detection", "abstract": "Self-supervised music foundation models underperform on key detection, which requires pitch-sensitive representations. In this work, we present the first systematic study showing that the design of self-supervised pretraining directly impacts pitch sensitivity, and demonstrate that masked contrastive embeddings uniquely enable state-of-the-art (SOTA) performance in key detection in the supervised setting. First, we discover that linear evaluation after masking-based contrastive pretraining on Mel spectrograms leads to competitive performance on music key detection out of the box. This leads us to train shallow but wide multi-layer perceptrons (MLPs) on features extracted from our base model, leading to SOTA performance without the need for sophisticated data augmentation policies. We further analyze robustness and show empirically that the learned representations naturally encode common augmentations. Our study establishes self-supervised pretraining as an effective approach for pitch-sensitive MIR tasks and provides insights for designing and probing music foundation models.", "date": "2026-04-11", "categories": []}
{"id": "2604.10503", "title": "Cross-Cultural Bias in Mel-Scale Representations: Evidence and Alternatives from Speech and Music", "abstract": "Modern audio systems universally employ mel-scale representations derived from 1940s Western psychoacoustic studies, potentially encoding cultural biases that create systematic performance disparities. We present a comprehensive evaluation of cross-cultural bias in audio front-ends, comparing mel-scale features with learnable alternatives (LEAF, SincNet) and psychoacoustic variants (ERB, Bark, CQT) across speech recognition (11 languages), music analysis (6 collections), and European acoustic scene classification (10 European cities). Our controlled experiments isolate front-end contributions while holding architecture and training protocols minimal and constant. Results demonstrate that mel-scale features yield 31.2% WER for tonal languages compared to 18.7% for non-tonal languages (12.5% gap), and show 15.7% F1 degradation between Western and non-Western music. Alternative representations significantly reduce these disparities: LEAF reduces the speech gap by 34% through adaptive frequency allocation, CQT achieves 52% reduction in music performance gaps, and ERB-scale filtering cuts disparities by 31% with only 1% computational overhead. We also release FairAudioBench, enabling cross-cultural evaluation, and demonstrate that adaptive frequency decomposition offers practical paths toward equitable audio processing. These findings reveal how foundational signal processing choices propagate bias, providing crucial guidance for developing inclusive audio systems.", "date": "2026-04-12", "categories": []}
{"id": "2604.10905", "title": "Audio Flamingo Next: Next-Generation Open Audio-Language Models for Speech, Sound, and Music", "abstract": "We present Audio Flamingo Next (AF-Next), the next-generation and most capable large audio-language model in the Audio Flamingo series, designed to advance understanding and reasoning over speech, environmental sounds and music. Compared to Audio Flamingo 3, AF-Next introduces: (i) a stronger foundational audio-language model that significantly improves accuracy across diverse audio understanding tasks; (ii) scalable strategies for constructing large-scale audio understanding and reasoning data beyond existing academic benchmarks; (iii) support for long and complex audio inputs up to 30 minutes; and (iv) Temporal Audio Chain-of-Thought, a new reasoning paradigm that explicitly grounds intermediate reasoning steps to timestamps in long audio, enabling fine-grained temporal alignment and improved interpretability. To enable these capabilities, we first conduct a systematic analysis of Audio Flamingo 3 to identify key gaps in audio understanding and reasoning. We then curate and scale new large-scale datasets totaling over 1 million hours to address these limitations and expand the existing AudioSkills-XL, LongAudio-XL, AF-Think and AF-Chat datasets. AF-Next is trained using a curriculum-based strategy spanning pre-training, mid-training and post-training stages. Extensive experiments across 20 audio understanding and reasoning benchmarks, including challenging long-audio tasks, show that AF-Next outperforms similarly sized open models by large margins and remains highly competitive with and sometimes surpasses, much larger open-weight and closed models. Beyond benchmark performance, AF-Next exhibits strong real-world utility and transfers well to unseen tasks, highlighting its robustness and generalization ability. In addition to all data, code and methods, we open-source 3 variants of AF-Next, including AF-Next-Instruct, AF-Next-Think and AF-Next-Captioner.", "date": "2026-04-13", "categories": []}
{"id": "2604.11560", "title": "bacpipe: a Python package to make bioacoustic deep learning models accessible", "abstract": "1. Natural sounds have been recorded for millions of hours over the previous decades using passive acoustic monitoring. Improvements in deep learning models have vastly accelerated the analysis of large portions of this data. While new models advance the state-of-the-art, accessing them using tools to harness their full potential is not always straightforward. Here we present bacpipe, a collection of bioacoustic deep learning models and evaluation pipelines accessible through a graphical and programming interface, designed for both ecologists and computer scientists. Bacpipe is a modular software package intended as a point of convergence for bioacoustic models.\n  2. Bacpipe streamlines the usage of state-of-the-art models on custom audio datasets, generating acoustic feature vectors (embeddings) and classifier predictions. A modular design allows evaluation and benchmarking of models through interactive visualizations, clustering and probing.\n  3. We believe that access to new deep learning models is important. By designing bacpipe to target a wide audience, researchers will be enabled to answer new ecological and evolutionary questions in bioacoustics.\n  4. In conclusion, we believe accessibility to developments in deep learning to a wider audience benefits the ecological questions we are trying to answer.", "date": "2026-04-13", "categories": []}
{"id": "2604.09021", "title": "Noise-Aware In-Context Learning for Hallucination Mitigation in ALLMs", "abstract": "Auditory large language models (ALLMs) have demonstrated strong general capabilities in audio understanding and reasoning tasks. However, their reliability is still undermined by hallucination issues. Existing hallucination evaluation methods are formulated as binary classification tasks, which are insufficient to characterize the more complex hallucination patterns that arise in generative tasks. Moreover, current hallucination mitigation strategies rely on fine-tuning, resulting in high computational costs. To address the above limitations, we propose a plug-and-play Noise-Aware In-Context Learning (NAICL) method. Specifically, we construct a noise prior library, retrieve noise examples relevant to the input audio, and incorporate them as contextual priors, thereby guiding the model to reduce speculative associations when acoustic evidence is insufficient and to adopt a more conservative generation strategy. In addition, we establish a hallucination benchmark for audio caption tasks including the construction of the Clotho-1K multi-event benchmark dataset, the definition of four types of auditory hallucinations, and the introduction of metrics such as hallucination type distribution to support fine-grained analysis. Experimental results show that all evaluated ALLMs exhibit same hallucination behaviors. Moreover, the proposed NAICL method reduces the overall hallucination rate from 26.53% to 16.98%.", "date": "2026-04-10", "categories": []}
{"id": "2604.09094", "title": "Few-Shot Contrastive Adaptation for Audio Abuse Detection in Low-Resource Indic Languages", "abstract": "Abusive speech detection is becoming increasingly important as social media shifts towards voice-based interaction, particularly in multilingual and low-resource settings. Most current systems rely on automatic speech recognition (ASR) followed by text-based hate speech classification, but this pipeline is vulnerable to transcription errors and discards prosodic information carried in speech. We investigate whether Contrastive Language-Audio Pre-training (CLAP) can support abusive speech detection directly from audio. Using the ADIMA dataset, we evaluate CLAP-based representations under few-shot supervised contrastive adaptation in cross-lingual and leave-one-language-out settings, with zero-shot prompting included as an auxiliary analysis. Our results show that CLAP yields strong cross-lingual audio representations across ten Indic languages, and that lightweight projection-only adaptation achieves competitive performance with respect to fully supervised systems trained on complete training data. However, the benefits of few-shot adaptation are language-dependent and not monotonic with shot size. These findings suggest that contrastive audio-text models provide a promising basis for cross-lingual audio abuse detection in low-resource settings, while also indicating that transfer remains incomplete and language-specific in important ways.", "date": "2026-04-10", "categories": []}
{"id": "2604.09222", "title": "GRM: Utility-Aware Jailbreak Attacks on Audio LLMs via Gradient-Ratio Masking", "abstract": "Audio large language models (ALLMs) enable rich speech-text interaction, but they also introduce jailbreak vulnerabilities in the audio modality. Existing audio jailbreak methods mainly optimize jailbreak success while overlooking utility preservation, as reflected in transcription quality and question answering performance. In practice, stronger attacks often come at the cost of degraded utility. To study this trade-off, we revisit existing attacks by varying their perturbation coverage in the frequency domain, from partial-band to full-band, and find that broader frequency coverage does not necessarily improve jailbreak performance, while utility consistently deteriorates. This suggests that concentrating perturbation on a subset of bands can yield a better attack-utility trade-off than indiscriminate full-band coverage. Based on this insight, we propose GRM, a utility-aware frequency-selective jailbreak framework. It ranks Mel bands by their attack contribution relative to utility sensitivity, perturbs only a selected subset of bands, and learns a reusable universal perturbation under a semantic-preservation objective. Experiments on four representative ALLMs show that GRM achieves an average Jailbreak Success Rate (JSR) of 88.46% while providing a better attack-utility trade-off than representative baselines. These results highlight the potential of frequency-selective perturbation for better balancing attack effectiveness and utility preservation in audio jailbreak. Content Warning: This paper includes harmful query examples and unsafe model responses.", "date": "2026-04-10", "categories": []}
{"id": "2504.13102", "title": "A Multi-task Learning Balanced Attention Convolutional Neural Network Model for Few-shot Underwater Acoustic Target Recognition", "abstract": "Underwater acoustic target recognition (UATR) is of great significance for the protection of marine diversity and national defense security. The development of deep learning provides new opportunities for UATR, but faces challenges brought by the scarcity of reference samples and complex environmental interference. To address these issues, we proposes a multi-task balanced channel attention convolutional neural network (MT-BCA-CNN). The method integrates a channel attention mechanism with a multi-task learning strategy, constructing a shared feature extractor and multi-task classifiers to jointly optimize target classification and feature reconstruction tasks. The channel attention mechanism dynamically enhances discriminative acoustic features such as harmonic structures while suppressing noise. Experiments on the Watkins Marine Life Dataset demonstrate that MT-BCA-CNN achieves 97\\% classification accuracy and 95\\% $F1$-score in 27-class few-shot scenarios, significantly outperforming traditional CNN and ACNN models, as well as popular state-of-the-art UATR methods. Ablation studies confirm the synergistic benefits of multi-task learning and attention mechanisms, while a dynamic weighting adjustment strategy effectively balances task contributions. This work provides an efficient solution for few-shot underwater acoustic recognition, advancing research in marine bioacoustics and sonar signal processing.", "date": "2025-04-17", "categories": []}
{"id": "2509.23727", "title": "AudioMoG: Guiding Audio Generation with Mixture-of-Guidance", "abstract": "The design of diffusion-based audio generation systems has been investigated from diverse perspectives, such as data space, network architecture, and conditioning techniques, while most of these innovations require model re-training. In sampling, classifier-free guidance (CFG) has been uniformly adopted to enhance generation quality by strengthening condition alignment. However, CFG often compromises diversity, resulting in suboptimal performance. Although the recent autoguidance (AG) method proposes another direction of guidance that maintains diversity, its direct application in audio generation has so far underperformed CFG. In this work, we introduce AudioMoG, an improved sampling method that enhances text-to-audio (T2A) and video-to-audio (V2A) generation quality without requiring extensive training resources. We start with an analysis of both CFG and AG, examining their respective advantages and limitations for guiding diffusion models. Building upon our insights, we introduce a mixture-of-guidance framework that integrates diverse guidance signals with their interaction terms (e.g., the unconditional bad version of the model) to maximize cumulative advantages. Experiments show that, given the same inference speed, our approach consistently outperforms single guidance in T2A generation across sampling steps, concurrently showing advantages in V2A, text-to-music, and image generation. Demo samples are available at: https://audiomog.github.io.", "date": "2025-09-28", "categories": []}
{"id": "2603.01482", "title": "A SUPERB-Style Benchmark of Self-Supervised Speech Models for Audio Deepfake Detection", "abstract": "Self-supervised learning (SSL) has transformed speech processing, with benchmarks such as SUPERB establishing fair comparisons across diverse downstream tasks. Despite it's security-critical importance, Audio deepfake detection has remained outside these efforts. In this work, we introduce Spoof-SUPERB, a benchmark for audio deepfake detection that systematically evaluates 20 SSL models spanning generative, discriminative, and spectrogram-based architectures. We evaluated these models on multiple in-domain and out-of-domain datasets. Our results reveal that large-scale discriminative models such as XLS-R, UniSpeech-SAT, and WavLM Large consistently outperform other models, benefiting from multilingual pretraining, speaker-aware objectives, and model scale. We further analyze the robustness of these models under acoustic degradations, showing that generative approaches degrade sharply, while discriminative models remain resilient. This benchmark establishes a reproducible baseline and provides practical insights into which SSL representations are most reliable for securing speech systems against audio deepfakes.", "date": "2026-03-02", "categories": []}
{"id": "2604.07612", "title": "Towards Real-Time Human-AI Musical Co-Performance: Accompaniment Generation with Latent Diffusion Models and MAX/MSP", "abstract": "We present a framework for real-time human-AI musical co-performance, in which a latent diffusion model generates instrumental accompaniment in response to a live stream of context audio. The system combines a MAX/MSP front-end-handling real-time audio input, buffering, and playback-with a Python inference server running the generative model, communicating via OSC/UDP messages. This allows musicians to perform in MAX/MSP - a well-established, real-time capable environment - while interacting with a large-scale Python-based generative model, overcoming the fundamental disconnect between real-time music tools and state-of-the-art AI models. We formulate accompaniment generation as a sliding-window look-ahead protocol, training the model to predict future audio from partial context, where system latency is a critical constraint. To reduce latency, we apply consistency distillation to our diffusion model, achieving a 5.4x reduction in sampling time, with both models achieving real-time operation. Evaluated on musical coherence, beat alignment, and audio quality, both models achieve strong performance in the Retrospective regime and degrade gracefully as look-ahead increases. These results demonstrate the feasibility of diffusion-based real-time accompaniment and expose the fundamental trade-off between model latency, look-ahead depth, and generation quality that any such system must navigate.", "date": "2026-04-08", "categories": []}
{"id": "2604.08087", "title": "DeepForestSound: a multi-species automatic detector for passive acoustic monitoring in African tropical forests, a case study in Kibale National Park", "abstract": "Passive Acoustic Monitoring (PAM) is widely used for biodiversity assessment. Its application in African tropical forests is limited by scarce annotated data, reducing the performance of general-purpose ecoacoustic models on underrepresented taxa. In this study, we introduce DeepForestSound (DFS), a multi-species automatic detection model designed for PAM in African tropical forests. DFS relies on a semi-supervised pipeline combining clustering of unannotated recordings with manual validation, followed by supervised fine-tuning of an Audio Spectrogram Transformer (AST) using low-rank adaptation, which is compared to a frozen-backbone linear baseline (DFS-Linear). The framework supports the detection of multiple taxonomic groups, including birds, primates, and elephants, from long-term acoustic recordings. DFS was trained on acoustic data collected in the Sebitoli area, in Kibale National Park, Uganda, and evaluated on an independent dataset recorded two years later at different locations within the same forest. This evaluation therefore assesses generalization across time and recording sites within a single tropical forest ecosystem. Across 8 out of 12 taxons, DFS outperforms existing automatic detection tools, particularly for non-avian taxa, achieving average AP values of 0.964 for primates and 0.961 for elephants. Results further show that LoRA-based fine-tuning substantially outperforms linear probing across taxa. Overall, these results demonstrate that task-oriented, region-specific training substantially improves detection performance in acoustically complex tropical environments, and highlight the potential of DFS as a practical tool for biodiversity monitoring and conservation in African rainforests.", "date": "2026-04-09", "categories": []}
{"id": "2604.08104", "title": "Quantum Vision Theory Applied to Audio Classification for Deepfake Speech Detection", "abstract": "We propose Quantum Vision (QV) theory as a new perspective for deep learning-based audio classification, applied to deepfake speech detection. Inspired by particle-wave duality in quantum physics, QV theory is based on the idea that data can be represented not only in its observable, collapsed form, but also as information waves. In conventional deep learning, models are trained directly on these collapsed representations, such as images. In QV theory, inputs are first transformed into information waves using a QV block, and then fed into deep learning models for classification. QV-based models improve performance in image classification compared to their non-QV counterparts. What if QV theory is applied speech spectrograms for audio classification tasks? This is the motivation and novelty of the proposed approach. In this work, Short-Time Fourier Transform (STFT), Mel-spectrograms, and Mel-Frequency Cepstral Coefficients (MFCC) of speech signals are converted into information waves using the proposed QV block and used to train QV-based Convolutional Neural Networks (QV-CNN) and QV-based Vision Transformers (QV-ViT). Extensive experiments are conducted on the ASVSpoof dataset for deepfake speech classification. The results show that QV-CNN and QV-ViT consistently outperform standard CNN and ViT models, achieving higher classification accuracy and improved robustness in distinguishing genuine and spoofed speech. Moreover, the QV-CNN model using MFCC features achieves the best overall performance on the ASVspoof dataset, with an accuracy of 94.20% and an EER of 9.04%, while the QV-CNN with Mel-spectrograms attains the highest accuracy of 94.57%. These findings demonstrate that QV theory is an effective and promising approach for audio deepfake detection and opens new directions for quantum-inspired learning in audio perception tasks.", "date": "2026-04-09", "categories": []}
{"id": "2604.08184", "title": "AT-ADD: All-Type Audio Deepfake Detection Challenge Evaluation Plan", "abstract": "The rapid advancement of Audio Large Language Models (ALLMs) has enabled cost-effective, high-fidelity generation and manipulation of both speech and non-speech audio, including sound effects, singing voices, and music. While these capabilities foster creativity and content production, they also introduce significant security and trust challenges, as realistic audio deepfakes can now be generated and disseminated at scale. Existing audio deepfake detection (ADD) countermeasures (CMs) and benchmarks, however, remain largely speech-centric, often relying on speech-specific artifacts and exhibiting limited robustness to real-world distortions, as well as restricted generalization to heterogeneous audio types and emerging spoofing techniques. To address these gaps, we propose the All-Type Audio Deepfake Detection (AT-ADD) Grand Challenge for ACM Multimedia 2026, designed to bridge controlled academic evaluation with practical multimedia forensics. AT-ADD comprises two tracks: (1) Robust Speech Deepfake Detection, which evaluates detectors under real-world scenarios and against unseen, state-of-the-art speech generation methods; and (2) All-Type Audio Deepfake Detection, which extends detection beyond speech to diverse, unknown audio types and promotes type-agnostic generalization across speech, sound, singing, and music. By providing standardized datasets, rigorous evaluation protocols, and reproducible baselines, AT-ADD aims to accelerate the development of robust and generalizable audio forensic technologies, supporting secure communication, reliable media verification, and responsible governance in an era of pervasive synthetic audio.", "date": "2026-04-09", "categories": []}
{"id": "2604.08412", "title": "Selective Attention System (SAS): Device-Addressed Speech Detection for Real-Time On-Device Voice AI", "abstract": "We study device-addressed speech detection under pre-ASR edge deployment constraints, where systems must decide whether to forward audio before transcription under strict latency and compute limits. We show that, in multi-speaker environments with temporally ambiguous utterances, this task is more effectively modelled as a sequential routing problem over interaction history than as an utterance-local classification task. We formalize this as Sequential Device-Addressed Routing (SDAR) and present the Selective Attention System (SAS), an on-device implementation that instantiates this formulation.\n  On a held-out 60-hour multi-speaker English test set, the primary audio-only configuration achieves F1=0.86 (precision=0.89, recall=0.83); with an optional camera, audio+video fusion raises F1 to 0.95 (precision=0.97, recall=0.93). Removing causal interaction history (Stage~3) reduced F1 from 0.95 to 0.57+/-0.03 in the audio+video configuration under our evaluation protocol. Among the tested components, this was the largest observed ablation effect, indicating that short-horizon interaction history carries substantial decision-relevant information in the evaluated setting. SAS runs fully on-device on ARM Cortex-A class hardware (<150 ms latency, <20 MB footprint). All results are from internal evaluation on a proprietary dataset evaluated primarily in English; a 5-hour evaluation subset may be shared for independent verification (Section 8.8).", "date": "2026-04-09", "categories": []}
{"id": "2604.05011", "title": "YMIR: A new Benchmark Dataset and Model for Arabic Yemeni Music Genre Classification Using Convolutional Neural Networks", "abstract": "Automatic music genre classification is a major task in music information retrieval; however, most current benchmarks and models have been developed primarily for Western music, leaving culturally specific traditions underrepresented. In this paper, we introduce the Yemeni Music Information Retrieval (YMIR) dataset, which contains 1,475 carefully selected audio clips covering five traditional Yemeni genres: Sanaani, Hadhrami, Lahji, Tihami, and Adeni. The dataset was labeled by five Yemeni music experts following a clear and structured protocol, resulting in strong inter-annotator agreement (Fleiss kappa = 0.85). We also propose the Yemeni Music Classification Model (YMCM), a convolutional neural network (CNN)-based system designed to classify music genres from time-frequency features. Using a consistent preprocessing pipeline, we perform a systematic comparison across six experimental groups and five different architectures, resulting in a total of 30 experiments. Specifically, we evaluate several feature representations, including Mel-spectrograms, Chroma, FilterBank, and MFCCs with 13, 20, and 40 coefficients, and benchmark YMCM against standard models (AlexNet, VGG16, MobileNet, and a baseline CNN) under the same experimental conditions. The experimental findings reveal that YMCM is the most effective, achieving the highest accuracy of 98.8% with Mel-spectrogram features. The results also provide practical insights into the relationship between feature representation and model capacity. The findings establish YMIR as a useful benchmark and YMCM as a strong baseline for classifying Yemeni music genres.", "date": "2026-04-06", "categories": []}
{"id": "2604.05526", "title": "Controllable Singing Style Conversion with Boundary-Aware Information Bottleneck", "abstract": "This paper presents the submission of the S4 team to the Singing Voice Conversion Challenge 2025 (SVCC2025)-a novel singing style conversion system that advances fine-grained style conversion and control within in-domain settings. To address the critical challenges of style leakage, dynamic rendering, and high-fidelity generation with limited data, we introduce three key innovations: a boundary-aware Whisper bottleneck that pools phoneme-span representations to suppress residual source style while preserving linguistic content; an explicit frame-level technique matrix, enhanced by targeted F0 processing during inference, for stable and distinct dynamic style rendering; and a perceptually motivated high-frequency band completion strategy that leverages an auxiliary standard 48kHz SVC model to augment the high-frequency spectrum, thereby overcoming data scarcity without overfitting. In the official SVCC2025 subjective evaluation, our system achieves the best naturalness performance among all submissions while maintaining competitive results in speaker similarity and technique control, despite using significantly less extra singing data than other top-performing systems. Audio samples are available online.", "date": "2026-04-07", "categories": []}
{"id": "2604.05731", "title": "FoleyDesigner: Immersive Stereo Foley Generation with Precise Spatio-Temporal Alignment for Film Clips", "abstract": "Foley art plays a pivotal role in enhancing immersive auditory experiences in film, yet manual creation of spatio-temporally aligned audio remains labor-intensive. We propose FoleyDesigner, a novel framework inspired by professional Foley workflows, integrating film clip analysis, spatio-temporally controllable Foley generation, and professional audio mixing capabilities. FoleyDesigner employs a multi-agent architecture for precise spatio-temporal analysis. It achieves spatio-temporal alignment through latent diffusion models trained on spatio-temporal cues extracted from video frames, combined with large language model (LLM)-driven hybrid mechanisms that emulate post-production practices in film industry. To address the lack of high-quality stereo audio datasets in film, we introduce FilmStereo, the first professional stereo audio dataset containing spatial metadata, precise timestamps, and semantic annotations for eight common Foley categories. For applications, the framework supports interactive user control while maintaining seamless integration with professional pipelines, including 5.1-channel Dolby Atmos systems compliant with ITU-R BS.775 standards, thereby offering extensive creative flexibility. Extensive experiments demonstrate that our method achieves superior spatio-temporal alignment compared to existing baselines, with seamless compatibility with professional film production standards. The project page is available at https://gekiii996.github.io/FoleyDesigner/ .", "date": "2026-04-07", "categories": []}
{"id": "2602.00681", "title": "Audio-to-Image Bird Species Retrieval without Audio-Image Pairs via Text Distillation", "abstract": "Audio-to-image retrieval offers an interpretable alternative to audio-only classification for bioacoustic species recognition, but learning aligned audio-image representations is challenging due to the scarcity of paired audio-image data. We propose a simple and data-efficient approach that enables audio-to-image retrieval without any audio-image supervision. Our proposed method uses text as a semantic intermediary: we distill the text embedding space of a pretrained image-text model (BioCLIP-2), which encodes rich visual and taxonomic structure, into a pretrained audio-text model (BioLingual) by fine-tuning its audio encoder with a contrastive objective. This distillation transfers visually grounded semantics into the audio representation, inducing emergent alignment between audio and image embeddings without using images during training. We evaluate the resulting model on multiple bioacoustic benchmarks. The distilled audio encoder preserves audio discriminative power while substantially improving audio-text alignment on focal recordings and soundscape datasets. Most importantly, on the SSW60 benchmark, the proposed approach achieves strong audio-to-image retrieval performance exceeding baselines based on zero-shot model combinations or learned mappings between text embeddings, despite not training on paired audio-image data. These results demonstrate that indirect semantic transfer through text is sufficient to induce meaningful audio-image alignment, providing a practical solution for visually grounded species recognition in data-scarce bioacoustic settings.", "date": "2026-01-31", "categories": []}
{"id": "2604.04348", "title": "OmniSonic: Towards Universal and Holistic Audio Generation from Video and Text", "abstract": "In this paper, we propose Universal Holistic Audio Generation (UniHAGen), a task for synthesizing comprehensive auditory scenes that include both on-screen and off-screen sounds across diverse domains (e.g., ambient events, musical instruments, and human speech). Prior video-conditioned audio generation models typically focus on producing on-screen environmental sounds that correspond to visible sounding events, neglecting off-screen auditory events. While recent holistic joint text-video-to-audio generation models aim to produce auditory scenes with both on- and off-screen sound but they are limited to non-speech sounds, lacking the ability to generate or integrate human speech. To overcome these limitations, we introduce OmniSonic, a flow-matching-based diffusion framework jointly conditioned on video and text. It features a TriAttn-DiT architecture that performs three cross-attention operations to process on-screen environmental sound, off-screen environmental sound, and speech conditions simultaneously, with a Mixture-of-Experts (MoE) gating mechanism that adaptively balances their contributions during generation. Furthermore, we construct UniHAGen-Bench, a new benchmark with over one thousand samples covering three representative on/off-screen speech-environment scenarios. Extensive experiments show that OmniSonic consistently outperforms state-of-the-art approaches on both objective metrics and human evaluations, establishing a strong baseline for universal and holistic audio generation. Project page: https://weiguopian.github.io/OmniSonic_webpage/", "date": "2026-04-06", "categories": []}
{"id": "2604.02913", "title": "Split and Conquer Partial Deepfake Speech", "abstract": "Partial deepfake speech detection requires identifying manipulated regions that may occur within short temporal portions of an otherwise bona fide utterance, making the task particularly challenging for conventional utterance-level classifiers. We propose a split-and-conquer framework that decomposes the problem into two stages: boundary detection and segment-level classification. A dedicated boundary detector first identifies temporal transition points, allowing the audio signal to be divided into segments that are expected to contain acoustically consistent content. Each resulting segment is then evaluated independently to determine whether it corresponds to bona fide or fake speech.\n  This formulation simplifies the learning objective by explicitly separating temporal localization from authenticity assessment, allowing each component to focus on a well-defined task. To further improve robustness, we introduce a reflection-based multi-length training strategy that converts variable-duration segments into several fixed input lengths, producing diverse feature-space representations. Each stage is trained using multiple configurations with different feature extractors and augmentation strategies, and their complementary predictions are fused to obtain improved final models.\n  Experiments on the PartialSpoof benchmark demonstrate state-of-the-art performance across multiple temporal resolutions as well as at the utterance level, with substantial improvements in the accurate detection and localization of spoofed regions. In addition, the proposed method achieves state-of-the-art performance on the Half-Truth dataset, further confirming the robustness and generalization capability of the framework.", "date": "2026-04-03", "categories": []}
{"id": "2604.01330", "title": "Evolutionary Multi-Objective Fusion of Deepfake Speech Detectors", "abstract": "While deepfake speech detectors built on large self-supervised learning (SSL) models achieve high accuracy, employing standard ensemble fusion to further enhance robustness often results in oversized systems with diminishing returns. To address this, we propose an evolutionary multi-objective score fusion framework that jointly minimizes detection error and system complexity. We explore two encodings optimized by NSGA-II: binary-coded detector selection for score averaging and a real-valued scheme that optimizes detector weights for a weighted sum. Experiments on the ASVspoof 5 dataset with 36 SSL-based detectors show that the obtained Pareto fronts outperform simple averaging and logistic regression baselines. The real-valued variant achieves 2.37% EER (0.0684 minDCF) and identifies configurations that match state-of-the-art performance while significantly reducing system complexity, requiring only half the parameters. Our method also provides a diverse set of trade-off solutions, enabling deployment choices that balance accuracy and computational cost.", "date": "2026-04-01", "categories": []}
{"id": "2511.20224", "title": "DuoTok: Source-Aware Dual-Track Tokenization for Multi-Track Music Language Modeling", "abstract": "Audio tokenization bridges continuous waveforms and multi-track music language models. In dual-track modeling, tokens should preserve three properties at once: high-fidelity reconstruction, strong predictability under a language model, and cross-track correspondence. We introduce DuoTok, a source-aware dual-track tokenizer that addresses this trade-off through staged disentanglement. DuoTok first pretrains a semantic encoder, then regularizes it with multi-task supervision, freezes the encoder, and applies hard dual-codebook routing while keeping auxiliary objectives on quantized codes. A diffusion decoder reconstructs high-frequency details, allowing tokens to focus on structured information for sequence modeling. On standard benchmarks, DuoTok achieves a favorable predictability-fidelity trade-off, reaching the lowest cnBPT while maintaining competitive reconstruction at 0.75 kbps. Under a held-constant dual-track language modeling protocol, enBPT also improves, indicating gains beyond codebook size effects. Controlled diagnostics show larger predictability costs under cross-track corruption and larger gains from longer context, suggesting that models trained on DuoTok tokens use cross-track structure and non-local history.", "date": "2025-11-25", "categories": []}
{"id": "2604.01083", "title": "TRACE: Training-Free Partial Audio Deepfake Detection via Embedding Trajectory Analysis of Speech Foundation Models", "abstract": "Partial audio deepfakes, where synthesized segments are spliced into genuine recordings, are particularly deceptive because most of the audio remains authentic. Existing detectors are supervised: they require frame-level annotations, overfit to specific synthesis pipelines, and must be retrained as new generative models emerge. We argue that this supervision is unnecessary. We hypothesize that speech foundation models implicitly encode a forensic signal: genuine speech forms smooth, slowly varying embedding trajectories, while splice boundaries introduce abrupt disruptions in frame-level transitions. Building on this, we propose TRACE (Training-free Representation-based Audio Countermeasure via Embedding dynamics), a training-free framework that detects partial audio deepfakes by analyzing the first-order dynamics of frozen speech foundation model representations without any training, labeled data, or architectural modification. We evaluate TRACE on four benchmarks that span two languages using six speech foundation models. In PartialSpoof, TRACE achieves 8.08% EER, competitive with fine-tuned supervised baselines. In LlamaPartialSpoof, the most challenging benchmark featuring LLM-driven commercial synthesis, TRACE surpasses a supervised baseline outright (24.12% vs. 24.49% EER) without any target-domain data. These results show that temporal dynamics in speech foundation models provide an effective, generalize signal for training-free audio forensics.", "date": "2026-04-01", "categories": []}
{"id": "2511.06458", "title": "EchoMark: Perceptual Acoustic Environment Transfer with Watermark-Embedded Room Impulse Response", "abstract": "Acoustic Environment Matching (AEM) is the task of transferring clean audio into a target acoustic environment, enabling engaging applications such as audio dubbing and auditory immersive virtual reality (VR). Recovering similar room impulse response (RIR) directly from reverberant speech offers more accessible and flexible AEM solution. However, this capability also introduces vulnerabilities of arbitrary ``relocation\" if misused by malicious user, such as facilitating advanced voice spoofing attacks or undermining the authenticity of recorded evidence. To address this issue, we propose EchoMark, the first deep learning-based AEM framework that generates perceptually similar RIRs with embedded watermark. Our design tackle the challenges posed by variable RIR characteristics, such as different durations and energy decays, by operating in the latent domain. By jointly optimizing the model with a perceptual loss for RIR reconstruction and a loss for watermark detection, EchoMark achieves both high-quality environment transfer and reliable watermark recovery. Experiments on diverse datasets validate that EchoMark achieves room acoustic parameter matching performance comparable to FiNS, the state-of-the-art RIR estimator. Furthermore, a high Mean Opinion Score (MOS) of 4.22 out of 5, watermark detection accuracy exceeding 99\\%, and bit error rates (BER) below 0.3\\% collectively demonstrate the effectiveness of EchoMark in preserving perceptual quality while ensuring reliable watermark embedding.", "date": "2025-11-09", "categories": []}
{"id": "2603.29541", "title": "Can LLM Agents Identify Spoken Dialects like a Linguist?", "abstract": "Due to the scarcity of labeled dialectal speech, audio dialect classification is a challenging task for most languages, including Swiss German. In this work, we explore the ability of large language models (LLMs) as agents in understanding the dialects and whether they can show comparable performance to models such as HuBERT in dialect classification. In addition, we provide an LLM baseline and a human linguist one. Our approach uses phonetic transcriptions produced by ASR systems and combines them with linguistic resources such as dialect feature maps, vowel history, and rules. Our findings indicate that, when linguistic information is provided, the LLM predictions improve. The human baseline shows that automatically generated transcriptions can be beneficial for such classifications, but also presents opportunities for improvement.", "date": "2026-03-31", "categories": []}
{"id": "2508.01277", "title": "Foundation Models for Bioacoustics -- a Comparative Review", "abstract": "Automated bioacoustic analysis is essential for biodiversity monitoring and conservation, requiring advanced deep learning models that can adapt to diverse bioacoustic tasks. This article presents a comprehensive review of large-scale pretrained bioacoustic foundation models and systematically investigates their transferability across multiple bioacoustic classification tasks. We overview bioacoustic representation learning by analysing pretraining data sources and benchmarks. On this basis, we review bioacoustic foundation models, dissecting the models' training data, preprocessing, augmentations, architecture, and training paradigm. Additionally, we conduct an extensive empirical study of selected models on the BEANS and BirdSet benchmarks, evaluating generalisability under linear and attentive probing. Our experimental analysis reveals that Perch~2.0 achieves the highest BirdSet score (restricted evaluation) and the strongest linear probing result on BEANS, building on diverse multi-taxa supervised pretraining; that BirdMAE is the best model among probing-based strategies on BirdSet and second on BEANS after BEATs$_{NLM}$, the encoder of NatureLM-audio; that attentive probing is beneficial to extract the full performance of transformer-based models; and that general-purpose audio models trained with self-supervised learning on AudioSet outperform many specialised bird sound models on BEANS when evaluated with attentive probing. These findings provide valuable guidance for practitioners selecting appropriate models to adapt them to new bioacoustic classification tasks via probing.", "date": "2025-08-02", "categories": []}
{"id": "2602.19778", "title": "Enhancing Automatic Chord Recognition via Pseudo-Labeling and Knowledge Distillation", "abstract": "Automatic Chord Recognition (ACR) is constrained by the scarcity of aligned chord labels, as well-aligned annotations are costly to acquire. At the same time, open-weight pre-trained models are currently more accessible than their proprietary training data. In this work, we present a two-stage training pipeline that leverages pre-trained models together with unlabeled audio. The proposed method decouples training into two stages. In the first stage, we use a pre-trained BTC model as a teacher to generate pseudo-labels for over 1,000 hours of diverse unlabeled audio and train a student model solely on these pseudo-labels. In the second stage, the student is continually trained on ground-truth labels as they become available. To prevent catastrophic forgetting of the representations learned in the first stage, we apply selective knowledge distillation (KD) from the teacher as a regularizer. In our experiments, two models (BTC, 2E1D) were used as students. In stage 1, using only pseudo-labels, the BTC student achieves over 99% of the teacher's performance, while the 2E1D model achieves about 97% across seven standard mir_eval metrics. After a single training run for both students in stage 2, the resulting BTC student model surpasses the traditional supervised learning baseline by 2.5% and the original pre-trained teacher model by 1.1-3.2% across all metrics. The resulting 2E1D student model improves over the traditional supervised learning baseline by 2.67% on average and achieves almost the same performance as the teacher. Both cases show large gains on rare chord qualities.", "date": "2026-02-23", "categories": []}
{"id": "2603.26856", "title": "AFSS: Artifact-Focused Self-Synthesis for Mitigating Bias in Audio Deepfake Detection", "abstract": "The rapid advancement of generative models has enabled highly realistic audio deepfakes, yet current detectors suffer from a critical bias problem, leading to poor generalization across unseen datasets. This paper proposes Artifact-Focused Self-Synthesis (AFSS), a method designed to mitigate this bias by generating pseudo-fake samples from real audio via two mechanisms: self-conversion and self-reconstruction. The core insight of AFSS lies in enforcing same-speaker constraints, ensuring that real and pseudo-fake samples share identical speaker identity and semantic content. This forces the detector to focus exclusively on generation artifacts rather than irrelevant confounding factors. Furthermore, we introduce a learnable reweighting loss to dynamically emphasize synthetic samples during training. Extensive experiments across 7 datasets demonstrate that AFSS achieves state-of-the-art performance with an average EER of 5.45\\%, including a significant reduction to 1.23\\% on WaveFake and 2.70\\% on In-the-Wild, all while eliminating the dependency on pre-collected fake datasets. Our code is publicly available at https://github.com/NguyenLeHaiSonGit/AFSS.", "date": "2026-03-27", "categories": []}
{"id": "2603.27001", "title": "PHONOS: PHOnetic Neutralization for Online Streaming Applications", "abstract": "Speaker anonymization (SA) systems modify timbre while leaving regional or non-native accents intact, which is problematic because accents can narrow the anonymity set. To address this issue, we present PHONOS, a streaming module for real-time SA that neutralizes non-native accent to sound native-like. Our approach pre-generates golden speaker utterances that preserve source timbre and rhythm but replace foreign segmentals with native ones using silence-aware DTW alignment and zero-shot voice conversion. These utterances supervise a causal accent translator that maps non-native content tokens to native equivalents with at most 40ms look-ahead, trained using joint cross-entropy and CTC losses. Our evaluations show an 81% reduction in non-native accent confidence, with listening-test ratings consistent with this shift, and reduced speaker linkability as accent-neutralized utterances move away from the original speaker in embedding space while having latency under 241 ms on single GPU.", "date": "2026-03-27", "categories": []}
{"id": "2603.27237", "title": "Can pre-trained Deep Learning models predict groove ratings?", "abstract": "This study explores the extent to which deep learning models can predict groove and its related perceptual dimensions directly from audio signals. We critically examine the effectiveness of seven state-of-the-art deep learning models in predicting groove ratings and responses to groove-related queries through the extraction of audio embeddings. Additionally, we compare these predictions with traditional handcrafted audio features. To better understand the underlying mechanics, we extend this methodology to analyze predictions based on source-separated instruments, thereby isolating the contributions of individual musical elements. Our analysis reveals a clear separation of groove characteristics driven by the underlying musical style of the tracks (funk, pop, and rock). These findings indicate that deep audio representations can successfully encode complex, style-dependent groove components that traditional features often miss. Ultimately, this work highlights the capacity of advanced deep learning models to capture the multifaceted concept of groove, demonstrating the strong potential of representation learning to advance predictive Music Information Retrieval methodologies.", "date": "2026-03-28", "categories": []}
{"id": "2603.27877", "title": "HumMusQA: A Human-written Music Understanding QA Benchmark Dataset", "abstract": "The evaluation of music understanding in Large Audio-Language Models (LALMs) requires a rigorously defined benchmark that truly tests whether models can perceive and interpret music, a standard that current data methodologies frequently fail to meet. This paper introduces a meticulously structured approach to music evaluation, proposing a new dataset of 320 hand-written questions curated and validated by experts with musical training, arguing that such focused, manual curation is superior for probing complex audio comprehension. To demonstrate the use of the dataset, we benchmark six state-of-the-art LALMs and additionally test their robustness to uni-modal shortcuts.", "date": "2026-03-29", "categories": []}
{"id": "2508.00307", "title": "Acoustic Imaging for UAV Detection: Dense Beamformed Energy Maps and U-Net SELD", "abstract": "We introduce a U-net model for 360{\\deg} acoustic source localization formulated as a spherical semantic segmentation task. Rather than regressing discrete direction-of-arrival (DoA) angles, our model segments beamformed audio maps (azimuth & elevation) into regions of active sound presence. Using delay-and-sum (DAS) beamforming on a custom 24-microphone array, we generate signals aligned with drone GPS telemetry to create binary supervision masks. A modified U-Net, trained on frequency-domain representations of these maps, learns to identify spatially distributed source regions while addressing class imbalance via the Tversky loss. Because the network operates on beamformed energy maps, the approach is inherently array-independent and can adapt to different microphone configurations and can be transferred to different microphone configurations with minimal adaptation. The segmentation outputs are post-processed by computing centroids over activated regions, enabling robust DoA estimates. Our dataset includes real-world open-field recordings of a DJI Air 3 drone, synchronized with 360{\\deg} video and flight logs across multiple dates and locations. Experimental results show that U-net generalizes across environments, providing improved angular precision, offering a new paradigm for dense spatial audio understanding beyond traditional Sound Source Localization (SSL). We additionally validate the same beamforming-plus-segmentation formulation on the DCASE 2019 TAU Spatial Sound Events benchmark, showing that the approach generalizes beyond drone acoustics to multiclass Sound Event Localization and Detection (SELD) scenarios.", "date": "2025-08-01", "categories": []}
{"id": "2603.25767", "title": "Unlocking Strong Supervision: A Data-Centric Study of General-Purpose Audio Pre-Training Methods", "abstract": "Current audio pre-training seeks to learn unified representations for broad audio understanding tasks, but it remains fragmented and is fundamentally bottlenecked by its reliance on weak, noisy, and scale-limited labels. Drawing lessons from vision's foundational pre-training blueprint, we argue that the audio field must first establish its own large-scale, strong supervision framework. We introduce a new data-centric pipeline that leverages a high-fidelity captioner to create SOTA-quality captions and the first Unified Tag System (UTS) that bridges speech, music, and environmental sounds. We then conduct a systematic comparative study of different pre-training objectives on these strong source data. Our experiments suggest that data quality and coverage are the primary drivers of performance, while the choice of objective dictates downstream task specialization.", "date": "2026-03-26", "categories": []}
{"id": "2603.26098", "title": "A Human-Inspired Decoupled Architecture for Efficient Audio Representation Learning", "abstract": "While self-supervised learning (SSL) has revolutionized audio representation, the excessive parameterization and quadratic computational cost of standard Transformers limit their deployment on resource-constrained devices. To address this bottleneck, we propose HEAR (Human-inspired Efficient Audio Representation), a novel decoupled architecture. Inspired by the human cognitive ability to isolate local acoustic features from global context, HEAR splits the processing pipeline into two dedicated modules: an Acoustic Model for local feature extraction and a Task Model for global semantic integration. Coupled with an Acoustic Tokenizer trained via knowledge distillation, our approach enables robust Masked Audio Modeling (MAM). Extensive experiments demonstrate that HEAR requires only 15M parameters and 9.47 GFLOPs for inference, operating at a fraction of the computational cost of conventional foundation models (which typically require 85M-94M parameters). Despite this high efficiency, HEAR achieves highly competitive performance across diverse audio classification benchmarks. The code and pre-trained models are available at https://github.com/HarunoriKawano/HEAR", "date": "2026-03-27", "categories": []}
{"id": "2603.26135", "title": "TinyML for Acoustic Anomaly Detection in IoT Sensor Networks", "abstract": "Tiny Machine Learning enables real-time, energy-efficient data processing directly on microcontrollers, making it ideal for Internet of Things sensor networks. This paper presents a compact TinyML pipeline for detecting anomalies in environmental sound within IoT sensor networks. Acoustic monitoring in IoT systems can enhance safety and context awareness, yet cloud-based processing introduces challenges related to latency, power usage, and privacy. Our pipeline addresses these issues by extracting Mel Frequency Cepstral Coefficients from sound signals and training a lightweight neural network classifier optimized for deployment on edge devices. The model was trained and evaluated using the UrbanSound8K dataset, achieving a test accuracy of 91% and balanced F1-scores of 0.91 across both normal and anomalous sound classes. These results demonstrate the feasibility and reliability of embedded acoustic anomaly detection for scalable and responsive IoT deployments.", "date": "2026-03-27", "categories": []}
{"id": "2507.14237", "title": "U-DREAM: Unsupervised Dereverberation guided by a Reverberation Model", "abstract": "This paper explores the outcome of training state-of-the-art dereverberation models with supervision settings ranging from weakly-supervised to virtually unsupervised, relying solely on reverberant signals and an acoustic model for training. Most of the existing deep learning approaches typically require paired dry and reverberant data, which are difficult to obtain in practice. We develop instead a sequential learning strategy motivated by a maximum-likelihood formulation of the dereverberation problem, wherein acoustic parameters and dry signals are estimated from reverberant inputs using deep neural networks, guided by a reverberation matching loss. Our most data-efficient variant requires only 100 reverberation-parameter-labeled samples to outperform an unsupervised baseline, demonstrating the effectiveness and practicality of the proposed method in low-resource scenarios.", "date": "2025-07-17", "categories": []}
{"id": "2603.24343", "title": "Enhancing Efficiency and Performance in Deepfake Audio Detection through Neuron-level Dropin & Neuroplasticity Mechanisms", "abstract": "Current audio deepfake detection has achieved remarkable performance using diverse deep learning architectures such as ResNet, and has seen further improvements with the introduction of large models (LMs) like Wav2Vec. The success of large language models (LLMs) further demonstrates the benefits of scaling model parameters, but also highlights one bottleneck where performance gains are constrained by parameter counts. Simply stacking additional layers, as done in current LLMs, is computationally expensive and requires full retraining. Furthermore, existing low-rank adaptation methods are primarily applied to attention-based architectures, which limits their scope. Inspired by the neuronal plasticity observed in mammalian brains, we propose novel algorithms, dropin and further plasticity, that dynamically adjust the number of neurons in certain layers to flexibly modulate model parameters. We evaluate these algorithms on multiple architectures, including ResNet, Gated Recurrent Neural Networks, and Wav2Vec. Experimental results using the widely recognised ASVSpoof2019 LA, PA, and FakeorReal dataset demonstrate consistent improvements in computational efficiency with the dropin approach and a maximum of around 39% and 66% relative reduction in Equal Error Rate with the dropin and plasticity approach among these dataset, respectively. The code and supplementary material are available at Github link.", "date": "2026-03-25", "categories": []}
{"id": "2603.25570", "title": "TAAC: A gate into Trustable Audio Affective Computing", "abstract": "With the emergence of AI techniques for depression diagnosis, the conflict between high demand and limited supply for depression screening has been significantly alleviated. Among various modal data, audio-based depression diagnosis has received increasing attention from both academia and industry since audio is the most common carrier of emotion transmission. Unfortunately, audio data also contains User-sensitive Identity Information (ID), which is extremely vulnerable and may be maliciously used during the smart diagnosis process. Among previous methods, the clarification between depression features and sensitive features has always serve as a barrier. It is also critical to the problem for introducing a safe encryption methodology that only encrypts the sensitive features and a powerful classifier that can correctly diagnose the depression. To track these challenges, by leveraging adversarial loss-based Subspace Decomposition, we propose a first practical framework \\name presented for Trustable Audio Affective Computing, to perform automated depression detection through audio within a trustable environment. The key enablers of TAAC are Differentiating Features Subspace Decompositor (DFSD), Flexible Noise Encryptor (FNE) and Staged Training Paradigm, used for decomposition, ID encryption and performance enhancement, respectively. Extensive experiments with existing encryption methods demonstrate our framework's preeminent performance in depression detection, ID reservation and audio reconstruction. Meanwhile, the experiments across various setting demonstrates our model's stability under different encryption strengths. Thus proving our framework's excellence in Confidentiality, Accuracy, Traceability, and Adjustability.", "date": "2026-03-26", "categories": []}
{"id": "2510.17662", "title": "DELULU: Discriminative Embedding Learning Using Latent Units for Speaker-Aware Self-Trained Speech Foundational Model", "abstract": "Self-supervised speech models have achieved remarkable success on content-driven tasks, yet they remain limited in capturing speaker-discriminative features critical for verification, diarization, and profiling applications. We introduce \\textsc{DELULU}, a speaker-aware self-trained foundational model that addresses this limitation by incorporating speaker-informed structure into pseudo-label generation. DELULU leverages frame-level embeddings from ReDimNet, a state-of-the-art speaker verification model, to guide k-means clustering during pre-training, introducing a speaker-discriminative inductive bias that aligns representation learning with speaker identity. DELULU significantly outperforms prior SSL models across a range of speaker-centric tasks, achieving up to \\textbf{62\\% relative improvement} in equal error rate (EER) for speaker verification and consistent gains on zero-shot profiling tasks including gender, age, accent, and speaker counting; notably surpassing even its teacher model on zero-shot evaluations. Our findings demonstrate that \\textbf{DELULU is a strong universal encoder for speaker-aware speech processing}, enabling superior performance without task-specific fine-tuning.", "date": "2025-10-20", "categories": []}
{"id": "2603.23667", "title": "Echoes: A semantically-aligned music deepfake detection dataset", "abstract": "We introduce Echoes, a new dataset for music deepfake detection designed for training and benchmarking detectors under realistic and provider-diverse conditions. Echoes comprises 3,577 tracks (110 hours of audio) spanning multiple genres (pop, rock, electronic), and includes content generated by ten popular AI music generation systems. To prevent shortcut learning and promote robust generalization, the dataset is deliberately constructed to be challenging, enforcing semantic-level alignment between spoofed audio and bona fide references. This alignment is achieved by conditioning generated audio samples directly on bona-fide waveforms or song descriptors. We evaluate Echoes in a cross-dataset setting against three existing AI-generated music datasets using state-of-the-art Wav2Vec2 XLS-R 2B representations. Results show that (i) Echoes is the hardest in-domain dataset; (ii) detectors trained on existing datasets transfer poorly to Echoes; (iii) training on Echoes yields the strongest generalization performance. These findings suggest that provider diversity and semantic alignment help learn more transferable detection cues.", "date": "2026-03-24", "categories": []}
{"id": "2603.23947", "title": "Variable-Length Audio Fingerprinting", "abstract": "Audio fingerprinting converts audio to much lower-dimensional representations, allowing distorted recordings to still be recognized as their originals through similar fingerprints. Existing deep learning approaches rigidly fingerprint fixed-length audio segments, thereby neglecting temporal dynamics during segmentation. To address limitations due to this rigidity, we propose Variable-Length Audio FingerPrinting (VLAFP), a novel method that supports variable-length fingerprinting. To the best of our knowledge, VLAFP is the first deep audio fingerprinting model capable of processing audio of variable length, for both training and testing. Our experiments show that VLAFP outperforms existing state-of-the-arts in live audio identification and audio retrieval across three real-world datasets.", "date": "2026-03-25", "categories": []}
{"id": "2603.24283", "title": "Bridging Biological Hearing and Neuromorphic Computing: End-to-End Time-Domain Audio Signal Processing with Reservoir Computing", "abstract": "Despite the advancements in cutting-edge technologies, audio signal processing continues to pose challenges and lacks the precision of a human speech processing system. To address these challenges, we propose a novel approach to simplify audio signal processing by leveraging time-domain techniques and reservoir computing. Through our research, we have developed a real-time audio signal processing system by simplifying audio signal processing through the utilization of reservoir computers, which are significantly easier to train.\n  Feature extraction is a fundamental step in speech signal processing, with Mel Frequency Cepstral Coefficients (MFCCs) being a dominant choice due to their perceptual relevance to human hearing. However, conventional MFCC extraction relies on computationally intensive time-frequency transformations, limiting efficiency in real-time applications. To address this, we propose a novel approach that leverages reservoir computing to streamline MFCC extraction. By replacing traditional frequency-domain conversions with convolution operations, we eliminate the need for complex transformations while maintaining feature discriminability. We present an end-to-end audio processing framework that integrates this method, demonstrating its potential for efficient and real-time speech analysis. Our results contribute to the advancement of energy-efficient audio processing technologies, enabling seamless deployment in embedded systems and voice-driven applications. This work bridges the gap between biologically inspired feature extraction and modern neuromorphic computing, offering a scalable solution for next-generation speech recognition systems.", "date": "2026-03-25", "categories": []}
{"id": "2511.03334", "title": "UniAVGen: Unified Audio and Video Generation with Asymmetric Cross-Modal Interactions", "abstract": "Due to the lack of effective cross-modal modeling, existing open-source audio-video generation methods often exhibit compromised lip synchronization and insufficient semantic consistency. To mitigate these drawbacks, we propose UniAVGen, a unified framework for joint audio and video generation. UniAVGen is anchored in a dual-branch joint synthesis architecture, incorporating two parallel Diffusion Transformers (DiTs) to build a cohesive cross-modal latent space. At its heart lies an Asymmetric Cross-Modal Interaction mechanism, which enables bidirectional, temporally aligned cross-attention, thus ensuring precise spatiotemporal synchronization and semantic consistency. Furthermore, this cross-modal interaction is augmented by a Face-Aware Modulation module, which dynamically prioritizes salient regions in the interaction process. To enhance generative fidelity during inference, we additionally introduce Modality-Aware Classifier-Free Guidance, a novel strategy that explicitly amplifies cross-modal correlation signals. Notably, UniAVGen's robust joint synthesis design enables seamless unification of pivotal audio-video tasks within a single model, such as joint audio-video generation and continuation, video-to-audio dubbing, and audio-driven video synthesis. Comprehensive experiments validate that, with far fewer training samples (1.3M vs. 30.1M), UniAVGen delivers overall advantages in audio-video synchronization, timbre consistency, and emotion consistency.", "date": "2025-11-05", "categories": []}
{"id": "2512.19703", "title": "ASK: Adaptive Self-improving Knowledge Framework for Audio Text Retrieval", "abstract": "The dominant paradigm for Audio-Text Retrieval (ATR) relies on dual-encoder architectures optimized via mini-batch contrastive learning. However, restricting optimization to local in-batch samples creates a fundamental limitation we term the Gradient Locality Bottleneck (GLB), which prevents the resolution of acoustic ambiguities and hinders the learning of rare long-tail concepts. While external knowledge injection can break this bottleneck, it often triggers a problem called Representation-Drift Mismatch (RDM), where a static knowledge base becomes misaligned with evolving encoders, degrading guidance into noise. To address these intertwined challenges, we propose the Adaptive Self-improving Knowledge (ASK) framework. ASK breaks the GLB via multi-grained knowledge injection and mitigates RDM through a dynamic refinement strategy that synchronizes the knowledge base with the model. Additionally, an adaptive reliability weighting scheme is employed to filter retrieval noise based on cross-modal consistency. Extensive experiments across multiple benchmarks demonstrate that ASK consistently achieves new state-of-the-art performance across various backbones.", "date": "2025-12-11", "categories": []}
{"id": "2603.22677", "title": "MuQ-Eval: An Open-Source Per-Sample Quality Metric for AI Music Generation Evaluation", "abstract": "Distributional metrics such as Fr\\'echet Audio Distance cannot score individual music clips and correlate poorly with human judgments, while the only per-sample learned metric achieving high human correlation is closed-source. We introduce MUQ-EVAL, an open-source per-sample quality metric for AIgenerated music built by training lightweight prediction heads on frozen MuQ-310M features using MusicEval, a dataset of generated clips from 31 text-to-music systems with expert quality ratings. Our simplest model, frozen features with attention pooling and a two-layer MLP, achieves system-level SRCC = 0.957 and utterance-level SRCC = 0.838 with human mean opinion scores. A systematic ablation over training objectives and adaptation strategies shows that no addition meaningfully improves the frozen baseline, indicating that frozen MuQ representations already capture quality-relevant information. Encoder choice is the dominant design factor, outweighing all architectural and training decisions. LoRA-adapted models trained on as few as 150 clips already achieve usable correlation, enabling personalized quality evaluators from individual listener annotations. A controlled degradation analysis reveals selective sensitivity to signal-level artifacts but insensitivity to musical-structural distortions. Our metric, MUQ-EVAL, is fully open-source, outperforms existing open per-sample metrics, and runs in real time on a single consumer GPU. Code, model weights, and evaluation scripts are available at https://github.com/dgtql/MuQ-Eval.", "date": "2026-03-24", "categories": []}
{"id": "2602.07077", "title": "CALM: Class-Conditional Sparse Attention Vectors for Large Audio-Language Models", "abstract": "Large audio-language models (LALMs) exhibit strong zero-shot capabilities in multiple downstream tasks, such as audio question answering (AQA) and abstract reasoning; however, these models still lag behind specialized models for certain discriminative tasks (e.g., audio classification). Recent studies show that sparse subsets of attention heads within an LALM can serve as strong discriminative feature extractors for downstream tasks such as classification via simple voting schemes. However, these methods assign uniform weights to all selected heads, implicitly assuming that each head contributes equally across all semantic categories. In this work, we propose Class-Conditional Sparse Attention Vectors for Large Audio-Language Models, a few-shot classification method that learns class-dependent importance weights over attention heads. This formulation allows individual heads to specialize in distinct semantic categories and to contribute to ensemble predictions proportionally to their estimated reliability. Experiments on multiple few-shot audio and audiovisual classification benchmarks and tasks demonstrate that our method consistently outperforms state-of-the-art uniform voting-based approaches by up to 14.52%, 1.53%, 8.35% absolute gains for audio classification, audio-visual classification, and spoofing detection respectively.", "date": "2026-02-06", "categories": []}
{"id": "2603.14803", "title": "VorTEX: Various overlap ratio for Target speech EXtraction", "abstract": "Target speech extraction (TSE) aims to recover a target speaker's voice from a mixture. While recent text-prompted approaches have shown promise, most approaches assume fully overlapped mixtures, limiting insight into behavior across realistic overlap ratios. We introduce VorTEX (Various overlap ratio for Target speech EXtraction), a text-prompted TSE architecture with a Decoupled Adaptive Multi-branch (DAM) Fusion block that separates primary extraction from auxiliary regularization pathways. To enable controlled analysis, we construct PORTE, a two-speaker dataset spanning overlap ratios from 0% to 100%. We further propose Suppression Ratio on Energy (SuRE), a diagnostic metric that detects suppression behavior not captured by conventional measures. Experiments show that existing models exhibit suppression or residual interference under overlap, whereas VorTEX achieves the highest separation fidelity across 20-100% overlap (e.g., 5.50 dB at 20% and 2.04 dB at 100%) while maintaining zero SuRE, indicating robust extraction without suppression-driven artifacts.", "date": "2026-03-16", "categories": []}
{"id": "2603.20255", "title": "Abjad-Kids: An Arabic Speech Classification Dataset for Primary Education", "abstract": "Speech-based AI educational applications have gained significant interest in recent years, particularly for children. However, children speech research remains limited due to the lack of publicly available datasets, especially for low-resource languages such as Arabic.This paper presents Abjad-Kids, an Arabic speech dataset designed for kindergarten and primary education, focusing on fundamental learning of alphabets, numbers, and colors. The dataset consists of 46397 audio samples collected from children aged 3 - 12 years, covering 141 classes. All samples were recorded under controlled specifications to ensure consistency in duration, sampling rate, and format. To address high intra-class similarity among Arabic phonemes and the limited samples per class, we propose a hierarchical audio classification based on CNN-LSTM architectures. Our proposed methodology decomposes alphabet recognition into a two-stage process: an initial grouping classification model followed by specialized classifiers for each group. Both strategies: static linguistic-based grouping and dynamic clustering-based grouping, were evaluated. Experimental results demonstrate that static linguistic-based grouping achieves superior performance. Comparisons between traditional machine learning with deep learning approaches, highlight the effectiveness of CNN-LSTM models combined with data augmentation. Despite achieving promising results, most of our experiments indicate a challenge with overfitting, which is likely due to the limited number of samples, even after data augmentation and model regularization. Thus, future work may focus on collecting additional data to address this issue. Abjad-Kids will be publicly available. We hope that Abjad-Kids enrich children representation in speech dataset, and be a good resource for future research in Arabic speech classification for kids.", "date": "2026-03-11", "categories": []}
{"id": "2603.21316", "title": "HELIX: Scaling Raw Audio Understanding with Hybrid Mamba-Attention Beyond the Quadratic Limit", "abstract": "Audio representation learning typically evaluates design choices such as input frontend, sequence backbone, and sequence length in isolation. We show that these axes are coupled, and conclusions from one setting often do not transfer to others. We introduce HELIX, a controlled framework comparing pure Mamba, pure attention, and a minimal hybrid with a single attention bottleneck. All models are parameter-matched at about 8.3M parameters to isolate architectural effects. Across six datasets, we find that the preferred input representation depends on the backbone, and that attention hurts performance on short, stationary audio but becomes important at longer sequence lengths. On a 5-minute speaker identification task with 30,000 tokens, pure attention fails with out-of-memory errors, while HELIX closes an 11.5-point gap over pure Mamba.", "date": "2026-03-22", "categories": []}
{"id": "2603.21684", "title": "LipsAM: Lipschitz-Continuous Amplitude Modifier for Audio Signal Processing and its Application to Plug-and-Play Dereverberation", "abstract": "The robustness of deep neural networks (DNNs) can be certified through their Lipschitz continuity, which has made the construction of Lipschitz-continuous DNNs an active research field. However, DNNs for audio processing have not been a major focus due to their poor compatibility with existing results. In this paper, we consider the amplitude modifier (AM), a popular architecture for handling audio signals, and propose its Lipschitz-continuous variants, which we refer to as LipsAM. We prove a sufficient condition for an AM to be Lipschitz continuous and propose two architectures as examples of LipsAM. The proposed architectures were applied to a Plug-and-Play algorithm for speech dereverberation, and their improved stability is demonstrated through numerical experiments.", "date": "2026-03-23", "categories": []}
{"id": "2603.21875", "title": "Disentangling Speaker Traits for Deepfake Source Verification via Chebyshev Polynomial and Riemannian Metric Learning", "abstract": "Speech deepfake source verification systems aims to determine whether two synthetic speech utterances originate from the same source generator, often assuming that the resulting source embeddings are independent of speaker traits. However, this assumption remains unverified. In this paper, we first investigate the impact of speaker factors on source verification. We propose a speaker-disentangled metric learning (SDML) framework incorporating two novel loss functions. The first leverages Chebyshev polynomial to mitigate gradient instability during disentanglement optimization. The second projects source and speaker embeddings into hyperbolic space, leveraging Riemannian metric distances to reduce speaker information and learn more discriminative source features. Experimental results on MLAAD benchmark, evaluated under four newly proposed protocols designed for source-speaker disentanglement scenarios, demonstrate the effectiveness of SDML framework. The code, evaluation protocols and demo website are available at https://github.com/xxuan-acoustics/RiemannSD-Net.", "date": "2026-03-23", "categories": []}
{"id": "2509.24773", "title": "VSSFlow: Unifying Video-conditioned Sound and Speech Generation via Joint Learning", "abstract": "Video-conditioned audio generation, including Video-to-Sound (V2S) and Visual Text-to-Speech (VisualTTS), has traditionally been treated as distinct tasks, leaving the potential for a unified generative framework largely underexplored. In this paper, we bridge this gap with VSSFlow, a unified flow-matching framework that seamlessly solve both problems. To effectively handle multiple input signals within a Diffusion Transformer (DiT) architecture, we propose a disentangled condition aggregation mechanism leveraging distinct intrinsic properties of attention layers: cross-attention for semantic conditions, and self-attention for temporally-intensive conditions. Besides, contrary to the prevailing belief that joint training for the two tasks leads to performance degradation, we demonstrate that VSSFlow maintains superior performance during end-to-end joint learning process. Furthermore, we use a straightforward feature-level data synthesis method, demonstrating that our framework provides a robust foundation that easily adapts to joint sound and speech generation using synthetic data. Extensive experiments on V2S, VisualTTS and joint generation benchmarks show that VSSFlow effectively unifies these tasks and surpasses state-of-the-art domain-specific baselines, underscoring the critical potential of unified generative models. Project page: https://vasflow1.github.io/vasflow/", "date": "2025-09-29", "categories": []}
{"id": "2603.15597", "title": "AC-Foley: Reference-Audio-Guided Video-to-Audio Synthesis with Acoustic Transfer", "abstract": "Existing video-to-audio (V2A) generation methods predominantly rely on text prompts alongside visual information to synthesize audio. However, two critical bottlenecks persist: semantic granularity gaps in training data, such as conflating acoustically distinct sounds under coarse labels, and textual ambiguity in describing micro-acoustic features. These bottlenecks make it difficult to perform fine-grained sound synthesis using text-controlled modes. To address these limitations, we propose AC-Foley, an audio-conditioned V2A model that directly leverages reference audio to achieve precise and fine-grained control over generated sounds. This approach enables fine-grained sound synthesis, timbre transfer, zero-shot sound generation, and improved audio quality. By directly conditioning on audio signals, our approach bypasses the semantic ambiguities of text descriptions while enabling precise manipulation of acoustic attributes. Empirically, AC-Foley achieves state-of-the-art performance for Foley generation when conditioned on reference audio, while remaining competitive with state-of-the-art video-to-audio methods even without audio conditioning. Code and demo are available at: https://ff2416.github.io/AC-Foley-Page", "date": "2026-03-16", "categories": []}
{"id": "2603.18048", "title": "DEAF: A Benchmark for Diagnostic Evaluation of Acoustic Faithfulness in Audio Language Models", "abstract": "Recent Audio Multimodal Large Language Models (Audio MLLMs) demonstrate impressive performance on speech benchmarks, yet it remains unclear whether these models genuinely process acoustic signals or rely on text-based semantic inference. To systematically study this question, we introduce DEAF (Diagnostic Evaluation of Acoustic Faithfulness), a benchmark of over 2,700 conflict stimuli spanning three acoustic dimensions: emotional prosody, background sounds, and speaker identity. Then, we design a controlled multi-level evaluation framework that progressively increases textual influence, ranging from semantic conflicts in the content to misleading prompts and their combination, allowing us to disentangle content-driven bias from prompt-induced sycophancy. We further introduce diagnostic metrics to quantify model reliance on textual cues over acoustic signals. Our evaluation of seven Audio MLLMs reveals a consistent pattern of text dominance: models are sensitive to acoustic variations, yet predictions are predominantly driven by textual inputs, revealing a gap between high performance on standard speech benchmarks and genuine acoustic understanding.", "date": "2026-03-17", "categories": []}
{"id": "2603.19615", "title": "CAF-Score: Calibrating CLAP with LALMs for Reference-free Audio Captioning Evaluation", "abstract": "While Large Audio-Language Models (LALMs) have advanced audio captioning, robust evaluation remains difficult. Reference-based metrics are expensive and often fail to assess acoustic fidelity, while Contrastive Language-Audio Pretraining (CLAP)-based approaches frequently overlook syntactic errors and fine-grained details. We propose CAF-Score, a reference-free metric that calibrates CLAP's coarse-grained semantic alignment with the fine-grained comprehension and syntactic awareness of LALMs. By combining contrastive audio-text embeddings with LALM reasoning, CAF-Score effectively detects syntactic inconsistencies and subtle hallucinations. Experiments on the BRACE benchmark demonstrate that our approach achieves the highest correlation with human judgments, even outperforming reference-based baselines in challenging scenarios. These results highlight the efficacy of CAF-Score for reference-free audio captioning evaluation. Code and results are available at https://github.com/inseong00/CAF-Score.", "date": "2026-03-20", "categories": []}
{"id": "2507.02768", "title": "DeSTA2.5-Audio: Toward General-Purpose Large Audio Language Model with Self-Generated Cross-Modal Alignment", "abstract": "We introduce DeSTA2.5-Audio, a general-purpose Large Audio Language Model (LALM) designed for robust auditory perception and instruction-following. Recent LALMs augment Large Language Models (LLMs) with auditory capabilities by training on large-scale audio-instruction datasets. However, existing LALMs have often suffered from the catastrophic forgetting of the LLM's original abilities. Therefore, balancing knowledge retention and audio perception has become a critical challenge. To address this, we revisit the data construction pipeline and propose a self-generated cross-modal alignment strategy in which the backbone LLM generates its own training targets, named DeSTA. This approach aims at preserving the LLM's native language proficiency thereby enabling zero-shot generalization without task-specific tuning. We construct DeSTA-AQA5M, a large-scale, task-agnostic dataset containing 5 million training samples derived from 7,000 hours of audio spanning 50 diverse datasets, including speech, environmental sounds, and music. DeSTA2.5-Audio achieves state-of-the-art or competitive performance across a wide range of audio-language benchmarks, including Dynamic-SUPERB, MMAU, SAKURA, Speech-IFEval, and VoiceBench. Comprehensive comparative studies demonstrate that our self-generated strategy outperforms existing training strategies. Our findings underscore the importance of carefully designed data construction in LALM development and offer practical insights for building robust, general-purpose LALMs.", "date": "2025-07-03", "categories": []}
{"id": "2509.11075", "title": "Machine Learning Framework for Audio-Based Equipment Condition Monitoring: A Comparative Study of Classification Algorithms", "abstract": "Audio-based equipment condition monitoring suffers from a lack of standardized methodologies for algorithm selection, hindering reproducible research. This paper addresses this gap by introducing a comprehensive framework for the systematic and statistically rigorous evaluation of machine learning models. Leveraging a rich 127-feature set across time, frequency, and time-frequency domains, our methodology is validated on both synthetic and real-world datasets. Results demonstrate that an ensemble method achieves superior performance (94.2% accuracy, 0.942 F1-score), with statistical testing confirming its significant outperformance of individual algorithms by 8-15%. Ultimately, this work provides a validated benchmarking protocol and practical guidelines for selecting robust monitoring solutions in industrial settings.", "date": "2025-09-14", "categories": []}
{"id": "2603.19176", "title": "Few-shot Acoustic Synthesis with Multimodal Flow Matching", "abstract": "Generating audio that is acoustically consistent with a scene is essential for immersive virtual environments. Recent neural acoustic field methods enable spatially continuous sound rendering but remain scene-specific, requiring dense audio measurements and costly training for each environment. Few-shot approaches improve scalability across rooms but still rely on multiple recordings and, being deterministic, fail to capture the inherent uncertainty of scene acoustics under sparse context. We introduce flow-matching acoustic generation (FLAC), a probabilistic method for few-shot acoustic synthesis that models the distribution of plausible room impulse responses (RIRs) given minimal scene context. FLAC leverages a diffusion transformer trained with a flow-matching objective to generate RIRs at arbitrary positions in novel scenes, conditioned on spatial, geometric, and acoustic cues. FLAC outperforms state-of-the-art eight-shot baselines with one-shot on both the AcousticRooms and Hearing Anything Anywhere datasets. To complement standard perceptual metrics, we further introduce AGREE, a joint acoustic-geometry embedding, enabling geometry-consistent evaluation of generated RIRs through retrieval and distributional metrics. This work is the first to apply generative flow matching to explicit RIR synthesis, establishing a new direction for robust and data-efficient acoustic synthesis.", "date": "2026-03-19", "categories": []}
{"id": "2603.19195", "title": "How Auditory Knowledge in LLM Backbones Shapes Audio Language Models: A Holistic Evaluation", "abstract": "Large language models (LLMs) have been widely used as knowledge backbones of Large Audio Language Models (LALMs), yet how much auditory knowledge they encode through text-only pre-training and how this affects downstream performance remains unclear. We study this gap by comparing different LLMs under two text-only and one audio-grounded setting: (1) direct probing on AKB-2000, a curated benchmark testing the breadth and depth of auditory knowledge; (2) cascade evaluation, where LLMs reason over text descriptions from an audio captioner; and (3) audio-grounded evaluation, where each LLM is fine-tuned into a Large Audio Language Model (LALM) with an audio encoder. Our findings reveal that auditory knowledge varies substantially across families, and text-only results are strongly correlated with audio performance. Our work provides empirical grounding for a comprehensive understanding of LLMs in audio research.", "date": "2026-03-19", "categories": []}
{"id": "2603.17025", "title": "Shared Representation Learning for Reference-Guided Targeted Sound Detection", "abstract": "Human listeners exhibit the remarkable ability to segregate a desired sound from complex acoustic scenes through selective auditory attention, motivating the study of Targeted Sound Detection (TSD). The task requires detecting and localizing a target sound in a mixture when a reference audio of that sound is provided. Prior approaches, rely on generating a sound-discriminative conditional embedding vector for the reference and pairing it with a mixture encoder, jointly optimized with a multi-task learning approach. In this work, we propose a unified encoder architecture that processes both the reference and mixture audio within a shared representation space, promoting stronger alignment while reducing architectural complexity. This design choice not only simplifies the overall framework but also enhances generalization to unseen classes. Following the multi-task training paradigm, our method achieves substantial improvements over prior approaches, surpassing existing methods and establishing a new state-of-the-art benchmark for targeted sound detection, with a segment-level F1 score of 83.15% and an overall accuracy of 95.17% on the URBAN-SED dataset.", "date": "2026-03-17", "categories": []}
{"id": "2512.07209", "title": "Coherent Audio-Visual Editing via Conditional Audio Generation Following Video Edits", "abstract": "We introduce a novel pipeline for joint audio-visual editing that enhances the coherence between edited video and its accompanying audio. Our approach first applies state-of-the-art video editing techniques to produce the target video, then performs audio editing to align with the visual changes. To achieve this, we present a new video-to-audio generation model that conditions on the source audio, target video, and a text prompt. We extend the model architecture to incorporate conditional audio input and propose a data augmentation strategy that improves training efficiency. Furthermore, our model dynamically adjusts the influence of the source audio based on the complexity of the edits, preserving the original audio structure where possible. Experimental results demonstrate that our method outperforms existing approaches in maintaining audio-visual alignment and content integrity.", "date": "2025-12-08", "categories": []}
{"id": "2603.13952", "title": "LLM-Guided Reinforcement Learning for Audio-Visual Speech Enhancement", "abstract": "In existing Audio-Visual Speech Enhancement (AVSE) methods, objectives such as Scale-Invariant Signal-to-Noise Ratio (SI-SNR) and Mean Squared Error (MSE) are widely used; however, they often correlate poorly with perceptual quality and provide limited interpretability for optimization. This work proposes a reinforcement learning-based AVSE framework with a Large Language Model (LLM)-based interpretable reward model. An audio LLM generates natural language descriptions of enhanced speech, which are converted by a sentiment analysis model into a 1-5 rating score serving as the PPO reward for fine-tuning a pretrained AVSE model. Compared with scalar metrics, LLM-generated feedback is semantically rich and explicitly describes improvements in speech quality. Experiments on the 4th COG-MHEAR AVSE Challenge (AVSEC-4) dataset show that the proposed method outperforms a supervised baseline and a DNSMOS-based RL baseline in PESQ, STOI, neural quality metrics, and subjective listening tests.", "date": "2026-03-14", "categories": []}
{"id": "2603.15688", "title": "PulmoVec: A Two-Stage Stacking Meta-Learning Architecture Built on the HeAR Foundation Model for Multi-Task Classification of Pediatric Respiratory Sounds", "abstract": "Background: Respiratory diseases are a leading cause of childhood morbidity and mortality, yet lung auscultation remains subjective and limited by inter-listener variability, particularly in pediatric populations. Existing AI approaches are further constrained by small datasets and single-task designs. We developed PulmoVec, a multi-task framework built on the Health Acoustic Representations (HeAR) foundation model for classification of pediatric respiratory sounds. Methods: In this retrospective analysis of the SPRSound database, 24,808 event-level annotated segments from 1,652 pediatric patients were analyzed. Three task-specific classifiers were trained for screening, sound-pattern recognition, and disease-group prediction. Their out-of-fold probability outputs were combined with demographic metadata in a LightGBM stacking meta-model, and event-level predictions were aggregated to the patient level using ensemble voting. Results: At the event level, the screening model achieved an ROC-AUC of 0.96 (95% CI, 0.95-0.97), the sound-pattern recognition model a macro ROC-AUC of 0.96 (95% CI, 0.96-0.97), and the disease-group prediction model a macro ROC-AUC of 0.94 (95% CI, 0.93-0.94). At the patient level, disease-group classification yielded an accuracy of 0.74 (95% CI, 0.71-0.77), a weighted F1-score of 0.73, and a macro ROC-AUC of 0.91 (95% CI, 0.90-0.93). Stacking improved performance across all tasks compared with base models alone. Conclusions: PulmoVec links event-level acoustic phenotyping with patient-level clinical classification, supporting the potential of foundation-model-based digital auscultation in pediatric respiratory medicine. Multi-center external validation across devices and real-world conditions remains essential.", "date": "2026-03-15", "categories": []}
{"id": "2603.16201", "title": "Robust Generative Audio Quality Assessment: Disentangling Quality from Spurious Correlations", "abstract": "The rapid proliferation of AI-Generated Content (AIGC) has necessitated robust metrics for perceptual quality assessment. However, automatic Mean Opinion Score (MOS) prediction models are often compromised by data scarcity, predisposing them to learn spurious correlations-- such as dataset-specific acoustic signatures-- rather than generalized quality features. To address this, we leverage domain adversarial training (DAT) to disentangle true quality perception from these nuisance factors. Unlike prior works that rely on static domain priors, we systematically investigate domain definition strategies ranging from explicit metadata-driven labels to implicit data-driven clusters. Our findings reveal that there is no \"one-size-fits-all\" domain definition; instead, the optimal strategy is highly dependent on the specific MOS aspect being evaluated. Experimental results demonstrate that our aspect-specific domain strategy effectively mitigates acoustic biases, significantly improving correlation with human ratings and achieving superior generalization on unseen generative scenarios.", "date": "2026-03-17", "categories": []}
{"id": "2509.20397", "title": "Variational Low-Rank Adaptation for Personalized Impaired Speech Recognition", "abstract": "Speech impairments resulting from congenital disorders, such as cerebral palsy, down syndrome, or apert syndrome, as well as acquired brain injuries due to stroke, traumatic accidents, or tumors, present major challenges to automatic speech recognition (ASR) systems. Despite recent advancements, state-of-the-art ASR models like Whisper still struggle with non-normative speech due to limited training data availability and high acoustic variability. Moreover, collecting and annotating non-normative speech is burdensome: speaking is effortful for many affected individuals, while laborious annotation often requires caregivers familiar with the speaker. This work introduces a novel ASR personalization method based on Bayesian Low-rank Adaptation for data-efficient fine-tuning. We validate our method on the English UA-Speech dataset and a newly collected German speech dataset, BF-Sprache, from a child with structural speech impairment. The dataset and approach are designed to reflect the challenges of low-resource settings that include individuals with speech impairments. Our method significantly improves ASR accuracy for impaired speech while maintaining data and annotation efficiency, offering a practical path toward inclusive ASR.", "date": "2025-09-23", "categories": []}
{"id": "2510.17512", "title": "AWARE: Audio Watermarking with Adversarial Resistance to Edits", "abstract": "Prevailing practice in learning-based audio watermarking is to pursue robustness by expanding the set of simulated distortions during training. However, such surrogates are narrow and prone to overfitting. This paper presents AWARE (Audio Watermarking with Adversarial Resistance to Edits), an alternative approach that avoids reliance on attack-simulation stacks and handcrafted differentiable distortions. Embedding is obtained through adversarial optimization in the time-frequency domain under a level-proportional perceptual budget. Detection employs a time-order-agnostic detector with a Bitwise Readout Head (BRH) that aggregates temporal evidence into one score per watermark bit, enabling reliable watermark decoding even under desynchronization and temporal cuts. Empirically, AWARE attains high audio quality and speech intelligibility (PESQ/STOI) and consistently low BER across various audio edits, often surpassing representative state-of-the-art learning-based systems.", "date": "2025-10-20", "categories": []}
{"id": "2602.09823", "title": "Covo-Audio Technical Report", "abstract": "In this work, we present Covo-Audio, a 7B-parameter end-to-end LALM that directly processes continuous audio inputs and generates audio outputs within a single unified architecture. Through large-scale curated pretraining and targeted post-training, Covo-Audio achieves state-of-the-art or competitive performance among models of comparable scale across a broad spectrum of tasks, including speech-text modeling, spoken dialogue, speech understanding, audio understanding, and full-duplex voice interaction. Extensive evaluations demonstrate that the pretrained foundation model exhibits strong speech-text comprehension and semantic reasoning capabilities on multiple benchmarks, outperforming representative open-source models of comparable scale. Furthermore, Covo-Audio-Chat, the dialogue-oriented variant, demonstrates strong spoken conversational abilities, including understanding, contextual reasoning, instruction following, and generating contextually appropriate and empathetic responses, validating its applicability to real-world conversational assistant scenarios. Covo-Audio-Chat-FD, the evolved full-duplex model, achieves substantially superior performance on both spoken dialogue capabilities and full-duplex interaction behaviors, demonstrating its competence in practical robustness. To mitigate the high cost of deploying end-to-end LALMs for natural conversational systems, we propose an intelligence-speaker decoupling strategy that separates dialogue intelligence from voice rendering, enabling flexible voice customization with minimal text-to-speech (TTS) data while preserving dialogue performance. Overall, our results highlight the strong potential of 7B-scale models to integrate sophisticated audio intelligence with high-level semantic reasoning, and suggest a scalable path toward more capable and versatile LALMs.", "date": "2026-02-10", "categories": []}
{"id": "2603.05310", "title": "Latent-Mark: An Audio Watermark Robust to Neural Resynthesis", "abstract": "While existing audio watermarking techniques have achieved strong robustness against traditional digital signal processing (DSP) attacks, they remain vulnerable to neural resynthesis. This occurs because modern neural audio codecs act as semantic filters and discard the imperceptible waveform variations used in prior watermarking methods. To address this limitation, we propose Latent-Mark, the first zero-bit audio watermarking framework designed to survive semantic compression. Our key insight is that robustness to the encode-decode process requires embedding the watermark within the codec's invariant latent space. We achieve this by optimizing the audio waveform to induce a detectable directional shift in its encoded latent representation, while constraining perturbations to align with the natural audio manifold to ensure imperceptibility. To prevent overfitting to a single codec's quantization rules, we introduce Cross-Codec Optimization, jointly optimizing the waveform across multiple surrogate codecs to target shared latent invariants. Extensive evaluations demonstrate robust zero-shot transferability to unseen neural codecs, achieving state-of-the-art resilience against traditional DSP attacks while preserving perceptual imperceptibility. Our work inspires future research into universal watermarking frameworks capable of maintaining integrity across increasingly complex and diverse generative distortions.", "date": "2026-03-05", "categories": []}
{"id": "2603.12840", "title": "DAST: A Dual-Stream Voice Anonymization Attacker with Staged Training", "abstract": "Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10\\% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.", "date": "2026-03-13", "categories": []}
{"id": "2603.13362", "title": "Patient-Level Multimodal Question Answering from Multi-Site Auscultation Recordings", "abstract": "Auscultation is a vital diagnostic tool, yet its utility is often limited by subjective interpretation. While general-purpose Audio-Language Models (ALMs) excel in general domains, they struggle with the nuances of physiological signals. We propose a framework that aligns multi-site auscultation recordings directly with a frozen Large Language Model (LLM) embedding space via gated cross-attention. By leveraging the LLM's latent world knowledge, our approach moves beyond isolated classification toward holistic, patient-level assessment. On the CaReSound benchmark, our model achieves a state-of-the-art 0.865 F1-macro and 0.952 BERTScore. We demonstrate that lightweight, domain-specific encoders rival large-scale ALMs and that multi-site aggregation provides spatial redundancy that mitigates temporal truncation. This alignment of medical acoustics with text foundations offers a scalable path for bridging signal processing and clinical assessment.", "date": "2026-03-09", "categories": []}
{"id": "2603.13749", "title": "Sub-Band Spectral Matching with Localized Score Aggregation for Robust Anomalous Sound Detection", "abstract": "Detecting subtle deviations in noisy acoustic environments is central to anomalous sound detection (ASD). A common training-free ASD pipeline temporally pools frame-level representations into a band-preserving feature vector and scores anomalies using a single nearest-neighbor match. However, this global matching can inflate normal-score variance through two effects. First, when normal sounds exhibit band-wise variability, a single global neighbor forces all bands to share the same reference, increasing band-level mismatch. Second, cosine-based matching is energy-coupled, allowing a few high-energy bands to dominate score computation under normal energy fluctuations and further increase variance. We propose BEAM, which stores temporally pooled sub-band vectors in a memory bank, retrieves neighbors per sub-band, and uniformly aggregates scores to reduce normal-score variability and improve discriminability. We further introduce a parameter-free adaptive fusion to better handle diverse temporal dynamics in sub-band responses. Experiments on multiple DCASE Task 2 benchmarks show strong performance without task-specific training, robustness to noise and domain shifts, and complementary gains when combined with encoder fine-tuning.", "date": "2026-03-14", "categories": []}
{"id": "2603.14033", "title": "What Counts as Real? Speech Restoration and Voice Quality Conversion Pose New Challenges to Deepfake Detection", "abstract": "Audio anti-spoofing systems are typically formulated as binary classifiers distinguishing bona fide from spoofed speech. This assumption fails under layered generative processing, where benign transformations introduce distributional shifts that are misclassified as spoofing. We show that phonation-modifying voice conversion and speech restoration are treated as out-of-distribution despite preserving speaker authenticity. Using a multi-class setup separating bona fide, converted, spoofed, and converted-spoofed speech, we analyse model behaviour through self-supervised learning (SSL) embeddings and acoustic correlates. The benign transformations induce a drift in the SSL space, compressing bona fide and spoofed speech and reducing classifier separability. Reformulating anti-spoofing as a multi-class problem improves robustness to benign shifts while preserving spoof detection, suggesting binary systems model the distribution of raw speech rather than authenticity itself.", "date": "2026-03-14", "categories": []}
{"id": "2603.14456", "title": "PARSA-Bench: A Comprehensive Persian Audio-Language Model Benchmark", "abstract": "Persian poses unique audio understanding challenges through its classical poetry, traditional music, and pervasive code-switching - none captured by existing benchmarks. We introduce PARSA-Bench (Persian Audio Reasoning and Speech Assessment Benchmark), the first benchmark for evaluating large audio-language models on Persian language and culture, comprising 16 tasks and over 8,000 samples across speech understanding, paralinguistic analysis, and cultural audio understanding. Ten tasks are newly introduced, including poetry meter and style detection, traditional Persian music understanding, and code-switching detection. Text-only baselines consistently outperform audio counterparts, suggesting models may not leverage audio-specific information beyond what transcription alone provides. Culturally-grounded tasks expose a qualitatively distinct failure mode: all models perform near random chance on vazn detection regardless of scale, suggesting prosodic perception remains beyond the reach of current models. The dataset is publicly available at https://huggingface.co/datasets/MohammadJRanjbar/PARSA-Bench", "date": "2026-03-15", "categories": []}
{"id": "2603.14636", "title": "Nudging Hidden States: Training-Free Model Steering for Chain-of-Thought Reasoning in Large Audio-Language Models", "abstract": "Chain-of-thought (CoT) prompting has been extended to large audio-language models (LALMs) to elicit reasoning, yet enhancing its effectiveness without training remains challenging. We study inference-time model steering as a training-free approach to improve LALM reasoning. We introduce three strategies using diverse information sources and evaluate them across four LALMs and four benchmarks. Results show general accuracy gains up to 4.4% over CoT prompting. Notably, we identify a cross-modal transfer where steering vectors derived from few text samples effectively guide speech-based reasoning, demonstrating high data efficiency. We also examine hyperparameter sensitivity to understand the robustness of these approaches. Our findings position model steering as a practical direction for strengthening LALM reasoning.", "date": "2026-03-15", "categories": []}
{"id": "2603.14767", "title": "Investigating the Impact of Speech Enhancement on Audio Deepfake Detection in Noisy Environments", "abstract": "Logical Access (LA) attacks, also known as audio deepfake attacks, use Text-to-Speech (TTS) or Voice Conversion (VC) methods to generate spoofed speech data. This can represent a serious threat to Automatic Speaker Verification (ASV) systems, as intruders can use such attacks to bypass voice biometric security. In this study, we investigate the correlation between speech quality and the performance of audio spoofing detection systems (i.e., LA task). For that, the performance of two enhancement algorithms is evaluated based on two perceptual speech quality measures, namely Perceptual Evaluation of Speech Quality (PESQ) and Speech-to-Reverberation Modulation Ratio (SRMR), and in respect to their impact on the audio spoofing detection system. We adopted the LA dataset, provided in the ASVspoof 2019 Challenge, and corrupted its test set with different Signal-to-Noise Ratio (SNR) levels, while leaving the training data untouched. Enhancement was applied to attenuate the detrimental effects of noisy speech, and the performances of two models, Speech Enhancement Generative Adversarial Network (SEGAN) and Metric-Optimized Generative Adversarial Network Plus (MetricGAN+), were compared. Although we expect that speech quality will correlate well with speech applications' performance, it can also have as a side effect on downstream tasks if unwanted artifacts are introduced or relevant information is removed from the speech signal. Our results corroborate with this hypothesis, as we found that the enhancement algorithm leading to the highest speech quality scores, MetricGAN+, provided the lowest Equal Error Rate (EER) on the audio spoofing detection task, whereas the enhancement method with the lowest speech quality scores, SEGAN, led to the lowest EER, thus leading to better performance on the LA task.", "date": "2026-03-16", "categories": []}
{"id": "2603.14917", "title": "Spectrogram features for audio and speech analysis", "abstract": "Spectrogram-based representations have grown to dominate the feature space for deep learning audio analysis systems, and are often adopted for speech analysis also. Initially, the primary motivator for spectrogram-based representations was their ability to present sound as a two dimensional signal in the time-frequency plane, which not only provides an interpretable physical basis for analysing sound, but also unlocks the use of a wide range of machine learning techniques such as convolutional neural networks, that had been developed for image processing. A spectrogram is a matrix characterised by the resolution and span of its two dimensions, as well as by the representation and scaling of each element. Many possibilities for these three characteristics have been explored by researchers across numerous application areas, with different settings showing affinity for various tasks. This paper reviews the use of spectrogram-based representations and surveys the state-of-the-art to question how front-end feature representation choice allies with back-end classifier architecture for different tasks.", "date": "2026-03-16", "categories": []}
{"id": "2603.15440", "title": "Music Genre Classification: A Comparative Analysis of Classical Machine Learning and Deep Learning Approaches", "abstract": "Automatic music genre classification is a long-standing challenge in Music Information Retrieval (MIR); work on non-Western music traditions remains scarce. Nepali music encompasses culturally rich and acoustically diverse genres--from the call-and-response duets of Lok Dohori to the rhythmic poetry of Deuda and the distinctive melodies of Tamang Selo--that have not been addressed by existing classification systems. In this paper, we construct a novel dataset of approximately 8,000 labeled 30-second audio clips spanning eight Nepali music genres and conduct a systematic comparison of nine classification models across two paradigms. Five classical machine learning classifiers (Logistic Regression, SVM, KNN, Random Forest, and XGBoost) are trained on 51 hand-crafted audio features extracted via Librosa, while four deep learning architectures (CNN, RNN, parallel CNN-RNN, and sequential CNN followed by RNN) operate on Mel spectrograms of dimension 640 x 128. Our experiments reveal that the sequential Convolutional Recurrent Neural Network (CRNN)--in which convolutional layers feed into an LSTM--achieves the highest accuracy of 84%, substantially outperforming both the best classical models (Logistic Regression and XGBoost, both at 71%) and all other deep architectures. We provide per-class precision, recall, F1-score, confusion matrices, and ROC analysis for every model, and offer a culturally grounded interpretation of misclassification patterns that reflects genuine overlaps in Nepal's musical traditions.", "date": "2026-03-16", "categories": []}
{"id": "2509.26471", "title": "On Deepfake Voice Detection -- It's All in the Presentation", "abstract": "While the technologies empowering malicious audio deepfakes have dramatically evolved in recent years due to generative AI advances, the same cannot be said of global research into spoofing (deepfake) countermeasures. This paper highlights how current deepfake datasets and research methodologies led to systems that failed to generalize to real world application. The main reason is due to the difference between raw deepfake audio, and deepfake audio that has been presented through a communication channel, e.g. by phone. We propose a new framework for data creation and research methodology, allowing for the development of spoofing countermeasures that would be more effective in real-world scenarios. By following the guidelines outlined here we improved deepfake detection accuracy by 39% in more robust and realistic lab setups, and by 57% on a real-world benchmark. We also demonstrate how improvement in datasets would have a bigger impact on deepfake detection accuracy than the choice of larger SOTA models would over smaller models; that is, it would be more important for the scientific community to make greater investment on comprehensive data collection programs than to simply train larger models with higher computational demands.", "date": "2025-09-30", "categories": []}
{"id": "2603.12837", "title": "Mask2Flow-TSE: Two-Stage Target Speaker Extraction with Masking and Flow Matching", "abstract": "Target speaker extraction (TSE) extracts the target speaker's voice from overlapping speech mixtures given a reference utterance. Existing approaches typically fall into two categories: discriminative and generative. Discriminative methods apply time-frequency masking for fast inference but often over-suppress the target signal, while generative methods synthesize high-quality speech at the cost of numerous iterative steps. We propose Mask2Flow-TSE, a two-stage framework combining the strengths of both paradigms. The first stage applies discriminative masking for coarse separation, and the second stage employs flow matching to refine the output toward target speech. Unlike generative approaches that synthesize speech from Gaussian noise, our method starts from the masked spectrogram, enabling high-quality reconstruction in a single inference step. Experiments show that Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.", "date": "2026-03-13", "categories": []}
{"id": "2505.16211", "title": "AudioTrust: Benchmarking the Multifaceted Trustworthiness of Audio Large Language Models", "abstract": "The rapid development and widespread adoption of Audio Large Language Models (ALLMs) demand rigorous evaluation of their trustworthiness. However, existing evaluation frameworks are primarily designed for text and fail to capture vulnerabilities introduced by the acoustic properties of audio. We find that significant trustworthiness risks in ALLMs arise from non-semantic acoustic cues, such as timbre, accent, and background noise, which can be exploited to manipulate model behavior. To address this gap, we propose AudioTrust, the first large-scale and systematic framework for evaluating ALLM trustworthiness under audio-specific risks. AudioTrust covers six key dimensions: fairness, hallucination, safety, privacy, robustness, and authenticition. It includes 26 sub-tasks and a curated dataset of more than 4,420 audio samples collected from real-world scenarios, including daily conversations, emergency calls, and voice assistant interactions, and is specifically designed to probe trustworthiness across multiple dimensions. Our comprehensive evaluation spans 18 experimental settings and uses human-validated automated pipelines to enable objective and scalable assessment of model outputs. Experimental results on 14 state-of-the-art open-source and closed-source ALLMs reveal important limitations and failure boundaries under diverse high-risk audio scenarios, providing critical insights for the secure and trustworthy deployment of future audio models. Our platform and benchmark are publicly available at https://github.com/JusperLee/AudioTrust.", "date": "2025-05-22", "categories": []}
{"id": "2603.10005", "title": "SENS-ASR: Semantic Embedding injection in Neural-transducer for Streaming Automatic Speech Recognition", "abstract": "Many Automatic Speech Recognition (ASR) applications require streaming processing of the audio data. In streaming mode, ASR systems need to start transcribing the input stream before it is complete, i.e., the systems have to process a stream of inputs with a limited (or no) future context. Compared to offline mode, this reduction of the future context degrades the performance of Streaming-ASR systems, especially while working with low-latency constraint. In this work, we present SENS-ASR, an approach to enhance the transcription quality of Streaming-ASR by reinforcing the acoustic information with semantic information. This semantic information is extracted from the available past frame-embeddings by a context module. This module is trained using knowledge distillation from a sentence embedding Language Model fine-tuned on the training dataset transcriptions. Experiments on standard datasets show that SENS-ASR significantly improves the Word Error Rate on small-chunk streaming scenarios.", "date": "2026-02-17", "categories": []}
{"id": "2603.10725", "title": "Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning", "abstract": "The modern generative audio models can be used by an adversary in an unlawful manner, specifically, to impersonate other people to gain access to private information. To mitigate this issue, speech deepfake detection (SDD) methods started to evolve. Unfortunately, current SDD methods generally suffer from the lack of generalization to new audio domains and generators. More than that, they lack interpretability, especially human-like reasoning that would naturally explain the attribution of a given audio to the bona fide or spoof class and provide human-perceptible cues. In this paper, we propose HIR-SDD, a novel SDD framework that combines the strengths of Large Audio Language Models (LALMs) with the chain-of-thought reasoning derived from the novel proposed human-annotated dataset. Experimental evaluation demonstrates both the effectiveness of the proposed method and its ability to provide reasonable justifications for predictions.", "date": "2026-03-11", "categories": []}
{"id": "2603.11168", "title": "Huntington Disease Automatic Speech Recognition with Biomarker Supervision", "abstract": "Automatic speech recognition (ASR) for pathological speech remains underexplored, especially for Huntington's disease (HD), where irregular timing, unstable phonation, and articulatory distortion challenge current models. We present a systematic HD-ASR study using a high-fidelity clinical speech corpus not previously used for end-to-end ASR training. We compare multiple ASR families under a unified evaluation, analyzing WER as well as substitution, deletion, and insertion patterns. HD speech induces architecture-specific error regimes, with Parakeet-TDT outperforming encoder-decoder and CTC baselines. HD-specific adaptation reduces WER from 6.99% to 4.95% and we also propose a method for using biomarker-based auxiliary supervision and analyze how error behavior is reshaped in severity-dependent ways rather than uniformly improving WER. We open-source all code and models.", "date": "2026-03-11", "categories": []}
{"id": "2603.11589", "title": "Toward Complex-Valued Neural Networks for Waveform Generation", "abstract": "Neural vocoders have recently advanced waveform generation, yielding natural and expressive audio. Among these approaches, iSTFT-based vocoders have recently gained attention. They predict a complex-valued spectrogram and then synthesize the waveform via iSTFT, thereby avoiding learned upsampling stages that can increase computational cost. However, current approaches use real-valued networks that process the real and imaginary parts independently. This separation limits their ability to capture the inherent structure of complex spectrograms. We present ComVo, a Complex-valued neural Vocoder whose generator and discriminator use native complex arithmetic. This enables an adversarial training framework that provides structured feedback in complex-valued representations. To guide phase transformations in a structured manner, we introduce phase quantization, which discretizes phase values and regularizes the training process. Finally, we propose a block-matrix computation scheme to improve training efficiency by reducing redundant operations. Experiments demonstrate that ComVo achieves higher synthesis quality than comparable real-valued baselines, and that its block-matrix scheme reduces training time by 25%. Audio samples and code are available at https://hs-oh-prml.github.io/ComVo/.", "date": "2026-03-12", "categories": []}
{"id": "2603.11947", "title": "Resurfacing Paralinguistic Awareness in Large Audio Language Models", "abstract": "Large Audio Language Models (LALMs) have expanded the interaction with human to speech modality, which introduces great interactive potential, due to the paralinguistic cues implicitly indicating the user context. However, building on the current content-centred paradigm, LALMs usually neglect such paralinguistic cues and respond solely based on query content. In this work, to resurface the paralinguistic awareness in LALMs, we introduce five diverse layer-wise analyses to jointly identify paralinguistic layers and semantic understanding layers. Based on these insights, we propose a paralinguistic-enhanced fine-tuning (PE-FT) protocol accordingly to equip LALMs with paralinguistic-aware capabilities, including (1) selective-layer fine-tuning, and (2) an auxiliary dual-level classification head. Our experiments demonstrate that PE-FT protocol efficiently and effectively resurfaces the paralinguistic awareness, even surpassing the performance of the all-layer fine-tuning strategy.", "date": "2026-03-12", "categories": []}
{"id": "2509.23610", "title": "Efficient Audio-Visual Speech Separation with Discrete Lip Semantics and Multi-Scale Global-Local Attention", "abstract": "Audio-visual speech separation (AVSS) methods leverage visual cues to extract target speech and have demonstrated strong separation quality in noisy acoustic environments. However, these methods usually involve a large number of parameters and require high computational cost, which is unacceptable in many applications where speech separation serves as only a preprocessing step for further speech processing. To address this issue, we propose an efficient AVSS method, named Dolphin. For visual feature extraction, we develop DP-LipCoder, a dual-path lightweight video encoder that transforms lip-motion into discrete audio-aligned semantic tokens. For audio separation, we construct a lightweight encoder-decoder separator, in which each layer incorporates a global-local attention (GLA) block to efficiently capture multi-scale dependencies. Experiments on three benchmark datasets showed that Dolphin not only surpassed the current state-of-the-art (SOTA) model in separation quality but also achieved remarkable improvements in efficiency: over 50% fewer parameters, more than 2.4x reduction in MACs, and over 6x faster GPU inference speed. These results indicate that Dolphin offers a practical and deployable solution for high-performance AVSS in real-world scenarios. Our code and demo page are publicly available at http://cslikai.cn/Dolphin/.", "date": "2025-09-28", "categories": []}
{"id": "2603.10175", "title": "Calibration-Reasoning Framework for Descriptive Speech Quality Assessment", "abstract": "Explainable speech quality assessment requires moving beyond Mean Opinion Scores (MOS) to analyze underlying perceptual dimensions. To address this, we introduce a novel post-training method that tailors the foundational Audio Large Language Model for multidimensional reasoning, detection and classification of audio artifacts. First, a calibration stage aligns the model to predict predefined perceptual dimensions. Second, a reinforcement learning stage leverages Group Relative Policy Optimization (GRPO) with dimension-specific rewards to heavily enhance accuracy of descriptions and temporal localization of quality issues. With this approach we reach state-of-the-art results of 0.71 mean PCC score on the multidimensional QualiSpeech benchmark and 13% improvement in MOS prediction driven by RL-based reasoning. Furthermore, our fine-grained GRPO rewards substantially advance the model's ability to pinpoint and classify audio artifacts in time.", "date": "2026-03-10", "categories": []}
{"id": "2603.10465", "title": "MoXaRt: Audio-Visual Object-Guided Sound Interaction for XR", "abstract": "In Extended Reality (XR), complex acoustic environments often overwhelm users, compromising both scene awareness and social engagement due to entangled sound sources. We introduce MoXaRt, a real-time XR system that uses audio-visual cues to separate these sources and enable fine-grained sound interaction. MoXaRt's core is a cascaded architecture that performs coarse, audio-only separation in parallel with visual detection of sources (e.g., faces, instruments). These visual anchors then guide refinement networks to isolate individual sources, separating complex mixes of up to 5 concurrent sources (e.g., 2 voices + 3 instruments) with ~2 second processing latency. We validate MoXaRt through a technical evaluation on a new dataset of 30 one-minute recordings featuring concurrent speech and music, and a 22-participant user study. Empirical results indicate that our system significantly enhances speech intelligibility, yielding a 36.2% (p < 0.01) increase in listening comprehension within adversarial acoustic environments while substantially reducing cognitive load (p < 0.001), thereby paving the way for more perceptive and socially adept XR experiences.", "date": "2026-03-11", "categories": []}
{"id": "2603.10623", "title": "Geo-ATBench: A Benchmark for Geospatial Audio Tagging with Geospatial Semantic Context", "abstract": "Environmental sound understanding in computational auditory scene analysis (CASA) is often formulated as an audio-only recognition problem. This formulation leaves a persistent drawback in multi-label audio tagging (AT): acoustic similarity can make certain events difficult to separate from waveforms alone. In such cases, disambiguating cues often lie outside the waveform. Geospatial semantic context (GSC), derived from geographic information system data, e.g., points of interest (POI), provides location-tied environmental priors that can help reduce this ambiguity. A systematic study of this direction is enabled through the proposed geospatial audio tagging (Geo-AT) task, which conditions multi-label sound event tagging on GSC alongside audio. To benchmark Geo-AT, Geo-ATBench is introduced as a polyphonic audio benchmark with geographical annotations, containing 10.71 hours of audio across 28 event categories; each clip is paired with a GSC representation from 11 semantic context categories. GeoFusion-AT is proposed as a unified geo-audio fusion framework that evaluates feature-, representation-, and decision-level fusion on representative audio backbones, with audio- and GSC-only baselines. Results show that incorporating GSC improves AT performance, especially on acoustically confounded labels, indicating geospatial semantics provide effective priors beyond audio alone. A crowdsourced listening study with 10 participants on 579 samples shows that there is no significant difference in performance between models on Geo-ATBench labels and aggregated human labels, supporting Geo-ATBench as a human-aligned benchmark. The Geo-AT task, benchmark Geo-ATBench, and reproducible geo-audio fusion framework GeoFusion-AT provide a foundation for studying AT with geospatial semantic context within the CASA community. Dataset, code, models are on homepage (https://github.com/WuYanru2002/Geo-ATBench).", "date": "2026-03-11", "categories": []}
{"id": "2603.10701", "title": "AlphaFlowTSE: One-Step Generative Target Speaker Extraction via Conditional AlphaFlow", "abstract": "In target speaker extraction (TSE), we aim to recover target speech from a multi-talker mixture using a short enrollment utterance as reference. Recent studies on diffusion and flow-matching generators have improved target-speech fidelity. However, multi-step sampling increases latency, and one-step solutions often rely on a mixture-dependent time coordinate that can be unreliable for real-world conversations. We present AlphaFlowTSE, a one-step conditional generative model trained with a Jacobian-vector product (JVP)-free AlphaFlow objective. AlphaFlowTSE learns mean-velocity transport along a mixture-to-target trajectory starting from the observed mixture, eliminating auxiliary mixing-ratio prediction, and stabilizes training by combining flow matching with an interval-consistency teacher-student target. Experiments on Libri2Mix and REAL-T confirm that AlphaFlowTSE improves target-speaker similarity and real-mixture generalization for downstream automatic speech recognition (ASR).", "date": "2026-03-11", "categories": []}
{"id": "2603.02030", "title": "TCG CREST System Description for the DISPLACE-M Challenge", "abstract": "This report presents the TCG CREST system description for Track 1 (Speaker Diarization) of the DISPLACE-M challenge, focusing on naturalistic medical conversations in noisy rural-healthcare scenarios. Our study evaluates the impact of various voice activity detection (VAD) methods and advanced clustering algorithms on overall speaker diarization (SD) performance. We compare and analyze two SD frameworks: a modular pipeline utilizing SpeechBrain with ECAPA-TDNN embeddings, and a state-of-the-art (SOTA) hybrid end-to-end neural diarization system, Diarizen, built on top of a pre-trained WavLM. With these frameworks, we explore diverse clustering techniques, including agglomerative hierarchical clustering (AHC), and multiple novel variants of spectral clustering, such as SC-adapt, SC-PNA, and SC-MK. Experimental results demonstrate that the Diarizen system provides an approximate $39\\%$ relative improvement in the diarization error rate (DER) on the post-evaluation analysis of Phase~I compared to the SpeechBrain baseline. Our best-performing submitted system employing the Diarizen baseline with AHC employing a median filtering with a larger context window of $29$ achieved a DER of 10.37\\% on the development and 9.21\\% on the evaluation sets, respectively. Our team ranked fifth out of the 11 participating teams after the Phase~I evaluation.", "date": "2026-03-02", "categories": []}
{"id": "2603.08759", "title": "EDMFormer: Genre-Specific Self-Supervised Learning for Music Structure Segmentation", "abstract": "Music structure segmentation is a key task in audio analysis, but existing models perform poorly on Electronic Dance Music (EDM). This problem exists because most approaches rely on lyrical or harmonic similarity, which works well for pop music but not for EDM. EDM structure is instead defined by changes in energy, rhythm, and timbre, with different sections such as buildup, drop, and breakdown. We introduce EDMFormer, a transformer model that combines self-supervised audio embeddings using an EDM-specific dataset and taxonomy. We release this dataset as EDM-98: a group of 98 professionally annotated EDM tracks. EDMFormer improves boundary detection and section labelling compared to existing models, particularly for drops and buildups. The results suggest that combining learned representations with genre-specific data and structural priors is effective for EDM and could be applied to other specialized music genres or broader audio domains.", "date": "2026-03-08", "categories": []}
{"id": "2603.09332", "title": "TimberAgent: Gram-Guided Retrieval for Executable Music Effect Control", "abstract": "Digital audio workstations expose rich effect chains, yet a semantic gap remains between perceptual user intent and low-level signal-processing parameters. We study retrieval-grounded audio effect control, where the output is an editable plugin configuration rather than a finalized waveform. Our focus is Texture Resonance Retrieval (TRR), an audio representation built from Gram matrices of projected mid-level Wav2Vec2 activations. This design preserves texture-relevant co-activation structure. We evaluate TRR on a guitar-effects benchmark with 1,063 candidate presets and 204 queries. The evaluation follows Protocol-A, a cross-validation scheme that prevents train-test leakage. We compare TRR against CLAP and internal retrieval baselines (Wav2Vec-RAG, Text-RAG, FeatureNN-RAG), using min-max normalized metrics grounded in physical DSP parameter ranges. Ablation studies validate TRR's core design choices: projection dimensionality, layer selection, and projection type. A near-duplicate sensitivity analysis confirms that results are robust to trivial knowledge-base matches. TRR achieves the lowest normalized parameter error among evaluated methods. A multiple-stimulus listening study with 26 participants provides complementary perceptual evidence. We interpret these results as benchmark evidence that texture-aware retrieval is useful for editable audio effect control, while broader personalization and real-audio robustness claims remain outside the verified evidence presented here.", "date": "2026-03-10", "categories": []}
{"id": "2603.09556", "title": "ALARM: Audio-Language Alignment for Reasoning Models", "abstract": "Large audio language models (ALMs) extend LLMs with auditory understanding. A common approach freezes the LLM and trains only an adapter on self-generated targets. However, this fails for reasoning LLMs (RLMs) whose built-in chain-of-thought traces expose the textual surrogate input, yielding unnatural responses. We propose self-rephrasing, converting self-generated responses into audio-understanding variants compatible with RLMs while preserving distributional alignment. We further fuse and compress multiple audio encoders for stronger representations. For training, we construct a 6M-instance multi-task corpus (2.5M unique prompts) spanning 19K hours of speech, music, and sound. Our 4B-parameter ALM outperforms similarly sized models and surpasses most larger ALMs on related audio-reasoning benchmarks, while preserving textual capabilities with a low training cost. Notably, we achieve the best open-source result on the MMAU-speech and MMSU benchmarks and rank third among all the models.", "date": "2026-03-10", "categories": []}
{"id": "2603.09714", "title": "MUGEN: Evaluating and Improving Multi-audio Understanding of Large Audio-Language Models", "abstract": "While multi-audio understanding is critical for large audio-language models (LALMs), it remains underexplored. We introduce MUGEN, a comprehensive benchmark evaluating this capability across speech, general audio, and music. Our experiments reveal consistent weaknesses in multi-audio settings, and performance degrades sharply as the number of concurrent audio inputs increases, identifying input scaling as a fundamental bottleneck. We further investigate training-free strategies and observe that Audio-Permutational Self-Consistency, which diversifies the order of audio candidates, helps models form more robust aggregated predictions, yielding up to 6.28% accuracy gains. Combining this permutation strategy with Chain-of-Thought further improves performance to 6.74%. These results expose blind spots in current LALMs and provide a foundation for evaluating complex auditory comprehension.", "date": "2026-03-10", "categories": []}
{"id": "2603.09809", "title": "RA-SSU: Towards Fine-Grained Audio-Visual Learning with Region-Aware Sound Source Understanding", "abstract": "Audio-Visual Learning (AVL) is one fundamental task of multi-modality learning and embodied intelligence, displaying the vital role in scene understanding and interaction. However, previous researchers mostly focus on exploring downstream tasks from a coarse-grained perspective (e.g., audio-visual correspondence, sound source localization, and audio-visual event localization). Considering providing more specific scene perception details, we newly define a fine-grained Audio-Visual Learning task, termed Region-Aware Sound Source Understanding (RA-SSU), which aims to achieve region-aware, frame-level, and high-quality sound source understanding. To support this goal, we innovatively construct two corresponding datasets, i.e. fine-grained Music (f-Music) and fine-grained Lifescene (f-Lifescene), each containing annotated sound source masks and frame-by-frame textual descriptions. The f-Music dataset includes 3,976 samples across 22 scene types related to specific application scenarios, focusing on music scenes with complex instrument mixing. The f-Lifescene dataset contains 6,156 samples across 61 types representing diverse sounding objects in life scenarios. Moreover, we propose SSUFormer, a Sound-Source Understanding TransFormer benchmark that facilitates both the sound source segmentation and sound region description with a multi-modal input and multi-modal output architecture. Specifically, we design two modules for this framework, Mask Collaboration Module (MCM) and Mixture of Hierarchical-prompted Experts (MoHE), to respectively enhance the accuracy and enrich the elaboration of the sound source description. Extensive experiments are conducted on our two datasets to verify the feasibility of the task, evaluate the availability of the datasets, and demonstrate the superiority of the SSUFormer, which achieves SOTA performance on the Sound Source Understanding benchmark.", "date": "2026-03-10", "categories": []}
{"id": "2603.09853", "title": "SCENEBench: An Audio Understanding Benchmark Grounded in Assistive and Industrial Use Cases", "abstract": "Advances in large language models (LLMs) have enabled significant capabilities in audio processing, resulting in state-of-the-art models now known as Large Audio Language Models (LALMs). However, minimal work has been done to measure audio understanding beyond automatic speech recognition (ASR). This paper closes that gap by proposing a benchmark suite, SCENEBench (Spatial, Cross-lingual, Environmental, Non-speech Evaluation), that targets a broad form of audio comprehension across four real-world categories: background sound understanding, noise localization, cross-linguistic speech understanding, and vocal characterizer recognition. These four categories are selected based on understudied needs from accessibility technology and industrial noise monitoring. In addition to performance, we also measure model latency. The purpose of this benchmark suite is to assess audio beyond just what words are said - rather, how they are said and the non-speech components of the audio. Because our audio samples are synthetically constructed (e.g., by overlaying two natural audio samples), we further validate our benchmark against 20 natural audio items per task, sub-sampled from existing datasets to match our task criteria, to assess ecological validity. We assess five state-of-the-art LALMs and find critical gaps: performance varies across tasks, with some tasks performing below random chance and others achieving high accuracy. These results provide direction for targeted improvements in model capabilities.", "date": "2026-03-10", "categories": []}
{"id": "2505.07365", "title": "Multi-Domain Audio Question Answering Benchmark Toward Acoustic Content Reasoning", "abstract": "We present Task 5 of the DCASE 2025 Challenge: an Audio Question Answering (AQA) benchmark spanning multiple domains of sound understanding. This task defines three QA subsets (Bioacoustics, Temporal Soundscapes, and Complex QA) to test audio-language models on interactive question-answering over diverse acoustic scenes. We describe the dataset composition (from marine mammal calls to soundscapes and complex real-world clips), the evaluation protocol (top-1 accuracy with answer-shuffling robustness), and baseline systems (Qwen2-Audio-7B, AudioFlamingo 2, Gemini-2-Flash). Preliminary results on the development set are compared, showing strong variation across models and subsets. This challenge aims to advance the audio understanding and reasoning capabilities of audio-language models toward human-level acuity, which are crucial for enabling AI agents to perceive and interact about the world effectively.", "date": "2025-05-12", "categories": []}
{"id": "2601.20185", "title": "Improving X-Codec-2.0 for Multi-Lingual Speech: 25 Hz Latent Rate and 24 kHz Sampling", "abstract": "X-Codec-2.0 has shown strong performance in neural audio compression and multilingual speech modeling, operating at a 50 Hz latent rate and a 16 kHz sampling rate using frozen HuBERT features. While effective, this configuration limits temporal efficiency and audio fidelity. In this work, we explore a simple and effective modification by introducing additional pooling and increasing the decoder hop size. This reduces the latent rate from 50 Hz to 25 Hz and simultaneously raises the output sampling rate from 16 kHz to 24 kHz, improving efficiency and perceptual quality without altering the core architecture. Evaluated on the multilingual Common Voice 17 test set, the proposed configuration achieves a 0.29 MOS improvement over the original X-Codec-2.0 baseline based on UTMOSv2, and attains the best reported performance among all codecs operating at 25 Hz. The source code, checkpoints, and generation comparisons are released at \\href{https://huggingface.co/Scicom-intl/xcodec2-25TPS-24k}{https://huggingface.co/Scicom-intl/xcodec2-25TPS-24k}.", "date": "2026-01-28", "categories": []}
{"id": "2602.14612", "title": "LongAudio-RAG: Event-Grounded Question Answering over Multi-Hour Long Audio", "abstract": "Long-duration audio is increasingly common in industrial and consumer settings, yet reviewing multi-hour recordings is impractical, motivating systems that answer natural-language queries with precise temporal grounding and minimal hallucination. Existing audio-language models show promise, but long-audio question answering remains difficult due to context-length limits. We introduce LongAudio-RAG (LA-RAG), a hybrid framework that grounds Large Language Model outputs in retrieved, timestamped acoustic event detections rather than raw audio. Multi-hour streams are converted into structured event records stored in an SQL database, and at inference time the system resolves natural-language time references, classifies intent, retrieves only the relevant events, and generates answers using this constrained evidence. To evaluate performance, we construct a synthetic long-audio benchmark by concatenating recordings with preserved timestamps and generating template-based question-answer pairs for detection, counting, and summarization tasks. Finally, we demonstrate the practicality of our approach by deploying it in a hybrid edge-cloud environment, where the audio grounding model runs on-device on IoT-class hardware while the LLM is hosted on a GPU-backed server. This architecture enables low-latency event extraction at the edge and high-quality language reasoning in the cloud. Experiments show that structured, event-level retrieval significantly improves accuracy compared to vanilla Retrieval-Augmented Generation (RAG) or text-to-SQL approaches.", "date": "2026-02-16", "categories": []}
{"id": "2602.21772", "title": "UniWhisper: Efficient Continual Multi-task Training for Robust Universal Audio Representation", "abstract": "A universal audio representation should capture fine-grained speech cues and high-level semantics for environmental sounds and music in a single encoder. Existing encoders often excel in one domain but degrade in others. We propose UniWhisper, an efficient continual multi-task training framework that casts heterogeneous audio tasks into a unified instruction and answer format. This enables standard next-token training without task-specific heads and losses. We train it on 38k hours of public audio and assess the encoder using shallow MLP probes and k-nearest neighbors (kNN) on 20 tasks spanning speech, environmental sound, and music. UniWhisper reaches normalized weighted averages of 0.81 with MLP probes and 0.61 with kNN, compared to 0.64 and 0.46 for Whisper, while retaining strong speech performance.", "date": "2026-02-25", "categories": []}
{"id": "2603.06991", "title": "Adaptive Discovery of Interpretable Audio Attributes with Multimodal LLMs for Low-Resource Classification", "abstract": "In predictive modeling for low-resource audio classification, extracting high-accuracy and interpretable attributes is critical. Particularly in high-reliability applications, interpretable audio attributes are indispensable. While human-driven attribute discovery is effective, its low throughput becomes a bottleneck. We propose a method for adaptively discovering interpretable audio attributes using Multimodal Large Language Models (MLLMs). By replacing humans in the AdaFlock framework with MLLMs, our method achieves significantly faster attribute discovery. Our method dynamically identifies salient acoustic characteristics via prompting and constructs an attribute-based ensemble classifier. Experimental results across various audio tasks demonstrate that our method outperforms direct MLLM prediction in the majority of evaluated cases. The entire training completes within 11 minutes, proving it a practical, adaptive solution that surpasses conventional human-reliant approaches.", "date": "2026-03-07", "categories": []}
{"id": "2603.07285", "title": "Fast and Flexible Audio Bandwidth Extension via Vocos", "abstract": "We propose a Vocos-based bandwidth extension model that enhances audio at 8-48 kHz by generating missing high-frequency content. Inputs are resampled to 48 kHz and processed by a neural vocoder backbone, enabling a single network to support arbitrary upsampling ratios. A lightweight Linkwitz-Riley-inspired refiner merges the original low band with the generated high frequencies via a smooth crossover. On validation, the model achieves competitive log-spectral distance while running at a real-time factor of 0.0001 on an NVIDIA A100 GPU and 0.0053 on an 8-core CPU, demonstrating practical, high-quality BWE at extreme throughput.", "date": "2026-03-07", "categories": []}
{"id": "2603.08126", "title": "Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows", "abstract": "Coordinated audio generation based on video inputs typically requires a strict audio-visual (AV) alignment, where both semantics and rhythmics of the generated audio segments shall correspond to those in the video frames. Previous studies leverage a two-stage design where the AV encoders are firstly aligned via contrastive learning, then the encoded video representations guide the audio generation process. We observe that both contrastive learning and global video guidance are effective in aligning overall AV semantics while limiting temporally rhythmic synchronization. In this work, we propose FoleyFlow to first align unimodal AV encoders via masked modeling training, where the masked audio segments are recovered under the guidance of the corresponding video segments. After training, the AV encoders which are separately pretrained using only unimodal data are aligned with semantic and rhythmic consistency. Then, we develop a dynamic conditional flow for the final audio generation. Built upon the efficient velocity flow generation framework, our dynamic conditional flow utilizes temporally varying video features as the dynamic condition to guide corresponding audio segment generations. To this end, we extract coherent semantic and rhythmic representations during masked AV alignment, and use this representation of video segments to guide audio generation temporally. Our audio results are evaluated on the standard benchmarks and largely surpass existing results under several metrics. The superior performance indicates that FoleyFlow is effective in generating coordinated audios that are both semantically and rhythmically coherent to various video sequences.", "date": "2026-03-09", "categories": []}
{"id": "2603.05577", "title": "Koopman Regularized Deep Speech Disentanglement for Speaker Verification", "abstract": "Human speech contains both linguistic content and speaker dependent characteristics making speaker verification a key technology in identity critical applications. Modern deep learning speaker verification systems aim to learn speaker representations that are invariant to semantic content and nuisance factors such as ambient noise. However, many existing approaches depend on labelled data, textual supervision or large pretrained models as feature extractors, limiting scalability and practical deployment, raising sustainability concerns. We propose Deep Koopman Speech Disentanglement Autoencoder (DKSD-AE), a structured autoencoder that combines a novel multi-step Koopman operator learning module with instance normalization to disentangle speaker and content dynamics. Quantitative experiments across multiple datasets demonstrate that DKSD-AE achieves improved or competitive speaker verification performance compared to state-of-the-art baselines while maintaining high content EER, confirming effective disentanglement. These results are obtained with substantially fewer parameters and without textual supervision. Moreover, performance remains stable under increased evaluation scale, highlighting representation robustness and generalization. Our findings suggest that Koopman-based temporal modelling, when combined with instance normalization, provides an efficient and principled solution for speaker-focused representation learning.", "date": "2026-03-05", "categories": []}
{"id": "2603.06079", "title": "StreamVoiceAnon+: Emotion-Preserving Streaming Speaker Anonymization via Frame-Level Acoustic Distillation", "abstract": "We address the challenge of preserving emotional content in streaming speaker anonymization (SA). Neural audio codec language models trained for audio continuation tend to degrade source emotion: content tokens discard emotional information, and the model defaults to dominant acoustic patterns rather than preserving paralinguistic attributes. We propose supervised finetuning with neutral-emotion utterance pairs from the same speaker, combined with frame-level emotion distillation on acoustic token hidden states. All modifications are confined to finetuning, which takes less than 2 hours on 4 GPUs and adds zero inference latency overhead, while maintaining a competitive 180ms streaming latency. On the VoicePrivacy 2024 protocol, our approach achieves a 49.2% UAR (emotion preservation) with 5.77% WER (intelligibility), a +24% relative UAR improvement over the baseline (39.7%->49.2%) and +10% over the emotion-prompt variant (44.6% UAR), while maintaining strong privacy (EER 49.0%). Demo and code are available: https://anonymous3842031239.github.io/", "date": "2026-03-06", "categories": []}
{"id": "2603.06164", "title": "Do Compact SSL Backbones Matter for Audio Deepfake Detection? A Controlled Study with RAPTOR", "abstract": "Self-supervised learning (SSL) underpins modern audio deepfake detection, yet most prior work centers on a single large wav2vec2-XLSR backbone, leaving compact under studied. We present RAPTOR, Representation Aware Pairwise-gated Transformer for Out-of-domain Recognition a controlled study of compact SSL backbones from the HuBERT and WavLM within a unified pairwise-gated fusion detector, evaluated across 14 cross-domain benchmarks. We show that multilingual HuBERT pre-training is the primary driver of cross-domain robustness, enabling 100M models to match larger and commercial systems. Beyond EER, we introduce a test-time augmentation protocol with perturbation-based aleatoric uncertainty to expose calibration differences invisible to standard metrics: WavLM variants exhibit overconfident miscalibration under perturbation, whereas iterative mHuBERT remains stable. These findings indicate that SSL pre-training trajectory, not model scale, drives reliable audio deepfake detection.", "date": "2026-03-06", "categories": []}
{"id": "2603.06193", "title": "Whisper-CD: Accurate Long-Form Speech Recognition using Multi-Negative Contrastive Decoding", "abstract": "Long-form speech recognition with large encoder-decoder models such as Whisper often exhibit hallucinations, repetition loops, and content omissions. These errors can accumulate and be further amplified when the previous segment's transcription is used as decoding context. We propose Whisper-CD, a training-free contrastive decoding framework that contrasts clean-audio logits against negative logits computed from three acoustically motivated perturbations: Gaussian noise injection, silence signal, and audio temporal shift. We aggregate these negatives via the log-sum-exp operator, building a unified multi-negative objective for token-by-token decoding. Across five English long-form benchmarks, Whisper-CD reduces WER by up to 24.3pp on CORAAL and shows 48% faster token generation throughput than beam search. Because Whisper-CD operates purely at inference time, it can be applied as a drop-in replacement to already-deployed Whisper systems without retraining.", "date": "2026-03-06", "categories": []}
{"id": "2603.06220", "title": "Word-Anchored Temporal Forgery Localization", "abstract": "Current temporal forgery localization (TFL) approaches typically rely on temporal boundary regression or continuous frame-level anomaly detection paradigms to derive candidate forgery proposals. However, they suffer not only from feature granularity misalignment but also from costly computation. To address these issues, we propose word-anchored temporal forgery localization (WAFL), a novel paradigm that shifts the TFL task from temporal regression and continuous localization to discrete word-level binary classification. Specifically, we first analyze the essence of temporal forgeries and identify the minimum meaningful forgery units, word tokens, and then align data preprocessing with the natural linguistic boundaries of speech. To adapt powerful pre-trained foundation backbones for feature extraction, we introduce the forensic feature realignment (FFR) module, mapping representations from the pre-trained semantic space to a discriminative forensic manifold. This allows subsequent lightweight linear classifiers to efficiently perform binary classification and accomplish the TFL task. Furthermore, to overcome the extreme class imbalance inherent to forgery detection, we design the artifact-centric asymmetric (ACA) loss, which breaks the standard precision-recall trade-off by dynamically suppressing overwhelming authentic gradients while asymmetrically prioritizing subtle forensic artifacts. Extensive experiments demonstrate that WAFL significantly outperforms state-of-the-art approaches in localization performance under both in- and cross-dataset settings, while requiring substantially fewer learnable parameters and operating at high computational efficiency.", "date": "2026-03-06", "categories": []}
{"id": "2509.21739", "title": "Noise-to-Notes: Diffusion-based Generation and Refinement for Automatic Drum Transcription", "abstract": "Automatic drum transcription (ADT) is traditionally formulated as a discriminative task to predict drum events from audio spectrograms. In this work, we redefine ADT as a conditional generative task and introduce Noise-to-Notes (N2N), a framework leveraging diffusion modeling to transform audio-conditioned Gaussian noise into drum events with associated velocities. This generative diffusion approach offers distinct advantages, including a flexible speed-accuracy trade-off and strong inpainting capabilities. However, the generation of binary onset and continuous velocity values presents a challenge for diffusion models, and to overcome this, we introduce an Annealed Pseudo-Huber loss to facilitate effective joint optimization. Finally, to augment low-level spectrogram features, we propose incorporating features extracted from music foundation models (MFMs), which capture high-level semantic information and enhance robustness to out-of-domain drum audio. Experimental results demonstrate that including MFM features significantly improves robustness and N2N establishes a new state-of-the-art performance across multiple ADT benchmarks.", "date": "2025-09-26", "categories": []}
{"id": "2603.00395", "title": "Fine-grained Soundscape Control for Augmented Hearing", "abstract": "Hearables are becoming ubiquitous, yet their sound controls remain blunt: users can either enable global noise suppression or focus on a single target sound. Real-world acoustic scenes, however, contain many simultaneous sources that users may want to adjust independently. We introduce Aurchestra, the first system to provide fine-grained, real-time soundscape control on resource-constrained hearables. Our system has two key components: (1) a dynamic interface that surfaces only active sound classes and (2) a real-time, on-device multi-output extraction network that generates separate streams for each selected class, achieving robust performance for upto 5 overlapping target sounds, and letting users mix their environment by customizing per-class volumes, much like an audio engineer mixes tracks. We optimize the model architecture for multiple compute-limited platforms and demonstrate real-time performance on 6 ms streaming audio chunks. Across real-world environments in previously unseen indoor and outdoor scenarios, our system enables expressive per-class sound control and achieves substantial improvements in target-class enhancement and interference suppression. Our results show that the world need not be heard as a single, undifferentiated stream: with Aurchestra, the soundscape becomes truly programmable.", "date": "2026-02-28", "categories": []}
{"id": "2603.04809", "title": "WhisperAlign: Word-Boundary-Aware ASR and WhisperX-Anchored Pyannote Diarization for Long-Form Bengali Speech", "abstract": "This paper presents our solution for the DL Sprint 4.0, addressing the dual challenges of Bengali Long-Form Speech Recognition (Task 1) and Speaker Diarization (Task 2). Processing long-form, multi-speaker Bengali audio introduces significant hurdles in voice activity detection, overlapping speech, and context preservation. To solve the long-form transcription challenge, we implemented a robust audio chunking strategy utilizing whisper-timestamped, allowing us to feed precise, context-aware segments into our fine-tuned acoustic model for high-accuracy transcription. For the diarization task, we developed an integrated pipeline leveraging pyannote.audio and WhisperX. A key contribution of our approach is the domain-specific fine-tuning of the Pyannote segmentation model on the competition dataset. This adaptation allowed the model to better capture the nuances of Bengali conversational dynamics and accurately resolve complex, overlapping speaker boundaries. Our methodology demonstrates that applying intelligent timestamped chunking to ASR and targeted segmentation fine-tuning to diarization significantly drives down Word Error Rate (WER) and Diarization Error Rate (DER), in low-resource settings.", "date": "2026-03-05", "categories": []}
{"id": "2603.04032", "title": "Multi-Stage Music Source Restoration with BandSplit-RoFormer Separation and HiFi++ GAN", "abstract": "Music Source Restoration (MSR) targets recovery of original, unprocessed instrument stems from fully mixed and mastered audio, where production effects and distribution artifacts violate common linear-mixture assumptions. This technical report presents the CP-JKU team's system for the MSR ICASSP Challenge 2025. Our approach decomposes MSR into separation and restoration. First, a single BandSplit-RoFormer separator predicts eight stems plus an auxiliary other stem, and is trained with a three-stage curriculum that progresses from 4-stem warm-start fine-tuning (with LoRA) to 8-stem extension via head expansion. Second, we apply a HiFi++ GAN waveform restorer trained as a generalist and then specialized into eight instrument-specific experts.", "date": "2026-03-04", "categories": []}
{"id": "2603.04122", "title": "FastWave: Optimized Diffusion Model for Audio Super-Resolution", "abstract": "Audio Super-Resolution is a set of techniques aimed at high-quality estimation of the given signal as if it would be sampled with higher sample rate. Among suggested methods there are diffusion and flow models (which are considered slower), generative adversarial networks (which are considered faster), however both approaches are currently presented by high-parametric networks, requiring high computational costs both for training and inference. We propose a solution to both these problems by re-considering the recent advances in the training of diffusion models and applying them to super-resolution from any to 48 kHz sample rate. Our approach shows better results than NU-Wave 2 and is comparable to state-of-the-art models. Our model called FastWave has around 50 GFLOPs of computational complexity and 1.3 M parameters and can be trained with less resources and significantly faster than the majority of recently proposed diffusion- and flow-based solutions. The code has been made publicly available.", "date": "2026-03-04", "categories": []}
{"id": "2603.04366", "title": "Low-Resource Guidance for Controllable Latent Audio Diffusion", "abstract": "Generative audio requires fine-grained controllable outputs, yet most existing methods require model retraining on specific controls or inference-time controls (\\textit{e.g.}, guidance) that can also be computationally demanding. By examining the bottlenecks of existing guidance-based controls, in particular their high cost-per-step due to decoder backpropagation, we introduce a guidance-based approach through selective TFG and Latent-Control Heads (LatCHs), which enables controlling latent audio diffusion models with low computational overhead. LatCHs operate directly in latent space, avoiding the expensive decoder step, and requiring minimal training resources (7M parameters and $\\approx$ 4 hours of training). Experiments with Stable Audio Open demonstrate effective control over intensity, pitch, and beats (and a combination of those) while maintaining generation quality. Our method balances precision and audio fidelity with far lower computational costs than standard end-to-end guidance. Demo examples can be found at https://zacharynovack.github.io/latch/latch.html.", "date": "2026-03-04", "categories": []}
{"id": "2602.06823", "title": "AI-Generated Music Detection in Broadcast Monitoring", "abstract": "AI music generators have advanced to the point where their outputs are often indistinguishable from human compositions. While detection methods have emerged, they are typically designed and validated in music streaming contexts with clean, full-length tracks. Broadcast audio, however, poses a different challenge: music appears as short excerpts, often masked by dominant speech, conditions under which existing detectors fail. In this work, we introduce AI-OpenBMAT, the first dataset tailored to broadcast-style AI-music detection. It contains 3,294 one-minute audio excerpts (54.9 hours) that follow the duration patterns and loudness relations of real television audio, combining human-made production music with stylistically matched continuations generated with Suno v3.5. We benchmark a CNN baseline and state-of-the-art SpectTTTra models to assess SNR and duration robustness, and evaluate on a full broadcast scenario. Across all settings, models that excel in streaming scenarios suffer substantial degradation, with F1-scores dropping below 60% when music is in the background or has a short duration. These results highlight speech masking and short music length as critical open challenges for AI music detection, and position AI-OpenBMAT as a benchmark for developing detectors capable of meeting industrial broadcast requirements.", "date": "2026-02-06", "categories": []}
{"id": "2603.02022", "title": "CodecFlow: Efficient Bandwidth Extension via Conditional Flow Matching in Neural Codec Latent Space", "abstract": "Speech Bandwidth Extension improves clarity and intelligibility by restoring/inferring appropriate high-frequency content for low-bandwidth speech. Existing methods often rely on spectrogram or waveform modeling, which can incur higher computational cost and have limited high-frequency fidelity. Neural audio codecs offer compact latent representations that better preserve acoustic detail, yet accurately recovering high-resolution latent information remains challenging due to representation mismatch. We present CodecFlow, a neural codec-based BWE framework that performs efficient speech reconstruction in a compact latent space. CodecFlow employs a voicing-aware conditional flow converter on continuous codec embeddings and a structure-constrained residual vector quantizer to improve latent alignment stability. Optimized end-to-end, CodecFlow achieves strong spectral fidelity and enhanced perceptual quality on 8 kHz to 16 kHz and 44.1 kHz speech BWE tasks.", "date": "2026-03-02", "categories": []}
{"id": "2603.02724", "title": "Single Microphone Own Voice Detection based on Simulated Transfer Functions for Hearing Aids", "abstract": "This paper presents a simulation-based approach to own voice detection (OVD) in hearing aids using a single microphone. While OVD can significantly improve user comfort and speech intelligibility, existing solutions often rely on multiple microphones or additional sensors, increasing device complexity and cost. To enable ML-based OVD without requiring costly transfer-function measurements, we propose a data augmentation strategy based on simulated acoustic transfer functions (ATFs) that expose the model to a wide range of spatial propagation conditions. A transformer-based classifier is first trained on analytically generated ATFs and then progressively fine-tuned using numerically simulated ATFs, transitioning from a rigid-sphere model to a detailed head-and-torso representation. This hierarchical adaptation enabled the model to refine its spatial understanding while maintaining generalization. Experimental results show 95.52% accuracy on simulated head-and-torso test data. Under short-duration conditions, the model maintained 90.02% accuracy with one-second utterances. On real hearing aid recordings, the model achieved 80% accuracy without fine-tuning, aided by lightweight test-time feature compensation. This highlights the model's ability to generalize from simulated to real-world conditions, demonstrating practical viability and pointing toward a promising direction for future hearing aid design.", "date": "2026-03-03", "categories": []}
{"id": "2509.11128", "title": "ERIS: Evolutionary Real-world Interference Scheme for Jailbreaking Audio Large Models", "abstract": "Existing Audio Large Models (ALMs) alignment focuses on clean inputs, neglecting security risks in complex environments. We propose ERIS, a framework transforming real-world interference into a strategically optimized carrier for jailbreaking ALMs. Unlike methods relying on manually designed acoustic patterns, ERIS uses a genetic algorithm to optimize the selection and synthesis of naturalistic signals. Through population initialization, crossover fusion, and probabilistic mutation, it evolves audio fusing malicious instructions with real-world interference. To humans and safety filters, these samples present as natural speech with harmless background noise, yet bypass alignment. Evaluations on multiple ALMs show ERIS significantly outperforms both text and audio jailbreak baselines. Our findings reveal that seemingly innocuous real-world interference can be leveraged to circumvent safety constraints, providing new insights for defensive mechanisms in complex acoustic scenarios.", "date": "2025-09-14", "categories": []}
{"id": "2510.26818", "title": "GACA-DiT: Diffusion-based Dance-to-Music Generation with Genre-Adaptive Rhythm and Context-Aware Alignment", "abstract": "Dance-to-music (D2M) generation aims to automatically compose music that is rhythmically and temporally aligned with dance movements. Existing methods typically rely on coarse rhythm embeddings, such as global motion features or binarized joint-based rhythm values, which discard fine-grained motion cues and result in weak rhythmic alignment. Moreover, temporal mismatches introduced by feature downsampling further hinder precise synchronization between dance and music. To address these problems, we propose \\textbf{GACA-DiT}, a diffusion transformer-based framework with two novel modules for rhythmically consistent and temporally aligned music generation. First, a \\textbf{genre-adaptive rhythm extraction} module combines multi-scale temporal wavelet analysis and spatial phase histograms with adaptive joint weighting to capture fine-grained, genre-specific rhythm patterns. Second, a \\textbf{context-aware temporal alignment} module resolves temporal mismatches using learnable context queries to align music latents with relevant dance rhythm features. Extensive experiments on the AIST++ and TikTok datasets demonstrate that GACA-DiT outperforms state-of-the-art methods in both objective metrics and human evaluation. Project page: https://beria-moon.github.io/GACA-DiT/.", "date": "2025-10-28", "categories": []}
{"id": "2603.00355", "title": "StethoLM: Audio Language Model for Cardiopulmonary Analysis Across Clinical Tasks", "abstract": "Listening to heart and lung sounds - auscultation - is one of the first and most fundamental steps in a clinical examination. Despite being fast and non-invasive, it demands years of experience to interpret subtle audio cues. Recent deep learning methods have made progress in automating cardiopulmonary sound analysis, yet most are restricted to simple classification and offer little clinical interpretability or decision support. We present StethoLM, the first audio-language model specialized for cardiopulmonary auscultation, capable of performing instruction-driven clinical tasks across the full spectrum of auscultation analysis. StethoLM integrates audio encoding with a medical language model backbone and is trained on StethoBench, a comprehensive benchmark comprising 77,027 instruction-response pairs synthesized from 16,125 labeled cardiopulmonary recordings spanning seven clinical task categories: binary classification, detection, reporting, reasoning, differential diagnosis, comparison, and location-based analysis. Through multi-stage training that combines supervised fine-tuning and direct preference optimization, StethoLM achieves substantial gains in performance and robustness on out-of-distribution data. Our work establishes a foundation for instruction-following AI systems in clinical auscultation.", "date": "2026-02-27", "categories": []}
{"id": "2602.15909", "title": "Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis", "abstract": "Deep learning-based respiratory auscultation is currently hindered by two fundamental challenges: (i) inherent information loss, as converting signals into spectrograms discards transient acoustic events and clinical context; (ii) limited data availability, exacerbated by severe class imbalance. To bridge these gaps, we present Resp-Agent, an autonomous multimodal system orchestrated by a novel Active Adversarial Curriculum Agent (Thinker-A$^2$CA). Unlike static pipelines, Thinker-A$^2$CA serves as a central controller that actively identifies diagnostic weaknesses and schedules targeted synthesis in a closed loop. To address the representation gap, we introduce a modality-weaving Diagnoser that weaves clinical text with audio tokens via strategic global attention and sparse audio anchors, capturing both long-range clinical context and millisecond-level transients. To address the data gap, we design a flow matching Generator that adapts a text-only Large Language Model (LLM) via modality injection, decoupling pathological content from acoustic style to synthesize hard-to-diagnose samples. As a foundation for this work, we introduce Resp-229k, a benchmark corpus of 229k recordings paired with LLM-distilled clinical narratives. Extensive experiments demonstrate that Resp-Agent consistently outperforms prior approaches across diverse evaluation settings, improving diagnostic robustness under data scarcity and long-tailed class imbalance. Our code and data are available at https://github.com/zpforlove/Resp-Agent.", "date": "2026-02-16", "categories": []}
{"id": "2602.23387", "title": "Hello-Chat: Towards Realistic Social Audio Interactions", "abstract": "Recent advancements in Large Audio Language Models (LALMs) have demonstrated exceptional performance in speech recognition and translation. However, existing models often suffer from a disconnect between perception and expression, resulting in a robotic \"read-speech\" style that lacks the spontaneity and emotional resonance of real human interaction. In this report, we introduce Hello-Chat, an end-to-end audio language model designed for realistic social scenarios. By leveraging a massive dataset of real-life conversations and employing a modality-interleaved training strategy, Hello-Chat achieves a breakthrough in anthropomorphic generation. Experimental results show that our model not only reaches state-of-the-art (SOTA) performance on specific audio understanding tasks but also significantly outperforms existing baselines in prosodic naturalness and emotional alignment, paving the way for the next generation of empathetic AI agents.", "date": "2026-02-16", "categories": []}
{"id": "2602.23649", "title": "AudioCapBench: Quick Evaluation on Audio Captioning across Sound, Music, and Speech", "abstract": "We introduce AudioCapBench, a benchmark for evaluating audio captioning capabilities of large multimodal models. \\method covers three distinct audio domains, including environmental sound, music, and speech, with 1,000 curated evaluation samples drawn from established datasets. We evaluate 13 models across two providers (OpenAI, Google Gemini) using both reference-based metrics (METEOR, BLEU, ROUGE-L) and an LLM-as-Judge framework that scores predictions on three orthogonal dimensions: \\textit{accuracy} (semantic correctness), \\textit{completeness} (coverage of reference content), and \\textit{hallucination} (absence of fabricated content). Our results reveal that Gemini models generally outperform OpenAI models on overall captioning quality, with Gemini~3~Pro achieving the highest overall score (6.00/10), while OpenAI models exhibit lower hallucination rates. All models perform best on speech captioning and worst on music captioning. We release the benchmark as well as evaluation code to facilitate reproducible audio understanding research.", "date": "2026-02-27", "categories": []}
{"id": "2602.23003", "title": "Scattering Transform for Auditory Attention Decoding", "abstract": "The use of hearing aids will increase in the coming years due to demographic change. One open problem that remains to be solved by a new generation of hearing aids is the cocktail party problem. A possible solution is electroencephalography-based auditory attention decoding. This has been the subject of several studies in recent years, which have in common that they use the same preprocessing methods in most cases. In this work, in order to achieve an advantage, the use of a scattering transform is proposed as an alternative to these preprocessing methods. The two-layer scattering transform is compared with a regular filterbank, the synchrosqueezing short-time Fourier transform and the common preprocessing. To demonstrate the performance, the known and the proposed preprocessing methods are compared for different classification tasks on two widely used datasets, provided by the KU Leuven (KUL) and the Technical University of Denmark (DTU). Both established and new neural-network-based models, CNNs, LSTMs, and recent Transformer/graph-based models are used for classification. Various evaluation strategies were compared, with a focus on the task of classifying speakers who are unknown from the training. We show that the two-layer scattering transform can significantly improve the performance for subject-related conditions, especially on the KUL dataset. However, on the DTU dataset, this only applies to some of the models, or when larger amounts of training data are provided, as in 10-fold cross-validation. This suggests that the scattering transform is capable of extracting additional relevant information.", "date": "2026-02-26", "categories": []}
{"id": "2509.14659", "title": "Aligning Audio Captions with Human Preferences", "abstract": "Current audio captioning relies on supervised learning with paired audio-caption data, which is costly to curate and may not reflect human preferences in real-world scenarios. To address this, we propose a preference-aligned audio captioning framework based on Reinforcement Learning from Human Feedback (RLHF). To capture nuanced preferences, we train a Contrastive Language-Audio Pretraining (CLAP) based reward model using human-labeled pairwise preference data. This reward model is integrated into an RL framework to fine-tune any baseline captioning system without ground-truth annotations. Extensive human evaluations across multiple datasets show that our method produces captions preferred over baseline models, particularly when baselines fail to provide correct and natural captions. Furthermore, our framework achieves performance comparable to supervised approaches with ground-truth data, demonstrating effective alignment with human preferences and scalability in real-world use.", "date": "2025-09-18", "categories": []}
{"id": "2602.21476", "title": "A Knowledge-Driven Approach to Music Segmentation, Music Source Separation and Cinematic Audio Source Separation", "abstract": "We propose a knowledge-driven, model-based approach to segmenting audio into single-category and mixed-category chunks with applications to source separation. \"Knowledge\" here denotes information associated with the data, such as music scores. \"Model\" here refers to tool that can be used for audio segmentation and recognition, such as hidden Markov models. In contrast to conventional learning that often relies on annotated data with given segment categories and their corresponding boundaries to guide the learning process, the proposed framework does not depend on any pre-segmented training data and learns directly from the input audio and its related knowledge sources to build all necessary models autonomously. Evaluation on simulation data shows that score-guided learning achieves very good music segmentation and separation results. Tested on movie track data for cinematic audio source separation also shows that utilizing sound category knowledge achieves better separation results than those obtained with data-driven techniques without using such information.", "date": "2026-02-25", "categories": []}
{"id": "2507.03043", "title": "K-Function: Joint Pronunciation Transcription and Feedback for Evaluating Kids Language Function", "abstract": "Evaluating young children's language is challenging for automatic speech recognizers due to high-pitched voices, prolonged sounds, and limited data. We introduce K-Function, a framework that combines accurate sub-word transcription with objective, Large Language Model (LLM)-driven scoring. Its core, Kids-Weighted Finite State Transducer (K-WFST), merges an acoustic phoneme encoder with a phoneme-similarity model to capture child-specific speech errors while remaining fully interpretable. K-WFST achieves a 1.39 % phoneme error rate on MyST and 8.61 % on Multitudes-an absolute improvement of 10.47 % and 7.06 % over a greedy-search decoder. These high-quality transcripts are used by an LLM to grade verbal skills, developmental milestones, reading, and comprehension, with results that align closely with human evaluators. Our findings show that precise phoneme recognition is essential for creating an effective assessment framework, enabling scalable language screening for children.", "date": "2025-07-03", "categories": []}
{"id": "2511.08261", "title": "Uncertainty Calibration of Multi-Label Bird Sound Classifiers", "abstract": "Passive acoustic monitoring enables large-scale biodiversity assessment, but reliable classification of bioacoustic sounds requires not only high accuracy but also well-calibrated uncertainty estimates to ground decision-making. In bioacoustics, calibration is challenged by overlapping vocalisations, long-tailed species distributions, and distribution shifts between training and deployment data. The calibration of multi-label deep learning classifiers within the domain of bioacoustics has not yet been assessed. We systematically benchmark the calibration of four state-of-the-art multi-label bird sound classifiers on the BirdSet benchmark, evaluating both global, per-dataset and per-class calibration using threshold-free calibration metrics (ECE, MCS) alongside discrimination metrics (cmAP). Model calibration varies significantly across datasets and classes. While Perch v2 and ConvNeXt$_{BS}$ show better global calibration, results vary between datasets. Both models indicate consistent underconfidence, while AudioProtoPNet and BirdMAE are mostly overconfident. Surprisingly, calibration seems to be better for less frequent classes. Using simple post hoc calibration methods we demonstrate a straightforward way to improve calibration. A small labelled calibration set is sufficient to significantly improve calibration with Platt scaling, while global calibration parameters suffer from dataset variability. Our findings highlight the importance of evaluating and improving uncertainty calibration in bioacoustic classifiers.", "date": "2025-11-11", "categories": []}
{"id": "2602.20744", "title": "Voices of the Mountains: Deep Learning-Based Vocal Error Detection System for Kurdish Maqams", "abstract": "Maqam, a singing type, is a significant component of Kurdish music. A maqam singer receives training in a traditional face-to-face or through self-training. Automatic Singing Assessment (ASA) uses machine learning (ML) to provide the accuracy of singing styles and can help learners to improve their performance through error detection. Currently, the available ASA tools follow Western music rules. The musical composition requires all notes to stay within their expected pitch range from start to finish. The system fails to detect micro-intervals and pitch bends, so it identifies Kurdish maqam singing as incorrect even though the singer performs according to traditional rules. Kurdish maqam requires recognizing performance errors within microtonal spaces, which is beyond Western equal temperament. This research is the first attempt to address the mentioned gap. While many error types happen during singing, our focus is on pitch, rhythm, and modal stability errors in the context of Bayati-Kurd. We collected 50 songs from 13 vocalists ( 2-3 hours) and annotated 221 error spans (150 fine pitch, 46 rhythm, 25 modal drift). The data was segmented into 15,199 overlapping windows and converted to log-mel spectrograms. We developed a two-headed CNN-BiLSTM with attention mode to decide whether a window contains an error and to classify it based on the chosen errors. Trained for 20 epochs with early stopping at epoch 10, the model reached a validation macro-F1 of 0.468. On the full 50-song evaluation at a 0.750 threshold, recall was 39.4% and precision 25.8% . Within detected windows, type macro-F1 was 0.387, with F1 of 0.492 (fine pitch), 0.536 (rhythm), and 0.133 (modal drift); modal drift recall was 8.0%. The better performance on common error types shows that the method works, while the poor modal-drift recall shows that more data and balancing are needed.", "date": "2026-02-24", "categories": []}
{"id": "2503.23377", "title": "JavisDiT: Joint Audio-Video Diffusion Transformer with Hierarchical Spatio-Temporal Prior Synchronization", "abstract": "This paper introduces JavisDiT, a novel Joint Audio-Video Diffusion Transformer designed for synchronized audio-video generation (JAVG). Based on the powerful Diffusion Transformer (DiT) architecture, JavisDiT simultaneously generates high-quality audio and video content from open-ended user prompts in a unified framework. To ensure audio-video synchronization, we introduce a fine-grained spatio-temporal alignment mechanism through a Hierarchical Spatial-Temporal Synchronized Prior (HiST-Sypo) Estimator. This module extracts both global and fine-grained spatio-temporal priors, guiding the synchronization between the visual and auditory components. Furthermore, we propose a new benchmark, JavisBench, which consists of 10,140 high-quality text-captioned sounding videos and focuses on synchronization evaluation in diverse and complex real-world scenarios. Further, we specifically devise a robust metric for measuring the synchrony between generated audio-video pairs in real-world content. Experimental results demonstrate that JavisDiT significantly outperforms existing methods by ensuring both high-quality generation and precise synchronization, setting a new standard for JAVG tasks. Our code, model, and data are available at https://javisverse.github.io/JavisDiT-page/.", "date": "2025-03-30", "categories": []}
{"id": "2602.05220", "title": "Bagpiper: Solving Open-Ended Audio Tasks via Rich Captions", "abstract": "Current audio foundation models typically rely on rigid, task-specific supervision, addressing isolated factors of audio rather than the whole. In contrast, human intelligence processes audio holistically, seamlessly bridging physical signals with abstract cognitive concepts to execute complex tasks. Grounded in this philosophy, we introduce Bagpiper, an 8B audio foundation model that interprets physical audio via rich captions, i.e., comprehensive natural language descriptions that encapsulate the critical cognitive concepts inherent in the signal (e.g., transcription, audio events). By pre-training on a massive corpus of 600B tokens, the model establishes a robust bidirectional mapping between raw audio and this high-level conceptual space. During fine-tuning, Bagpiper adopts a caption-then-process workflow, simulating an intermediate cognitive reasoning step to solve diverse tasks without task-specific priors. Experimentally, Bagpiper outperforms Qwen-2.5-Omni on MMAU and AIRBench for audio understanding and surpasses CosyVoice3 and TangoFlux in generation quality, capable of synthesizing arbitrary compositions of speech, music, and sound effects. To the best of our knowledge, Bagpiper is among the first works that achieve unified understanding generation for general audio. Model, data, and code are available at Bagpiper Home Page.", "date": "2026-02-05", "categories": []}
{"id": "2602.15082", "title": "S-PRESSO: Ultra Low Bitrate Sound Effect Compression With Diffusion Autoencoders And Offline Quantization", "abstract": "Neural audio compression models have recently achieved extreme compression rates, enabling efficient latent generative modeling. Conversely, latent generative models have been applied to compression, pushing the limits of continuous and discrete approaches. However, existing methods remain constrained to low-resolution audio and degrade substantially at very low bitrates, where audible artifacts are prominent. In this paper, we present S-PRESSO, a 48kHz sound effect compression model that produces both continuous and discrete embeddings at ultra-low bitrates, down to 0.096 kbps, via offline quantization. Our model relies on a pretrained latent diffusion model to decode compressed audio embeddings learned by a latent encoder. Leveraging the generative priors of the diffusion decoder, we achieve extremely low frame rates, down to 1Hz (750x compression rate), producing convincing and realistic reconstructions at the cost of exact fidelity. Despite operating at high compression rates, we demonstrate that S-PRESSO outperforms both continuous and discrete baselines in audio quality, acoustic similarity and reconstruction metrics.", "date": "2026-02-16", "categories": []}
{"id": "2602.17749", "title": "Detection and Classification of Cetacean Echolocation Clicks using Image-based Object Detection Methods applied to Advanced Wavelet-based Transformations", "abstract": "A challenge in marine bioacoustic analysis is the detection of animal signals, like calls, whistles and clicks, for behavioral studies. Manual labeling is too time-consuming to process sufficient data to get reasonable results. Thus, an automatic solution to overcome the time-consuming data analysis is necessary. Basic mathematical models can detect events in simple environments, but they struggle with complex scenarios, like differentiating signals with a low signal-to-noise ratio or distinguishing clicks from echoes. Deep Learning Neural Networks, such as ANIMAL-SPOT, are better suited for such tasks. DNNs process audio signals as image representations, often using spectrograms created by Short-Time Fourier Transform. However, spectrograms have limitations due to the uncertainty principle, which creates a tradeoff between time and frequency resolution. Alternatives like the wavelet, which provides better time resolution for high frequencies and improved frequency resolution for low frequencies, may offer advantages for feature extraction in complex bioacoustic environments. This thesis shows the efficacy of CLICK-SPOT on Norwegian Killer whale underwater recordings provided by the cetacean biologist Dr. Vester. Keywords: Bioacoustics, Deep Learning, Wavelet Transformation", "date": "2026-02-19", "categories": []}
{"id": "2602.18104", "title": "MeanVoiceFlow: One-step Nonparallel Voice Conversion with Mean Flows", "abstract": "In voice conversion (VC) applications, diffusion and flow-matching models have exhibited exceptional speech quality and speaker similarity performances. However, they are limited by slow conversion owing to their iterative inference. Consequently, we propose MeanVoiceFlow, a novel one-step nonparallel VC model based on mean flows, which can be trained from scratch without requiring pretraining or distillation. Unlike conventional flow matching that uses instantaneous velocity, mean flows employ average velocity to more accurately compute the time integral along the inference path in a single step. However, training the average velocity requires its derivative to compute the target velocity, which can cause instability. Therefore, we introduce a structural margin reconstruction loss as a zero-input constraint, which moderately regularizes the input-output behavior of the model without harmful statistical averaging. Furthermore, we propose conditional diffused-input training in which a mixture of noise and source data is used as input to the model during both training and inference. This enables the model to effectively leverage source information while maintaining consistency between training and inference. Experimental results validate the effectiveness of these techniques and demonstrate that MeanVoiceFlow achieves performance comparable to that of previous multi-step and distillation-based models, even when trained from scratch. Audio samples are available at https://www.kecl.ntt.co.jp/people/kaneko.takuhiro/projects/meanvoiceflow/.", "date": "2026-02-20", "categories": []}
{"id": "2602.16008", "title": "MAEB: Massive Audio Embedding Benchmark", "abstract": "We introduce the Massive Audio Embedding Benchmark (MAEB), a large-scale benchmark covering 30 tasks across speech, music, environmental sounds, and cross-modal audio-text reasoning in 100+ languages. We evaluate 50+ models and find that no single model dominates across all tasks: contrastive audio-text models excel at environmental sound classification (e.g., ESC50) but score near random on multilingual speech tasks (e.g., SIB-FLEURS), while speech-pretrained models show the opposite pattern. Clustering remains challenging for all models, with even the best-performing model achieving only modest results. We observe that models excelling on acoustic understanding often perform poorly on linguistic tasks, and vice versa. We also show that the performance of audio encoders on MAEB correlates highly with their performance when used in audio large language models. MAEB is derived from MAEB+, a collection of 98 tasks. MAEB is designed to maintain task diversity while reducing evaluation cost, and it integrates into the MTEB ecosystem for unified evaluation across text, image, and audio modalities. We release MAEB and all 98 tasks along with code and a leaderboard at https://github.com/embeddings-benchmark/mteb.", "date": "2026-02-17", "categories": []}
{"id": "2602.16334", "title": "Spatial Audio Question Answering and Reasoning on Dynamic Source Movements", "abstract": "Spatial audio understanding aims to enable machines to interpret complex auditory scenes, particularly when sound sources move over time. In this work, we study Spatial Audio Question Answering (Spatial AQA) with a focus on movement reasoning, where a model must infer object motion, position, and directional changes directly from stereo audio. First, we introduce a movement-centric spatial audio augmentation framework that synthesizes diverse motion patterns from isolated mono audio events, enabling controlled and scalable training data generation. Second, we propose an end-to-end multimodal finetuning approach with a thinking mode, which allows audio-language models to produce explicit intermediate reasoning steps before predicting an answer. Third, we investigate the impact of query-conditioned source separation as a preprocessing stage and compare three inference regimes: no masking, an audio grounding model (AGM), and ground-truth masks. Our results show that reasoning amplifies the benefits of source separation, with thinking mode showing significant improvement of +5.1% when a single event is present in the question. These findings highlight the interplay between movement modeling, reasoning, and separation quality, offering new insights for advancing spatial audio understanding.", "date": "2026-02-18", "categories": []}
{"id": "2602.16343", "title": "How to Label Resynthesized Audio: The Dual Role of Neural Audio Codecs in Audio Deepfake Detection", "abstract": "Since Text-to-Speech systems typically don't produce waveforms directly, recent spoof detection studies use resynthesized waveforms from vocoders and neural audio codecs to simulate an attacker. Unlike vocoders, which are specifically designed for speech synthesis, neural audio codecs were originally developed for compressing audio for storage and transmission. However, their ability to discretize speech also sparked interest in language-modeling-based speech synthesis. Owing to this dual functionality, codec resynthesized data may be labeled as either bonafide or spoof. So far, very little research has addressed this issue. In this study, we present a challenging extension of the ASVspoof 5 dataset constructed for this purpose. We examine how different labeling choices affect detection performance and provide insights into labeling strategies.", "date": "2026-02-18", "categories": []}
{"id": "2602.16687", "title": "Scaling Open Discrete Audio Foundation Models with Interleaved Semantic, Acoustic, and Text Tokens", "abstract": "Current audio language models are predominantly text-first, either extending pre-trained text LLM backbones or relying on semantic-only audio tokens, limiting general audio modeling. This paper presents a systematic empirical study of native audio foundation models that apply next-token prediction to audio at scale, jointly modeling semantic content, acoustic details, and text to support both general audio generation and cross-modal capabilities. We provide comprehensive empirical insights for building such models: (1) We systematically investigate design choices -- data sources, text mixture ratios, and token composition -- establishing a validated training recipe. (2) We conduct the first scaling law study for discrete audio models via IsoFLOP analysis on 64 models spanning $3{\\times}10^{18}$ to $3{\\times}10^{20}$ FLOPs, finding that optimal data grows 1.6$\\times$ faster than optimal model size. (3) We apply these lessons to train SODA (Scaling Open Discrete Audio), a suite of models from 135M to 4B parameters on 500B tokens, comparing against our scaling predictions and existing models. SODA serves as a flexible backbone for diverse audio/text tasks -- we demonstrate this by fine-tuning for voice-preserving speech-to-speech translation, using the same unified architecture.", "date": "2026-02-18", "categories": []}
{"id": "2507.08333", "title": "Token-Based Audio Inpainting via Discrete Diffusion", "abstract": "Audio inpainting seeks to restore missing segments in degraded recordings. Previous diffusion-based methods exhibit impaired performance when the missing region is large. We introduce the first approach that applies discrete diffusion over tokenized music representations from a pre-trained audio tokenizer, enabling stable and semantically coherent restoration of long gaps. Our method further incorporates two training approaches: a derivative-based regularization loss that enforces smooth temporal dynamics, and a span-based absorbing transition that provides structured corruption during diffusion. Experiments on the MusicNet and MAESTRO datasets with gaps up to 750 ms show that our approach consistently outperforms strong baselines across range of gap lengths, for gaps of 150 ms and above. This work advances musical audio restoration and introduces new directions for discrete diffusion model training. Visit our project page for examples and code.", "date": "2025-07-11", "categories": []}
{"id": "2508.14949", "title": "XAI-Driven Spectral Analysis of Cough Sounds for Respiratory Disease Characterization", "abstract": "This paper proposes an eXplainable Artificial Intelligence (XAI)-driven methodology to enhance the understanding of cough sound analysis for respiratory disease management. We employ occlusion maps to highlight relevant spectral regions in cough spectrograms processed by a Convolutional Neural Network (CNN). Subsequently, spectral analysis of spectrograms weighted by these occlusion maps reveals significant differences between disease groups, particularly in patients with COPD, where cough patterns appear more variable in the identified spectral regions of interest. This contrasts with the lack of significant differences observed when analyzing raw spectrograms. The proposed approach extracts and analyzes several spectral features, demonstrating the potential of XAI techniques to uncover disease-specific acoustic signatures and improve the diagnostic capabilities of cough sound analysis by providing more interpretable results.", "date": "2025-08-20", "categories": []}
{"id": "2510.10509", "title": "MARS-Sep: Multimodal-Aligned Reinforced Sound Separation", "abstract": "Universal sound separation faces a fundamental misalignment: models optimized for low-level signal metrics often produce semantically contaminated outputs, failing to suppress perceptually salient interference from acoustically similar sources. We introduce a preference alignment perspective, analogous to aligning LLMs with human intent. To address this, we introduce MARS-Sep, a reinforcement learning framework that reformulates separation as decision making. Instead of simply regressing ground-truth masks, MARS-Sep learns a factorized Beta mask policy that is steered by a preference reward model and optimized by a stable, clipped trust-region surrogate. The reward, derived from a progressively-aligned audio-text-vision encoder, directly incentivizes semantic consistency with query prompts. Extensive experiments on multiple benchmarks demonstrate consistent gains in Text-, Audio-, and Image-Queried separation, with notable improvements in signal metrics and semantic quality. Our code is available at https://github.com/mars-sep/MARS-Sep. Sound separation samples are available at https://mars-sep.github.io/.", "date": "2025-10-12", "categories": []}
{"id": "2602.15491", "title": "The Equalizer: Introducing Shape-Gain Decomposition in Neural Audio Codecs", "abstract": "Neural audio codecs (NACs) typically encode the short-term energy (gain) and normalized structure (shape) of speech/audio signals jointly within the same latent space. As a result, they are poorly robust to a global variation of the input signal level in the sense that such variation has strong influence on the embedding vectors at the output of the encoder and their quantization. This methodology is inherently inefficient, leading to codebook redundancy and suboptimal bitrate-distortion performance. To address these limitations, we propose to introduce shape-gain decomposition, widely used in classical speech/audio coding, into the NAC framework. The principle of the proposed Equalizer methodology is to decompose the input signal -- before the NAC encoder -- into gain and normalized shape vector on a short-term basis. The shape vector is processed by the NAC, while the gain is quantized with scalar quantization and transmitted separately. The output (decoded) signal is reconstructed from the normalized output of the NAC and the quantized gain. Our experiments conducted on speech signals show that this general methodology, easily applicable to any NAC, enables a substantial gain in bitrate-distortion performance, as well as a massive reduction in complexity.", "date": "2026-02-17", "categories": []}
{"id": "2602.13685", "title": "AuTAgent: A Reinforcement Learning Framework for Tool-Augmented Audio Reasoning", "abstract": "Large Audio Language Models (LALMs) excel at perception but struggle with complex reasoning requiring precise acoustic measurements. While external tools can extract fine-grained features like exact tempo or pitch, effective integration remains challenging: naively using all tools causes information overload, while prompt-based selection fails to assess context-dependent utility. To address this, we propose AuTAgent (Audio Tool Agent), a reinforcement learning framework that learns when and which tools to invoke. By employing a sparse-feedback training strategy with a novel Differential Reward mechanism, the agent learns to filter out irrelevant tools and invokes external assistance only when it yields a net performance gain over the base model. Experimental results confirm that AuTAgent complements the representation bottleneck of LALMs by providing verifiable acoustic evidence. It improves accuracy by 4.20% / 6.20% and 9.80% / 8.00% for open-source and closed-source backbones on the MMAU Test-mini and the MMAR benchmarks, respectively. In addition, further experiments demonstrate exceptional transferability. We highlight the complementary role of external tools in augmenting audio model reasoning.", "date": "2026-02-14", "categories": []}
{"id": "2602.13928", "title": "voice2mode: Phonation Mode Classification in Singing using Self-Supervised Speech Models", "abstract": "We present voice2mode, a method for classification of four singing phonation modes (breathy, neutral (modal), flow, and pressed) using embeddings extracted from large self-supervised speech models. Prior work on singing phonation has relied on handcrafted signal features or task-specific neural nets; this work evaluates the transferability of speech foundation models to singing phonation classification. voice2mode extracts layer-wise representations from HuBERT and two wav2vec2 variants, applies global temporal pooling, and classifies the pooled embeddings with lightweight classifiers (SVM, XGBoost). Experiments on a publicly available soprano dataset (763 sustained vowel recordings, four labels) show that foundation-model features substantially outperform conventional spectral baselines (spectrogram, mel-spectrogram, MFCC). HuBERT embeddings obtained from early layers yield the best result (~95.7% accuracy with SVM), an absolute improvement of ~12-15% over the best traditional baseline. We also show layer-wise behaviour: lower layers, which retain acoustic/phonetic detail, are more effective than top layers specialized for Automatic Speech Recognition (ASR).", "date": "2026-02-14", "categories": []}
{"id": "2602.13954", "title": "Eureka-Audio: Triggering Audio Intelligence in Compact Language Models", "abstract": "We present Eureka-Audio, a compact yet high-performance audio language model that achieves competitive performance against models that are 4 to 18 times larger across a broad range of audio understanding benchmarks. Despite containing only 1.7B parameters, Eureka-Audio demonstrates strong performance on automatic speech recognition (ASR), audio understanding, and dense audio captioning, matching or surpassing multiple 7B to 30B audio and omni-modal baselines. The model adopts a unified end-to-end architecture composed of a lightweight language backbone, a Whisper-based audio encoder, and a sparsely activated Mixture-of-Experts (MoE) adapter that explicitly accounts for audio heterogeneity and alleviates cross-modal optimization conflicts under limited capacity. To further enhance paralinguistic reasoning, we introduce DataFlux, a closed loop audio instruction data synthesis and verification pipeline that constructs high quality, logically consistent supervision from raw audio. Extensive evaluations across ASR, knowledge reasoning, safety, instruction following, and paralinguistic benchmarks, demonstrate that Eureka-Audio achieves an efficient balance between computational cost and performance. These results establish Eureka Audio as a strong and practical baseline for lightweight audio understanding models.", "date": "2026-02-15", "categories": []}
{"id": "2602.14224", "title": "The Interspeech 2026 Audio Reasoning Challenge: Evaluating Reasoning Process Quality for Audio Reasoning Models and Agents", "abstract": "Recent Large Audio Language Models (LALMs) excel in understanding but often lack transparent reasoning. To address this \"black-box\" limitation, we organized the Audio Reasoning Challenge at Interspeech 2026, the first shared task dedicated to evaluating Chain-of-Thought (CoT) quality in the audio domain. The challenge introduced MMAR-Rubrics, a novel instance-level protocol assessing the factuality and logic of reasoning chains. Featured Single Model and Agent tracks, the competition attracting 156 teams from 18 countries and regions. Results show agent systems currently lead in reasoning quality, utilizing iterative tool orchestration and cross-modal analysis. Besides, single models are rapidly advancing via reinforcement learning and sophisticated data pipeline. We details the challenge design, methodology, and a comprehensive analysis of state-of-the-art systems, providing new insights for explainable audio intelligence.", "date": "2026-02-15", "categories": []}
{"id": "2602.14785", "title": "SA-SSL-MOS: Self-supervised Learning MOS Prediction with Spectral Augmentation for Generalized Multi-Rate Speech Assessment", "abstract": "Designing a speech quality assessment (SQA) system for estimating mean-opinion-score (MOS) of multi-rate speech with varying sampling frequency (16-48 kHz) is a challenging task. The challenge arises due to the limited availability of a MOS-labeled training dataset comprising multi-rate speech samples. While self-supervised learning (SSL) models have been widely adopted in SQA to boost performance, a key limitation is that they are pretrained on 16 kHz speech and therefore discard high-frequency information present in higher sampling rates. To address this issue, we propose a spectrogram-augmented SSL method that incorporates high-frequency features (up to 48 kHz sampling rate) through a parallel-branch architecture. We further introduce a two-step training scheme: the model is first pre-trained on a large 48 kHz dataset and then fine-tuned on a smaller multi-rate dataset. Experimental results show that leveraging high-frequency information overlooked by SSL features is crucial for accurate multi-rate SQA, and that the proposed two-step training substantially improves generalization when multi-rate data is limited.", "date": "2026-02-16", "categories": []}
{"id": "2601.07969", "title": "Tuberculosis Screening from Cough Audio: Baseline Models, Clinical Variables, and Uncertainty Quantification", "abstract": "In this paper, we propose a standardized framework for automatic tuberculosis (TB) detection from cough audio and routinely collected clinical data using machine learning. While TB screening from audio has attracted growing interest, progress is difficult to measure because existing studies vary substantially in datasets, cohort definitions, feature representations, model families, validation protocols, and reported metrics. Consequently, reported gains are often not directly comparable, and it remains unclear whether improvements stem from modeling advances or from differences in data and evaluation. We address this gap by establishing a strong, well-documented baseline for TB prediction using cough recordings and accompanying clinical metadata from a recently compiled dataset from several countries. Our pipeline is reproducible end-to-end, covering feature extraction, multimodal fusion, cougher-independent evaluation, and uncertainty quantification, and it reports a consistent suite of clinically relevant metrics to enable fair comparison. We further quantify performance for cough audio-only and fused (audio + clinical metadata) models, and release the full experimental protocol to facilitate benchmarking. This baseline is intended to serve as a common reference point and to reduce methodological variance that currently holds back progress in the field.", "date": "2026-01-12", "categories": []}
{"id": "2510.03728", "title": "Lightweight and Generalizable Acoustic Scene Representations via Contrastive Fine-Tuning and Distillation", "abstract": "Acoustic scene classification (ASC) models on edge devices typically operate under fixed class assumptions, lacking the transferability needed for real-world applications that require adaptation to new or refined acoustic categories. We propose ContrastASC, which learns generalizable acoustic scene representations by structuring the embedding space to preserve semantic relationships between scenes, enabling adaptation to unseen categories without retraining. Our approach combines supervised contrastive fine-tuning of pre-trained models with contrastive representation distillation to transfer this structured knowledge to compact student models. Our evaluation shows that ContrastASC demonstrates improved few-shot adaptation to unseen categories while maintaining strong closed-set performance.", "date": "2025-10-04", "categories": []}
{"id": "2509.04345", "title": "AUDETER: A Large-scale Dataset for Deepfake Audio Detection in Open Worlds", "abstract": "Speech synthesis systems can now produce highly realistic vocalisations that pose significant authenticity challenges. Despite substantial progress in deepfake detection models, their real-world effectiveness is often undermined by evolving distribution shifts between training and test data, driven by the complexity of human speech and the rapid evolution of synthesis systems. Existing datasets suffer from limited real speech diversity, insufficient coverage of recent synthesis systems, and heterogeneous mixtures of deepfake sources, which hinder systematic evaluation and open-world model training. To address these issues, we introduce AUDETER (AUdio DEepfake TEst Range), a large-scale and highly diverse deepfake audio dataset comprising over 4,500 hours of synthetic audio generated by 11 recent TTS models and 10 vocoders, totalling 3 million clips. We further observe that most existing detectors default to binary supervised training, which can induce negative transfer across synthesis sources when the training data contains highly diverse deepfake patterns, impacting overall generalisation. As a complementary contribution, we propose an effective curriculum-learning-based approach to mitigate this effect. Extensive experiments show that existing detection models struggle to generalise to novel deepfakes and human speech in AUDETER, whereas XLR-based detectors trained on AUDETER achieve strong cross-domain performance across multiple benchmarks, achieving an EER of 1.87% on In-the-Wild. AUDETER is available on GitHub.", "date": "2025-09-04", "categories": []}
{"id": "2509.17143", "title": "MaskVCT: Masked Voice Codec Transformer for Zero-Shot Voice Conversion With Increased Controllability via Multiple Guidances", "abstract": "We introduce MaskVCT, a zero-shot voice conversion (VC) model that offers multi-factor controllability through multiple classifier-free guidances (CFGs). While previous VC models rely on a fixed conditioning scheme, MaskVCT integrates diverse conditions in a single model. To further enhance robustness and control, the model can leverage continuous or quantized linguistic features to enhance intelligibility and speaker similarity, and can use or omit pitch contour to control prosody. These choices allow users to seamlessly balance speaker identity, linguistic content, and prosodic factors in a zero-shot VC setting. Extensive experiments demonstrate that MaskVCT achieves the best target speaker and accent similarities while obtaining competitive word and character error rates compared to existing baselines. Audio samples are available at https://maskvct.github.io/.", "date": "2025-09-21", "categories": []}
{"id": "2602.10230", "title": "Frame-Level Internal Tool Use for Temporal Grounding in Audio LMs", "abstract": "Large audio language models are increasingly used for complex audio understanding tasks, but they struggle with temporal tasks that require precise temporal grounding, such as word alignment and speaker diarization. The standard approach, where we generate timestamps as sequences of text tokens, is computationally expensive and prone to hallucination, especially when processing audio lengths outside the model's training distribution. In this work, we propose frame-level internal tool use, a method that trains audio LMs to use their own internal audio representations to perform temporal grounding directly. We introduce a lightweight prediction mechanism trained via two objectives: a binary frame classifier and a novel inhomogeneous Poisson process (IHP) loss that models temporal event intensity. Across word localization, speaker diarization, and event localization tasks, our approach outperforms token-based baselines. Most notably, it achieves a >50x inference speedup and demonstrates robust length generalization, maintaining high accuracy on out-of-distribution audio durations where standard token-based models collapse completely.", "date": "2026-02-10", "categories": []}
{"id": "2602.10439", "title": "AudioRouter: Data Efficient Audio Understanding via RL based Dual Reasoning", "abstract": "Large Audio Language Models (LALMs) have demonstrated strong capabilities in audio understanding and reasoning. However, their performance on fine grained auditory perception remains unreliable, and existing approaches largely rely on data intensive training to internalize perceptual abilities. We propose AudioRouter, a reinforcement learning framework that enables LALMs to improve audio understanding by learning when and how to use external audio tools. Rather than tightly coupling tool usage with audio reasoning, AudioRouter formulates tool use as an explicit decision making problem and optimizes a lightweight routing policy while keeping the underlying reasoning model frozen. Experimental results show that AudioRouter achieves substantial improvements on standard audio understanding benchmarks while requiring up to 600x less training data to learn tool usage compared with conventional training paradigms. These findings suggest that learning effective tool usage offers a data efficient and scalable alternative to internalizing perceptual abilities in LALMs.", "date": "2026-02-11", "categories": []}
{"id": "2508.13786", "title": "DegDiT: Controllable Audio Generation with Dynamic Event Graph Guided Diffusion Transformer", "abstract": "Controllable text-to-audio generation aims to synthesize audio from textual descriptions while satisfying user-specified constraints, including event types, temporal sequences, and onset and offset timestamps. This enables precise control over both the content and temporal structure of the generated audio. Despite recent progress, existing methods still face inherent trade-offs among accurate temporal localization, open-vocabulary scalability, and practical efficiency. To address these challenges, we propose DegDiT, a novel dynamic event graph-guided diffusion transformer framework for open-vocabulary controllable audio generation. DegDiT encodes the events in the description as structured dynamic graphs. The nodes in each graph are designed to represent three aspects: semantic features, temporal attributes, and inter-event connections. A graph transformer is employed to integrate these nodes and produce contextualized event embeddings that serve as guidance for the diffusion model. To ensure high-quality and diverse training data, we introduce a quality-balanced data selection pipeline that combines hierarchical event annotation with multi-criteria quality scoring, resulting in a curated dataset with semantic diversity. Furthermore, we present consensus preference optimization, facilitating audio generation through consensus among multiple reward signals. Extensive experiments on AudioCondition, DESED, and AudioTime datasets demonstrate that DegDiT achieves state-of-the-art performances across a variety of objective and subjective evaluation metrics.", "date": "2025-08-19", "categories": []}
{"id": "2510.21280", "title": "WhaleVAD-BPN: Improving Baleen Whale Call Detection with Boundary Proposal Networks and Post-processing Optimisation", "abstract": "While recent sound event detection (SED) systems can identify baleen whale calls in marine audio, challenges related to false positive and minority-class detection persist. We propose the boundary proposal network (BPN), which extends an existing lightweight SED system. The BPN is inspired by work in image object detection and aims to reduce the number of false positive detections. It achieves this by using intermediate latent representations computed within the backbone classification model to gate the final output. When added to an existing SED system, the BPN achieves a 16.8 % absolute increase in precision, as well as 21.3 % and 9.4 % improvements in the F1-score for minority-class d-calls and bp-calls, respectively. We further consider two approaches to the selection of post-processing hyperparameters: a forward-search and a backward-search. By separately optimising event-level and frame-level hyperparameters, these two approaches lead to considerable performance improvements over parameters selected using empirical methods. The complete WhaleVAD-BPN system achieves a cross-validated development F1-score of 0.475, which is a 9.8 % absolute improvement over the baseline.", "date": "2025-10-24", "categories": []}
{"id": "2602.07143", "title": "Massive Sound Embedding Benchmark (MSEB)", "abstract": "Audio is a critical component of multimodal perception, and any truly intelligent system must demonstrate a wide range of auditory capabilities. These capabilities include transcription, classification, retrieval, reasoning, segmentation, clustering, reranking, and reconstruction. Fundamentally, each task involves transforming a raw audio signal into a meaningful 'embedding' - be it a single vector, a sequence of continuous or discrete representations, or another structured form - which then serves as the basis for generating the task's final response. To accelerate progress towards robust machine auditory intelligence, we present the Massive Sound Embedding Benchmark (MSEB): an extensible framework designed to evaluate the auditory components of any multimodal system. In its first release, MSEB offers a comprehensive suite of eight core tasks, with more planned for the future, supported by diverse datasets, including the new, large-scale Simple Voice Questions (SVQ) dataset. Our initial experiments establish clear performance headrooms, highlighting the significant opportunity to improve real-world multimodal experiences where audio is a core signal. We encourage the research community to use MSEB to assess their algorithms and contribute to its growth. The library is publicly hosted at github.", "date": "2026-02-06", "categories": []}
{"id": "2602.06180", "title": "STACodec: Semantic Token Assignment for Balancing Acoustic Fidelity and Semantic Information in Audio Codecs", "abstract": "Neural audio codecs are widely used for audio compression and can be integrated into token-based language models. Traditional codecs preserve acoustic details well but lack semantic information. Recent hybrid codecs attempt to incorporate semantic information through distillation, but this often degrades reconstruction performance, making it difficult to achieve both. To address this limitation, we introduce STACodec, a unified codec that integrates semantic information from self-supervised learning (SSL) models into the first layer of residual vector quantization (RVQ-1) via semantic token assignment (STA). To further eliminate reliance on SSL-based semantic tokenizers and improve efficiency during inference, we propose a semantic pre-distillation (SPD) module, which predicts semantic tokens directly for assignment to the first RVQ layer during inference. Experimental results show that STACodec outperforms existing hybrid codecs in both audio reconstruction and downstream semantic tasks, demonstrating a better balance between acoustic fidelity and semantic capability.", "date": "2026-02-05", "categories": []}
{"id": "2602.06917", "title": "Automatic Detection and Analysis of Singing Mistakes for Music Pedagogy", "abstract": "The advancement of machine learning in audio analysis has opened new possibilities for technology-enhanced music education. This paper introduces a framework for automatic singing mistake detection in the context of music pedagogy, supported by a newly curated dataset. The dataset comprises synchronized teacher learner vocal recordings, with annotations marking different types of mistakes made by learners. Using this dataset, we develop different deep learning models for mistake detection and benchmark them. To compare the efficacy of mistake detection systems, a new evaluation methodology is proposed. Experiments indicate that the proposed learning-based methods are superior to rule-based methods. A systematic study of errors and a cross-teacher study reveal insights into music pedagogy that can be utilised for various music applications. This work sets out new directions of research in music pedagogy. The codes and dataset are publicly available.", "date": "2026-02-06", "categories": []}
{"id": "2505.14814", "title": "GraphemeAug: A Systematic Approach to Synthesized Hard Negative Keyword Spotting Examples", "abstract": "Spoken Keyword Spotting (KWS) is the task of distinguishing between the presence and absence of a keyword in audio. The accuracy of a KWS model hinges on its ability to correctly classify examples close to the keyword and non-keyword boundary. These boundary examples are often scarce in training data, limiting model performance. In this paper, we propose a method to systematically generate adversarial examples close to the decision boundary by making insertion/deletion/substitution edits on the keyword's graphemes. We evaluate this technique on held-out data for a popular keyword and show that the technique improves AUC on a dataset of synthetic hard negatives by 61% while maintaining quality on positives and ambient negative audio data.", "date": "2025-05-20", "categories": []}
{"id": "2510.00771", "title": "UniverSR: Unified and Versatile Audio Super-Resolution via Vocoder-Free Flow Matching", "abstract": "In this paper, we present a vocoder-free framework for audio super-resolution that employs a flow matching generative model to capture the conditional distribution of complex-valued spectral coefficients. Unlike conventional two-stage diffusion-based approaches that predict a mel-spectrogram and then rely on a pre-trained neural vocoder to synthesize waveforms, our method directly reconstructs waveforms via the inverse Short-Time Fourier Transform (iSTFT), thereby eliminating the dependence on a separate vocoder. This design not only simplifies end-to-end optimization but also overcomes a critical bottleneck of two-stage pipelines, where the final audio quality is fundamentally constrained by vocoder performance. Experiments show that our model consistently produces high-fidelity 48 kHz audio across diverse upsampling factors, achieving state-of-the-art performance on both speech and general audio datasets.", "date": "2025-10-01", "categories": []}
{"id": "2601.13948", "title": "Stream-Voice-Anon: Enhancing Utility of Real-Time Speaker Anonymization via Neural Audio Codec and Language Models", "abstract": "Protecting speaker identity is crucial for online voice applications, yet streaming speaker anonymization (SA) remains underexplored. Recent research has demonstrated that neural audio codec (NAC) provides superior speaker feature disentanglement and linguistic fidelity. NAC can also be used with causal language models (LM) to enhance linguistic fidelity and prompt control for streaming tasks. However, existing NAC-based online LM systems are designed for voice conversion (VC) rather than anonymization, lacking the techniques required for privacy protection. Building on these advances, we present Stream-Voice-Anon, which adapts modern causal LM-based NAC architectures specifically for streaming SA by integrating anonymization techniques. Our anonymization approach incorporates pseudo-speaker representation sampling, a speaker embedding mixing and diverse prompt selection strategies for LM conditioning that leverage the disentanglement properties of quantized content codes to prevent speaker information leakage. Additionally, we compare dynamic and fixed delay configurations to explore latency-privacy trade-offs in real-time scenarios. Under the VoicePrivacy 2024 Challenge protocol, Stream-Voice-Anon achieves substantial improvements in intelligibility (up to 46% relative WER reduction) and emotion preservation (up to 28% UAR relative) compared to the previous state-of-the-art streaming method DarkStream while maintaining comparable latency (180ms vs 200ms) and privacy protection against lazy-informed attackers, though showing 15% relative degradation against semi-informed attackers.", "date": "2026-01-20", "categories": []}
{"id": "2506.00934", "title": "GRAM: Spatial general-purpose audio representation models for real-world applications", "abstract": "Audio foundation models learn general-purpose audio representations that facilitate a wide range of downstream tasks. While the performance of these models has greatly increased for conventional single-channel, dry audio clips, their success in real-world acoustic environments with reverberation and noise is limited. Furthermore, most audio foundation models ignore the spatial dimension of real-world acoustic environments, ruling out tasks involving sound localization. To address these limitations, we propose GRAM: a general-purpose real-world audio model that employs a multi-channel masked autoencoder to efficiently learn spatial audio representations. We evaluated GRAM and other audio foundation models in a standardized manner on high-quality simulations of naturalistic, spatial acoustic environments as well as recordings of real-world environments and release these two complementary benchmark task suites: NatHEAR and RealSELD. Our results demonstrate that GRAM outperforms all state-of-the-art self-supervised audio foundation models on NatHEAR and the clean, single-channel version HEAR, while using only a fraction of the training data. GRAM also shows state-of-the-art localization performance in simulated environments and generalizes efficiently to real-world recordings in RealSELD. Taken together, GRAM presents a significant advance toward robust spatial audio foundation models for real-world environments.", "date": "2025-06-01", "categories": []}
{"id": "2601.23174", "title": "Beyond Fixed Frames: Dynamic Character-Aligned Speech Tokenization", "abstract": "Neural audio codecs are at the core of modern conversational speech technologies, converting continuous speech into sequences of discrete tokens that can be processed by LLMs. However, existing codecs typically operate at fixed frame rates, allocating tokens uniformly in time and producing unnecessarily long sequences. In this work, we introduce DyCAST, a Dynamic Character-Aligned Speech Tokenizer that enables variable-frame-rate tokenization through soft character-level alignment and explicit duration modeling. DyCAST learns to associate tokens with character-level linguistic units during training and supports alignment-free inference with direct control over token durations at decoding time. To improve speech resynthesis quality at low frame rates, we further introduce a retrieval-augmented decoding mechanism that enhances reconstruction fidelity without increasing bitrate. Experiments show that DyCAST achieves competitive speech resynthesis quality and downstream performance while using significantly fewer tokens than fixed-frame-rate codecs. Code and checkpoints will be released publicly at https://github.com/lucadellalib/dycast.", "date": "2026-01-30", "categories": []}
{"id": "2602.03762", "title": "Conditional Flow Matching for Visually-Guided Acoustic Highlighting", "abstract": "Visually-guided acoustic highlighting seeks to rebalance audio in alignment with the accompanying video, creating a coherent audio-visual experience. While visual saliency and enhancement have been widely studied, acoustic highlighting remains underexplored, often leading to misalignment between visual and auditory focus. Existing approaches use discriminative models, which struggle with the inherent ambiguity in audio remixing, where no natural one-to-one mapping exists between poorly-balanced and well-balanced audio mixes. To address this limitation, we reframe this task as a generative problem and introduce a Conditional Flow Matching (CFM) framework. A key challenge in iterative flow-based generation is that early prediction errors -- in selecting the correct source to enhance -- compound over steps and push trajectories off-manifold. To address this, we introduce a rollout loss that penalizes drift at the final step, encouraging self-correcting trajectories and stabilizing long-range flow integration. We further propose a conditioning module that fuses audio and visual cues before vector field regression, enabling explicit cross-modal source selection. Extensive quantitative and qualitative evaluations show that our method consistently surpasses the previous state-of-the-art discriminative approach, establishing that visually-guided audio remixing is best addressed through generative modeling.", "date": "2026-02-03", "categories": []}
{"id": "2602.04085", "title": "BASS: Benchmarking Audio LMs for Musical Structure and Semantic Reasoning", "abstract": "Music understanding is a complex task that often requires reasoning over both structural and semantic elements of audio. We introduce BASS, designed to evaluate music understanding and reasoning in audio language models across four broad categories: structural segmentation, lyric transcription, musicological analysis, and artist collaboration. BASS comprises 2658 questions spanning 12 tasks, 1993 unique songs and covering over 138 hours of music from a wide range of genres and tracks, crafted to assess musicological knowledge and reasoning in real-world scenarios. We evaluate 14 open-source and frontier multimodal LMs, finding that even state-of-the-art models struggle on higher-level reasoning tasks such as structural segmentation and artist collaboration, while performing best on lyric transcription. Our analysis reveals that current models leverage linguistic priors effectively but remain limited in reasoning over musical structure, vocal, and musicological attributes. BASS provides an evaluation framework with widespread applications in music recommendation and search and has the potential to guide the development of audio LMs.", "date": "2026-02-03", "categories": []}
{"id": "2602.04680", "title": "Audio ControlNet for Fine-Grained Audio Generation and Editing", "abstract": "We study the fine-grained text-to-audio (T2A) generation task. While recent models can synthesize high-quality audio from text descriptions, they often lack precise control over attributes such as loudness, pitch, and sound events. Unlike prior approaches that retrain models for specific control types, we propose to train ControlNet models on top of pre-trained T2A backbones to achieve controllable generation over loudness, pitch, and event roll. We introduce two designs, T2A-ControlNet and T2A-Adapter, and show that the T2A-Adapter model offers a more efficient structure with strong control ability. With only 38M additional parameters, T2A-Adapter achieves state-of-the-art performance on the AudioSet-Strong in both event-level and segment-level F1 scores. We further extend this framework to audio editing, proposing T2A-Editor for removing and inserting audio events at time locations specified by instructions. Models, code, dataset pipelines, and benchmarks will be released to support future research on controllable audio generation and editing.", "date": "2026-02-04", "categories": []}
{"id": "2510.18190", "title": "Joint Estimation of Piano Dynamics and Metrical Structure with a Multi-task Multi-Scale Network", "abstract": "Estimating piano dynamic from audio recordings is a fundamental challenge in computational music analysis. In this paper, we propose an efficient multi-task network that jointly predicts dynamic levels, change points, beats, and downbeats from a shared latent representation. These four targets form the metrical structure of dynamics in the music score. Inspired by recent vocal dynamic research, we use a multi-scale network as the backbone, which takes Bark-scale specific loudness as the input feature. Compared to log-Mel as input, this reduces model size from 14.7 M to 0.5 M, enabling long sequential input. We use a 60-second audio length in audio segmentation, which doubled the length of beat tracking commonly used. Evaluated on the public MazurkaBL dataset, our model achieves state-of-the-art results across all tasks. This work sets a new benchmark for piano dynamic estimation and delivers a powerful and compact tool, paving the way for large-scale, resource-efficient analysis of musical expression.", "date": "2025-10-21", "categories": []}
{"id": "2512.01834", "title": "Mitigating Gender Bias in Depression Detection via Counterfactual Inference", "abstract": "Audio-based depression detection models have demonstrated promising performance but often suffer from gender bias due to imbalanced training data. Epidemiological statistics show a higher prevalence of depression in females, leading models to learn spurious correlations between gender and depression. Consequently, models tend to over-diagnose female patients while underperforming on male patients, raising significant fairness concerns. To address this, we propose a novel Counterfactual Debiasing Framework grounded in causal inference. We construct a causal graph to model the decision-making process and identify gender bias as the direct causal effect of gender on the prediction. During inference, we employ counterfactual inference to estimate and subtract this direct effect, ensuring the model relies primarily on authentic acoustic pathological features. Extensive experiments on the DAIC-WOZ dataset using two advanced acoustic backbones demonstrate that our framework not only significantly reduces gender bias but also improves overall detection performance compared to existing debiasing strategies.", "date": "2025-12-01", "categories": []}
{"id": "2602.02955", "title": "Synthetic Data Augmentation for Medical Audio Classification: A Preliminary Evaluation", "abstract": "Medical audio classification remains challenging due to low signal-to-noise ratios, subtle discriminative features, and substantial intra-class variability, often compounded by class imbalance and limited training data. Synthetic data augmentation has been proposed as a potential strategy to mitigate these constraints; however, prior studies report inconsistent methodological approaches and mixed empirical results. In this preliminary study, we explore the impact of synthetic augmentation on respiratory sound classification using a baseline deep convolutional neural network trained on a moderately imbalanced dataset (73%:27%). Three generative augmentation strategies (variational autoencoders, generative adversarial networks, and diffusion models) were assessed under controlled experimental conditions. The baseline model without augmentation achieved an F1-score of 0.645. Across individual augmentation strategies, performance gains were not observed, with several configurations demonstrating neutral or degraded classification performance. Only an ensemble of augmented models yielded a modest improvement in F1-score (0.664). These findings suggest that, for medical audio classification, synthetic augmentation may not consistently enhance performance when applied to a standard CNN classifier. Future work should focus on delineating task-specific data characteristics, model-augmentation compatibility, and evaluation frameworks necessary for synthetic augmentation to be effective in medical audio applications.", "date": "2026-02-03", "categories": []}
{"id": "2602.03355", "title": "PACE: Pretrained Audio Continual Learning", "abstract": "Audio is a fundamental modality for analyzing speech, music, and environmental sounds. Although pretrained audio models have significantly advanced audio understanding, they remain fragile in real-world settings where data distributions shift over time. In this work, we present the first systematic benchmark for audio continual learning (CL) with pretrained models (PTMs), together with a comprehensive analysis of its unique challenges. Unlike in vision, where parameter-efficient fine-tuning (PEFT) has proven effective for CL, directly transferring such strategies to audio leads to poor performance. This stems from a fundamental property of audio backbones: they focus on low-level spectral details rather than structured semantics, causing severe upstream-downstream misalignment. Through extensive empirical study, we identify analytic classifiers with first-session adaptation (FSA) as a promising direction, but also reveal two major limitations: representation saturation in coarse-grained scenarios and representation drift in fine-grained scenarios. To address these challenges, we propose PACE, a novel method that enhances FSA via a regularized analytic classifier and enables multi-session adaptation through adaptive subspace-orthogonal PEFT for improved semantic alignment. In addition, we introduce spectrogram-based boundary-aware perturbations to mitigate representation overlap and improve stability. Experiments on six diverse audio CL benchmarks demonstrate that PACE substantially outperforms state-of-the-art baselines, marking an important step toward robust and scalable audio continual learning with PTMs.", "date": "2026-02-03", "categories": []}
{"id": "2602.03817", "title": "Adaptive Evidence Weighting for Audio-Spatiotemporal Fusion", "abstract": "Many machine learning systems have access to multiple sources of evidence for the same prediction target, yet these sources often differ in reliability and informativeness across inputs. In bioacoustic classification, species identity may be inferred both from the acoustic signal and from spatiotemporal context such as location and season; while Bayesian inference motivates multiplicative evidence combination, in practice we typically only have access to discriminative predictors rather than calibrated generative models. We introduce \\textbf{F}usion under \\textbf{IN}dependent \\textbf{C}onditional \\textbf{H}ypotheses (\\textbf{FINCH}), an adaptive log-linear evidence fusion framework that integrates a pre-trained audio classifier with a structured spatiotemporal predictor. FINCH learns a per-sample gating function that estimates the reliability of contextual information from uncertainty and informativeness statistics. The resulting fusion family \\emph{contains} the audio-only classifier as a special case and explicitly bounds the influence of contextual evidence, yielding a risk-contained hypothesis class with an interpretable audio-only fallback. Across benchmarks, FINCH consistently outperforms fixed-weight fusion and audio-only baselines, improving robustness and error trade-offs even when contextual information is weak in isolation. We achieve state-of-the-art performance on CBI and competitive or improved performance on several subsets of BirdSet using a lightweight, interpretable, evidence-based approach. Code is available: \\texttt{\\href{https://anonymous.4open.science/r/birdnoise-85CD/README.md}{anonymous-repository}}", "date": "2026-02-03", "categories": []}
{"id": "2506.10423", "title": "PAL: Probing Audio Encoders via LLMs -- Audio Information Transfer into LLMs", "abstract": "Integration of audio perception into large language models (LLMs) is an emerging research area for enabling machine listening applications, yet efficient transfer of rich audio semantics from audio encoders to LLMs remains underexplored. The most widely used integration paradigm projects audio-encoder output tokens into the LLM input space (e.g., via an MLP or a Q-Former) and then prepends or inserts them into the text token sequence. We refer to this generic scheme as Prepend to the LLM's input token space (PLITS) integration. We propose an efficient alternative, Lightweight Audio LLM Integration (LAL). LAL injects audio representations solely through the attention mechanism at selected LLM layers, bypassing the feed-forward module. It encodes rich audio semantics at an appropriate level of abstraction for integration into different transformer blocks, substantially reducing computational overhead compared to existing approaches. We further introduce PAL, a hybrid integration approach for efficiently Probing Audio encoders via LLM. PAL applies PLITS only to a compact set of summary tokens while integrating the full audio token sequence via LAL. Under an identical training curriculum, LAL consistently matches or outperforms existing integration approaches across multiple base LLMs and tasks, with improvements of up to 30% over a strong PLITS baseline, while reducing memory usage by about 60% and increasing throughput by about 190%. Moreover, PAL matches or exceeds PLITS performance while offering substantially better computational and memory efficiency.", "date": "2025-06-12", "categories": []}
{"id": "2509.21033", "title": "SupCLAP: Controlling Optimization Trajectory Drift in Audio-Text Contrastive Learning with Support Vector Regularization", "abstract": "Contrastive language-audio pretraining, which aims to unify multimodal representations in a shared embedding space, serves as a cornerstone for building a wide range of applications, from cross-modal retrieval to cutting-edge multimodal large language models. However, we find that the perpendicular component of the pushing force from negative samples in contrastive learning is a double-edged sword: it contains rich supplementary information from negative samples, yet its unconstrained nature causes optimization trajectory drift and training instability. To address this, we propose Support Vector Regularization (SVR), a method that introduces an auxiliary support vector to control this perpendicular component, aiming to harness its rich information while mitigating the associated trajectory drift. The efficacy of SVR is critically governed by its semantic radius, for which we explore two unsupervised modeling strategies: direct parameterization and an adaptive radius predictor module enhanced with constraints to improve its predicting accuracy. Extensive experimental results demonstrate that our method surpasses widely used baselines like InfoNCE and SigLIP loss across classification, monolingual retrieval, and multilingual retrieval on standard audio-text datasets. Both the theoretical analysis and the experimental results on optimizing trajectory drift validate the correctness and effectiveness of our SVR method. Notably, our method is highly efficient, it operates without the need for extra training data or inference computation, and adds only a negligible overhead to the training.", "date": "2025-09-25", "categories": []}
{"id": "2601.05011", "title": "Leveraging Prediction Entropy for Automatic Prompt Weighting in Zero-Shot Audio-Language Classification", "abstract": "Audio-language models have recently demonstrated strong zero-shot capabilities by leveraging natural-language supervision to classify audio events without labeled training data. Yet, their performance is highly sensitive to the wording of text prompts, with small variations leading to large fluctuations in accuracy. Prior work has mitigated this issue through prompt learning or prompt ensembling. However, these strategies either require annotated data or fail to account for the fact that some prompts may negatively impact performance. In this work, we present an entropy-guided prompt weighting approach that aims to find a robust combination of prompt contributions to maximize prediction confidence. To this end, we formulate a tailored objective function that minimizes prediction entropy to yield new prompt weights, utilizing low-entropy as a proxy for high confidence. Our approach can be applied to individual samples or a batch of audio samples, requiring no additional labels and incurring negligible computational overhead. Experiments on five audio classification datasets covering environmental, urban, and vocal sounds, demonstrate consistent gains compared to classical prompt ensembling methods in a zero-shot setting, with accuracy improvements 5-times larger across the whole benchmark.", "date": "2026-01-08", "categories": []}
{"id": "2602.00295", "title": "Multi-Speaker Conversational Audio Deepfake: Taxonomy, Dataset and Pilot Study", "abstract": "The rapid advances in text-to-speech (TTS) technologies have made audio deepfakes increasingly realistic and accessible, raising significant security and trust concerns. While existing research has largely focused on detecting single-speaker audio deepfakes, real-world malicious applications with multi-speaker conversational settings is also emerging as a major underexplored threat. To address this gap, we propose a conceptual taxonomy of multi-speaker conversational audio deepfakes, distinguishing between partial manipulations (one or multiple speakers altered) and full manipulations (entire conversations synthesized). As a first step, we introduce a new Multi-speaker Conversational Audio Deepfakes Dataset (MsCADD) of 2,830 audio clips containing real and fully synthetic two-speaker conversations, generated using VITS and SoundStorm-based NotebookLM models to simulate natural dialogue with variations in speaker gender, and conversational spontaneity. MsCADD is limited to text-to-speech (TTS) types of deepfake. We benchmark three neural baseline models; LFCC-LCNN, RawNet2, and Wav2Vec 2.0 on this dataset and report performance in terms of F1 score, accuracy, true positive rate (TPR), and true negative rate (TNR). Results show that these baseline models provided a useful benchmark, however, the results also highlight that there is a significant gap in multi-speaker deepfake research in reliably detecting synthetic voices under varied conversational dynamics. Our dataset and benchmarks provide a foundation for future research on deepfake detection in conversational scenarios, which is a highly underexplored area of research but also a major area of threat to trustworthy information in audio settings. The MsCADD dataset is publicly available to support reproducibility and benchmarking by the research community.", "date": "2026-01-30", "categories": []}
{"id": "2602.01032", "title": "HierCon: Hierarchical Contrastive Attention for Audio Deepfake Detection", "abstract": "Audio deepfakes generated by modern TTS and voice conversion systems are increasingly difficult to distinguish from real speech, raising serious risks for security and online trust. While state-of-the-art self-supervised models provide rich multi-layer representations, existing detectors treat layers independently and overlook temporal and hierarchical dependencies critical for identifying synthetic artefacts. We propose HierCon, a hierarchical layer attention framework combined with margin-based contrastive learning that models dependencies across temporal frames, neighbouring layers, and layer groups, while encouraging domain-invariant embeddings. Evaluated on ASVspoof 2021 DF and In-the-Wild datasets, our method achieves state-of-the-art performance (1.93% and 6.87% EER), improving over independent layer weighting by 36.6% and 22.5% respectively. The results and attention visualisations confirm that hierarchical modelling enhances generalisation to cross-domain generation techniques and recording conditions.", "date": "2026-02-01", "categories": []}
{"id": "2602.01060", "title": "TLDiffGAN: A Latent Diffusion-GAN Framework with Temporal Information Fusion for Anomalous Sound Detection", "abstract": "Existing generative models for unsupervised anomalous sound detection are limited by their inability to fully capture the complex feature distribution of normal sounds, while the potential of powerful diffusion models in this domain remains largely unexplored. To address this challenge, we propose a novel framework, TLDiffGAN, which consists of two complementary branches. One branch incorporates a latent diffusion model into the GAN generator for adversarial training, thereby making the discriminator's task more challenging and improving the quality of generated samples. The other branch leverages pretrained audio model encoders to extract features directly from raw audio waveforms for auxiliary discrimination. This framework effectively captures feature representations of normal sounds from both raw audio and Mel spectrograms. Moreover, we introduce a TMixup spectrogram augmentation technique to enhance sensitivity to subtle and localized temporal patterns that are often overlooked. Extensive experiments on the DCASE 2020 Challenge Task 2 dataset demonstrate the superior detection performance of TLDiffGAN, as well as its strong capability in anomalous time-frequency localization.", "date": "2026-02-01", "categories": []}
{"id": "2506.10754", "title": "BNMusic: Blending Environmental Noises into Personalized Music", "abstract": "While being disturbed by environmental noises, the acoustic masking technique is a conventional way to reduce the annoyance in audio engineering that seeks to cover up the noises with other dominant yet less intrusive sounds. However, misalignment between the dominant sound and the noise-such as mismatched downbeats-often requires an excessive volume increase to achieve effective masking. Motivated by recent advances in cross-modal generation, in this work, we introduce an alternative method to acoustic masking, aiming to reduce the noticeability of environmental noises by blending them into personalized music generated based on user-provided text prompts. Following the paradigm of music generation using mel-spectrogram representations, we propose a Blending Noises into Personalized Music (BNMusic) framework with two key stages. The first stage synthesizes a complete piece of music in a mel-spectrogram representation that encapsulates the musical essence of the noise. In the second stage, we adaptively amplify the generated music segment to further reduce noise perception and enhance the blending effectiveness, while preserving auditory quality. Our experiments with comprehensive evaluations on MusicBench, EPIC-SOUNDS, and ESC-50 demonstrate the effectiveness of our framework, highlighting the ability to blend environmental noise with rhythmically aligned, adaptively amplified, and enjoyable music segments, minimizing the noticeability of the noise, thereby improving overall acoustic experiences. Project page: https://d-fas.github.io/BNMusic_page/.", "date": "2025-06-12", "categories": []}
{"id": "2509.11168", "title": "An Entropy-Guided Curriculum Learning Strategy for Data-Efficient Acoustic Scene Classification under Domain Shift", "abstract": "Acoustic Scene Classification (ASC) faces challenges in generalizing across recording devices, particularly when labeled data is limited. The DCASE 2024 Challenge Task 1 highlights this issue by requiring models to learn from small labeled subsets recorded on a few devices. These models need to then generalize to recordings from previously unseen devices under strict complexity constraints. While techniques such as data augmentation and the use of pre-trained models are well-established for improving model generalization, optimizing the training strategy represents a complementary yet less-explored path that introduces no additional architectural complexity or inference overhead. Among various training strategies, curriculum learning offers a promising paradigm by structuring the learning process from easier to harder examples. In this work, we propose an entropy-guided curriculum learning strategy to address the domain shift problem in data-efficient ASC. Specifically, we quantify the uncertainty of device domain predictions for each training sample by computing the Shannon entropy of the device posterior probabilities estimated by an auxiliary domain classifier. Using entropy as a proxy for domain invariance, the curriculum begins with high-entropy samples and gradually incorporates low-entropy, domain-specific ones to facilitate the learning of generalizable representations. Experimental results on multiple DCASE 2024 ASC baselines demonstrate that our strategy effectively mitigates domain shift, particularly under limited labeled data conditions. Our strategy is architecture-agnostic and introduces no additional inference cost, making it easily integrable into existing ASC baselines and offering a practical solution to domain shift.", "date": "2025-09-14", "categories": []}
{"id": "2510.17345", "title": "DDSC: Dynamic Dual-Signal Curriculum for Data-Efficient Acoustic Scene Classification under Domain Shift", "abstract": "Acoustic scene classification (ASC) suffers from device-induced domain shift, especially when labels are limited. Prior work focuses on curriculum-based training schedules that structure data presentation by ordering or reweighting training examples from easy-to-hard to facilitate learning; however, existing curricula are static, fixing the ordering or the weights before training and ignoring that example difficulty and marginal utility evolve with the learned representation. To overcome this limitation, we propose the Dynamic Dual-Signal Curriculum (DDSC), a training schedule that adapts the curriculum online by combining two signals computed each epoch: a domain-invariance signal and a learning-progress signal. A time-varying scheduler fuses these signals into per-example weights that prioritize domain-invariant examples in early epochs and progressively emphasize device-specific cases. DDSC is lightweight, architecture-agnostic, and introduces no additional inference overhead. Under the official DCASE 2024 Task~1 protocol, DDSC consistently improves cross-device performance across diverse ASC baselines and label budgets, with the largest gains on unseen-device splits.", "date": "2025-10-20", "categories": []}
{"id": "2601.23066", "title": "Towards Explicit Acoustic Evidence Perception in Audio LLMs for Speech Deepfake Detection", "abstract": "Speech deepfake detection (SDD) focuses on identifying whether a given speech signal is genuine or has been synthetically generated. Existing audio large language model (LLM)-based methods excel in content understanding; however, their predictions are often biased toward semantically correlated cues, which results in fine-grained acoustic artifacts being overlooked during the decisionmaking process. Consequently, fake speech with natural semantics can bypass detectors despite harboring subtle acoustic anomalies; this suggests that the challenge stems not from the absence of acoustic data, but from its inadequate accessibility when semantic-dominant reasoning prevails. To address this issue, we investigate SDD within the audio LLM paradigm and introduce SDD with Auditory Perception-enhanced Audio Large Language Model (SDD-APALLM), an acoustically enhanced framework designed to explicitly expose fine-grained time-frequency evidence as accessible acoustic cues. By combining raw audio with structured spectrograms, the proposed framework empowers audio LLMs to more effectively capture subtle acoustic inconsistencies without compromising their semantic understanding. Experimental results indicate consistent gains in detection accuracy and robustness, especially in cases where semantic cues are misleading. Further analysis reveals that these improvements stem from a coordinated utilization of semantic and acoustic information, as opposed to simple modality aggregation.", "date": "2026-01-30", "categories": []}
{"id": "2601.23161", "title": "DIFFA-2: A Practical Diffusion Large Language Model for General Audio Understanding", "abstract": "Autoregressive (AR) large audio language models (LALMs) such as Qwen-2.5-Omni have achieved strong performance on audio understanding and interaction, but scaling them remains costly in data and computation, and strictly sequential decoding limits inference efficiency. Diffusion large language models (dLLMs) have recently been shown to make effective use of limited training data, and prior work on DIFFA indicates that replacing an AR backbone with a diffusion counterpart can substantially improve audio understanding under matched settings, albeit at a proof-of-concept scale without large-scale instruction tuning, preference alignment, or practical decoding schemes. We introduce DIFFA-2, a practical diffusion-based LALM for general audio understanding. DIFFA-2 upgrades the speech encoder, employs dual semantic and acoustic adapters, and is trained with a four-stage curriculum that combines semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization, using only fully open-source corpora. Experiments on MMSU, MMAU, and MMAR show that DIFFA-2 consistently improves over DIFFA and is competitive to strong AR LALMs under practical training budgets, supporting diffusion-based modeling is a viable backbone for large-scale audio understanding. Our code is available at https://github.com/NKU-HLT/DIFFA.git.", "date": "2026-01-30", "categories": []}
{"id": "2509.14934", "title": "Mitigating data replication in text-to-audio generative diffusion models through anti-memorization guidance", "abstract": "A persistent challenge in generative audio models is data replication, where the model unintentionally generates parts of its training data during inference. In this work, we address this issue in text-to-audio diffusion models by exploring the use of anti-memorization strategies. We adopt Anti-Memorization Guidance (AMG), a technique that modifies the sampling process of pre-trained diffusion models to discourage memorization. Our study explores three types of guidance within AMG, each designed to reduce replication while preserving generation quality. We use Stable Audio Open as our backbone, leveraging its fully open-source architecture and training dataset. Our comprehensive experimental analysis suggests that AMG significantly mitigates memorization in diffusion-based text-to-audio generation without compromising audio fidelity or semantic alignment.", "date": "2025-09-18", "categories": []}
{"id": "2510.13244", "title": "MotionBeat: Motion-Aligned Music Representation via Embodied Contrastive Learning and Bar-Equivariant Contact-Aware Encoding", "abstract": "Music is both an auditory and an embodied phenomenon, closely linked to human motion and naturally expressed through dance. However, most existing audio representations neglect this embodied dimension, limiting their ability to capture rhythmic and structural cues that drive movement. We propose MotionBeat, a framework for motion-aligned music representation learning. MotionBeat is trained with two newly proposed objectives: the Embodied Contrastive Loss (ECL), an enhanced InfoNCE formulation with tempo-aware and beat-jitter negatives to achieve fine-grained rhythmic discrimination, and the Structural Rhythm Alignment Loss (SRAL), which ensures rhythm consistency by aligning music accents with corresponding motion events. Architecturally, MotionBeat introduces bar-equivariant phase rotations to capture cyclic rhythmic patterns and contact-guided attention to emphasize motion events synchronized with musical accents. Experiments show that MotionBeat outperforms state-of-the-art audio encoders in music-to-dance generation and transfers effectively to beat tracking, music tagging, genre and instrument classification, emotion recognition, and audio-visual retrieval. Our project demo page: https://motionbeat2025.github.io/.", "date": "2025-10-15", "categories": []}
{"id": "2511.15131", "title": "CASTELLA: Long Audio Dataset with Captions and Temporal Boundaries", "abstract": "We introduce CASTELLA, a human-annotated audio benchmark for the task of audio moment retrieval (AMR). Although AMR has various useful potential applications, there is still no established benchmark with real-world data. The initial study of AMR trained the models solely on synthetic datasets. Moreover, the evaluation is based on an annotated dataset of fewer than 100 samples. This resulted in less reliable reported performance. To ensure performance for applications in real-world environments, we present CASTELLA, a large-scale manually annotated AMR dataset. CASTELLA consists of 1009, 213, and 640 audio recordings for training, validation, and test splits, respectively, which is 24 times larger than the previous dataset. We also establish a baseline model for AMR using CASTELLA. Our experiments demonstrate that a model fine-tuned on CASTELLA after pre-training on the synthetic data outperformed a model trained solely on the synthetic data by 10.4 points in Recall1@0.7. CASTELLA is publicly available in https://h-munakata.github.io/CASTELLA-demo/.", "date": "2025-11-19", "categories": []}
{"id": "2512.17209", "title": "Do Foundational Audio Encoders Understand Music Structure?", "abstract": "In music information retrieval (MIR) research, the use of pretrained foundational audio encoders (FAEs) has recently become a trend. FAEs pretrained on large amounts of music and audio data have been shown to improve performance on MIR tasks such as music tagging and automatic music transcription. However, their use for music structure analysis (MSA) remains underexplored: only a small subset of FAEs has been examined for MSA, and the impact of factors such as learning methods, training data, and model context length on MSA performance remains unclear. In this study, we conduct comprehensive experiments on 11 types of FAEs to investigate how these factors affect MSA performance. Our results demonstrate that FAEs using self-supervised learning with masked language modeling on music data are particularly effective for MSA. These findings pave the way for future research in FAE and MSA.", "date": "2025-12-19", "categories": []}
{"id": "2601.15017", "title": "SpatialV2A: Visual-Guided High-fidelity Spatial Audio Generation", "abstract": "While video-to-audio generation has achieved remarkable progress in semantic and temporal alignment, most existing studies focus solely on these aspects, paying limited attention to the spatial perception and immersive quality of the synthesized audio. This limitation stems largely from current models' reliance on mono audio datasets, which lack the binaural spatial information needed to learn visual-to-spatial audio mappings. To address this gap, we introduce two key contributions: we construct BinauralVGGSound, the first large-scale video-binaural audio dataset designed to support spatially aware video-to-audio generation; and we propose a end-to-end spatial audio generation framework guided by visual cues, which explicitly models spatial features. Our framework incorporates a visual-guided audio spatialization module that ensures the generated audio exhibits realistic spatial attributes and layered spatial depth while maintaining semantic and temporal alignment. Experiments show that our approach substantially outperforms state-of-the-art models in spatial fidelity and delivers a more immersive auditory experience, without sacrificing temporal or semantic consistency. The demo page can be accessed at https://github.com/renlinjie868-web/SpatialV2A.", "date": "2026-01-21", "categories": []}
{"id": "2601.19946", "title": "MK-SGC-SC: Multiple Kernel Guided Sparse Graph Construction in Spectral Clustering for Unsupervised Speaker Diarization", "abstract": "Speaker diarization aims to segment audio recordings into regions corresponding to individual speakers. Although unsupervised speaker diarization is inherently challenging, the prospect of identifying speaker regions without pretraining or weak supervision motivates research on clustering techniques. In this work, we share the notable observation that measuring multiple kernel similarities of speaker embeddings to thereafter craft a sparse graph for spectral clustering in a principled manner is sufficient to achieve state-of-the-art performances in a fully unsupervised setting. Specifically, we consider four polynomial kernels and a degree one arccosine kernel to measure similarities in speaker embeddings, using which sparse graphs are constructed in a principled manner to emphasize local similarities. Experiments show the proposed approach excels in unsupervised speaker diarization over a variety of challenging environments in the DIHARD-III, AMI, and VoxConverse corpora. To encourage further research, our implementations are available at https://github.com/nikhilraghav29/MK-SGC-SC.", "date": "2026-01-24", "categories": []}
{"id": "2601.21612", "title": "Representation-Regularized Convolutional Audio Transformer for Audio Understanding", "abstract": "Bootstrap-based Self-Supervised Learning (SSL) has achieved remarkable progress in audio understanding. However, existing methods typically operate at a single level of granularity, limiting their ability to model the diverse temporal and spectral structures inherent in complex audio signals. Furthermore, bootstrapping representations from scratch is computationally expensive, often requiring extensive training to converge. In this work, we propose the Convolutional Audio Transformer (CAT), a unified framework designed to address these challenges. First, to capture hierarchical audio features, CAT incorporates a Multi-resolution Block that aggregates information across varying granularities. Second, to enhance training efficiency, we introduce a Representation Regularization objective. Drawing inspiration from generative modeling, this auxiliary task guides the student model by aligning its predictions with high-quality semantic representations from frozen, pre-trained external encoders. Experimental results demonstrate that CAT significantly outperforms baselines on audio understanding benchmarks. Notably, it achieves competitive performance on the AudioSet 20k dataset with 5 times faster convergence than existing methods. Codes and checkpoints will be released soon at https://github.com/realzhouchushu/CAT.", "date": "2026-01-29", "categories": []}
{"id": "2509.20682", "title": "Addressing Gradient Misalignment in Data-Augmented Training for Robust Speech Deepfake Detection", "abstract": "In speech deepfake detection (SDD), data augmentation (DA) is commonly used to improve model generalization across varied speech conditions and spoofing attacks. However, during training, the backpropagated gradients from original and augmented inputs may misalign, which can result in conflicting parameter updates. These conflicts could hinder convergence and push the model toward suboptimal solutions, thereby reducing the benefits of DA. To investigate and address this issue, we design a dual-path data-augmented (DPDA) training framework with gradient alignment for SDD. In our framework, each training utterance is processed through two input paths: one using the original speech and the other with its augmented version. This design allows us to compare and align their backpropagated gradient directions to reduce optimization conflicts. Our analysis shows that approximately 25% of training iterations exhibit gradient conflicts between the original inputs and their augmented counterparts when using RawBoost augmentation. By resolving these conflicts with gradient alignment, our method accelerates convergence by reducing the number of training epochs and achieves up to an 18.69% relative reduction in Equal Error Rate on the In-the-Wild dataset compared to the baseline.", "date": "2025-09-25", "categories": []}
{"id": "2510.05305", "title": "WaveSP-Net: Learnable Wavelet-Domain Sparse Prompt Tuning for Speech Deepfake Detection", "abstract": "Modern front-end design for speech deepfake detection relies on full fine-tuning of large pre-trained models like XLSR. However, this approach is not parameter-efficient and may lead to suboptimal generalization to realistic, in-the-wild data types. To address these limitations, we introduce a new family of parameter-efficient front-ends that fuse prompt-tuning with classical signal processing transforms. These include FourierPT-XLSR, which uses the Fourier Transform, and two variants based on the Wavelet Transform: WSPT-XLSR and Partial-WSPT-XLSR. We further propose WaveSP-Net, a novel architecture combining a Partial-WSPT-XLSR front-end and a bidirectional Mamba-based back-end. This design injects multi-resolution features into the prompt embeddings, which enhances the localization of subtle synthetic artifacts without altering the frozen XLSR parameters. Experimental results demonstrate that WaveSP-Net outperforms several state-of-the-art models on two new and challenging benchmarks, Deepfake-Eval-2024 and SpoofCeleb, with low trainable parameters and notable performance gains. The code and models are available at https://github.com/xxuan-acoustics/WaveSP-Net.", "date": "2025-10-06", "categories": []}
{"id": "2510.23530", "title": "Learning Linearity in Audio Consistency Autoencoders via Implicit Regularization", "abstract": "Audio autoencoders learn useful, compressed audio representations, but their non-linear latent spaces prevent intuitive algebraic manipulation such as mixing or scaling. We introduce a simple training methodology to induce linearity in a high-compression Consistency Autoencoder (CAE) by using data augmentation, thereby inducing homogeneity (equivariance to scalar gain) and additivity (the decoder preserves addition) without altering the model's architecture or loss function. When trained with our method, the CAE exhibits linear behavior in both the encoder and decoder while preserving reconstruction fidelity. We test the practical utility of our learned space on music source composition and separation via simple latent arithmetic. This work presents a straightforward technique for constructing structured latent spaces, enabling more intuitive and efficient audio processing.", "date": "2025-10-27", "categories": []}
{"id": "2601.17517", "title": "EuleroDec: A Complex-Valued RVQ-VAE for Efficient and Robust Audio Coding", "abstract": "Audio codecs power discrete music generative modelling, music streaming and immersive media by shrinking PCM audio to bandwidth-friendly bit-rates. Recent works have gravitated towards processing in the spectral domain; however, spectrogram-domains typically struggle with phase modeling which is naturally complex-valued. Most frequency-domain neural codecs either disregard phase information or encode it as two separate real-valued channels, limiting spatial fidelity. This entails the need to introduce adversarial discriminators at the expense of convergence speed and training stability to compensate for the inadequate representation power of the audio signal. In this work we introduce an end-to-end complex-valued RVQ-VAE audio codec that preserves magnitude-phase coupling across the entire analysis-quantization-synthesis pipeline and removes adversarial discriminators and diffusion post-filters. Without GANs or diffusion we match or surpass much longer-trained baselines in-domain and reach SOTA out-of-domain performance. Compared to standard baselines that train for hundreds of thousands of steps, our model reducing training budget by an order of magnitude is markedly more compute-efficient while preserving high perceptual quality.", "date": "2026-01-24", "categories": []}
{"id": "2601.20142", "title": "Mind the Shift: Using Delta SSL Embeddings to Enhance Child ASR", "abstract": "Self-supervised learning (SSL) models have achieved impressive results across many speech tasks, yet child automatic speech recognition (ASR) remains challenging due to limited data and pretraining domain mismatch. Fine-tuning SSL models on child speech induces shifts in the representation space. We hypothesize that delta SSL embeddings, defined as the differences between embeddings from a finetuned model and those from its pretrained counterpart, encode task-specific information that complements finetuned features from another SSL model. We evaluate multiple fusion strategies on the MyST childrens corpus using different models. Results show that delta embedding fusion with WavLM yields up to a 10 percent relative WER reduction for HuBERT and a 4.4 percent reduction for W2V2, compared to finetuned embedding fusion. Notably, fusing WavLM with delta W2V2 embeddings achieves a WER of 9.64, setting a new state of the art among SSL models on the MyST corpus. These findings demonstrate the effectiveness of delta embeddings and highlight feature fusion as a promising direction for advancing child ASR.", "date": "2026-01-28", "categories": []}
{"id": "2601.20510", "title": "Audio Deepfake Detection in the Age of Advanced Text-to-Speech models", "abstract": "Recent advances in Text-to-Speech (TTS) systems have substantially increased the realism of synthetic speech, raising new challenges for audio deepfake detection. This work presents a comparative evaluation of three state-of-the-art TTS models--Dia2, Maya1, and MeloTTS--representing streaming, LLM-based, and non-autoregressive architectures. A corpus of 12,000 synthetic audio samples was generated using the Daily-Dialog dataset and evaluated against four detection frameworks, including semantic, structural, and signal-level approaches. The results reveal significant variability in detector performance across generative mechanisms: models effective against one TTS architecture may fail against others, particularly LLM-based synthesis. In contrast, a multi-view detection approach combining complementary analysis levels demonstrates robust performance across all evaluated models. These findings highlight the limitations of single-paradigm detectors and emphasize the necessity of integrated detection strategies to address the evolving landscape of audio deepfake threats.", "date": "2026-01-28", "categories": []}
{"id": "2509.18561", "title": "SoundCompass: Navigating Target Sound Extraction With Effective Directional Clue Integration In Complex Acoustic Scenes", "abstract": "Recent advances in target sound extraction (TSE) utilize directional clues derived from direction of arrival (DoA), which represent an inherent spatial property of sound available in any acoustic scene. However, previous DoA-based methods rely on hand-crafted features or discrete encodings, which lose fine-grained spatial information and limit adaptability. We propose SoundCompass, an effective directional clue integration framework centered on a Spectral Pairwise INteraction (SPIN) module that captures cross-channel spatial correlations in the complex spectrogram domain to preserve full spatial information in multichannel signals. The input feature expressed in terms of spatial correlations is fused with a DoA clue represented as spherical harmonics (SH) encoding. The fusion is carried out across overlapping frequency subbands, inheriting the benefits reported in the previous band-split architectures. We also incorporate the iterative refinement strategy, chain-of-inference (CoI), in the TSE framework, which recursively fuses DoA with sound event activation estimated from the previous inference stage. Experiments demonstrate that SoundCompass, combining SPIN, SH embedding, and CoI, robustly extracts target sources across diverse signal classes and spatial configurations.", "date": "2025-09-23", "categories": []}
{"id": "2510.01812", "title": "SingMOS-Pro: An Comprehensive Benchmark for Singing Quality Assessment", "abstract": "Singing voice generation progresses rapidly, yet evaluating singing quality remains a critical challenge. Human subjective assessment, typically in the form of listening tests, is costly and time consuming, while existing objective metrics capture only limited perceptual aspects. In this work, we introduce SingMOS-Pro, a dataset for automatic singing quality assessment. Building on our preview version SingMOS, which provides only overall ratings, SingMOS-Pro extends the annotations of the additional data to include lyrics, melody, and overall quality, offering broader coverage and greater diversity. The dataset contains 7,981 singing clips generated by 41 models across 12 datasets, spanning from early systems to recent state-of-the-art approaches. Each clip is rated by at least five experienced annotators to ensure reliability and consistency. Furthermore, we investigate strategies for effectively utilizing MOS data annotated under heterogeneous standards and benchmark several widely used evaluation methods from related tasks on SingMOS-Pro, establishing strong baselines and practical references for future research. The dataset is publicly available at https://huggingface.co/datasets/TangRain/SingMOS-Pro.", "date": "2025-10-02", "categories": []}
{"id": "2601.18908", "title": "Enhancing Speech Emotion Recognition using Dynamic Spectral Features and Kalman Smoothing", "abstract": "Speech Emotion Recognition systems often use static features like Mel-Frequency Cepstral Coefficients (MFCCs), Zero Crossing Rate (ZCR), and Root Mean Square Energy (RMSE). Because of this, they can misclassify emotions when there is acoustic noise in vocal signals. To address this, we added dynamic features using Dynamic Spectral features (Deltas and Delta-Deltas) along with the Kalman Smoothing algorithm. This approach reduces noise and improves emotion classification. Since emotion changes over time, the Kalman Smoothing filter also helped make the classifier outputs more stable. Tests on the RAVDESS dataset showed that this method achieved a state-of-the-art accuracy of 87\\% and reduced misclassification between emotions with similar acoustic features", "date": "2026-01-26", "categories": []}
{"id": "2601.19194", "title": "SE-DiCoW: Self-Enrolled Diarization-Conditioned Whisper", "abstract": "Speaker-attributed automatic speech recognition (ASR) in multi-speaker environments remains a major challenge. While some approaches achieve strong performance when fine-tuned on specific domains, few systems generalize well across out-of-domain datasets. Our prior work, Diarization-Conditioned Whisper (DiCoW), leverages speaker diarization outputs as conditioning information and, with minimal fine-tuning, demonstrated strong multilingual and multi-domain performance. In this paper, we address a key limitation of DiCoW: ambiguity in Silence-Target-Non-target-Overlap (STNO) masks, where two or more fully overlapping speakers may have nearly identical conditioning despite differing transcriptions. We introduce SE-DiCoW (Self-Enrolled Diarization-Conditioned Whisper), which uses diarization output to locate an enrollment segment anywhere in the conversation where the target speaker is most active. This enrollment segment is used as fixed conditioning via cross-attention at each encoder layer. We further refine DiCoW with improved data segmentation, model initialization, and augmentation. Together, these advances yield substantial gains: SE-DiCoW reduces macro-averaged tcpWER by 52.4% relative to the original DiCoW on the EMMA MT-ASR benchmark.", "date": "2026-01-27", "categories": []}
{"id": "2601.19702", "title": "SAM Audio Judge: A Unified Multimodal Framework for Perceptual Evaluation of Audio Separation", "abstract": "The performance evaluation remains a complex challenge in audio separation, and existing evaluation metrics are often misaligned with human perception, course-grained, relying on ground truth signals. On the other hand, subjective listening tests remain the gold standard for real-world evaluation, but they are expensive, time-consuming, and difficult to scale. This paper addresses the growing need for automated systems capable of evaluating audio separation without human intervention. The proposed evaluation metric, SAM Audio Judge (SAJ), is a multimodal fine-grained reference-free objective metric, which shows highly alignment with human perceptions. SAJ supports three audio domains (speech, music and general sound events) and three prompt inputs (text, visual and span), covering four different dimensions of evaluation (recall, percision, faithfulness, and overall). SAM Audio Judge also shows potential applications in data filtering, pseudo-labeling large datasets and reranking in audio separation models. We release our code and pre-trained models at: https://github.com/facebookresearch/sam-audio.", "date": "2026-01-27", "categories": []}
{"id": "2403.20130", "title": "Sound event localization and classification using WASN in Outdoor Environment", "abstract": "Deep learning-based sound event localization and classification is an emerging research area within wireless acoustic sensor networks. However, current methods for sound event localization and classification typically rely on a single microphone array, making them susceptible to signal attenuation and environmental noise, which limits their monitoring range. Moreover, methods using multiple microphone arrays often focus solely on source localization, neglecting the aspect of sound event classification. In this paper, we propose a deep learning-based method that employs multiple features and attention mechanisms to estimate the location and class of sound source. We introduce a Soundmap feature to capture spatial information across multiple frequency bands. We also use the Gammatone filter to generate acoustic features more suitable for outdoor environments. Furthermore, we integrate attention mechanisms to learn channel-wise relationships and temporal dependencies within the acoustic features. To evaluate our proposed method, we conduct experiments using simulated datasets with different levels of noise and size of monitoring areas, as well as different arrays and source positions. The experimental results demonstrate the superiority of our proposed method over state-of-the-art methods in both sound event classification and sound source localization tasks. And we provide further analysis to explain the reasons for the observed errors.", "date": "2024-03-29", "categories": []}
{"id": "2601.17080", "title": "PC-MCL: Patient-Consistent Multi-Cycle Learning with multi-label bias correction for respiratory sound classification", "abstract": "Automated respiratory sound classification supports the diagnosis of pulmonary diseases. However, many deep models still rely on cycle-level analysis and suffer from patient-specific overfitting. We propose PC-MCL (Patient-Consistent Multi-Cycle Learning) to address these limitations by utilizing three key components: multi-cycle concatenation, a 3-label formulation, and a patient-matching auxiliary task. Our work resolves a multi-label distributional bias in respiratory sound classification, a critical issue inherent to applying multi-cycle concatenation with the conventional 2-label formulation (crackle, wheeze). This bias manifests as a systematic loss of normal signal information when normal and abnormal cycles are combined. Our proposed 3-label formulation (normal, crackle, wheeze) corrects this by preserving information from all constituent cycles in mixed samples. Furthermore, the patient-matching auxiliary task acts as a multi-task regularizer, encouraging the model to learn more robust features and improving generalization. On the ICBHI 2017 benchmark, PC-MCL achieves an ICBHI Score of 65.37%, outperforming existing baselines. Ablation studies confirm that all three components are essential, working synergistically to improve the detection of abnormal respiratory events.", "date": "2026-01-23", "categories": []}
{"id": "2601.17611", "title": "ToS: A Team of Specialists ensemble framework for Stereo Sound Event Localization and Detection with distance estimation in Video", "abstract": "Sound event localization and detection with distance estimation (3D SELD) in video involves identifying active sound events at each time frame while estimating their spatial coordinates. This multimodal task requires joint reasoning across semantic, spatial, and temporal dimensions, a challenge that single models often struggle to address effectively. To tackle this, we introduce the Team of Specialists (ToS) ensemble framework, which integrates three complementary sub-networks: a spatio-linguistic model, a spatio-temporal model, and a tempo-linguistic model. Each sub-network specializes in a unique pair of dimensions, contributing distinct insights to the final prediction, akin to a collaborative team with diverse expertise. ToS has been benchmarked against state-of-the-art audio-visual models for 3D SELD on the DCASE2025 Task 3 Stereo SELD development set, consistently outperforming existing methods across key metrics. Future work will extend this proof of concept by strengthening the specialists with appropriate tasks, training, and pre-training curricula.", "date": "2026-01-24", "categories": []}
{"id": "2601.17679", "title": "BanglaRobustNet: A Hybrid Denoising-Attention Architecture for Robust Bangla Speech Recognition", "abstract": "Bangla, one of the most widely spoken languages, remains underrepresented in state-of-the-art automatic speech recognition (ASR) research, particularly under noisy and speaker-diverse conditions. This paper presents BanglaRobustNet, a hybrid denoising-attention framework built on Wav2Vec-BERT, designed to address these challenges. The architecture integrates a diffusion-based denoising module to suppress environmental noise while preserving Bangla-specific phonetic cues, and a contextual cross-attention module that conditions recognition on speaker embeddings for robustness across gender, age, and dialects. Trained end-to-end with a composite objective combining CTC loss, phonetic consistency, and speaker alignment, BanglaRobustNet achieves substantial reductions in word error rate (WER) and character error rate (CER) compared to Wav2Vec-BERT and Whisper baselines. Evaluations on Mozilla Common Voice Bangla and augmented noisy speech confirm the effectiveness of our approach, establishing BanglaRobustNet as a robust ASR system tailored to low-resource, noise-prone linguistic settings.", "date": "2026-01-25", "categories": []}
{"id": "2601.17690", "title": "Segment Length Matters: A Study of Segment Lengths on Audio Fingerprinting Performance", "abstract": "Audio fingerprinting provides an identifiable representation of acoustic signals, which can be later used for identification and retrieval systems. To obtain a discriminative representation, the input audio is usually segmented into shorter time intervals, allowing local acoustic features to be extracted and analyzed. Modern neural approaches typically operate on short, fixed-duration audio segments, yet the choice of segment duration is often made heuristically and rarely examined in depth. In this paper, we study how segment length affects audio fingerprinting performance. We extend an existing neural fingerprinting architecture to adopt various segment lengths and evaluate retrieval accuracy across different segment lengths and query durations. Our results show that short segment lengths (0.5-second) generally achieve better performance. Moreover, we evaluate LLM capacity in recommending the best segment length, which shows that GPT-5-mini consistently gives the best suggestions across five considerations among three studied LLMs. Our findings provide practical guidance for selecting segment duration in large-scale neural audio retrieval systems.", "date": "2026-01-25", "categories": []}
{"id": "2601.17711", "title": "CaSNet: Compress-and-Send Network Based Multi-Device Speech Enhancement Model for Distributed Microphone Arrays", "abstract": "Distributed microphone array (DMA) is a promising next-generation platform for speech interaction, where speech enhancement (SE) is still required to improve the speech quality in noisy cases. Existing SE methods usually first gather raw waveforms at a fusion center (FC) from all devices and then design a multi-microphone model, causing high bandwidth and energy costs. In this work, we propose a \\emph{Compress-and-Send Network (CaSNet)} for resource-constrained DMAs, where one microphone serves as the FC and reference. Each of other devices encodes the measured raw data into a feature matrix, which is then compressed by singular value decomposition (SVD) to produce a more compact representation. The received features at the FC are aligned via cross window query with respect to the reference, followed by neural decoding to yield spatially coherent enhanced speech. Experiments on multiple datasets show that the proposed CaSNet can save the data amount with a negligible impact on the performance compared to the uncompressed case. The reproducible code is available at https://github.com/Jokejiangv/CaSNet.", "date": "2026-01-25", "categories": []}
{"id": "2601.18335", "title": "Analytic Incremental Learning For Sound Source Localization With Imbalance Rectification", "abstract": "Sound source localization (SSL) demonstrates remarkable results in controlled settings but struggles in real-world deployment due to dual imbalance challenges: intra-task imbalance arising from long-tailed direction-of-arrival (DoA) distributions, and inter-task imbalance induced by cross-task skews and overlaps. These often lead to catastrophic forgetting, significantly degrading the localization accuracy. To mitigate these issues, we propose a unified framework with two key innovations. Specifically, we design a GCC-PHAT-based data augmentation (GDA) method that leverages peak characteristics to alleviate intra-task distribution skews. We also propose an Analytic dynamic imbalance rectifier (ADIR) with task-adaption regularization, which enables analytic updates that adapt to inter-task dynamics. On the SSLR benchmark, our proposal achieves state-of-the-art (SoTA) results of 89.0% accuracy, 5.3{\\deg} mean absolute error, and 1.6 backward transfer, demonstrating robustness to evolving imbalances without exemplar storage.", "date": "2026-01-26", "categories": []}
{"id": "2601.18766", "title": "Learning to Discover: A Generalized Framework for Raga Identification without Forgetting", "abstract": "Raga identification in Indian Art Music (IAM) remains challenging due to the presence of numerous rarely performed Ragas that are not represented in available training datasets. Traditional classification models struggle in this setting, as they assume a closed set of known categories and therefore fail to recognise or meaningfully group previously unseen Ragas. Recent works have tried categorizing unseen Ragas, but they run into a problem of catastrophic forgetting, where the knowledge of previously seen Ragas is diminished. To address this problem, we adopt a unified learning framework that leverages both labeled and unlabeled audio, enabling the model to discover coherent categories corresponding to the unseen Ragas, while retaining the knowledge of previously known ones. We test our model on benchmark Raga Identification datasets and demonstrate its performance in categorizing previously seen, unseen, and all Raga classes. The proposed approach surpasses the previous NCD-based pipeline even in discovering the unseen Raga categories, offering new insights into representation learning for IAM tasks.", "date": "2026-01-26", "categories": []}
{"id": "2504.17586", "title": "A Machine Learning Approach for Denoising and Upsampling HRTFs", "abstract": "The demand for realistic virtual immersive audio continues to grow, with Head-Related Transfer Functions (HRTFs) playing a key role. HRTFs capture how sound reaches our ears, reflecting unique anatomical features and enhancing spatial perception. It has been shown that personalized HRTFs improve localization accuracy, but their measurement remains time-consuming and requires a noise-free environment. Although machine learning has been shown to reduce the required measurement points and, thus, the measurement time, a controlled environment is still necessary. This paper proposes a method to address this constraint by presenting a novel technique that can upsample sparse, noisy HRTF measurements. The proposed approach combines an HRTF Denoisy U-Net for denoising and an Autoencoding Generative Adversarial Network (AE-GAN) for upsampling from three measurement points. The proposed method achieves a log-spectral distortion (LSD) error of 5.41 dB and a cosine similarity loss of 0.0070, demonstrating the method's effectiveness in HRTF upsampling.", "date": "2025-04-24", "categories": []}
{"id": "2509.21463", "title": "Enhanced Generative Machine Listener", "abstract": "We present GMLv2, a reference-based model designed for the prediction of subjective audio quality as measured by MUSHRA scores. GMLv2 introduces a Beta distribution-based loss to model the listener ratings and incorporates additional neural audio coding (NAC) subjective datasets to extend its generalization and applicability. Extensive evaluations on diverse testset demonstrate that proposed GMLv2 consistently outperforms widely used metrics, such as PEAQ and ViSQOL, both in terms of correlation with subjective scores and in reliably predicting these scores across diverse content types and codec configurations. Consequently, GMLv2 offers a scalable and automated framework for perceptual audio quality evaluation, poised to accelerate research and development in modern audio coding technologies.", "date": "2025-09-25", "categories": []}
{"id": "2601.16675", "title": "I Guess That's Why They Call it the Blues: Causal Analysis for Audio Classifiers", "abstract": "It is well-known that audio classifiers often rely on non-musically relevant features and spurious correlations to classify audio. Hence audio classifiers are easy to manipulate or confuse, resulting in wrong classifications. While inducing a misclassification is not hard, until now the set of features that the classifiers rely on was not well understood.\n  In this paper we introduce a new method that uses causal reasoning to discover features of the frequency space that are sufficient and necessary for a given classification. We describe an implementation of this algorithm in the tool FreqReX and provide experimental results on a number of standard benchmark datasets. Our experiments show that causally sufficient and necessary subsets allow us to manipulate the outputs of the models in a variety of ways by changing the input very slightly. Namely, a change to one out of 240,000 frequencies results in a change in classification 58% of the time, and the change can be so small that it is practically inaudible. These results show that causal analysis is useful for understanding the reasoning process of audio classifiers and can be used to successfully manipulate their outputs.", "date": "2026-01-23", "categories": []}
{"id": "2509.07526", "title": "Competitive Audio-Language Models with Data-Efficient Single-Stage Training on Public Data", "abstract": "Large language models (LLMs) have transformed NLP, yet their integration with audio remains underexplored despite audio's centrality to human communication. We introduce Falcon3-Audio, a family of Audio-Language Models (ALMs) built on instruction-tuned LLMs and Whisper encoders. Using a remarkably small amount of public audio data, less than 30K hours (5K unique), Falcon3-Audio-7B matches the best reported performance among open-weight models on the MMAU benchmark, with a score of 64.14, matching R1-AQA, while distinguishing itself through superior data and parameter efficiency, single-stage training, and transparency. Notably, our smallest 1B model remains competitive with larger open models ranging from 2B to 13B parameters. Through extensive ablations, we find that common complexities such as curriculum learning, multiple audio encoders, and intricate cross-attention connectors are not required for strong performance, even compared to models trained on over 500K hours of data.", "date": "2025-09-09", "categories": []}
{"id": "2511.13273", "title": "AudioMotionBench: Evaluating Auditory Motion Perception in Audio LLMs", "abstract": "Large Audio-Language Models (LALMs) have recently shown impressive progress in speech recognition, audio captioning, and auditory question answering. Yet, whether these models can perceive spatial dynamics, particularly the motion of sound sources, remains unclear. In this work, we uncover a systematic motion perception deficit in current ALLMs. To investigate this issue, we introduce AudioMotionBench, the first benchmark explicitly designed to evaluate auditory motion understanding. AudioMotionBench introduces a controlled question-answering benchmark designed to evaluate whether Audio-Language Models (LALMs) can infer the direction and trajectory of moving sound sources from binaural audio. Comprehensive quantitative and qualitative analyses reveal that current models struggle to reliably recognize motion cues or distinguish directional patterns. The average accuracy remains below 50\\%, underscoring a fundamental limitation in auditory spatial reasoning. Our study highlights a fundamental gap between human and model auditory spatial reasoning, providing both a diagnostic tool and new insight for enhancing spatial cognition in future Audio-Language Models.", "date": "2025-11-17", "categories": []}
{"id": "2601.15118", "title": "WavLink: Compact Audio-Text Embeddings with a Global Whisper Token", "abstract": "Whisper has become the de-facto encoder for extracting general-purpose audio features in large audio-language models, where a 30-second clip is typically represented by 1500 frame features projected into an LLM. In contrast, audio-text embedding models like CLAP-based models have largely relied on alternative audio encoders (e.g., HTS-AT, PaSST), and have not leveraged Whisper effectively. We present WavLink, a compact audio-text embedding model that augments Whisper encoder with a learnable global token, trained jointly with a text encoder. Through a systematic study of design choices, including pretrained text encoders, loss functions, training modes, and data mixtures, we identify configurations that yield state-of-the-art retrieval performance. Our two-stage training recipe across three model sizes, combined with Matryoshka-style supervision, improves scalability, enabling 8x smaller embeddings with minimal performance drop. WavLink also demonstrates competitive performance on AIR-Bench with MCQs and zero-shot classification.", "date": "2026-01-21", "categories": []}
{"id": "2601.15676", "title": "Bridging the Perception Gap: A Lightweight Coarse-to-Fine Architecture for Edge Audio Systems", "abstract": "Deploying Audio-Language Models (Audio-LLMs) on edge infrastructure exposes a persistent tension between perception depth and computational efficiency. Lightweight local models tend to produce passive perception - generic summaries that miss the subtle evidence required for multi-step audio reasoning - while indiscriminate cloud offloading incurs unacceptable latency, bandwidth cost, and privacy risk. We propose CoFi-Agent (Tool-Augmented Coarse-to-Fine Agent), a hybrid architecture targeting edge servers and gateways. It performs fast local perception and triggers conditional forensic refinement only when uncertainty is detected. CoFi-Agent runs an initial single-pass on a local 7B Audio-LLM, then a cloud controller gates difficult cases and issues lightweight plans for on-device tools such as temporal re-listening and local ASR. On the MMAR benchmark, CoFi-Agent improves accuracy from 27.20% to 53.60%, while achieving a better accuracy-efficiency trade-off than an always-on investigation pipeline. Overall, CoFi-Agent bridges the perception gap via tool-enabled, conditional edge-cloud collaboration under practical system constraints.", "date": "2026-01-22", "categories": []}
{"id": "2505.24486", "title": "Rehearsal with Auxiliary-Informed Sampling for Audio Deepfake Detection", "abstract": "The performance of existing audio deepfake detection frameworks degrades when confronted with new deepfake attacks. Rehearsal-based continual learning (CL), which updates models using a limited set of old data samples, helps preserve prior knowledge while incorporating new information. However, existing rehearsal techniques don't effectively capture the diversity of audio characteristics, introducing bias and increasing the risk of forgetting. To address this challenge, we propose Rehearsal with Auxiliary-Informed Sampling (RAIS), a rehearsal-based CL approach for audio deepfake detection. RAIS employs a label generation network to produce auxiliary labels, guiding diverse sample selection for the memory buffer. Extensive experiments show RAIS outperforms state-of-the-art methods, achieving an average Equal Error Rate (EER) of 1.953 % across five experiences. The code is available at: https://github.com/falihgoz/RAIS.", "date": "2025-05-30", "categories": []}
{"id": "2510.15231", "title": "Extending Audio Context for Long-Form Understanding in Large Audio-Language Models", "abstract": "Large Audio-Language Models (LALMs) are often constrained by short audio context windows, even when their text backbones support long contexts, limiting long-form audio understanding. Prior work has introduced context-extension methods (e.g. YaRN) on unimodal LLMs, yet their application to LALMs remains unexplored. First, building on RoPE-based context extension, we introduce Partial YaRN, a training-free, modality-decoupled extension method that modifies only audio token positions, leaving text positions intact to preserve the base LLM's text capabilities. Second, we propose Virtual Longform Audio Training (VLAT), a training strategy that extends Partial YaRN into a training-time positional augmentation. VLAT simulates diverse audio lengths during training, enabling generalization to inputs far longer than those seen in training. Our experiments on SALMONN and Qwen2-Audio confirm that Partial YaRN outperforms the original models across wide range of settings, and VLAT provides substantial performance improvement on long audio of unseen lengths.", "date": "2025-10-17", "categories": []}
{"id": "2601.14304", "title": "Guided by the Plan: Enhancing Faithful Autoregressive Text-to-Audio Generation with Guided Decoding", "abstract": "Autoregressive (AR) models excel at generating temporally coherent audio by producing tokens sequentially, yet they often falter in faithfully following complex textual prompts, especially those describing complex sound events. We uncover a surprising capability in AR audio generators: their early prefix tokens implicitly encode global semantic attributes of the final output, such as event count and sound-object category, revealing a form of implicit planning. Building on this insight, we propose Plan-Critic, a lightweight auxiliary model trained with a Generalized Advantage Estimation (GAE)-inspired objective to predict final instruction-following quality from partial generations. At inference time, Plan-Critic enables guided exploration: it evaluates candidate prefixes early, prunes low-fidelity trajectories, and reallocates computation to high-potential planning seeds. Our Plan-Critic-guided sampling achieves up to a 10-point improvement in CLAP score over the AR baseline-establishing a new state of the art in AR text-to-audio generation-while maintaining computational parity with standard best-of-N decoding. This work bridges the gap between causal generation and global semantic alignment, demonstrating that even strictly autoregressive models can plan ahead.", "date": "2026-01-18", "categories": []}
{"id": "2601.15083", "title": "Bangla Music Genre Classification Using Bidirectional LSTMS", "abstract": "Bangla music is enrich in its own music cultures. Now a days music genre classification is very significant because of the exponential increase in available music, both in digital and physical formats. It is necessary to index them accordingly to facilitate improved retrieval. Automatically classifying Bangla music by genre is essential for efficiently locating specific pieces within a vast and diverse music library. Prevailing methods for genre classification predominantly employ conventional machine learning or deep learning approaches. This work introduces a novel music dataset comprising ten distinct genres of Bangla music. For the task of audio classification, we utilize a recurrent neural network (RNN) architecture. Specifically, a Long Short-Term Memory (LSTM) network is implemented to train the model and perform the classification. Feature extraction represents a foundational stage in audio data processing. This study utilizes Mel-Frequency Cepstral Coefficients (MFCCs) to transform raw audio waveforms into a compact and representative set of features. The proposed framework facilitates music genre classification by leveraging these extracted features. Experimental results demonstrate a classification accuracy of 78%, indicating the system's strong potential to enhance and streamline the organization of Bangla music genres.", "date": "2026-01-21", "categories": []}
{"id": "2208.03084", "title": "Deep Feature Learning for Medical Acoustics", "abstract": "The purpose of this paper is to compare different learnable frontends in medical acoustics tasks. A framework has been implemented to classify human respiratory sounds and heartbeats in two categories, i.e. healthy or affected by pathologies. After obtaining two suitable datasets, we proceeded to classify the sounds using two learnable state-of-art frontends -- LEAF and nnAudio -- plus a non-learnable baseline frontend, i.e. Mel-filterbanks. The computed features are then fed into two different CNN models, namely VGG16 and EfficientNet. The frontends are carefully benchmarked in terms of the number of parameters, computational resources, and effectiveness.\n  This work demonstrates how the integration of learnable frontends in neural audio classification systems may improve performance, especially in the field of medical acoustics. However, the usage of such frameworks makes the needed amount of data even larger. Consequently, they are useful if the amount of data available for training is adequately large to assist the feature learning process.", "date": "2022-08-05", "categories": []}
{"id": "2506.00462", "title": "XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark", "abstract": "Recent advances in audio generation led to an increasing number of deepfakes, making the general public more vulnerable to financial scams, identity theft, and misinformation. Audio deepfake detectors promise to alleviate this issue, with many recent studies reporting accuracy rates close to 99%. However, these methods are typically tested in an in-domain setup, where the deepfake samples from the training and test sets are produced by the same generative models. To this end, we introduce XMAD-Bench, a large-scale cross-domain multilingual audio deepfake benchmark comprising 668.8 hours of real and deepfake speech. In our novel dataset, the speakers, the generative methods, and the real audio sources are distinct across training and test splits. This leads to a challenging cross-domain evaluation setup, where audio deepfake detectors can be tested \"in the wild\". Our in-domain and cross-domain experiments indicate a clear disparity between the in-domain performance of deepfake detectors, which is usually as high as 100%, and the cross-domain performance of the same models, which is sometimes similar to random chance. Our benchmark highlights the need for the development of robust audio deepfake detectors, which maintain their generalization capacity across different languages, speakers, generative methods, and data sources. Our benchmark is publicly released at https://github.com/ristea/xmad-bench/.", "date": "2025-05-31", "categories": []}
{"id": "2506.11350", "title": "GLAP: General contrastive audio-text pretraining across domains and languages", "abstract": "Contrastive Language Audio Pretraining (CLAP) is a widely-used method to bridge the gap between audio and text domains. Current CLAP methods enable sound and music retrieval in English, ignoring multilingual spoken content. To address this, we introduce general language audio pretraining (GLAP), which expands CLAP with multilingual and multi-domain abilities. GLAP demonstrates its versatility by achieving competitive performance on standard audio-text retrieval benchmarks like Clotho and AudioCaps, while significantly surpassing existing methods in speech retrieval and classification tasks. Additionally, GLAP achieves strong results on widely used sound-event zero-shot benchmarks, while simultaneously outperforming previous methods on speech content benchmarks. Further keyword spotting evaluations across 50 languages emphasize GLAP's advanced multilingual capabilities. Finally, multilingual sound and music understanding is evaluated across four languages. Checkpoints and Source: https://github.com/xiaomi-research/dasheng-glap.", "date": "2025-06-12", "categories": []}
{"id": "2509.16715", "title": "QASTAnet: A DNN-based Quality Metric for Spatial Audio", "abstract": "In the development of spatial audio technologies, reliable and shared methods for evaluating audio quality are essential. Listening tests are currently the standard but remain costly in terms of time and resources. Several models predicting subjective scores have been proposed, but they do not generalize well to real-world signals. In this paper, we propose QASTAnet (Quality Assessment for SpaTial Audio network), a new metric based on a deep neural network, specialized on spatial audio (ambisonics and binaural). As training data is scarce, we aim for the model to be trainable with a small amount of data. To do so, we propose to rely on expert modeling of the low-level auditory system and use a neurnal network to model the high-level cognitive function of the quality judgement. We compare its performance to two reference metrics on a wide range of content types (speech, music, ambiance, anechoic, reverberated) and focusing on codec artifacts. Results demonstrate that QASTAnet overcomes the aforementioned limitations of the existing methods. The strong correlation between the proposed metric prediction and subjective scores makes it a good candidate for comparing codecs in their development.", "date": "2025-09-20", "categories": []}
{"id": "2601.11768", "title": "Lightweight Self-Supervised Detection of Fundamental Frequency and Accurate Probability of Voicing in Monophonic Music", "abstract": "Reliable fundamental frequency (F 0) and voicing estimation is essential for neural synthesis, yet many pitch extractors depend on large labeled corpora and degrade under realistic recording artifacts. We propose a lightweight, fully self-supervised framework for joint F 0 estimation and voicing inference, designed for rapid single-instrument training from limited audio. Using transposition-equivariant learning on CQT features, we introduce an EM-style iterative reweighting scheme that uses Shift Cross-Entropy (SCE) consistency as a reliability signal to suppress uninformative noisy/unvoiced frames. The resulting weights provide confidence scores that enable pseudo-labeling for a separate lightweight voicing classifier without manual annotations. Trained on MedleyDB and evaluated on MDB-stem-synth ground truth, our method achieves competitive cross-corpus performance (RPA 95.84, RCA 96.24) and demonstrates cross-instrument generalization.", "date": "2026-01-16", "categories": []}
{"id": "2601.11846", "title": "The Third VoicePrivacy Challenge: Preserving Emotional Expressiveness and Linguistic Content in Voice Anonymization", "abstract": "We present results and analyses from the third VoicePrivacy Challenge held in 2024, which focuses on advancing voice anonymization technologies. The task was to develop a voice anonymization system for speech data that conceals a speaker's voice identity while preserving linguistic content and emotional state. We provide a systematic overview of the challenge framework, including detailed descriptions of the anonymization task and datasets used for both system development and evaluation. We outline the attack model and objective evaluation metrics for assessing privacy protection (concealing speaker voice identity) and utility (content and emotional state preservation). We describe six baseline anonymization systems and summarize the innovative approaches developed by challenge participants. Finally, we provide key insights and observations to guide the design of future VoicePrivacy challenges and identify promising directions for voice anonymization research.", "date": "2026-01-17", "categories": []}
{"id": "2601.12205", "title": "Do Neural Codecs Generalize? A Controlled Study Across Unseen Languages and Non-Speech Tasks", "abstract": "This paper investigates three crucial yet underexplored aspects of the generalization capabilities of neural audio codecs (NACs): (i) whether NACs can generalize to unseen languages during pre-training, (ii) whether speech-only pre-trained NACs can effectively generalize to non-speech applications such as environmental sounds, music, and animal vocalizations, and (iii) whether incorporating non-speech data during pre-training can improve performance on both speech and non-speech tasks. Existing studies typically rely on off-the-shelf NACs for comparison, which limits insight due to variations in implementation. In this work, we train NACs from scratch using strictly controlled configurations and carefully curated pre-training data to enable fair comparisons. We conduct a comprehensive evaluation of NAC performance on both signal reconstruction quality and downstream applications using 11 metrics. Our results show that NACs can generalize to unseen languages during pre-training, speech-only pre-trained NACs exhibit degraded performance on non-speech tasks, and incorporating non-speech data during pre-training improves performance on non-speech tasks while maintaining comparable performance on speech tasks.", "date": "2026-01-18", "categories": []}
{"id": "2601.12354", "title": "Bone-conduction Guided Multimodal Speech Enhancement with Conditional Diffusion Models", "abstract": "Single-channel speech enhancement models face significant performance degradation in extremely noisy environments. While prior work has shown that complementary bone-conducted speech can guide enhancement, effective integration of this noise-immune modality remains a challenge. This paper introduces a novel multimodal speech enhancement framework that integrates bone-conduction sensors with air-conducted microphones using a conditional diffusion model. Our proposed model significantly outperforms previously established multimodal techniques and a powerful diffusion-based single-modal baseline across a wide range of acoustic conditions.", "date": "2026-01-18", "categories": []}
{"id": "2601.12594", "title": "SLAP: Scalable Language-Audio Pretraining with Variable-Duration Audio and Multi-Objective Training", "abstract": "Contrastive language-audio pretraining (CLAP) has achieved notable success in learning semantically rich audio representations and is widely adopted for various audio-related tasks. However, current CLAP models face several key limitations. First, they are typically trained on relatively small datasets, often comprising a few million audio samples. Second, existing CLAP models are restricted to short and fixed duration, which constrains their usage in real-world scenarios with variable-duration audio. Third, the standard contrastive training objective operates on global representations, which may hinder the learning of dense, fine-grained audio features. To address these challenges, we introduce Scalable Language-Audio Pretraining (SLAP), which scales language-audio pretraining to 109 million audio-text pairs with variable audio durations and incorporates multiple training objectives. SLAP unifies contrastive loss with additional self-supervised and captioning losses in a single-stage training, facilitating the learning of richer dense audio representations. The proposed SLAP model achieves new state-of-the-art performance on audio-text retrieval and zero-shot audio classification tasks, demonstrating its effectiveness across diverse benchmarks.", "date": "2026-01-18", "categories": []}
{"id": "2601.12660", "title": "Toward Faithful Explanations in Acoustic Anomaly Detection", "abstract": "Interpretability is essential for user trust in real-world anomaly detection applications. However, deep learning models, despite their strong performance, often lack transparency. In this work, we study the interpretability of autoencoder-based models for audio anomaly detection, by comparing a standard autoencoder (AE) with a mask autoencoder (MAE) in terms of detection performance and interpretability. We applied several attribution methods, including error maps, saliency maps, SmoothGrad, Integrated Gradients, GradSHAP, and Grad-CAM. Although MAE shows a slightly lower detection, it consistently provides more faithful and temporally precise explanations, suggesting a better alignment with true anomalies. To assess the relevance of the regions highlighted by the explanation method, we propose a perturbation-based faithfulness metric that replaces them with their reconstructions to simulate normal input. Our findings, based on experiments in a real industrial scenario, highlight the importance of incorporating interpretability into anomaly detection pipelines and show that masked training improves explanation quality without compromising performance.", "date": "2026-01-19", "categories": []}
{"id": "2601.13464", "title": "Context and Transcripts Improve Detection of Deepfake Audios of Public Figures", "abstract": "Humans use context to assess the veracity of information. However, current audio deepfake detectors only analyze the audio file without considering either context or transcripts. We create and analyze a Journalist-provided Deepfake Dataset (JDD) of 255 public deepfakes which were primarily contributed by over 70 journalists since early 2024. We also generate a synthetic audio dataset (SYN) of dead public figures and propose a novel Context-based Audio Deepfake Detector (CADD) architecture. In addition, we evaluate performance on two large-scale datasets: ITW and P$^2$V. We show that sufficient context and/or the transcript can significantly improve the efficacy of audio deepfake detectors. Performance (measured via F1 score, AUC, and EER) of multiple baseline audio deepfake detectors and traditional classifiers can be improved by 5%-37.58% in F1-score, 3.77%-42.79% in AUC, and 6.17%-47.83% in EER. We additionally show that CADD, via its use of context and/or transcripts, is more robust to 5 adversarial evasion strategies, limiting performance degradation to an average of just -0.71% across all experiments. Code, models, and datasets are available at our project page: https://sites.northwestern.edu/nsail/cadd-context-based-audio-deepfake-detection (access restricted during review).", "date": "2026-01-19", "categories": []}
{"id": "2601.13647", "title": "Fusion Segment Transformer: Bi-Directional Attention Guided Fusion Network for AI-Generated Music Detection", "abstract": "With the rise of generative AI technology, anyone can now easily create and deploy AI-generated music, which has heightened the need for technical solutions to address copyright and ownership issues. While existing works mainly focused on short-audio, the challenge of full-audio detection, which requires modeling long-term structure and context, remains insufficiently explored. To address this, we propose an improved version of the Segment Transformer, termed the Fusion Segment Transformer. As in our previous work, we extract content embeddings from short music segments using diverse feature extractors. Furthermore, we enhance the architecture for full-audio AI-generated music detection by introducing a Gated Fusion Layer that effectively integrates content and structural information, enabling the capture of long-term context. Experiments on the SONICS and AIME datasets show that our approach outperforms the previous model and recent baselines, achieving state-of-the-art results in AI-generated music detection.", "date": "2026-01-20", "categories": []}
{"id": "2601.14012", "title": "MATE: Matryoshka Audio-Text Embeddings for Open-Vocabulary Keyword Spotting", "abstract": "Open-vocabulary keyword spotting (KWS) with text-based enrollment has emerged as a flexible alternative to fixed-phrase triggers. Prior utterance-level matching methods, from an embedding-learning standpoint, learn embeddings at a single fixed dimensionality. We depart from this design and propose Matryoshka Audio-Text Embeddings (MATE), a dual-encoder framework that encodes multiple embedding granularities within a single vector via nested sub-embeddings (\"prefixes\"). Specifically, we introduce a PCA-guided prefix alignment: PCA-compressed versions of the full text embedding for each prefix size serve as teacher targets to align both audio and text prefixes. This alignment concentrates salient keyword cues in lower-dimensional prefixes, while higher dimensions add detail. MATE is trained with standard deep metric learning objectives for audio-text KWS, and is loss-agnostic. To our knowledge, this is the first application of matryoshka-style embeddings to KWS, achieving state-of-the-art results on WSJ and LibriPhrase without any inference overhead.", "date": "2026-01-20", "categories": []}
{"id": "2601.11039", "title": "SonicBench: Dissecting the Physical Perception Bottleneck in Large Audio Language Models", "abstract": "Large Audio Language Models (LALMs) excel at semantic and paralinguistic tasks, yet their ability to perceive the fundamental physical attributes of audio such as pitch, loudness, and spatial location remains under-explored. To bridge this gap, we introduce SonicBench, a psychophysically grounded benchmark that systematically evaluates 12 core physical attributes across five perceptual dimensions. Unlike previous datasets, SonicBench uses a controllable generation toolbox to construct stimuli for two complementary paradigms: recognition (absolute judgment) and comparison (relative judgment). This design allows us to probe not only sensory precision but also relational reasoning capabilities, a domain where humans typically exhibit greater proficiency. Our evaluation reveals a substantial deficiency in LALMs' foundational auditory understanding; most models perform near random guessing and, contrary to human patterns, fail to show the expected advantage on comparison tasks. Furthermore, explicit reasoning yields minimal gains. However, our linear probing analysis demonstrates crucially that frozen audio encoders do successfully capture these physical cues (accuracy at least 60%), suggesting that the primary bottleneck lies in the alignment and decoding stages, where models fail to leverage the sensory signals they have already captured.", "date": "2026-01-16", "categories": []}
{"id": "2601.09603", "title": "Linear Complexity Self-Supervised Learning for Music Understanding with Random Quantizer", "abstract": "In recent years, foundation models have become very popular due to their exceptional performance, mainly in natural language (NLP) tasks where they were first introduced. These models usually consist of hundreds of millions, or even billions, of parameters, making them resource-intensive during training and in production systems, leading to increased costs. This paper focuses on the reduction of a foundation's model size when applied to music information retrieval (MIR) tasks. Our research combines the Branchformer architecture with SummaryMixing, which were first applied in speech recognition, along with a random quantization process. To facilitate reproducibility, we conduct pre-training on publicly available datasets, complemented by a proprietary dataset comparable in scale to other private datasets reported in the literature. We ensure robust evaluation by using a framework consisting of a variety of downstream MIR tasks. Our results show that our architecture achieves competitive performance when compared with other state-of-the-art models that use multi-head self-attention, while reducing the model size from 8.5% up to 12.3%.", "date": "2026-01-14", "categories": []}
{"id": "2601.07958", "title": "LJ-Spoof: A Generatively Varied Corpus for Audio Anti-Spoofing and Synthesis Source Tracing", "abstract": "Speaker-specific anti-spoofing and synthesis-source tracing are central challenges in audio anti-spoofing. Progress has been hampered by the lack of datasets that systematically vary model architectures, synthesis pipelines, and generative parameters. To address this gap, we introduce LJ-Spoof, a speaker-specific, generatively diverse corpus that systematically varies prosody, vocoders, generative hyperparameters, bona fide prompt sources, training regimes, and neural post-processing. The corpus spans one speakers-including studio-quality recordings-30 TTS families, 500 generatively variant subsets, 10 bona fide neural-processing variants, and more than 3 million utterances. This variation-dense design enables robust speaker-conditioned anti-spoofing and fine-grained synthesis-source tracing. We further position this dataset as both a practical reference training resource and a benchmark evaluation suite for anti-spoofing and source tracing.", "date": "2026-01-12", "categories": []}
{"id": "2505.20166", "title": "From Alignment to Advancement: Bootstrapping Audio-Language Alignment with Synthetic Data", "abstract": "Audio-aware large language models (ALLMs) have recently made great strides in understanding and processing audio inputs. These models are typically adapted from text-based large language models (LLMs) through additional training on audio-related tasks. This adaptation process presents two major limitations. First, ALLMs often suffer from catastrophic forgetting, where crucial textual capabilities like instruction-following are lost after training on audio data. In some cases, models may even hallucinate sounds that are not present in the input audio, raising concerns about reliability. Second, achieving cross-modal alignment between audio and language typically relies on large collections of task-specific question-answer pairs for instruction tuning, making it resource-intensive. To address these issues, previous works have leveraged the backbone LLMs to synthesize general-purpose, caption-style alignment data. In this paper, we propose a data generation framework that produces contrastive-like training data, designed to enhance ALLMs' ability to differentiate between present and absent sounds. We further extend our approach to multi-audio scenarios, enabling the model to either explain differences between audio inputs or produce unified captions that describe all inputs, thereby enhancing audio-language alignment. We refer to the entire ALLM training framework as bootstrapping audio-language alignment via synthetic data generation from backbone LLMs (BALSa). Experimental results indicate that our method effectively mitigates audio hallucinations while reliably maintaining strong performance on audio understanding and reasoning benchmarks, as well as instruction-following skills. Moreover, incorporating multi-audio training further enhances the model's comprehension and reasoning capabilities. Overall, BALSa offers an efficient and scalable approach to developing ALLMs.", "date": "2025-05-26", "categories": []}
{"id": "2507.04094", "title": "MMMOS: Multi-domain Multi-axis Audio Quality Assessment", "abstract": "Accurate audio quality estimation is essential for developing and evaluating audio generation, retrieval, and enhancement systems. Existing non-intrusive assessment models predict a single Mean Opinion Score (MOS) for speech, merging diverse perceptual factors and failing to generalize beyond speech. We propose MMMOS, a no-reference, multi-domain audio quality assessment system that estimates four orthogonal axes: Production Quality, Production Complexity, Content Enjoyment, and Content Usefulness across speech, music, and environmental sounds. MMMOS fuses frame-level embeddings from three pretrained encoders (WavLM, MuQ, and M2D) and evaluates three aggregation strategies with four loss functions. By ensembling the top eight models, MMMOS shows a 20-30% reduction in mean squared error and a 4-5% increase in Kendall's {\\tau} versus baseline, gains first place in six of eight Production Complexity metrics, and ranks among the top three on 17 of 32 challenge metrics.", "date": "2025-07-05", "categories": []}
{"id": "2601.07331", "title": "SEE: Signal Embedding Energy for Quantifying Noise Interference in Large Audio Language Models", "abstract": "Large Audio Language Models (LALMs) have been widely applied in real-time scenarios, such as in-car assistants and online meeting comprehension. In practice, audio inputs are often corrupted by device and environmental noise, leading to performance degradation. However, existing LALM studies on noise lack quantitative analysis and rely mainly on intuition and empirical observation, thus failing to understand practical robustness. To address this issue, we introduce Signal Embedding Energy (SEE), a method for quantifying the impact of noise intensity on LALM inputs, enabling the differentiation of LALM robustness in real-world deployments. SEE introduces a perspective based on structured activation subspaces derived from the model's internal representations, which more accurately captures its perception of noise than raw audio features. Across experiments, SEE exhibits a strong correlation with LALM performance, achieving a correlation of 0.98. Surprisingly, traditional audio denoising methods are only marginally effective for LALMs, and, in some cases, even increase SEE and impair performance. This suggests a mismatch between speech-centric denoising objectives and the noise sensitivity of modern LALMs. Therefore, we propose a mitigation strategy derived from SEE to denoise LALM inputs, outperforming existing denoising methods. This paper introduces a novel metric for noise quantification in LALMs, providing guidance for robustness improvements in real-world deployments.", "date": "2026-01-12", "categories": []}
{"id": "2504.06753", "title": "Detect All-Type Deepfake Audio: Wavelet Prompt Tuning for Enhanced Auditory Perception", "abstract": "The rapid advancement of audio generation technologies has escalated the risks of malicious deepfake audio across speech, sound, singing voice, and music, threatening multimedia security and trust. While existing countermeasures (CMs) perform well in single-type audio deepfake detection (ADD), their performance declines in cross-type scenarios. This paper is dedicated to studying the all-type ADD task. We are the first to comprehensively establish an all-type ADD benchmark to evaluate current CMs, incorporating cross-type deepfake detection across speech, sound, singing voice, and music. Then, we introduce the prompt tuning self-supervised learning (PT-SSL) training paradigm, which optimizes SSL front-end by learning specialized prompt tokens for ADD, requiring 458x fewer trainable parameters than fine-tuning (FT). Considering the auditory perception of different audio types, we propose the wavelet prompt tuning (WPT)-SSL method to capture type-invariant auditory deepfake information from the frequency domain without requiring additional training parameters, thereby enhancing performance over FT in the all-type ADD task. To achieve an universally CM, we utilize all types of deepfake audio for co-training. Experimental results demonstrate that WPT-XLSR-AASIST achieved the best performance, with an average EER of 3.58% across all evaluation sets.", "date": "2025-04-09", "categories": []}
{"id": "2601.01568", "title": "MM-Sonate: Multimodal Controllable Audio-Video Generation with Zero-Shot Voice Cloning", "abstract": "Joint audio-video generation aims to synthesize synchronized multisensory content, yet current unified models struggle with fine-grained acoustic control, particularly for identity-preserving speech. Existing approaches either suffer from temporal misalignment due to cascaded generation or lack the capability to perform zero-shot voice cloning within a joint synthesis framework. In this work, we present MM-Sonate, a multimodal flow-matching framework that unifies controllable audio-video joint generation with zero-shot voice cloning capabilities. Unlike prior works that rely on coarse semantic descriptions, MM-Sonate utilizes a unified instruction-phoneme input to enforce strict linguistic and temporal alignment. To enable zero-shot voice cloning, we introduce a timbre injection mechanism that effectively decouples speaker identity from linguistic content. Furthermore, addressing the limitations of standard classifier-free guidance in multimodal settings, we propose a noise-based negative conditioning strategy that utilizes natural noise priors to significantly enhance acoustic fidelity. Empirical evaluations demonstrate that MM-Sonate establishes new state-of-the-art performance in joint generation benchmarks, significantly outperforming baselines in lip synchronization and speech intelligibility, while achieving voice cloning fidelity comparable to specialized Text-to-Speech systems.", "date": "2026-01-04", "categories": []}
{"id": "2601.02967", "title": "MoE Adapter for Large Audio Language Models: Sparsity, Disentanglement, and Gradient-Conflict-Free", "abstract": "Extending the input modality of Large Language Models~(LLMs) to the audio domain is essential for achieving comprehensive multimodal perception. However, it is well-known that acoustic information is intrinsically \\textit{heterogeneous}, entangling attributes such as speech, music, and environmental context. Existing research is limited to a dense, parameter-shared adapter to model these diverse patterns, which induces \\textit{gradient conflict} during optimization, as parameter updates required for distinct attributes contradict each other. To address this limitation, we introduce the \\textit{\\textbf{MoE-Adapter}}, a sparse Mixture-of-Experts~(MoE) architecture designed to decouple acoustic information. Specifically, it employs a dynamic gating mechanism that routes audio tokens to specialized experts capturing complementary feature subspaces while retaining shared experts for global context, thereby mitigating gradient conflicts and enabling fine-grained feature learning. Comprehensive experiments show that the MoE-Adapter achieves superior performance on both audio semantic and paralinguistic tasks, consistently outperforming dense linear baselines with comparable computational costs. Furthermore, we will release the related code and models to facilitate future research.", "date": "2026-01-06", "categories": []}
{"id": "2601.04343", "title": "Summary of The Inaugural Music Source Restoration Challenge", "abstract": "Music Source Restoration (MSR) aims to recover original, unprocessed instrument stems from professionally mixed and degraded audio, requiring the reversal of both production effects and real-world degradations. We present the inaugural MSR Challenge, which features objective evaluation on studio-produced mixtures using Multi-Mel-SNR, Zimtohrli, and FAD-CLAP, alongside subjective evaluation on real-world degraded recordings. Five teams participated in the challenge. The winning system achieved 4.46 dB Multi-Mel-SNR and 3.47 MOS-Overall, corresponding to relative improvements of 91% and 18% over the second-place system, respectively. Per-stem analysis reveals substantial variation in restoration difficulty across instruments, with bass averaging 4.59 dB across all teams, while percussion averages only 0.29 dB. The dataset, evaluation protocols, and baselines are available at https://msrchallenge.com/.", "date": "2026-01-07", "categories": []}
{"id": "2512.14714", "title": "Improving Underwater Acoustic Classification Through Learnable Gabor Filter Convolution and Attention Mechanisms", "abstract": "Remotely detecting and classifying underwater acoustic targets is critical for environmental monitoring and defence. However, the complexity of ship-radiated and environmental noise poses significant challenges for accurate signal processing. While recent advancements in machine learning have improved classification accuracy, limited dataset availability and a lack of standardised experimentation hinder generalisation and robustness. This paper introduces GSE ResNeXt, a deep learning architecture integrating learnable Gabor convolutional layers with a ResNeXt backbone enhanced by squeeze-and-excitation attention. The Gabor filters serve as two-dimensional adaptive band-pass filters, extending the feature channel representation. Its combination with channel attention improves training stability and convergence while enhancing the model's ability to extract discriminative features. The model is evaluated using three training-test split strategies that reflect increasingly complex classification tasks, demonstrating how systematic evaluation design addresses issues such as data leakage, temporal separation, and taxonomy. Results show that GSE ResNeXt consistently outperforms baseline models like Xception, ResNet, and MobileNetV2, in terms of classification performance. Regarding stability and convergence, adding Gabor convolutions to the initial layers of the model reduced training time by up to 62%. During the evaluation of training-testing splits, temporal separation between subsets significantly affected performance, proving more influential than training data volume. These findings suggest that signal processing can enhance model reliability and generalisation under varying environmental conditions, particularly in data-limited underwater acoustic classification. Future developments should focus on mitigating environmental effects on input signals.", "date": "2025-12-09", "categories": []}
{"id": "2601.03443", "title": "Discriminating real and synthetic super-resolved audio samples using embedding-based classifiers", "abstract": "Generative adversarial networks (GANs) and diffusion models have recently achieved state-of-the-art performance in audio super-resolution (ADSR), producing perceptually convincing wideband audio from narrowband inputs. However, existing evaluations primarily rely on signal-level or perceptual metrics, leaving open the question of how closely the distributions of synthetic super-resolved and real wideband audio match. Here we address this problem by analyzing the separability of real and super-resolved audio in various embedding spaces. We consider both middle-band ($4\\to 16$~kHz) and full-band ($16\\to 48$~kHz) upsampling tasks for speech and music, training linear classifiers to distinguish real from synthetic samples based on multiple types of audio embeddings. Comparisons with objective metrics and subjective listening tests reveal that embedding-based classifiers achieve near-perfect separation, even when the generated audio attains high perceptual quality and state-of-the-art metric scores. This behavior is consistent across datasets and models, including recent diffusion-based approaches, highlighting a persistent gap between perceptual quality and true distributional fidelity in ADSR models.", "date": "2026-01-06", "categories": []}
{"id": "2601.03531", "title": "PALM-Bench: A Comprehensive Benchmark for Personalized Audio-Language Models", "abstract": "Large Audio-Language Models (LALMs) have demonstrated strong performance in audio understanding and generation. Yet, our extensive benchmarking reveals that their behavior is largely generic (e.g., summarizing spoken content) and fails to adequately support personalized question answering (e.g., summarizing what my best friend says). In contrast, human conditions their interpretation and decision-making on each individual's personal context. To bridge this gap, we formalize the task of Personalized LALMs (PALM) for recognizing personal concepts and reasoning within personal context. Moreover, we create the first benchmark (PALM-Bench) to foster the methodological advances in PALM and enable structured evaluation on several tasks across multi-speaker scenarios. Our extensive experiments on representative open-source LALMs, show that existing training-free prompting and supervised fine-tuning strategies, while yield improvements, remains limited in modeling personalized knowledge and transferring them across tasks robustly. Data and code will be released.", "date": "2026-01-07", "categories": []}
{"id": "2601.03610", "title": "Investigation into respiratory sound classification for an imbalanced data set using hybrid LSTM-KAN architectures", "abstract": "Respiratory sounds captured via auscultation contain critical clues for diagnosing pulmonary conditions. Automated classification of these sounds faces challenges due to subtle acoustic differences and severe class imbalance in clinical datasets. This study investigates respiratory sound classification with a focus on mitigating pronounced class imbalance. We propose a hybrid deep learning model that combines a Long Short-Term Memory (LSTM) network for sequential feature encoding with a Kolmogorov-Arnold Network (KAN) for classification. The model is integrated with a comprehensive feature extraction pipeline and targeted imbalance mitigation strategies. Experiments were conducted on a public respiratory sound database comprising six classes with a highly skewed distribution. Techniques such as focal loss, class-specific data augmentation, and Synthetic Minority Over-sampling Technique (SMOTE) were employed to enhance minority class recognition. The proposed Hybrid LSTM-KAN model achieves an overall accuracy of 94.6 percent and a macro-averaged F1 score of 0.703, despite the dominant COPD class accounting for over 86 percent of the data. Improved detection performance is observed for minority classes compared to baseline approaches, demonstrating the effectiveness of the proposed architecture for imbalanced respiratory sound classification.", "date": "2026-01-07", "categories": []}
{"id": "2601.03626", "title": "Learning from Limited Labels: Transductive Graph Label Propagation for Indian Music Analysis", "abstract": "Supervised machine learning frameworks rely on extensive labeled datasets for robust performance on real-world tasks. However, there is a lack of large annotated datasets in audio and music domains, as annotating such recordings is resource-intensive, laborious, and often require expert domain knowledge. In this work, we explore the use of label propagation (LP), a graph-based semi-supervised learning technique, for automatically labeling the unlabeled set in an unsupervised manner. By constructing a similarity graph over audio embeddings, we propagate limited label information from a small annotated subset to a larger unlabeled corpus in a transductive, semi-supervised setting. We apply this method to two tasks in Indian Art Music (IAM): Raga identification and Instrument classification. For both these tasks, we integrate multiple public datasets along with additional recordings we acquire from Prasar Bharati Archives to perform LP. Our experiments demonstrate that LP significantly reduces labeling overhead and produces higher-quality annotations compared to conventional baseline methods, including those based on pretrained inductive models. These results highlight the potential of graph-based semi-supervised learning to democratize data annotation and accelerate progress in music information retrieval.", "date": "2026-01-07", "categories": []}
{"id": "2601.03783", "title": "HearSay Benchmark: Do Audio LLMs Leak What They Hear?", "abstract": "While Audio Large Language Models (ALLMs) have achieved remarkable progress in understanding and generation, their potential privacy implications remain largely unexplored. This paper takes the first step to investigate whether ALLMs inadvertently leak user privacy solely through acoustic voiceprints and introduces $\\textit{HearSay}$, a comprehensive benchmark constructed from over 22,000 real-world audio clips. To ensure data quality, the benchmark is meticulously curated through a rigorous pipeline involving automated profiling and human verification, guaranteeing that all privacy labels are grounded in factual records. Extensive experiments on $\\textit{HearSay}$ yield three critical findings: $\\textbf{Significant Privacy Leakage}$: ALLMs inherently extract private attributes from voiceprints, reaching 92.89% accuracy on gender and effectively profiling social attributes. $\\textbf{Insufficient Safety Mechanisms}$: Alarmingly, existing safeguards are severely inadequate; most models fail to refuse privacy-intruding requests, exhibiting near-zero refusal rates for physiological traits. $\\textbf{Reasoning Amplifies Risk}$: Chain-of-Thought (CoT) reasoning exacerbates privacy risks in capable models by uncovering deeper acoustic correlations. These findings expose critical vulnerabilities in ALLMs, underscoring the urgent need for targeted privacy alignment. The codes and dataset are available at https://github.com/JinWang79/HearSay_Benchmark", "date": "2026-01-07", "categories": []}
{"id": "2509.22461", "title": "CMDAR: A Chinese Multi-scene Dynamic Audio Reasoning Benchmark with Diverse Challenges", "abstract": "The ability to reason from audio, including speech, environmental sounds, and music, is essential for AI agents to interact effectively in real-world scenarios. Existing benchmarks mainly focus on static or single-scene settings and English audio data and do not fully capture scenarios where multiple speakers, unfolding events, and heterogeneous audio sources interact. To address these challenges, we introduce CMDAR, a Chinese benchmark for evaluating models on complex, multi-scene, and dynamically evolving audio reasoning tasks. CMDAR comprises 3,000 carefully curated question-answer pairs linked to diverse audio clips, covering five categories of complex reasoning and spanning three question types. We benchmark 26 state-of-the-art audio language models on CMDAR and observe that they exhibit limitations in complex reasoning tasks. In CMDAR-main, Qwen2.5-Omni achieves 76.67% accuracy, whereas GPT-4o Audio reaches 68.47%. However, GPT-4o Audio substantially outperforms Qwen2.5-Omni on the more challenging multiple-choice with multiple audios and open-ended tasks. And we provide detail analysis corresponding suggestions for the future development of large audio language models.", "date": "2025-09-26", "categories": []}
{"id": "2601.02391", "title": "WearVox: An Egocentric Multichannel Voice Assistant Benchmark for Wearables", "abstract": "Wearable devices such as AI glasses are transforming voice assistants into always-available, hands-free collaborators that integrate seamlessly with daily life, but they also introduce challenges like egocentric audio affected by motion and noise, rapid micro-interactions, and the need to distinguish device-directed speech from background conversations. Existing benchmarks largely overlook these complexities, focusing instead on clean or generic conversational audio. To bridge this gap, we present WearVox, the first benchmark designed to rigorously evaluate voice assistants in realistic wearable scenarios. WearVox comprises 3,842 multi-channel, egocentric audio recordings collected via AI glasses across five diverse tasks including Search-Grounded QA, Closed-Book QA, Side-Talk Rejection, Tool Calling, and Speech Translation, spanning a wide range of indoor and outdoor environments and acoustic conditions. Each recording is accompanied by rich metadata, enabling nuanced analysis of model performance under real-world constraints. We benchmark leading proprietary and open-source speech Large Language Models (SLLMs) and find that most real-time SLLMs achieve accuracies on WearVox ranging from 29% to 59%, with substantial performance degradation on noisy outdoor audio, underscoring the difficulty and realism of the benchmark. Additionally, we conduct a case study with two new SLLMs that perform inference with single-channel and multi-channel audio, demonstrating that multi-channel audio inputs significantly enhance model robustness to environmental noise and improve discrimination between device-directed and background speech. Our results highlight the critical importance of spatial audio cues for context-aware voice assistants and establish WearVox as a comprehensive testbed for advancing wearable voice AI research.", "date": "2025-12-25", "categories": []}
{"id": "2601.02776", "title": "UniSRCodec: Unified and Low-Bitrate Single Codebook Codec with Sub-Band Reconstruction", "abstract": "Neural Audio Codecs (NACs) can reduce transmission overhead by performing compact compression and reconstruction, which also aim to bridge the gap between continuous and discrete signals. Existing NACs can be divided into two categories: multi-codebook and single-codebook codecs. Multi-codebook codecs face challenges such as structural complexity and difficulty in adapting to downstream tasks, while single-codebook codecs, though structurally simpler, suffer from low-fidelity, ineffective modeling of unified audio, and an inability to support modeling of high-frequency audio. We propose the UniSRCodec, a single-codebook codec capable of supporting high sampling rate, low-bandwidth, high fidelity, and unified. We analyze the inefficiency of waveform-based compression and introduce the time and frequency compression method using the Mel-spectrogram, and cooperate with a Vocoder to recover the phase information of the original audio. Moreover, we propose a sub-band reconstruction technique to achieve high-quality compression across both low and high frequency bands. Subjective and objective experimental results demonstrate that UniSRCodec achieves state-of-the-art (SOTA) performance among cross-domain single-codebook codecs with only a token rate of 40, and its reconstruction quality is comparable to that of certain multi-codebook methods. Our demo page is available at https://wxzyd123.github.io/unisrcodec.", "date": "2026-01-06", "categories": []}
{"id": "2601.02983", "title": "Interpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency-Time Reinforcement Learning", "abstract": "Recent advances in audio large language models (ALLMs) have made high-quality synthetic audio widely accessible, increasing the risk of malicious audio deepfakes across speech, environmental sounds, singing voice, and music. Real-world audio deepfake detection (ADD) therefore requires all-type detectors that generalize across heterogeneous audio and provide interpretable decisions. Given the strong multi-task generalization ability of ALLMs, we first investigate their performance on all-type ADD under both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). However, SFT using only binary real/fake labels tends to reduce the model to a black-box classifier, sacrificing interpretability. Meanwhile, vanilla RFT under sparse supervision is prone to reward hacking and can produce hallucinated, ungrounded rationales. To address this, we propose an automatic annotation and polishing pipeline that constructs Frequency-Time structured chain-of-thought (CoT) rationales, producing ~340K cold-start demonstrations. Building on CoT data, we propose Frequency Time-Group Relative Policy Optimization (FT-GRPO), a two-stage training paradigm that cold-starts ALLMs with SFT and then applies GRPO under rule-based frequency-time constraints. Experiments demonstrate that FT-GRPO achieves state-of-the-art performance on all-type ADD while producing interpretable, FT-grounded rationales. The data and code are available online.", "date": "2026-01-06", "categories": []}
{"id": "2601.03227", "title": "The Sonar Moment: Benchmarking Audio-Language Models in Audio Geo-Localization", "abstract": "Geo-localization aims to infer the geographic origin of a given signal. In computer vision, geo-localization has served as a demanding benchmark for compositional reasoning and is relevant to public safety. In contrast, progress on audio geo-localization has been constrained by the lack of high-quality audio-location pairs. To address this gap, we introduce AGL1K, the first audio geo-localization benchmark for audio language models (ALMs), spanning 72 countries and territories. To extract reliably localizable samples from a crowd-sourced platform, we propose the Audio Localizability metric that quantifies the informativeness of each recording, yielding 1,444 curated audio clips. Evaluations on 16 ALMs show that ALMs have emerged with audio geo-localization capability. We find that closed-source models substantially outperform open-source models, and that linguistic clues often dominate as a scaffold for prediction. We further analyze ALMs' reasoning traces, regional bias, error causes, and the interpretability of the localizability metric. Overall, AGL1K establishes a benchmark for audio geo-localization and may advance ALMs with better geospatial reasoning capability.", "date": "2026-01-06", "categories": []}
{"id": "2508.04665", "title": "Perch 2.0: The Bittern Lesson for Bioacoustics", "abstract": "Perch is a performant pre-trained model for bioacoustics. It was trained in supervised fashion, providing both off-the-shelf classification scores for thousands of vocalizing species as well as strong embeddings for transfer learning. In this new release, Perch 2.0, we expand from training exclusively on avian species to a large multi-taxa dataset. The model is trained with self-distillation using a prototype-learning classifier as well as a new source-prediction training criterion. Perch 2.0 obtains state-of-the-art performance on the BirdSet and BEANS benchmarks. It also outperforms specialized marine models on marine transfer learning tasks, despite having almost no marine training data. We present hypotheses as to why fine-grained species classification is a particularly robust pre-training task for bioacoustics.", "date": "2025-08-06", "categories": []}
{"id": "2508.06372", "title": "SpeakerLM: End-to-End Versatile Speaker Diarization and Recognition with Multimodal Large Language Models", "abstract": "The Speaker Diarization and Recognition (SDR) task aims to predict \"who spoke when and what\" within an audio clip, which is a crucial task in various real-world multi-speaker scenarios such as meeting transcription and dialogue systems. Existing SDR systems typically adopt a cascaded framework, combining multiple modules such as speaker diarization (SD) and automatic speech recognition (ASR). The cascaded systems suffer from several limitations, such as error propagation, difficulty in handling overlapping speech, and lack of joint optimization for exploring the synergy between SD and ASR tasks. To address these limitations, we introduce SpeakerLM, a unified multimodal large language model for SDR that jointly performs SD and ASR in an end-to-end manner. Moreover, to facilitate diverse real-world scenarios, we incorporate a flexible speaker registration mechanism into SpeakerLM, enabling SDR under different speaker registration settings. SpeakerLM is progressively developed with a multi-stage training strategy on large-scale real data. Extensive experiments show that SpeakerLM demonstrates strong data scaling capability and generalizability, outperforming state-of-the-art cascaded baselines on both in-domain and out-of-domain public SDR benchmarks. Furthermore, experimental results show that the proposed speaker registration mechanism effectively ensures robust SDR performance of SpeakerLM across diverse speaker registration conditions and varying numbers of registered speakers.", "date": "2025-08-08", "categories": []}
{"id": "2509.05175", "title": "Room-acoustic simulations as an alternative to measurements for audio-algorithm evaluation", "abstract": "Audio-signal-processing and audio-machine-learning (ASP/AML) algorithms are ubiquitous in modern technology like smart devices, wearables, and entertainment systems. Development of such algorithms and models typically involves a formal evaluation to demonstrate their effectiveness and progress beyond the state-of-the-art. Ideally, a thorough evaluation should cover many diverse application scenarios and room-acoustic conditions. However, in practice, evaluation datasets are often limited in size and diversity because they rely on costly and time-consuming measurements. This paper explores how room-acoustic simulations can be used for evaluating ASP/AML algorithms. To this end, we evaluate three ASP/AML algorithms with room-acoustic measurements and data from different simulation engines, and assess the match between the evaluation results obtained from measurements and simulations. The presented investigation compares a numerical wave-based solver with two geometrical acoustics simulators. While numerical wave-based simulations yielded similar evaluation results as measurements for all three evaluated ASP/AML algorithms, geometrical acoustic simulations could not replicate the measured evaluation results as reliably.", "date": "2025-09-05", "categories": []}
{"id": "2512.24052", "title": "AHA: Aligning Large Audio-Language Models for Reasoning Hallucinations via Counterfactual Hard Negatives", "abstract": "Although Large Audio-Language Models (LALMs) deliver state-of-the-art (SOTA) performance, they frequently suffer from hallucinations, e.g. generating text not grounded in the audio input. We analyze these grounding failures and identify a distinct taxonomy: Event Omission, False Event Identity, Temporal Relation Error, and Quantitative Temporal Error. To address this, we introduce the AHA (Audio Hallucination Alignment) framework. By leveraging counterfactual hard negative mining, our pipeline constructs a high-quality preference dataset that forces models to distinguish strict acoustic evidence from linguistically plausible fabrications. Additionally, we establish AHA-Eval, a diagnostic benchmark designed to rigorously test these fine-grained temporal reasoning capabilities. We apply this data to align Qwen2.5-Omni. The resulting model, Qwen-Audio-AHA, achieves a 13.7% improvement on AHA-Eval. Crucially, this benefit generalizes beyond our diagnostic set. Our model shows substantial gains on public benchmarks, including 1.3% on MMAU-Test and 1.6% on MMAR, outperforming latest SOTA methods. The model and dataset are open-sourced at https://github.com/LLM-VLM-GSL/AHA.", "date": "2025-12-30", "categories": []}
{"id": "2601.01373", "title": "UltraEval-Audio: A Unified Framework for Comprehensive Evaluation of Audio Foundation Models", "abstract": "The development of audio foundation models has accelerated rapidly since the emergence of GPT-4o. However, the lack of comprehensive evaluation has become a critical bottleneck for further progress in the field, particularly in audio generation. Current audio evaluation faces three major challenges: (1) audio evaluation lacks a unified framework, with datasets and code scattered across various sources, hindering fair and efficient cross-model comparison;(2) audio codecs, as a key component of audio foundation models, lack a widely accepted and holistic evaluation methodology; (3) existing speech benchmarks are heavily reliant on English, making it challenging to objectively assess models' performance on Chinese. To address the first issue, we introduce UltraEval-Audio, a unified evaluation framework for audio foundation models, specifically designed for both audio understanding and generation tasks. UltraEval-Audio features a modular architecture, supporting 10 languages and 14 core task categories, while seamlessly integrating 24 mainstream models and 36 authoritative benchmarks. To enhance research efficiency, the framework provides a one-command evaluation feature, accompanied by real-time public leaderboards. For the second challenge, UltraEval-Audio adopts a novel comprehensive evaluation scheme for audio codecs, evaluating performance across three key dimensions: semantic accuracy, timbre fidelity, and acoustic quality. To address the third issue, we propose two new Chinese benchmarks, SpeechCMMLU and SpeechHSK, designed to assess Chinese knowledge proficiency and language fluency. We wish that UltraEval-Audio will provide both academia and industry with a transparent, efficient, and fair platform for comparison of audio models. Our code, benchmarks, and leaderboards are available at https://github.com/OpenBMB/UltraEval-Audio.", "date": "2026-01-04", "categories": []}
{"id": "2601.01392", "title": "SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning", "abstract": "Existing fraud detection methods predominantly rely on transcribed text, suffering from ASR errors and missing crucial acoustic cues like vocal tone and environmental context. This limits their effectiveness against complex deceptive strategies. To address these challenges, we first propose \\textbf{SAFE-QAQ}, an end-to-end comprehensive framework for audio-based slow-thinking fraud detection. First, the SAFE-QAQ framework eliminates the impact of transcription errors on detection performance. Secondly, we propose rule-based slow-thinking reward mechanisms that systematically guide the system to identify fraud-indicative patterns by accurately capturing fine-grained audio details, through hierarchical reasoning processes. Besides, our framework introduces a dynamic risk assessment framework during live calls, enabling early detection and prevention of fraud. Experiments on the TeleAntiFraud-Bench demonstrate that SAFE-QAQ achieves dramatic improvements over existing methods in multiple key dimensions, including accuracy, inference efficiency, and real-time processing capabilities. Currently deployed and analyzing over 70,000 calls daily, SAFE-QAQ effectively automates complex fraud detection, reducing human workload and financial losses. Code: https://anonymous.4open.science/r/SAFE-QAQ.", "date": "2026-01-04", "categories": []}
{"id": "2508.14556", "title": "Mamba2 Meets Silence: Robust Vocal Source Separation for Sparse Regions", "abstract": "We introduce a new music source separation model tailored for accurate vocal isolation. Unlike Transformer-based approaches, which often fail to capture intermittently occurring vocals, our model leverages Mamba2, a recent state space model, to better capture long-range temporal dependencies. To handle long input sequences efficiently, we combine a band-splitting strategy with a dual-path architecture. Experiments show that our approach outperforms recent state-of-the-art models, achieving a cSDR of 11.03 dB-the best reported to date-and delivering substantial gains in uSDR. Moreover, the model exhibits stable and consistent performance across varying input lengths and vocal occurrence patterns. These results demonstrate the effectiveness of Mamba-based models for high-resolution audio processing and open up new directions for broader applications in audio research.", "date": "2025-08-20", "categories": []}
{"id": "2509.17609", "title": "Audio Super-Resolution with Latent Bridge Models", "abstract": "Audio super-resolution (SR), i.e., upsampling the low-resolution (LR) waveform to the high-resolution (HR) version, has recently been explored with diffusion and bridge models, while previous methods often suffer from sub-optimal upsampling quality due to their uninformative generation prior. Towards high-quality audio super-resolution, we present a new system with latent bridge models (LBMs), where we compress the audio waveform into a continuous latent space and design an LBM to enable a latent-to-latent generation process that naturally matches the LR-toHR upsampling process, thereby fully exploiting the instructive prior information contained in the LR waveform. To further enhance the training results despite the limited availability of HR samples, we introduce frequency-aware LBMs, where the prior and target frequency are taken as model input, enabling LBMs to explicitly learn an any-to-any upsampling process at the training stage. Furthermore, we design cascaded LBMs and present two prior augmentation strategies, where we make the first attempt to unlock the audio upsampling beyond 48 kHz and empower a seamless cascaded SR process, providing higher flexibility for audio post-production. Comprehensive experimental results evaluated on the VCTK, ESC-50, Song-Describer benchmark datasets and two internal testsets demonstrate that we achieve state-of-the-art objective and perceptual quality for any-to-48kHz SR across speech, audio, and music signals, as well as setting the first record for any-to-192kHz audio SR. Demo at https://AudioLBM.github.io/.", "date": "2025-09-22", "categories": []}
{"id": "2511.18869", "title": "Hear: Hierarchically Enhanced Aesthetic Representations For Multidimensional Music Evaluation", "abstract": "Evaluating song aesthetics is challenging due to the multidimensional nature of musical perception and the scarcity of labeled data. We propose HEAR, a robust music aesthetic evaluation framework that combines: (1) a multi-source multi-scale representations module to obtain complementary segment- and track-level features, (2) a hierarchical augmentation strategy to mitigate overfitting, and (3) a hybrid training objective that integrates regression and ranking losses for accurate scoring and reliable top-tier song identification. Experiments demonstrate that HEAR consistently outperforms the baseline across all metrics on both tracks of the ICASSP 2026 SongEval benchmark. The code and trained model weights are available at https://github.com/Eps-Acoustic-Revolution-Lab/EAR_HEAR.", "date": "2025-11-24", "categories": []}
{"id": "2512.20407", "title": "AUDRON: A Deep Learning Framework with Fused Acoustic Signatures for Drone Type Recognition", "abstract": "Unmanned aerial vehicles (UAVs), commonly known as drones, are increasingly used across diverse domains, including logistics, agriculture, surveillance, and defense. While these systems provide numerous benefits, their misuse raises safety and security concerns, making effective detection mechanisms essential. Acoustic sensing offers a low-cost and non-intrusive alternative to vision or radar-based detection, as drone propellers generate distinctive sound patterns. This study introduces AUDRON (AUdio-based Drone Recognition Network), a hybrid deep learning framework for drone sound detection, employing a combination of Mel-Frequency Cepstral Coefficients (MFCC), Short-Time Fourier Transform (STFT) spectrograms processed with convolutional neural networks (CNNs), recurrent layers for temporal modeling, and autoencoder-based representations. Feature-level fusion integrates complementary information before classification. Experimental evaluation demonstrates that AUDRON effectively differentiates drone acoustic signatures from background noise, achieving high accuracy while maintaining generalizability across varying conditions. AUDRON achieves 98.51 percent and 97.11 percent accuracy in binary and multiclass classification. The results highlight the advantage of combining multiple feature representations with deep learning for reliable acoustic drone detection, suggesting the framework's potential for deployment in security and surveillance applications where visual or radar sensing may be limited.", "date": "2025-12-23", "categories": []}
{"id": "2512.23808", "title": "MiMo-Audio: Audio Language Models are Few-Shot Learners", "abstract": "Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demonstrates powerful speech continuation capabilities, capable of generating highly realistic talk shows, recitations, livestreaming and debates. At the post-training stage, we curate a diverse instruction-tuning corpus and introduce thinking mechanisms into both audio understanding and generation. MiMo-Audio-7B-Instruct achieves open-source SOTA on audio understanding benchmarks (MMSU, MMAU, MMAR, MMAU-Pro), spoken dialogue benchmarks (Big Bench Audio, MultiChallenge Audio) and instruct-TTS evaluations, approaching or surpassing closed-source models. Model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-Audio.", "date": "2025-12-29", "categories": []}
{"id": "2512.24628", "title": "AI-Driven Acoustic Voice Biomarker-Based Hierarchical Classification of Benign Laryngeal Voice Disorders from Sustained Vowels", "abstract": "Benign laryngeal voice disorders affect nearly one in five individuals and often manifest as dysphonia, while also serving as non-invasive indicators of broader physiological dysfunction. We introduce a clinically inspired hierarchical machine learning framework for automated classification of eight benign voice disorders alongside healthy controls, using acoustic features extracted from short, sustained vowel phonations. Experiments utilized 15,132 recordings from 1,261 speakers in the Saarbruecken Voice Database, covering vowels /a/, /i/, and /u/ at neutral, high, low, and gliding pitches. Mirroring clinical triage workflows, the framework operates in three sequential stages: Stage 1 performs binary screening of pathological versus non-pathological voices by integrating convolutional neural network-derived mel-spectrogram features with 21 interpretable acoustic biomarkers; Stage 2 stratifies voices into Healthy, Functional or Psychogenic, and Structural or Inflammatory groups using a cubic support vector machine; Stage 3 achieves fine-grained classification by incorporating probabilistic outputs from prior stages, improving discrimination of structural and inflammatory disorders relative to functional conditions. The proposed system consistently outperformed flat multi-class classifiers and pre-trained self-supervised models, including META HuBERT and Google HeAR, whose generic objectives are not optimized for sustained clinical phonation. By combining deep spectral representations with interpretable acoustic features, the framework enhances transparency and clinical alignment. These results highlight the potential of quantitative voice biomarkers as scalable, non-invasive tools for early screening, diagnostic triage, and longitudinal monitoring of vocal health.", "date": "2025-12-31", "categories": []}
{"id": "2512.24731", "title": "EchoFoley: Event-Centric Hierarchical Control for Video Grounded Creative Sound Generation", "abstract": "Sound effects build an essential layer of multimodal storytelling, shaping the emotional atmosphere and the narrative semantics of videos. Despite recent advancement in video-text-to-audio (VT2A), the current formulation faces three key limitations: First, an imbalance between visual and textual conditioning that leads to visual dominance; Second, the absence of a concrete definition for fine-grained controllable generation; Third, weak instruction understanding and following, as existing datasets rely on brief categorical tags. To address these limitations, we introduce EchoFoley, a new task designed for video-grounded sound generation with both event level local control and hierarchical semantic control. Our symbolic representation for sounding events specifies when, what, and how each sound is produced within a video or instruction, enabling fine-grained controls like sound generation, insertion, and editing. To support this task, we construct EchoFoley-6k, a large-scale, expert-curated benchmark containing over 6,000 video-instruction-annotation triplets. Building upon this foundation, we propose EchoVidia a sounding-event-centric agentic generation framework with slow-fast thinking strategy. Experiments show that EchoVidia surpasses recent VT2A models by 40.7% in controllability and 12.5% in perceptual quality.", "date": "2025-12-31", "categories": []}
{"id": "2512.22564", "title": "Geometry-Aware Optimization for Respiratory Sound Classification: Enhancing Sensitivity with SAM-Optimized Audio Spectrogram Transformers", "abstract": "Respiratory sound classification is hindered by the limited size, high noise levels, and severe class imbalance of benchmark datasets like ICBHI 2017. While Transformer-based models offer powerful feature extraction capabilities, they are prone to overfitting and often converge to sharp minima in the loss landscape when trained on such constrained medical data. To address this, we introduce a framework that enhances the Audio Spectrogram Transformer (AST) using Sharpness-Aware Minimization (SAM). Instead of merely minimizing the training loss, our approach optimizes the geometry of the loss surface, guiding the model toward flatter minima that generalize better to unseen patients. We also implement a weighted sampling strategy to handle class imbalance effectively. Our method achieves a state-of-the-art score of 68.10% on the ICBHI 2017 dataset, outperforming existing CNN and hybrid baselines. More importantly, it reaches a sensitivity of 68.31%, a crucial improvement for reliable clinical screening. Further analysis using t-SNE and attention maps confirms that the model learns robust, discriminative features rather than memorizing background noise.", "date": "2025-12-27", "categories": []}
{"id": "2512.22621", "title": "Chord Recognition with Deep Learning", "abstract": "Progress in automatic chord recognition has been slow since the advent of deep learning in the field. To understand why, I conduct experiments on existing methods and test hypotheses enabled by recent developments in generative models. Findings show that chord classifiers perform poorly on rare chords and that pitch augmentation boosts accuracy. Features extracted from generative models do not help and synthetic data presents an exciting avenue for future work. I conclude by improving the interpretability of model outputs with beat detection, reporting some of the best results in the field and providing qualitative analysis. Much work remains to solve automatic chord recognition, but I hope this thesis will chart a path for others to try.", "date": "2025-12-27", "categories": []}
{"id": "2512.21653", "title": "Semantic Codebooks as Effective Priors for Neural Speech Compression", "abstract": "Speech codecs are traditionally optimized for waveform fidelity, allocating bits to preserve acoustic detail even when much of it can be inferred from linguistic structure. This leads to inefficient compression and suboptimal performance on downstream recognition tasks. We propose SemDAC, a semantic-aware neural audio codec that leverages semantic codebooks as effective priors for speech compression. In SemDAC, the first quantizer in a residual vector quantization (RVQ) stack is distilled from HuBERT features to produce semantic tokens that capture phonetic content, while subsequent quantizers model residual acoustics. A FiLM-conditioned decoder reconstructs audio conditioned on the semantic tokens, improving efficiency in the use of acoustic codebooks. Despite its simplicity, this design proves highly effective: SemDAC outperforms DAC across perceptual metrics and achieves lower WER when running Whisper on reconstructed speech, all while operating at substantially lower bitrates (e.g., 0.95 kbps vs. 2.5 kbps for DAC). These results demonstrate that semantic codebooks provide an effective inductive bias for neural speech compression, producing compact yet recognition-friendly representations.", "date": "2025-12-25", "categories": []}
{"id": "2512.21702", "title": "Zero-Shot to Zero-Lies: Detecting Bengali Deepfake Audio through Transfer Learning", "abstract": "The rapid growth of speech synthesis and voice conversion systems has made deepfake audio a major security concern. Bengali deepfake detection remains largely unexplored. In this work, we study automatic detection of Bengali audio deepfakes using the BanglaFake dataset. We evaluate zeroshot inference with several pretrained models. These include Wav2Vec2-XLSR-53, Whisper, PANNsCNN14, WavLM and Audio Spectrogram Transformer. Zero-shot results show limited detection ability. The best model, Wav2Vec2-XLSR-53, achieves 53.80% accuracy, 56.60% AUC and 46.20% EER. We then f ine-tune multiple architectures for Bengali deepfake detection. These include Wav2Vec2-Base, LCNN, LCNN-Attention, ResNet18, ViT-B16 and CNN-BiLSTM. Fine-tuned models show strong performance gains. ResNet18 achieves the highest accuracy of 79.17%, F1 score of 79.12%, AUC of 84.37% and EER of 24.35%. Experimental results confirm that fine-tuning significantly improves performance over zero-shot inference. This study provides the first systematic benchmark of Bengali deepfake audio detection. It highlights the effectiveness of f ine-tuned deep learning models for this low-resource language.", "date": "2025-12-25", "categories": []}
{"id": "2512.18099", "title": "SAM Audio: Segment Anything in Audio", "abstract": "General audio source separation is a key capability for multimodal AI systems that can perceive and reason about sound. Despite substantial progress in recent years, existing separation models are either domain-specific, designed for fixed categories such as speech or music, or limited in controllability, supporting only a single prompting modality such as text. In this work, we present SAM Audio, a foundation model for general audio separation that unifies text, visual, and temporal span prompting within a single framework. Built on a diffusion transformer architecture, SAM Audio is trained with flow matching on large-scale audio data spanning speech, music, and general sounds, and can flexibly separate target sources described by language, visual masks, or temporal spans. The model achieves state-of-the-art performance across a diverse suite of benchmarks, including general sound, speech, music, and musical instrument separation in both in-the-wild and professionally produced audios, substantially outperforming prior general-purpose and specialized systems. Furthermore, we introduce a new real-world separation benchmark with human-labeled multimodal prompts and a reference-free evaluation model that correlates strongly with human judgment.", "date": "2025-12-19", "categories": []}
{"id": "2512.18797", "title": "Reliable Audio Deepfake Detection in Variable Conditions via Quantum-Kernel SVMs", "abstract": "Detecting synthetic speech is challenging when labeled data are scarce and recording conditions vary. Existing end-to-end deep models often overfit or fail to generalize, and while kernel methods can remain competitive, their performance heavily depends on the chosen kernel. Here, we show that using a quantum kernel in audio deepfake detection reduces falsepositive rates without increasing model size. Quantum feature maps embed data into high-dimensional Hilbert spaces, enabling the use of expressive similarity measures and compact classifiers. Building on this motivation, we compare quantum-kernel SVMs (QSVMs) with classical SVMs using identical mel-spectrogram preprocessing and stratified 5-fold cross-validation across four corpora (ASVspoof 2019 LA, ASVspoof 5 (2024), ADD23, and an In-the-Wild set). QSVMs achieve consistently lower equalerror rates (EER): 0.183 vs. 0.299 on ASVspoof 5 (2024), 0.081 vs. 0.188 on ADD23, 0.346 vs. 0.399 on ASVspoof 2019, and 0.355 vs. 0.413 In-the-Wild. At the EER operating point (where FPR equals FNR), these correspond to absolute false-positiverate reductions of 0.116 (38.8%), 0.107 (56.9%), 0.053 (13.3%), and 0.058 (14.0%), respectively. We also report how consistent the results are across cross-validation folds and margin-based measures of class separation, using identical settings for both models. The only modification is the kernel; the features and SVM remain unchanged, no additional trainable parameters are introduced, and the quantum kernel is computed on a conventional computer.", "date": "2025-12-21", "categories": []}
{"id": "2512.19687", "title": "Pushing the Frontier of Audiovisual Perception with Large-Scale Multimodal Correspondence Learning", "abstract": "We introduce Perception Encoder Audiovisual, PE-AV, a new family of encoders for audio and video understanding trained with scaled contrastive learning. Built on PE, PE-AV makes several key contributions to extend representations to audio, and natively support joint embeddings across audio-video, audio-text, and video-text modalities. PE-AV's unified cross-modal embeddings enable novel tasks such as speech retrieval, and set a new state of the art across standard audio and video benchmarks. We unlock this by building a strong audiovisual data engine that synthesizes high-quality captions for O(100M) audio-video pairs, enabling large-scale supervision consistent across modalities. Our audio data includes speech, music, and general sound effects-avoiding single-domain limitations common in prior work. We exploit ten pairwise contrastive objectives, showing that scaling cross-modality and caption-type pairs strengthens alignment and improves zero-shot performance. We further develop PE-A-Frame by fine-tuning PE-AV with frame-level contrastive objectives, enabling fine-grained audio-frame-to-text alignment for tasks such as sound event detection.", "date": "2025-12-22", "categories": []}
{"id": "2506.09707", "title": "Fine-Tuning Large Audio-Language Models with LoRA for Precise Temporal Localization of Prolonged Exposure Therapy Elements", "abstract": "Prolonged Exposure (PE) therapy is an effective treatment for post-traumatic stress disorder (PTSD), but evaluating therapist fidelity remains labor-intensive due to the need for manual review of session recordings. We present a method for the automatic temporal localization of key PE fidelity elements, identifying their start and stop times, directly from session audio and transcripts. Our approach fine-tunes a large pre-trained audio-language model, Qwen2-Audio, using Low-Rank Adaptation (LoRA) to process focused 30-second windows of audio-transcript input. Fidelity labels for three core protocol phases, therapist orientation (P1), imaginal exposure (P2), and post-imaginal processing (P3), are generated via LLM-based prompting and verified by trained raters. The model is trained to predict normalized boundary offsets using soft supervision guided by task-specific prompts. On a dataset of 308 real PE sessions, our best configuration (LoRA rank 8, 30s windows) achieves a mean absolute error (MAE) of 5.3s across tasks, within typical rater tolerance for timestamp review, enabling practical fidelity QC. We further analyze the effects of window size and LoRA rank, highlighting the importance of context granularity and model adaptation. This work introduces a privacy-preserving, scalable framework for fidelity tracking in PE therapy, with potential to support clinician training, supervision, and quality assurance.", "date": "2025-06-11", "categories": []}
{"id": "2508.20717", "title": "Unified Acoustic Representations for Screening Neurological and Respiratory Pathologies from Voice", "abstract": "Voice-based health assessment offers unprecedented opportunities for scalable, non-invasive disease screening, yet existing approaches typically focus on single conditions and fail to leverage the rich, multi-faceted information embedded in speech. We present MARVEL (Multi-task Acoustic Representations for Voice-based Health Analysis), a privacy-conscious multitask learning framework that simultaneously detects nine distinct neurological, respiratory, and voice disorders using only derived acoustic features, eliminating the need for raw audio transmission. Our dual-branch architecture employs specialized encoders with task-specific heads sharing a common acoustic backbone, enabling effective cross-condition knowledge transfer. Evaluated on the large-scale Bridge2AI-Voice v2.0 dataset, MARVEL achieves an overall AUROC of 0.78, with exceptional performance on neurological disorders (AUROC = 0.89), particularly for Alzheimer's disease/mild cognitive impairment (AUROC = 0.97). Our framework consistently outperforms single-modal baselines by 5-19% and surpasses state-of-the-art self-supervised models on 7 of 9 tasks, while correlation analysis reveals that the learned representations exhibit meaningful similarities with established acoustic features, indicating that the model's internal representations are consistent with clinically recognized acoustic patterns. By demonstrating that a single unified model can effectively screen for diverse conditions, this work establishes a foundation for deployable voice-based diagnostics in resource-constrained and remote healthcare settings.", "date": "2025-08-28", "categories": []}
{"id": "2512.17281", "title": "LibriVAD: A Scalable Open Dataset with Deep Learning Benchmarks for Voice Activity Detection", "abstract": "Robust Voice Activity Detection (VAD) remains a challenging task, especially under noisy, diverse, and unseen acoustic conditions. Beyond algorithmic development, a key limitation in advancing VAD research is the lack of large-scale, systematically controlled, and publicly available datasets. To address this, we introduce LibriVAD - a scalable open-source dataset derived from LibriSpeech and augmented with diverse real-world and synthetic noise sources. LibriVAD enables systematic control over speech-to-noise ratio, silence-to-speech ratio (SSR), and noise diversity, and is released in three sizes (15 GB, 150 GB, and 1.5 TB) with two variants (LibriVAD-NonConcat and LibriVAD-Concat) to support different experimental setups. We benchmark multiple feature-model combinations, including waveform, Mel-Frequency Cepstral Coefficients (MFCC), and Gammatone filter bank cepstral coefficients, and introduce the Vision Transformer (ViT) architecture for VAD. Our experiments show that ViT with MFCC features consistently outperforms established VAD models such as boosted deep neural network and convolutional long short-term memory deep neural network across seen, unseen, and out-of-distribution (OOD) conditions, including evaluation on the real-world VOiCES dataset. We further analyze the impact of dataset size and SSR on model generalization, experimentally showing that scaling up dataset size and balancing SSR noticeably and consistently enhance VAD performance under OOD conditions. All datasets, trained models, and code are publicly released to foster reproducibility and accelerate progress in VAD research.", "date": "2025-12-19", "categories": []}
{"id": "2503.20990", "title": "FinAudio: A Benchmark for Audio Large Language Models in Financial Applications", "abstract": "Audio Large Language Models (AudioLLMs) have received widespread attention and have significantly improved performance on audio tasks such as conversation, audio understanding, and automatic speech recognition (ASR). Despite these advancements, there is an absence of a benchmark for assessing AudioLLMs in financial scenarios, where audio data, such as earnings conference calls and CEO speeches, are crucial resources for financial analysis and investment decisions. In this paper, we introduce \\textsc{FinAudio}, the first benchmark designed to evaluate the capacity of AudioLLMs in the financial domain. We first define three tasks based on the unique characteristics of the financial domain: 1) ASR for short financial audio, 2) ASR for long financial audio, and 3) summarization of long financial audio. Then, we curate two short and two long audio datasets, respectively, and develop a novel dataset for financial audio summarization, comprising the \\textsc{FinAudio} benchmark. Then, we evaluate seven prevalent AudioLLMs on \\textsc{FinAudio}. Our evaluation reveals the limitations of existing AudioLLMs in the financial domain and offers insights for improving AudioLLMs. All datasets and codes will be released.", "date": "2025-03-26", "categories": []}
{"id": "2512.15532", "title": "A Conditioned UNet for Music Source Separation", "abstract": "In this paper we propose a conditioned UNet for Music Source Separation (MSS). MSS is generally performed by multi-output neural networks, typically UNets, with each output representing a particular stem from a predefined instrument vocabulary. In contrast, conditioned MSS networks accept an audio query related to a stem of interest alongside the signal from which that stem is to be extracted. Thus, a strict vocabulary is not required and this enables more realistic tasks in MSS. The potential of conditioned approaches for such tasks has been somewhat hidden due to a lack of suitable data, an issue recently addressed with the MoisesDb dataset. A recent method, Banquet, employs this dataset with promising results seen on larger vocabularies. Banquet uses Bandsplit RNN rather than a UNet and the authors state that UNets should not be suitable for conditioned MSS. We counter this argument and propose QSCNet, a novel conditioned UNet for MSS that integrates network conditioning elements in the Sparse Compressed Network for MSS. We find QSCNet to outperform Banquet by over 1dB SNR on a couple of MSS tasks, while using less than half the number of parameters.", "date": "2025-12-17", "categories": []}
{"id": "2512.16271", "title": "Domain-Agnostic Causal-Aware Audio Transformer for Infant Cry Classification", "abstract": "Accurate and interpretable classification of infant cry paralinguistics is essential for early detection of neonatal distress and clinical decision support. However, many existing deep learning methods rely on correlation-driven acoustic representations, which makes them vulnerable to noise, spurious cues, and domain shifts across recording environments. We propose DACH-TIC, a Domain-Agnostic Causal-Aware Hierarchical Audio Transformer for robust infant cry classification. The model integrates causal attention, hierarchical representation learning, multi-task supervision, and adversarial domain generalization within a unified framework.\n  DACH-TIC employs a structured transformer backbone with local token-level and global semantic encoders, augmented by causal attention masking and controlled perturbation training to approximate counterfactual acoustic variations. A domain-adversarial objective promotes environment-invariant representations, while multi-task learning jointly optimizes cry type recognition, distress intensity estimation, and causal relevance prediction. The model is evaluated on the Baby Chillanto and Donate-a-Cry datasets, with ESC-50 environmental noise overlays for domain augmentation.\n  Experimental results show that DACH-TIC outperforms state-of-the-art baselines, including HTS-AT and SE-ResNet Transformer, achieving improvements of 2.6 percent in accuracy and 2.2 points in macro-F1 score, alongside enhanced causal fidelity. The model generalizes effectively to unseen acoustic environments, with a domain performance gap of only 2.4 percent, demonstrating its suitability for real-world neonatal acoustic monitoring systems.", "date": "2025-12-18", "categories": []}
{"id": "2509.24793", "title": "Sparse Autoencoders Make Audio Foundation Models more Explainable", "abstract": "Audio pretrained models are widely employed to solve various tasks in speech processing, sound event detection, or music information retrieval. However, the representations learned by these models are unclear, and their analysis mainly restricts to linear probing of the hidden representations. In this work, we explore the use of Sparse Autoencoders (SAEs) to analyze the hidden representations of pretrained models, focusing on a case study in singing technique classification. We first demonstrate that SAEs retain both information about the original representations and class labels, enabling their internal structure to provide insights into self-supervised learning systems. Furthermore, we show that SAEs enhance the disentanglement of vocal attributes, establishing them as an effective tool for identifying the underlying factors encoded in the representations.", "date": "2025-09-29", "categories": []}
{"id": "2512.13998", "title": "Memo2496: Expert-Annotated Dataset and Dual-View Adaptive Framework for Music Emotion Recognition", "abstract": "Music Emotion Recogniser (MER) research faces challenges due to limited high-quality annotated datasets and difficulties in addressing cross-track feature drift. This work presents two primary contributions to address these issues. Memo2496, a large-scale dataset, offers 2496 instrumental music tracks with continuous valence arousal labels, annotated by 30 certified music specialists. Annotation quality is ensured through calibration with extreme emotion exemplars and a consistency threshold of 0.25, measured by Euclidean distance in the valence arousal space. Furthermore, the Dual-view Adaptive Music Emotion Recogniser (DAMER) is introduced. DAMER integrates three synergistic modules: Dual Stream Attention Fusion (DSAF) facilitates token-level bidirectional interaction between Mel spectrograms and cochleagrams via cross attention mechanisms; Progressive Confidence Labelling (PCL) generates reliable pseudo labels employing curriculum-based temperature scheduling and consistency quantification using Jensen Shannon divergence; and Style Anchored Memory Learning (SAML) maintains a contrastive memory queue to mitigate cross-track feature drift. Extensive experiments on the Memo2496, 1000songs, and PMEmo datasets demonstrate DAMER's state-of-the-art performance, improving arousal dimension accuracy by 3.43%, 2.25%, and 0.17%, respectively. Ablation studies and visualisation analyses validate each module's contribution. Both the dataset and source code are publicly available.", "date": "2025-12-16", "categories": []}
{"id": "2512.15180", "title": "BEAT2AASIST model with layer fusion for ESDD 2026 Challenge", "abstract": "Recent advances in audio generation have increased the risk of realistic environmental sound manipulation, motivating the ESDD 2026 Challenge as the first large-scale benchmark for Environmental Sound Deepfake Detection (ESDD). We propose BEAT2AASIST which extends BEATs-AASIST by splitting BEATs-derived representations along frequency or channel dimension and processing them with dual AASIST branches. To enrich feature representations, we incorporate top-k transformer layer fusion using concatenation, CNN-gated, and SE-gated strategies. In addition, vocoder-based data augmentation is applied to improve robustness against unseen spoofing methods. Experimental results on the official test sets demonstrate that the proposed approach achieves competitive performance across the challenge tracks.", "date": "2025-12-17", "categories": []}
{"id": "2512.13744", "title": "Toward Noise-Aware Audio Deepfake Detection: Survey, SNR-Benchmarks, and Practical Recipes", "abstract": "Deepfake audio detection has progressed rapidly with strong pre-trained encoders (e.g., WavLM, Wav2Vec2, MMS). However, performance in realistic capture conditions - background noise (domestic/office/transport), room reverberation, and consumer channels - often lags clean-lab results. We survey and evaluate robustness for state-of-the-art audio deepfake detection models and present a reproducible framework that mixes MS-SNSD noises with ASVspoof 2021 DF utterances to evaluate under controlled signal-to-noise ratios (SNRs). SNR is a measured proxy for noise severity used widely in speech; it lets us sweep from near-clean (35 dB) to very noisy (-5 dB) to quantify graceful degradation. We study multi-condition training and fixed-SNR testing for pretrained encoders (WavLM, Wav2Vec2, MMS), reporting accuracy, ROC-AUC, and EER on binary and four-class (authenticity x corruption) tasks. In our experiments, finetuning reduces EER by 10-15 percentage points at 10-0 dB SNR across backbones.", "date": "2025-12-15", "categories": []}
{"id": "2512.13880", "title": "Privacy-Enhancing Infant Cry Classification with Federated Transformers and Denoising Regularization", "abstract": "Infant cry classification can aid early assessment of infant needs. However, deployment of such solutions is limited by privacy concerns around audio data, sensitivity to background noise, and domain shift across recording environments. We present an end-to-end infant cry analysis pipeline that integrates a denoising autoencoder (DAE), a convolutional tokenizer, and a Transformer encoder trained using communication-efficient federated learning (FL). The system performs on-device denoising, adaptive segmentation, post hoc calibration, and energy-based out-of-distribution (OOD) abstention. Federated training employs a regularized control variate update with 8-bit adapter deltas under secure aggregation. Using the Baby Chillanto and Donate-a-Cry datasets with ESC-50 noise overlays, the model achieves a macro F1 score of 0.938, an AUC of 0.962, and an Expected Calibration Error (ECE) of 0.032, while reducing per-round client upload from approximately 36 to 42 MB to 3.3 MB. Real-time edge inference on an NVIDIA Jetson Nano (4 GB, TensorRT FP16) achieves 96 ms per one-second spectrogram frame. These results demonstrate a practical path toward privacy-preserving, noise-robust, and communication-efficient infant cry classification suitable for federated deployment.", "date": "2025-12-15", "categories": []}
{"id": "2512.14115", "title": "Joint Multimodal Contrastive Learning for Robust Spoken Term Detection and Keyword Spotting", "abstract": "Acoustic Word Embeddings (AWEs) improve the efficiency of speech retrieval tasks such as Spoken Term Detection (STD) and Keyword Spotting (KWS). However, existing approaches suffer from limitations, including unimodal supervision, disjoint optimization of audio-audio and audio-text alignment, and the need for task-specific models. To address these shortcomings, we propose a joint multimodal contrastive learning framework that unifies both acoustic and cross-modal supervision in a shared embedding space. Our approach simultaneously optimizes: (i) audio-text contrastive learning, inspired by the CLAP loss, to align audio and text representations and (ii) audio-audio contrastive learning, via Deep Word Discrimination (DWD) loss, to enhance intra-class compactness and inter-class separation. The proposed method outperforms existing AWE baselines on word discrimination task while flexibly supporting both STD and KWS. To our knowledge, this is the first comprehensive approach of its kind.", "date": "2025-12-16", "categories": []}
{"id": "2512.12365", "title": "Synthetic Swarm Mosquito Dataset for Acoustic Classification: A Proof of Concept", "abstract": "Mosquito-borne diseases pose a serious global health threat, causing over 700,000 deaths annually. This work introduces a proof-of-concept Synthetic Swarm Mosquito Dataset for Acoustic Classification, created to simulate realistic multi-species and noisy swarm conditions. Unlike conventional datasets that require labor-intensive recording of individual mosquitoes, the synthetic approach enables scalable data generation while reducing human resource demands. Using log-mel spectrograms, we evaluated lightweight deep learning architectures for the classification of mosquito species. Experiments show that these models can effectively identify six major mosquito vectors and are suitable for deployment on embedded low-power devices. The study demonstrates the potential of synthetic swarm audio datasets to accelerate acoustic mosquito research and enable scalable real-time surveillance solutions.", "date": "2025-12-13", "categories": []}
{"id": "2512.13632", "title": "StutterFuse: Mitigating Modality Collapse in Stuttering Detection with Jaccard-Weighted Metric Learning and Gated Fusion", "abstract": "Stuttering detection breaks down when disfluencies overlap. Existing parametric models struggle to distinguish complex, simultaneous disfluencies (e.g., a 'block' with a 'prolongation') due to the scarcity of these specific combinations in training data. While Retrieval-Augmented Generation (RAG) has revolutionized NLP by grounding models in external knowledge, this paradigm remains unexplored in pathological speech processing. To bridge this gap, we introduce StutterFuse, the first Retrieval-Augmented Classifier (RAC) for multi-label stuttering detection. By conditioning a Conformer encoder on a non-parametric memory bank of clinical examples, we allow the model to classify by reference rather than memorization. We further identify and solve \"Modality Collapse\", an \"Echo Chamber\" effect where naive retrieval boosts recall but degrades precision. We mitigate this using: (1) SetCon, a Jaccard-Weighted Metric Learning objective that optimizes for multi-label set similarity, and (2) a Gated Mixture-of-Experts fusion strategy that dynamically arbitrates between acoustic evidence and retrieved context. On the SEP-28k dataset, StutterFuse achieves a weighted F1-score of 0.65, outperforming strong baselines and demonstrating remarkable zero-shot cross-lingual generalization.", "date": "2025-12-15", "categories": []}
{"id": "2512.11545", "title": "Graph Embedding with Mel-spectrograms for Underwater Acoustic Target Recognition", "abstract": "Underwater acoustic target recognition (UATR) is extremely challenging due to the complexity of ship-radiated noise and the variability of ocean environments. Although deep learning (DL) approaches have achieved promising results, most existing models implicitly assume that underwater acoustic data lie in a Euclidean space. This assumption, however, is unsuitable for the inherently complex topology of underwater acoustic signals, which exhibit non-stationary, non-Gaussian, and nonlinear characteristics. To overcome this limitation, this paper proposes the UATR-GTransformer, a non-Euclidean DL model that integrates Transformer architectures with graph neural networks (GNNs). The model comprises three key components: a Mel patchify block, a GTransformer block, and a classification head. The Mel patchify block partitions the Mel-spectrogram into overlapping patches, while the GTransformer block employs a Transformer Encoder to capture mutual information between split patches to generate Mel-graph embeddings. Subsequently, a GNN enhances these embeddings by modeling local neighborhood relationships, and a feed-forward network (FFN) further performs feature transformation. Experiments results based on two widely used benchmark datasets demonstrate that the UATR-GTransformer achieves performance competitive with state-of-the-art methods. In addition, interpretability analysis reveals that the proposed model effectively extracts rich frequency-domain information, highlighting its potential for applications in ocean engineering.", "date": "2025-12-12", "categories": []}
{"id": "2505.13847", "title": "Forensic deepfake audio detection using segmental speech features", "abstract": "This study explores the potential of using acoustic features of segmental speech sounds to detect deepfake audio. These features are highly interpretable because of their close relationship with human articulatory processes and are expected to be more difficult for deepfake models to replicate. The results demonstrate that certain segmental features commonly used in forensic voice comparison (FVC) are effective in identifying deep-fakes, whereas some global features provide little value. These findings underscore the need to approach audio deepfake detection using methods that are distinct from those employed in traditional FVC, and offer a new perspective on leveraging segmental features for this purpose. In addition, the present study proposes a speaker-specific framework for deepfake detection, which differs fundamentally from the speaker-independent systems that dominate current benchmarks. While speaker-independent frameworks aim at broad generalization, the speaker-specific approach offers advantages in forensic contexts where case-by-case interpretability and sensitivity to individual phonetic realization are essential.", "date": "2025-05-20", "categories": []}
{"id": "2512.10170", "title": "Semantic-Aware Confidence Calibration for Automated Audio Captioning", "abstract": "Automated audio captioning models frequently produce overconfident predictions regardless of semantic accuracy, limiting their reliability in deployment. This deficiency stems from two factors: evaluation metrics based on n-gram overlap that fail to capture semantic correctness, and the absence of calibrated confidence estimation. We present a framework that addresses both limitations by integrating confidence prediction into audio captioning and redefining correctness through semantic similarity. Our approach augments a Whisper-based audio captioning model with a learned confidence prediction head that estimates uncertainty from decoder hidden states. We employ CLAP audio-text embeddings and sentence transformer similarities (FENSE) to define semantic correctness, enabling Expected Calibration Error (ECE) computation that reflects true caption quality rather than surface-level text overlap. Experiments on Clotho v2 demonstrate that confidence-guided beam search with semantic evaluation achieves dramatically improved calibration (CLAP-based ECE of 0.071) compared to greedy decoding baselines (ECE of 0.488), while simultaneously improving caption quality across standard metrics. Our results establish that semantic similarity provides a more meaningful foundation for confidence calibration in audio captioning than traditional n-gram metrics.", "date": "2025-12-11", "categories": []}
{"id": "2512.10403", "title": "BRACE: A Benchmark for Robust Audio Caption Quality Evaluation", "abstract": "Automatic audio captioning is essential for audio understanding, enabling applications such as accessibility and content indexing. However, evaluating the quality of audio captions remains a major challenge, especially in reference-free settings where high-quality ground-truth captions are unavailable. While CLAPScore is currently the most widely used reference-free Audio Caption Evaluation Metric(ACEM), its robustness under diverse conditions has not been systematically validated.\n  To address this gap, we introduce BRACE, a new benchmark designed to evaluate audio caption alignment quality in a reference-free setting. BRACE is primarily designed for assessing ACEMs, and can also be extended to measure the modality alignment abilities of Large Audio Language Model(LALM). BRACE consists of two sub-benchmarks: BRACE-Main for fine-grained caption comparison and BRACE-Hallucination for detecting subtle hallucinated content. We construct these datasets through high-quality filtering, LLM-based corruption, and human annotation.\n  Given the widespread adoption of CLAPScore as a reference-free ACEM and the increasing application of LALMs in audio-language tasks, we evaluate both approaches using the BRACE benchmark, testing CLAPScore across various CLAP model variants and assessing multiple LALMs.\n  Notably, even the best-performing CLAP-based ACEM achieves only a 70.01 F1-score on the BRACE-Main benchmark, while the best LALM reaches just 63.19.\n  By revealing the limitations of CLAP models and LALMs, our BRACE benchmark offers valuable insights into the direction of future research.", "date": "2025-12-11", "categories": []}
{"id": "2503.10287", "title": "MACS: Multi-source Audio-to-image Generation with Contextual Significance and Semantic Alignment", "abstract": "Propelled by the breakthrough in deep generative models, audio-to-image generation has emerged as a pivotal cross-modal task that converts complex auditory signals into rich visual representations. However, previous works only focus on single-source audio inputs for image generation, ignoring the multi-source characteristic in natural auditory scenes, thus limiting the performance in generating comprehensive visual content. To bridge this gap, we propose a method called MACS to conduct multi-source audio-to-image generation. To our best knowledge, this is the first work that explicitly separates multi-source audio to capture the rich audio components before image generation. MACS is a two-stage method. In the first stage, multi-source audio inputs are separated by a weakly supervised method, where the audio and text labels are semantically aligned by casting into a common space using the large pre-trained CLAP model. We introduce a ranking loss to consider the contextual significance of the separated audio signals. In the second stage, effective image generation is achieved by mapping the separated audio signals to the generation condition using only a trainable adapter and a MLP layer. We preprocess the LLP dataset as the first full multi-source audio-to-image generation benchmark. The experiments are conducted on multi-source, mixed-source, and single-source audio-to-image generation tasks. The proposed MACS outperforms the current state-of-the-art methods in 17 out of the 21 evaluation indexes on all tasks and delivers superior visual quality.", "date": "2025-03-13", "categories": []}
{"id": "2512.08973", "title": "Enhancing Automatic Speech Recognition Through Integrated Noise Detection Architecture", "abstract": "This research presents a novel approach to enhancing automatic speech recognition systems by integrating noise detection capabilities directly into the recognition architecture. Building upon the wav2vec2 framework, the proposed method incorporates a dedicated noise identification module that operates concurrently with speech transcription. Experimental validation using publicly available speech and environmental audio datasets demonstrates substantial improvements in transcription quality and noise discrimination. The enhanced system achieves superior performance in word error rate, character error rate, and noise detection accuracy compared to conventional architectures. Results indicate that joint optimization of transcription and noise classification objectives yields more reliable speech recognition in challenging acoustic conditions.", "date": "2025-12-02", "categories": []}
{"id": "2509.04166", "title": "Crossing the Species Divide: Transfer Learning from Speech to Animal Sounds", "abstract": "Self-supervised speech models have demonstrated impressive performance in speech processing, but their effectiveness on non-speech data remains underexplored. We study the transfer learning capabilities of such models on bioacoustic detection and classification tasks. We show that models such as HuBERT, WavLM, and XEUS can generate rich latent representations of animal sounds across taxa. We analyze the models properties with linear probing on time-averaged representations. We then extend the approach to account for the effect of time-wise information with other downstream architectures. Finally, we study the implication of frequency range and noise on performance. Notably, our results are competitive with fine-tuned bioacoustic pre-trained models and show the impact of noise-robust pre-training setups. These findings highlight the potential of speech-based self-supervised learning as an efficient framework for advancing bioacoustic research.", "date": "2025-09-04", "categories": []}
{"id": "2512.08238", "title": "SpeechQualityLLM: LLM-Based Multimodal Assessment of Speech Quality", "abstract": "Objective speech quality assessment is central to telephony, VoIP, and streaming systems, where large volumes of degraded audio must be monitored and optimized at scale. Classical metrics such as PESQ and POLQA approximate human mean opinion scores (MOS) but require carefully controlled conditions and expensive listening tests, while learning-based models such as NISQA regress MOS and multiple perceptual dimensions from waveforms or spectrograms, achieving high correlation with subjective ratings yet remaining rigid: they do not support interactive, natural-language queries and do not natively provide textual rationales. In this work, we introduce SpeechQualityLLM, a multimodal speech quality question-answering (QA) system that couples an audio encoder with a language model and is trained on the NISQA corpus using template-based question-answer pairs covering overall MOS and four perceptual dimensions (noisiness, coloration, discontinuity, and loudness) in both single-ended (degraded only) and double-ended (degraded plus clean reference) setups. Instead of directly regressing scores, our system is supervised to generate textual answers from which numeric predictions are parsed and evaluated with standard regression and ranking metrics; on held-out NISQA clips, the double-ended model attains a MOS mean absolute error (MAE) of 0.41 with Pearson correlation of 0.86, with competitive performance on dimension-wise tasks. Beyond these quantitative gains, it offers a flexible natural-language interface in which the language model acts as an audio quality expert: practitioners can query arbitrary aspects of degradations, prompt the model to emulate different listener profiles to capture human variability and produce diverse but plausible judgments rather than a single deterministic score, and thereby reduce reliance on large-scale crowdsourced tests and their monetary cost.", "date": "2025-12-09", "categories": []}
{"id": "2512.06040", "title": "Physics-Guided Deepfake Detection for Voice Authentication Systems", "abstract": "Voice authentication systems deployed at the network edge face dual threats: a) sophisticated deepfake synthesis attacks and b) control-plane poisoning in distributed federated learning protocols. We present a framework coupling physics-guided deepfake detection with uncertainty-aware in edge learning. The framework fuses interpretable physics features modeling vocal tract dynamics with representations coming from a self-supervised learning module. The representations are then processed via a Multi-Modal Ensemble Architecture, followed by a Bayesian ensemble providing uncertainty estimates. Incorporating physics-based characteristics evaluations and uncertainty estimates of audio samples allows our proposed framework to remain robust to both advanced deepfake attacks and sophisticated control-plane poisoning, addressing the complete threat model for networked voice authentication.", "date": "2025-12-04", "categories": []}
{"id": "2512.06304", "title": "Degrading Voice: A Comprehensive Overview of Robust Voice Conversion Through Input Manipulation", "abstract": "Identity, accent, style, and emotions are essential components of human speech. Voice conversion (VC) techniques process the speech signals of two input speakers and other modalities of auxiliary information such as prompts and emotion tags. It changes para-linguistic features from one to another, while maintaining linguistic contents. Recently, VC models have made rapid advancements in both generation quality and personalization capabilities. These developments have attracted considerable attention for diverse applications, including privacy preservation, voice-print reproduction for the deceased, and dysarthric speech recovery. However, these models only learn non-robust features due to the clean training data. Subsequently, it results in unsatisfactory performances when dealing with degraded input speech in real-world scenarios, including additional noise, reverberation, adversarial attacks, or even minor perturbation. Hence, it demands robust deployments, especially in real-world settings. Although latest researches attempt to find potential attacks and countermeasures for VC systems, there remains a significant gap in the comprehensive understanding of how robust the VC model is under input manipulation. here also raises many questions: For instance, to what extent do different forms of input degradation attacks alter the expected output of VC models? Is there potential for optimizing these attack and defense strategies? To answer these questions, we classify existing attack and defense methods from the perspective of input manipulation and evaluate the impact of degraded input speech across four dimensions, including intelligibility, naturalness, timbre similarity, and subjective perception. Finally, we outline open issues and future directions.", "date": "2025-12-06", "categories": []}
{"id": "2512.06999", "title": "Singing Timbre Popularity Assessment Based on Multimodal Large Foundation Model", "abstract": "Automated singing assessment is crucial for education and entertainment. However, existing systems face two fundamental limitations: reliance on reference tracks, which stifles creative expression, and the simplification of complex performances into non-diagnostic scores based solely on pitch and rhythm. We advocate for a shift from discriminative to descriptive evaluation, creating a complete ecosystem for reference-free, multi-dimensional assessment. First, we introduce Sing-MD, a large-scale dataset annotated by experts across four dimensions: breath control, timbre quality, emotional expression, and vocal technique. Our analysis reveals significant annotation inconsistencies among experts, challenging the validity of traditional accuracy-based metrics. Second, addressing the memory limitations of Multimodal Large Language Models (MLLMs) in analyzing full-length songs, we propose VocalVerse. This efficient hybrid architecture leverages a lightweight acoustic encoder to model global performance features and long-term dependencies. Third, to address automated metric shortcomings, we establish the H-TPR (Human-in-the-loop Tiered Perceptual Ranking) benchmark, which evaluates a model's ability to generate perceptually valid rankings rather than predicting noisy ground-truth scores.", "date": "2025-12-07", "categories": []}
{"id": "2512.07393", "title": "Empirical Results for Adjusting Truncated Backpropagation Through Time while Training Neural Audio Effects", "abstract": "This paper investigates the optimization of Truncated Backpropagation Through Time (TBPTT) for training neural networks in digital audio effect modeling, with a focus on dynamic range compression. The study evaluates key TBPTT hyperparameters -- sequence number, batch size, and sequence length -- and their influence on model performance. Using a convolutional-recurrent architecture, we conduct extensive experiments across datasets with and without conditionning by user controls. Results demonstrate that carefully tuning these parameters enhances model accuracy and training stability, while also reducing computational demands. Objective evaluations confirm improved performance with optimized settings, while subjective listening tests indicate that the revised TBPTT configuration maintains high perceptual quality.", "date": "2025-12-08", "categories": []}
{"id": "2512.03219", "title": "Perch 2.0 transfers 'whale' to underwater tasks", "abstract": "Perch 2.0 is a supervised bioacoustics foundation model pretrained on 14,597 species, including birds, mammals, amphibians, and insects, and has state-of-the-art performance on multiple benchmarks. Given that Perch 2.0 includes almost no marine mammal audio or classes in the training data, we evaluate Perch 2.0 performance on marine mammal and underwater audio tasks through few-shot transfer learning. We perform linear probing with the embeddings generated from this foundation model and compare performance to other pretrained bioacoustics models. In particular, we compare Perch 2.0 with previous multispecies whale, Perch 1.0, SurfPerch, AVES-bio, BirdAVES, and Birdnet V2.3 models, which have open-source tools for transfer-learning and agile modeling. We show that the embeddings from the Perch 2.0 model have consistently high performance for few-shot transfer learning, generally outperforming alternative embedding models on the majority of tasks, and thus is recommended when developing new linear classifiers for marine mammal classification with few labeled examples.", "date": "2025-12-02", "categories": []}
{"id": "2505.21025", "title": "Text-Queried Audio Source Separation via Hierarchical Modeling", "abstract": "Target audio source separation with natural language queries presents a promising paradigm for extracting arbitrary audio events through arbitrary text descriptions. Existing methods mainly face two challenges, the difficulty in jointly modeling acoustic-textual alignment and semantic-aware separation within a blindly-learned single-stage architecture, and the reliance on large-scale accurately-labeled training data to compensate for inefficient cross-modal learning and separation. To address these challenges, we propose a hierarchical decomposition framework, HSM-TSS, that decouples the task into global-local semantic-guided feature separation and structure-preserving acoustic reconstruction. Our approach introduces a dual-stage mechanism for semantic separation, operating on distinct global and local semantic feature spaces. We first perform global-semantic separation through a global semantic feature space aligned with text queries. A Q-Audio architecture is employed to align audio and text modalities, serving as pretrained global-semantic encoders. Conditioned on the predicted global feature, we then perform the second-stage local-semantic separation on AudioMAE features that preserve time-frequency structures, followed by acoustic reconstruction. We also propose an instruction processing pipeline to parse arbitrary text queries into structured operations, extraction or removal, coupled with audio descriptions, enabling flexible sound manipulation. Our method achieves state-of-the-art separation performance with data-efficient training while maintaining superior semantic consistency with queries in complex auditory scenes.", "date": "2025-05-27", "categories": []}
{"id": "2512.02669", "title": "SAND Challenge: Four Approaches for Dysartria Severity Classification", "abstract": "This paper presents a unified study of four distinct modeling approaches for classifying dysarthria severity in the Speech Analysis for Neurodegenerative Diseases (SAND) challenge. All models tackle the same five class classification task using a common dataset of speech recordings. We investigate: (1) a ViT-OF method leveraging a Vision Transformer on spectrogram images, (2) a 1D-CNN approach using eight 1-D CNN's with majority-vote fusion, (3) a BiLSTM-OF approach using nine BiLSTM models with majority vote fusion, and (4) a Hierarchical XGBoost ensemble that combines glottal and formant features through a two stage learning framework. Each method is described, and their performances on a validation set of 53 speakers are compared. Results show that while the feature-engineered XGBoost ensemble achieves the highest macro-F1 (0.86), the deep learning models (ViT, CNN, BiLSTM) attain competitive F1-scores (0.70) and offer complementary insights into the problem.", "date": "2025-12-02", "categories": []}
{"id": "2512.03036", "title": "ViSAudio: End-to-End Video-Driven Binaural Spatial Audio Generation", "abstract": "Despite progress in video-to-audio generation, the field focuses predominantly on mono output, lacking spatial immersion. Existing binaural approaches remain constrained by a two-stage pipeline that first generates mono audio and then performs spatialization, often resulting in error accumulation and spatio-temporal inconsistencies. To address this limitation, we introduce the task of end-to-end binaural spatial audio generation directly from silent video. To support this task, we present the BiAudio dataset, comprising approximately 97K video-binaural audio pairs spanning diverse real-world scenes and camera rotation trajectories, constructed through a semi-automated pipeline. Furthermore, we propose ViSAudio, an end-to-end framework that employs conditional flow matching with a dual-branch audio generation architecture, where two dedicated branches model the audio latent flows. Integrated with a conditional spacetime module, it balances consistency between channels while preserving distinctive spatial characteristics, ensuring precise spatio-temporal alignment between audio and the input video. Comprehensive experiments demonstrate that ViSAudio outperforms existing state-of-the-art methods across both objective metrics and subjective evaluations, generating high-quality binaural audio with spatial immersion that adapts effectively to viewpoint changes, sound-source motion, and diverse acoustic environments. Project website: https://kszpxxzmc.github.io/ViSAudio-project.", "date": "2025-12-02", "categories": []}
{"id": "2511.21237", "title": "3-Tracer: A Tri-level Temporal-Aware Framework for Audio Forgery Detection and Localization", "abstract": "Recently, partial audio forgery has emerged as a new form of audio manipulation. Attackers selectively modify partial but semantically critical frames while preserving the overall perceptual authenticity, making such forgeries particularly difficult to detect. Existing methods focus on independently detecting whether a single frame is forged, lacking the hierarchical structure to capture both transient and sustained anomalies across different temporal levels. To address these limitations, We identify three key levels relevant to partial audio forgery detection and present T3-Tracer, the first framework that jointly analyzes audio at the frame, segment, and audio levels to comprehensively detect forgery traces. T3-Tracer consists of two complementary core modules: the Frame-Audio Feature Aggregation Module (FA-FAM) and the Segment-level Multi-Scale Discrepancy-Aware Module (SMDAM). FA-FAM is designed to detect the authenticity of each audio frame. It combines both frame-level and audio-level temporal information to detect intra-frame forgery cues and global semantic inconsistencies. To further refine and correct frame detection, we introduce SMDAM to detect forgery boundaries at the segment level. It adopts a dual-branch architecture that jointly models frame features and inter-frame differences across multi-scale temporal windows, effectively identifying abrupt anomalies that appeared on the forged boundaries. Extensive experiments conducted on three challenging datasets demonstrate that our approach achieves state-of-the-art performance.", "date": "2025-11-26", "categories": []}
{"id": "2512.00563", "title": "Explainable Multi-Modal Deep Learning for Automatic Detection of Lung Diseases from Respiratory Audio Signals", "abstract": "Respiratory diseases remain major global health challenges, and traditional auscultation is often limited by subjectivity, environmental noise, and inter-clinician variability. This study presents an explainable multimodal deep learning framework for automatic lung-disease detection using respiratory audio signals. The proposed system integrates two complementary representations: a spectral-temporal encoder based on a CNN-BiLSTM Attention architecture, and a handcrafted acoustic-feature encoder capturing physiologically meaningful descriptors such as MFCCs, spectral centroid, spectral bandwidth, and zero-crossing rate. These branches are combined through late-stage fusion to leverage both data-driven learning and domain-informed acoustic cues. The model is trained and evaluated on the Asthma Detection Dataset Version 2 using rigorous preprocessing, including resampling, normalization, noise filtering, data augmentation, and patient-level stratified partitioning. The study achieved strong generalization with 91.21% accuracy, 0.899 macro F1-score, and 0.9866 macro ROC-AUC, outperforming all ablated variants. An ablation study confirms the importance of temporal modeling, attention mechanisms, and multimodal fusion. The framework incorporates Grad-CAM, Integrated Gradients, and SHAP, generating interpretable spectral, temporal, and feature-level explanations aligned with known acoustic biomarkers to build clinical transparency. The findings demonstrate the framework's potential for telemedicine, point-of-care diagnostics, and real-world respiratory screening.", "date": "2025-11-29", "categories": []}
{"id": "2512.00621", "title": "Melody or Machine: Detecting Synthetic Music with Dual-Stream Contrastive Learning", "abstract": "The rapid evolution of end-to-end AI music generation poses an escalating threat to artistic authenticity and copyright, demanding detection methods that can keep pace. While foundational, existing models like SpecTTTra falter when faced with the diverse and rapidly advancing ecosystem of new generators, exhibiting significant performance drops on out-of-distribution (OOD) content. This generalization failure highlights a critical gap: the need for more challenging benchmarks and more robust detection architectures. To address this, we first introduce Melody or Machine (MoM), a new large-scale benchmark of over 130,000 songs (6,665 hours). MoM is the most diverse dataset to date, built with a mix of open and closed-source models and a curated OOD test set designed specifically to foster the development of truly generalizable detectors. Alongside this benchmark, we introduce CLAM, a novel dual-stream detection architecture. We hypothesize that subtle, machine-induced inconsistencies between vocal and instrumental elements, often imperceptible in a mixed signal, offer a powerful tell-tale sign of synthesis. CLAM is designed to test this hypothesis by employing two distinct pre-trained audio encoders (MERT and Wave2Vec2) to create parallel representations of the audio. These representations are fused by a learnable cross-aggregation module that models their inter-dependencies. The model is trained with a dual-loss objective: a standard binary cross-entropy loss for classification, complemented by a contrastive triplet loss which trains the model to distinguish between coherent and artificially mismatched stream pairings, enhancing its sensitivity to synthetic artifacts without presuming a simple feature alignment. CLAM establishes a new state-of-the-art in synthetic music forensics. It achieves an F1 score of 0.925 on our challenging MoM benchmark.", "date": "2025-11-29", "categories": []}
{"id": "2507.04966", "title": "LAPS-Diff: A Diffusion-Based Framework for Singing Voice Synthesis With Language Aware Prosody-Style Guided Learning", "abstract": "The field of Singing Voice Synthesis (SVS) has seen significant advancements in recent years due to the rapid progress of diffusion-based approaches. However, capturing vocal style, genre-specific pitch inflections, and language-dependent characteristics remains challenging, particularly in low-resource scenarios. To address this, we propose LAPS-Diff, a diffusion model integrated with language-aware embeddings and a vocal-style guided learning mechanism, specifically designed for Bollywood Hindi singing style. We curate a Hindi SVS dataset and leverage pre-trained language models to extract word and phone-level embeddings for an enriched lyrics representation. Additionally, we incorporated a style encoder and a pitch extraction model to compute style and pitch losses, capturing features essential to the naturalness and expressiveness of the synthesized singing, particularly in terms of vocal style and pitch variations. Furthermore, we utilize MERT and IndicWav2Vec models to extract musical and contextual embeddings, serving as conditional priors to refine the acoustic feature generation process further. Based on objective and subjective evaluations, we demonstrate that LAPS-Diff significantly improves the quality of the generated samples compared to the considered state-of-the-art (SOTA) model for our constrained dataset that is typical of the low resource scenario.", "date": "2025-07-07", "categories": []}
{"id": "2510.24693", "title": "STAR-Bench: Probing Deep Spatio-Temporal Reasoning as Audio 4D Intelligence", "abstract": "Despite rapid progress in Multi-modal Large Language Models and Large Audio-Language Models, existing audio benchmarks largely test semantics that can be recovered from text captions, masking deficits in fine-grained perceptual reasoning. We formalize audio 4D intelligence that is defined as reasoning over sound dynamics in time and 3D space, and introduce STAR-Bench to measure it. STAR-Bench combines a Foundational Acoustic Perception setting (six attributes under absolute and relative regimes) with a Holistic Spatio-Temporal Reasoning setting that includes segment reordering for continuous and discrete processes and spatial tasks spanning static localization, multi-source relations, and dynamic trajectories. Our data curation pipeline uses two methods to ensure high-quality samples. For foundational tasks, we use procedurally synthesized and physics-simulated audio. For holistic data, we follow a four-stage process that includes human annotation and final selection based on human performance. Unlike prior benchmarks where caption-only answering reduces accuracy slightly, STAR-Bench induces far larger drops (-31.5\\% temporal, -35.2\\% spatial), evidencing its focus on linguistically hard-to-describe cues. Evaluating 19 models reveals substantial gaps compared with humans and a capability hierarchy: closed-source models are bottlenecked by fine-grained perception, while open-source models lag across perception, knowledge, and reasoning. Our STAR-Bench provides critical insights and a clear path forward for developing future models with a more robust understanding of the physical world.", "date": "2025-10-28", "categories": []}
{"id": "2511.21872", "title": "Advancing Marine Bioacoustics with Deep Generative Models: A Hybrid Augmentation Strategy for Southern Resident Killer Whale Detection", "abstract": "Automated detection and classification of marine mammals vocalizations is critical for conservation and management efforts but is hindered by limited annotated datasets and the acoustic complexity of real-world marine environments. Data augmentation has proven to be an effective strategy to address this limitation by increasing dataset diversity and improving model generalization without requiring additional field data. However, most augmentation techniques used to date rely on effective but relatively simple transformations, leaving open the question of whether deep generative models can provide additional benefits. In this study, we evaluate the potential of deep generative for data augmentation in marine mammal call detection including: Variational Autoencoders, Generative Adversarial Networks, and Denoising Diffusion Probabilistic Models. Using Southern Resident Killer Whale (Orcinus orca) vocalizations from two long-term hydrophone deployments in the Salish Sea, we compare these approaches against traditional augmentation methods such as time-shifting and vocalization masking. While all generative approaches improved classification performance relative to the baseline, diffusion-based augmentation yielded the highest recall (0.87) and overall F1-score (0.75). A hybrid strategy combining generative-based synthesis with traditional methods achieved the best overall performance with an F1-score of 0.81. We hope this study encourages further exploration of deep generative models as complementary augmentation strategies to advance acoustic monitoring of threatened marine mammal populations.", "date": "2025-11-26", "categories": []}
{"id": "2511.15848", "title": "Step-Audio-R1 Technical Report", "abstract": "Recent advances in reasoning models have demonstrated remarkable success in text and vision domains through extended chain-of-thought deliberation. However, a perplexing phenomenon persists in audio language models: they consistently perform better with minimal or no reasoning, raising a fundamental question - can audio intelligence truly benefit from deliberate thinking? We introduce Step-Audio-R1, the first audio reasoning model that successfully unlocks reasoning capabilities in the audio domain. Through our proposed Modality-Grounded Reasoning Distillation (MGRD) framework, Step-Audio-R1 learns to generate audio-relevant reasoning chains that genuinely ground themselves in acoustic features rather than hallucinating disconnected deliberations. Our model exhibits strong audio reasoning capabilities, surpassing Gemini 2.5 Pro and achieving performance comparable to the state-of-the-art Gemini 3 Pro across comprehensive audio understanding and reasoning benchmarks spanning speech, environmental sounds, and music. These results demonstrate that reasoning is a transferable capability across modalities when appropriately anchored, transforming extended deliberation from a liability into a powerful asset for audio intelligence. By establishing the first successful audio reasoning model, Step-Audio-R1 opens new pathways toward building truly multimodal reasoning systems that think deeply across all sensory modalities.", "date": "2025-11-19", "categories": []}
{"id": "2511.20973", "title": "Towards Audio Token Compression in Large Audio Language Models", "abstract": "Large Audio Language Models (LALMs) demonstrate impressive performance across diverse tasks, ranging from speech recognition to general audio understanding. However, their scalability is limited by the quadratic complexity of attention and the high token rates of audio signals. These challenges make it difficult to extend LALMs to long-form audio and to deploy them on resource-constrained platforms such as edge devices.\n  In this paper, we explore techniques such as unsupervised segmentation, uniform average pooling, etc., to reduce the number of audio tokens generated by the LALM's audio encoder but before they are consumed by the LLM decoder. To mitigate potential performance degradation introduced by the compressed representations, we employ low-rank adapters to finetune the model. We evaluate our proposed models on two tasks, automatic speech recognition and speech-to-speech translation tasks, that are dependent on effectively uncovering the underlying lexical content of the input signal and study the effect of downsampling on these tasks. Experimental results show that compressed LALMs can achieve performance closer to frame-level LALMs while reducing the input audio token count upto three times before the LLM backbone.", "date": "2025-11-26", "categories": []}
{"id": "2511.21146", "title": "AV-Edit: Multimodal Generative Sound Effect Editing via Audio-Visual Semantic Joint Control", "abstract": "Sound effect editing-modifying audio by adding, removing, or replacing elements-remains constrained by existing approaches that rely solely on low-level signal processing or coarse text prompts, often resulting in limited flexibility and suboptimal audio quality. To address this, we propose AV-Edit, a generative sound effect editing framework that enables fine-grained editing of existing audio tracks in videos by jointly leveraging visual, audio, and text semantics. Specifically, the proposed method employs a specially designed contrastive audio-visual masking autoencoder (CAV-MAE-Edit) for multimodal pre-training, learning aligned cross-modal representations. These representations are then used to train an editorial Multimodal Diffusion Transformer (MM-DiT) capable of removing visually irrelevant sounds and generating missing audio elements consistent with video content through a correlation-based feature gating training strategy. Furthermore, we construct a dedicated video-based sound editing dataset as an evaluation benchmark. Experiments demonstrate that the proposed AV-Edit generates high-quality audio with precise modifications based on visual content, achieving state-of-the-art performance in the field of sound effect editing and exhibiting strong competitiveness in the domain of audio generation.", "date": "2025-11-26", "categories": []}
{"id": "2511.21325", "title": "SONAR: Spectral-Contrastive Audio Residuals for Generalizable Deepfake Detection", "abstract": "Deepfake (DF) audio detectors still struggle to generalize to out of distribution inputs. A central reason is spectral bias, the tendency of neural networks to learn low-frequency structure before high-frequency (HF) details, which both causes DF generators to leave HF artifacts and leaves those same artifacts under-exploited by common detectors. To address this gap, we propose Spectral-cONtrastive Audio Residuals (SONAR), a frequency-guided framework that explicitly disentangles an audio signal into complementary representations. An XLSR encoder captures the dominant low-frequency content, while the same cloned path, preceded by learnable SRM, value-constrained high-pass filters, distills faint HF residuals. Frequency cross-attention reunites the two views for long- and short-range frequency dependencies, and a frequency-aware Jensen-Shannon contrastive loss pulls real content-noise pairs together while pushing fake embeddings apart, accelerating optimization and sharpening decision boundaries. Evaluated on the ASVspoof 2021 and in-the-wild benchmarks, SONAR attains state-of-the-art performance and converges four times faster than strong baselines. By elevating faint high-frequency residuals to first-class learning signals, SONAR unveils a fully data-driven, frequency-guided contrastive framework that splits the latent space into two disjoint manifolds: natural-HF for genuine audio and distorted-HF for synthetic audio, thereby sharpening decision boundaries. Because the scheme operates purely at the representation level, it is architecture-agnostic and, in future work, can be seamlessly integrated into any model or modality where subtle high-frequency cues are decisive.", "date": "2025-11-26", "categories": []}
{"id": "2511.21342", "title": "Generating Separated Singing Vocals Using a Diffusion Model Conditioned on Music Mixtures", "abstract": "Separating the individual elements in a musical mixture is an essential process for music analysis and practice. While this is generally addressed using neural networks optimized to mask or transform the time-frequency representation of a mixture to extract the target sources, the flexibility and generalization capabilities of generative diffusion models are giving rise to a novel class of solutions for this complicated task. In this work, we explore singing voice separation from real music recordings using a diffusion model which is trained to generate the solo vocals conditioned on the corresponding mixture. Our approach improves upon prior generative systems and achieves competitive objective scores against non-generative baselines when trained with supplementary data. The iterative nature of diffusion sampling enables the user to control the quality-efficiency trade-off, and also refine the output when needed. We present an ablation study of the sampling algorithm, highlighting the effects of the user-configurable parameters.", "date": "2025-11-26", "categories": []}
{"id": "2511.20470", "title": "Efficient and Fast Generative-Based Singing Voice Separation using a Latent Diffusion Model", "abstract": "Extracting individual elements from music mixtures is a valuable tool for music production and practice. While neural networks optimized to mask or transform mixture spectrograms into the individual source(s) have been the leading approach, the source overlap and correlation in music signals poses an inherent challenge. Also, accessing all sources in the mixture is crucial to train these systems, while complicated. Attempts to address these challenges in a generative fashion exist, however, the separation performance and inference efficiency remain limited. In this work, we study the potential of diffusion models to advance toward bridging this gap, focusing on generative singing voice separation relying only on corresponding pairs of isolated vocals and mixtures for training. To align with creative workflows, we leverage latent diffusion: the system generates samples encoded in a compact latent space, and subsequently decodes these into audio. This enables efficient optimization and faster inference. Our system is trained using only open data. We outperform existing generative separation systems, and level the compared non-generative systems on a list of signal quality measures and on interference removal. We provide a noise robustness study on the latent encoder, providing insights on its potential for the task. We release a modular toolkit for further research on the topic.", "date": "2025-11-25", "categories": []}
{"id": "2506.03425", "title": "A Data-Driven Diffusion-based Approach for Audio Deepfake Explanations", "abstract": "Evaluating explainability techniques, such as SHAP and LRP, in the context of audio deepfake detection is challenging due to lack of clear ground truth annotations. In the cases when we are able to obtain the ground truth, we find that these methods struggle to provide accurate explanations. In this work, we propose a novel data-driven approach to identify artifact regions in deepfake audio. We consider paired real and vocoded audio, and use the difference in time-frequency representation as the ground-truth explanation. The difference signal then serves as a supervision to train a diffusion model to expose the deepfake artifacts in a given vocoded audio. Experimental results on the VocV4 and LibriSeVoc datasets demonstrate that our method outperforms traditional explainability techniques, both qualitatively and quantitatively.", "date": "2025-06-03", "categories": []}
{"id": "2510.19368", "title": "AMAuT: A Flexible and Efficient Multiview Audio Transformer Framework Trained from Scratch", "abstract": "Recent foundational models, SSAST, EAT, HuBERT, Qwen-Audio, and Audio Flamingo, achieve top-tier results across standard audio benchmarks but are limited by fixed input rates and durations, hindering their reusability. This paper introduces the Augmentation-driven Multiview Audio Transformer (AMAuT), a training-from-scratch framework that eliminates the dependency on pre-trained weights while supporting arbitrary sample rates and audio lengths. AMAuT integrates four key components: (1) augmentation-driven multiview learning for robustness, (2) a conv1 + conv7 + conv1 one-dimensional CNN bottleneck for stable temporal encoding, (3) dual CLS + TAL tokens for bidirectional context representation, and (4) test-time adaptation/augmentation (TTA^2) to improve inference reliability. Experiments on five public benchmarks, AudioMNIST, SpeechCommands V1 & V2, VocalSound, and CochlScene, show that AMAuT achieves accuracies up to 99.8% while consuming less than 3% of the GPU hours required by comparable pre-trained models. Thus, AMAuT presents a highly efficient and flexible alternative to large pre-trained models, making state-of-the-art audio classification accessible in computationally constrained settings.", "date": "2025-10-22", "categories": []}
{"id": "2511.18487", "title": "InstructAudio: Unified speech and music generation with natural language instruction", "abstract": "Text-to-speech (TTS) and text-to-music (TTM) models face significant limitations in instruction-based control. TTS systems usually depend on reference audio for timbre, offer only limited text-level attribute control, and rarely support dialogue generation. TTM systems are constrained by input conditioning requirements that depend on expert knowledge annotations. The high heterogeneity of these input control conditions makes them difficult to joint modeling with speech synthesis. Despite sharing common acoustic modeling characteristics, these two tasks have long been developed independently, leaving open the challenge of achieving unified modeling through natural language instructions. We introduce InstructAudio, a unified framework that enables instruction-based (natural language descriptions) control of acoustic attributes including timbre (gender, age), paralinguistic (emotion, style, accent), and musical (genre, instrument, rhythm, atmosphere). It supports expressive speech, music, and dialogue generation in English and Chinese. The model employs joint and single diffusion transformer layers with a standardized instruction-phoneme input format, trained on 50K hours of speech and 20K hours of music data, enabling multi-task learning and cross-modal alignment. Fig. 1 visualizes performance comparisons with mainstream TTS and TTM models, demonstrating that InstructAudio achieves optimal results on most metrics. To our best knowledge, InstructAudio represents the first instruction-controlled framework unifying speech and music generation. Audio samples are available at: https://qiangchunyu.github.io/InstructAudio/", "date": "2025-11-23", "categories": []}
{"id": "2511.16757", "title": "Revisiting Audio-language Pretraining for Learning General-purpose Audio Representation", "abstract": "Audio-language pretraining holds promise for general-purpose audio understanding, yet remains underexplored compared to its vision counterpart. While vision-language models like CLIP serve as widely adopted foundations, existing audio-language models primarily excel at retrieval tasks with limited adoption as general-purpose encoders. We identify three key barriers: limited large-scale audio-text corpora, insufficient caption diversity, and lack of systematic exploration and evaluation. To this end, we introduce CaptionStew, a 10.7M caption dataset aggregating diverse open-source audio-text corpora across multiple domains and captioning styles. Using this resource, we conduct the first comprehensive evaluation comparing contrastive and captioning objectives for audio representation learning across speech, music, and environmental sound tasks. Our results demonstrate that audio-language pretraining yields competitive, transferable representations. Through systematic data-scaling experiments, we reveal complementary objective strengths: contrastive learning achieves superior data efficiency at smaller scales, while captioning demonstrates better scalability on language-involved audio understanding tasks. We also find that common supervised initialization practices provide diminishing returns at scale, challenging current approaches. These findings establish audio-language pretraining as a viable pathway toward general-purpose audio representations, guiding future research. To accelerate progress, we release data preparation recipes, training protocols, and pretrained models, paving the way toward universal audio understanding.", "date": "2025-11-20", "categories": []}
{"id": "2511.16848", "title": "Sex and age determination in European lobsters using AI-Enhanced bioacoustics", "abstract": "Monitoring aquatic species, especially elusive ones like lobsters, presents challenges. This study focuses on Homarus gammarus (European lobster), a key species for fisheries and aquaculture, and leverages non-invasive Passive Acoustic Monitoring (PAM). Understanding lobster habitats, welfare, reproduction, sex, and age is crucial for management and conservation. While bioacoustic emissions have classified various aquatic species using Artificial Intelligence (AI) models, this research specifically uses H. gammarus bioacoustics (buzzing/carapace vibrations) to classify lobsters by age (juvenile/adult) and sex (male/female).\n  The dataset was collected at Johnshaven, Scotland, using hydrophones in concrete tanks. We explored the efficacy of Deep Learning (DL) models (1D-CNN, 1D-DCNN) and six Machine Learning (ML) models (SVM, k-NN, Naive Bayes, Random Forest, XGBoost, MLP). Mel-frequency cepstral coefficients (MFCCs) were used as features.\n  For age classification (adult vs. juvenile), most models achieved over 97% accuracy (Naive Bayes: 91.31%). For sex classification, all models except Naive Bayes surpassed 93.23%. These strong results demonstrate the potential of supervised ML and DL to extract age- and sex-related features from lobster sounds. This research offers a promising non-invasive PAM approach for lobster conservation, detection, and management in aquaculture and fisheries, enabling real-world edge computing applications for underwater species.", "date": "2025-11-20", "categories": []}
{"id": "2511.17136", "title": "Device-Guided Music Transfer", "abstract": "Device-guided music transfer adapts playback across unseen devices for users who lack them. Existing methods mainly focus on modifying the timbre, rhythm, harmony, or instrumentation to mimic genres or artists, overlooking the diverse hardware properties of the playback device (i.e., speaker). Therefore, we propose DeMT, which processes a speaker's frequency response curve as a line graph using a vision-language model to extract device embeddings. These embeddings then condition a hybrid transformer via feature-wise linear modulation. Fine-tuned on a self-collected dataset, DeMT enables effective speaker-style transfer and robust few-shot adaptation for unseen devices, supporting applications like device-style augmentation and quality enhancement.", "date": "2025-11-21", "categories": []}
{"id": "2511.16639", "title": "Codec2Vec: Self-Supervised Speech Representation Learning Using Neural Speech Codecs", "abstract": "Recent advancements in neural audio codecs have not only enabled superior audio compression but also enhanced speech synthesis techniques. Researchers are now exploring their potential as universal acoustic feature extractors for a broader range of speech processing tasks. Building on this trend, we introduce Codec2Vec, the first speech representation learning framework that relies exclusively on discrete audio codec units. This approach offers several advantages, including improved data storage and transmission efficiency, faster training, and enhanced data privacy. We explore masked prediction with various training target derivation strategies to thoroughly understand the effectiveness of this framework. Evaluated on the SUPERB benchmark, Codec2Vec achieves competitive performance compared to continuous-input models while reducing storage requirements by up to 16.5x and training time by 2.3x, showcasing its scalability and efficiency.", "date": "2025-11-20", "categories": []}
{"id": "2511.14939", "title": "Fine-tuning Pre-trained Audio Models for COVID-19 Detection: A Technical Report", "abstract": "This technical report investigates the performance of pre-trained audio models on COVID-19 detection tasks using established benchmark datasets. We fine-tuned Audio-MAE and three PANN architectures (CNN6, CNN10, CNN14) on the Coswara and COUGHVID datasets, evaluating both intra-dataset and cross-dataset generalization. We implemented a strict demographic stratification by age and gender to prevent models from exploiting spurious correlations between demographic characteristics and COVID-19 status. Intra-dataset results showed moderate performance, with Audio-MAE achieving the strongest result on Coswara (0.82 AUC, 0.76 F1-score), while all models demonstrated limited performance on Coughvid (AUC 0.58-0.63). Cross-dataset evaluation revealed severe generalization failure across all models (AUC 0.43-0.68), with Audio-MAE showing strong performance degradation (F1-score 0.00-0.08). Our experiments demonstrate that demographic balancing, while reducing apparent model performance, provides more realistic assessment of COVID-19 detection capabilities by eliminating demographic leakage - a confounding factor that inflate performance metrics. Additionally, the limited dataset sizes after balancing (1,219-2,160 samples) proved insufficient for deep learning models that typically require substantially larger training sets. These findings highlight fundamental challenges in developing generalizable audio-based COVID-19 detection systems and underscore the importance of rigorous demographic controls for clinically robust model evaluation.", "date": "2025-11-18", "categories": []}
{"id": "2508.02175", "title": "Hidden in the Noise: Unveiling Backdoors in Audio LLMs Alignment through Latent Acoustic Pattern Triggers", "abstract": "As Audio Large Language Models (ALLMs) emerge as powerful tools for speech processing, their safety implications demand urgent attention. While considerable research has explored textual and vision safety, audio's distinct characteristics present significant challenges. This paper first investigates: Is ALLM vulnerable to backdoor attacks exploiting acoustic triggers? In response to this issue, we introduce Hidden in the Noise (HIN), a novel backdoor attack framework designed to exploit subtle, audio-specific features. HIN applies acoustic modifications to raw audio waveforms, such as alterations to temporal dynamics and strategic injection of spectrally tailored noise. These changes introduce consistent patterns that an ALLM's acoustic feature encoder captures, embedding robust triggers within the audio stream. To evaluate ALLM robustness against audio-feature-based triggers, we develop the AudioSafe benchmark, assessing nine distinct risk types. Extensive experiments on AudioSafe and three established safety datasets reveal critical vulnerabilities in existing ALLMs: (I) audio features like environment noise and speech rate variations achieve over 90% average attack success rate. (II) ALLMs exhibit significant sensitivity differences across acoustic features, particularly showing minimal response to volume as a trigger, and (III) poisoned sample inclusion causes only marginal loss curve fluctuations, highlighting the attack's stealth.", "date": "2025-08-04", "categories": []}
{"id": "2511.13936", "title": "Preference-Based Learning in Audio Applications: A Systematic Analysis", "abstract": "Despite the parallel challenges that audio and text domains face in evaluating generative model outputs, preference learning remains remarkably underexplored in audio applications. Through a PRISMA-guided systematic review of approximately 500 papers, we find that only 30 (6%) apply preference learning to audio tasks. Our analysis reveals a field in transition: pre-2021 works focused on emotion recognition using traditional ranking methods (rankSVM), while post-2021 studies have pivoted toward generation tasks employing modern RLHF frameworks. We identify three critical patterns: (1) the emergence of multi-dimensional evaluation strategies combining synthetic, automated, and human preferences; (2) inconsistent alignment between traditional metrics (WER, PESQ) and human judgments across different contexts; and (3) convergence on multi-stage training pipelines that combine reward signals. Our findings suggest that while preference learning shows promise for audio, particularly in capturing subjective qualities like naturalness and musicality, the field requires standardized benchmarks, higher-quality datasets, and systematic investigation of how temporal factors unique to audio impact preference learning frameworks.", "date": "2025-11-17", "categories": []}
{"id": "2511.14293", "title": "Segmentwise Pruning in Audio-Language Models", "abstract": "Recent audio-language models have shown impressive performance across a wide range of audio tasks and are increasingly capable of handling long audio inputs. However, the computing costs in these models heavily depend on sequence length, which can become very large given the nature of audio data. In the vision-language domain, token pruning methods have proven effective in reducing token counts while preserving strong performance on standard benchmarks. In this work, we investigate the relevance and effectiveness of such token selection strategies in the context of audio-language models. We also improve them by proposing a lightweight strategy that takes the time dimension into account. While retaining only a quarter of the initial tokens, our approach results in a relative maximum decrease of 2% in CIDEr on Clotho v2 and a relative maximum decrease of 4% in accuracy on MMAU.", "date": "2025-11-18", "categories": []}
{"id": "2511.14307", "title": "Audio Question Answering with GRPO-Based Fine-Tuning and Calibrated Segment-Level Predictions", "abstract": "In this report, we describe our submission to Track 5 of the DCASE 2025 Challenge for the task of Audio Question Answering(AQA). Our system leverages the SSL backbone BEATs to extract frame-level audio features, which are then processed by a classification head to generate segment-level predictions of acoustic events, following the Audioset ontology. These segment-level predictions are subsequently calibrated before producing event-level predictions. Finally, these predictions are incorporated into a structured prompt, along with the question and candidate answers. This prompt is then fed to a fine-tuned version of Qwen2.5-7B-Instruct, trained using the GRPO algorithm with a simple reward function. Our method achieves an accuracy of 62.6 % on the development set, demonstrating the effectiveness of combining acoustic event reasoning with instruction-tuned large language models for AQA.", "date": "2025-11-18", "categories": []}
{"id": "2511.11000", "title": "DialogGraph-LLM: Graph-Informed LLMs for End-to-End Audio Dialogue Intent Recognition", "abstract": "Recognizing speaker intent in long audio dialogues among speakers has a wide range of applications, but is a non-trivial AI task due to complex inter-dependencies in speaker utterances and scarce annotated data. To address these challenges, an end-to-end framework, namely DialogGraph-LLM, is proposed in the current work. DialogGraph-LLM combines a novel Multi-Relational Dialogue Attention Network (MR-DAN) architecture with multimodal foundation models (e.g., Qwen2.5-Omni-7B) for direct acoustic-to-intent inference. An adaptive semi-supervised learning strategy is designed using LLM with a confidence-aware pseudo-label generation mechanism based on dual-threshold filtering using both global and class confidences, and an entropy-based sample selection process that prioritizes high-information unlabeled instances. Extensive evaluations on the proprietary MarketCalls corpus and the publicly available MIntRec 2.0 benchmark demonstrate DialogGraph-LLM's superiority over strong audio and text-driven baselines. The framework demonstrates strong performance and efficiency in intent recognition in real world scenario audio dialogues, proving its practical value for audio-rich domains with limited supervision. Our code is available at https://github.com/david188888/DialogGraph-LLM.", "date": "2025-11-14", "categories": []}
{"id": "2511.11615", "title": "Lightweight Hopfield Neural Networks for Bioacoustic Detection and Call Monitoring of Captive Primates", "abstract": "Passive acoustic monitoring is a sustainable method of monitoring wildlife and environments that leads to the generation of large datasets and, currently, a processing backlog. Academic research into automating this process is focused on the application of resource intensive convolutional neural networks which require large pre-labelled datasets for training and lack flexibility in application. We present a viable alternative relevant in both wild and captive settings; a transparent, lightweight and fast-to-train associative memory AI model with Hopfield neural network (HNN) architecture. Adapted from a model developed to detect bat echolocation calls, this model monitors captive endangered black-and-white ruffed lemur Varecia variegata vocalisations. Lemur social calls of interest when monitoring welfare are stored in the HNN in order to detect other call instances across the larger acoustic dataset. We make significant model improvements by storing an additional signal caused by movement and achieve an overall accuracy of 0.94. The model can perform $340$ classifications per second, processing over 5.5 hours of audio data per minute, on a standard laptop running other applications. It has broad applicability and trains in milliseconds. Our lightweight solution reduces data-to-insight turnaround times and can accelerate decision making in both captive and wild settings.", "date": "2025-11-04", "categories": []}
{"id": "2511.06606", "title": "SPUR: A Plug-and-Play Framework for Integrating Spatial Audio Understanding and Reasoning into Large Audio-Language Models", "abstract": "Spatial perception is central to auditory intelligence, enabling accurate understanding of real-world acoustic scenes and advancing human-level perception of the world around us. While recent large audio-language models (LALMs) show strong reasoning over complex audios, most operate on monaural inputs and lack the ability to capture spatial cues such as direction, elevation, and distance. We introduce SPUR, a lightweight, plug-in approach that equips LALMs with spatial perception through minimal architectural changes. SPUR consists of: (i) a First-Order Ambisonics (FOA) encoder that maps (W, X, Y, Z) channels to rotation-aware, listener-centric spatial features, integrated into target LALMs via a multimodal adapter; and (ii) SPUR-Set, a spatial QA dataset combining open-source FOA recordings with controlled simulations, emphasizing relative direction, elevation, distance, and overlap for supervised spatial reasoning. Fine-tuning our model on the SPUR-Set consistently improves spatial QA and multi-speaker attribution while preserving general audio understanding. SPUR provides a simple recipe that transforms monaural LALMs into spatially aware models. Extensive ablations validate the effectiveness of our approach.", "date": "2025-11-10", "categories": []}
{"id": "2511.11006", "title": "MSMT-FN: Multi-segment Multi-task Fusion Network for Marketing Audio Classification", "abstract": "Audio classification plays an essential role in sentiment analysis and emotion recognition, especially for analyzing customer attitudes in marketing phone calls. Efficiently categorizing customer purchasing propensity from large volumes of audio data remains challenging. In this work, we propose a novel Multi-Segment Multi-Task Fusion Network (MSMT-FN) that is uniquely designed for addressing this business demand. Evaluations conducted on our proprietary MarketCalls dataset, as well as established benchmarks (CMU-MOSI, CMU-MOSEI, and MELD), show MSMT-FN consistently outperforms or matches state-of-the-art methods. Additionally, our newly curated MarketCalls dataset will be available upon request, and the code base is made accessible at GitHub Repository MSMT-FN, to facilitate further research and advancements in audio classification domain.", "date": "2025-11-14", "categories": []}
{"id": "2510.04688", "title": "A Study on the Data Distribution Gap in Music Emotion Recognition", "abstract": "Music Emotion Recognition (MER) is a task deeply connected to human perception, relying heavily on subjective annotations collected from contributors. Prior studies tend to focus on specific musical styles rather than incorporating a diverse range of genres, such as rock and classical, within a single framework. In this paper, we address the task of recognizing emotion from audio content by investigating five datasets with dimensional emotion annotations -- EmoMusic, DEAM, PMEmo, WTC, and WCMED -- which span various musical styles. We demonstrate the problem of out-of-distribution generalization in a systematic experiment. By closely looking at multiple data and feature sets, we provide insight into genre-emotion relationships in existing data and examine potential genre dominance and dataset biases in certain feature representations. Based on these experiments, we arrive at a simple yet effective framework that combines embeddings extracted from the Jukebox model with chroma features and demonstrate how, alongside a combination of several diverse training sets, this permits us to train models with substantially improved cross-dataset generalization capabilities.", "date": "2025-10-06", "categories": []}
{"id": "2510.21004", "title": "Can Current Detectors Catch Face-to-Voice Deepfake Attacks?", "abstract": "The rapid advancement of generative models has enabled the creation of increasingly stealthy synthetic voices, commonly referred to as audio deepfakes. A recent technique, FOICE [USENIX'24], demonstrates a particularly alarming capability: generating a victim's voice from a single facial image, without requiring any voice sample. By exploiting correlations between facial and vocal features, FOICE produces synthetic voices realistic enough to bypass industry-standard authentication systems, including WeChat Voiceprint and Microsoft Azure. This raises serious security concerns, as facial images are far easier for adversaries to obtain than voice samples, dramatically lowering the barrier to large-scale attacks. In this work, we investigate two core research questions: (RQ1) can state-of-the-art audio deepfake detectors reliably detect FOICE-generated speech under clean and noisy conditions, and (RQ2) whether fine-tuning these detectors on FOICE data improves detection without overfitting, thereby preserving robustness to unseen voice generators such as SpeechT5.\n  Our study makes three contributions. First, we present the first systematic evaluation of FOICE detection, showing that leading detectors consistently fail under both standard and noisy conditions. Second, we introduce targeted fine-tuning strategies that capture FOICE-specific artifacts, yielding significant accuracy improvements. Third, we assess generalization after fine-tuning, revealing trade-offs between specialization to FOICE and robustness to unseen synthesis pipelines. These findings expose fundamental weaknesses in today's defenses and motivate new architectures and training protocols for next-generation audio deepfake detection.", "date": "2025-10-23", "categories": []}
{"id": "2511.10289", "title": "Music Flamingo: Scaling Music Understanding in Audio Language Models", "abstract": "We introduce Music Flamingo, a novel large audio-language model designed to advance music (including song) understanding in foundational audio models. While audio-language research has progressed rapidly, music remains challenging due to its dynamic, layered, and information-dense nature. Progress has been further limited by the difficulty of scaling open audio understanding models, primarily because of the scarcity of high-quality music data and annotations. As a result, prior models are restricted to producing short, high-level captions, answering only surface-level questions, and showing limited generalization across diverse musical cultures. To address these challenges, we curate MF-Skills, a large-scale dataset labeled through a multi-stage pipeline that yields rich captions and question-answer pairs covering harmony, structure, timbre, lyrics, and cultural context. We fine-tune an enhanced Audio Flamingo 3 backbone on MF-Skills and further strengthen multiple skills relevant to music understanding. To improve the model's reasoning abilities, we introduce a post-training recipe: we first cold-start with MF-Think, a novel chain-of-thought dataset grounded in music theory, followed by GRPO-based reinforcement learning with custom rewards. Music Flamingo achieves state-of-the-art results across 10+ benchmarks for music understanding and reasoning, establishing itself as a generalist and musically intelligent audio-language model. Beyond strong empirical results, Music Flamingo sets a new standard for advanced music understanding by demonstrating how models can move from surface-level recognition toward layered, human-like perception of songs. We believe this work provides both a benchmark and a foundation for the community to build the next generation of models that engage with music as meaningfully as humans do.", "date": "2025-11-13", "categories": []}
{"id": "2503.09205", "title": "Quality Over Quantity? LLM-Based Curation for a Data-Efficient Audio-Video Foundation Model", "abstract": "Integrating audio and visual data for training multimodal foundational models remains a challenge. The Audio-Video Vector Alignment (AVVA) framework addresses this by considering AV scene alignment beyond mere temporal synchronization, and leveraging Large Language Models (LLMs) for data curation. AVVA implements a scoring mechanism for selecting aligned training data segments. It integrates Whisper, a speech-based foundation model, for audio and DINOv2 for video analysis in a dual-encoder structure with contrastive learning on AV pairs. Evaluations on AudioCaps, VALOR, and VGGSound demonstrate the effectiveness of the proposed model architecture and data curation approach. AVVA achieves a significant improvement in top-k accuracies for video-to-audio retrieval on all datasets compared to DenseAV, while using only 192 hrs of curated training data. Furthermore, an ablation study indicates that the data curation process effectively trades data quality for data quantity, yielding increases in top-k retrieval accuracies on AudioCaps, VALOR, and VGGSound, compared to training on the full spectrum of uncurated data.", "date": "2025-03-12", "categories": []}
{"id": "2511.07470", "title": "Slimmable NAM: Neural Amp Models with adjustable runtime computational cost", "abstract": "This work demonstrates \"slimmable Neural Amp Models\", whose size and computational cost can be changed without additional training and with negligible computational overhead, enabling musicians to easily trade off between the accuracy and compute of the models they are using. The method's performance is quantified against commonly-used baselines, and a real-time demonstration of the model in an audio effect plug-in is developed.", "date": "2025-11-08", "categories": []}
{"id": "2511.07493", "title": "Enabling Automatic Self-Talk Detection via Earables", "abstract": "Self-talk-an internal dialogue that can occur silently or be spoken aloud-plays a crucial role in emotional regulation, cognitive processing, and motivation, yet has remained largely invisible and unmeasurable in everyday life. In this paper, we present MutterMeter, a mobile system that automatically detects vocalized self-talk from audio captured by earable microphones in real-world settings. Detecting self-talk is technically challenging due to its diverse acoustic forms, semantic and grammatical incompleteness, and irregular occurrence patterns, which differ fundamentally from assumptions underlying conventional speech understanding models. To address these challenges, MutterMeter employs a hierarchical classification architecture that progressively integrates acoustic, linguistic, and contextual information through a sequential processing pipeline, adaptively balancing accuracy and computational efficiency. We build and evaluate MutterMeter using a first-of-its-kind dataset comprising 31.1 hours of audio collected from 25 participants. Experimental results demonstrate that MutterMeter achieves robust performance with a macro-averaged F1 score of 0.84, outperforming conventional approaches, including LLM-based and speech emotion recognition models.", "date": "2025-11-10", "categories": []}
{"id": "2511.08012", "title": "DOA Estimation with Lightweight Network on LLM-Aided Simulated Acoustic Scenes", "abstract": "Direction-of-Arrival (DOA) estimation is critical in spatial audio and acoustic signal processing, with wide-ranging applications in real-world. Most existing DOA models are trained on synthetic data by convolving clean speech with room impulse responses (RIRs), which limits their generalizability due to constrained acoustic diversity. In this paper, we revisit DOA estimation using a recently introduced dataset constructed with the assistance of large language models (LLMs), which provides more realistic and diverse spatial audio scenes. We benchmark several representative neural-based DOA methods on this dataset and propose LightDOA, a lightweight DOA estimation model based on depthwise separable convolutions, specifically designed for mutil-channel input in varying environments. Experimental results show that LightDOA achieves satisfactory accuracy and robustness across various acoustic scenes while maintaining low computational complexity. This study not only highlights the potential of spatial audio synthesized with the assistance of LLMs in advancing robust and efficient DOA estimation research, but also highlights LightDOA as efficient solution for resource-constrained applications.", "date": "2025-11-11", "categories": []}
{"id": "2511.05350", "title": "Perceptually Aligning Representations of Music via Noise-Augmented Autoencoders", "abstract": "We argue that training autoencoders to reconstruct inputs from noised versions of their encodings, when combined with perceptual losses, yields encodings that are structured according to a perceptual hierarchy. We demonstrate the emergence of this hierarchical structure by showing that, after training an audio autoencoder in this manner, perceptually salient information is captured in coarser representation structures than with conventional training. Furthermore, we show that such perceptual hierarchies improve latent diffusion decoding in the context of estimating surprisal in music pitches and predicting EEG-brain responses to music listening. Pretrained weights are available on github.com/CPJKU/pa-audioic.", "date": "2025-11-07", "categories": []}
{"id": "2511.05717", "title": "Persian Musical Instruments Classification Using Polyphonic Data Augmentation", "abstract": "Musical instrument classification is essential for music information retrieval (MIR) and generative music systems. However, research on non-Western traditions, particularly Persian music, remains limited. We address this gap by introducing a new dataset of isolated recordings covering seven traditional Persian instruments, two common but originally non-Persian instruments (i.e., violin, piano), and vocals. We propose a culturally informed data augmentation strategy that generates realistic polyphonic mixtures from monophonic samples. Using the MERT model (Music undERstanding with large-scale self-supervised Training) with a classification head, we evaluate our approach with out-of-distribution data which was obtained by manually labeling segments of traditional songs. On real-world polyphonic Persian music, the proposed method yielded the best ROC-AUC (0.795), highlighting complementary benefits of tonal and temporal coherence. These results demonstrate the effectiveness of culturally grounded augmentation for robust Persian instrument recognition and provide a foundation for culturally inclusive MIR and diverse music generation systems.", "date": "2025-11-07", "categories": []}
{"id": "2511.05399", "title": "Robust Neural Audio Fingerprinting using Music Foundation Models", "abstract": "The proliferation of distorted, compressed, and manipulated music on modern media platforms like TikTok motivates the development of more robust audio fingerprinting techniques to identify the sources of musical recordings. In this paper, we develop and evaluate new neural audio fingerprinting techniques with the aim of improving their robustness. We make two contributions to neural fingerprinting methodology: (1) we use a pretrained music foundation model as the backbone of the neural architecture and (2) we expand the use of data augmentation to train fingerprinting models under a wide variety of audio manipulations, including time streching, pitch modulation, compression, and filtering. We systematically evaluate our methods in comparison to two state-of-the-art neural fingerprinting models: NAFP and GraFPrint. Results show that fingerprints extracted with music foundation models (e.g., MuQ, MERT) consistently outperform models trained from scratch or pretrained on non-musical audio. Segment-level evaluation further reveals their capability to accurately localize fingerprint matches, an important practical feature for catalog management.", "date": "2025-11-07", "categories": []}
{"id": "2509.14912", "title": "Back to Ear: Perceptually Driven High Fidelity Music Reconstruction", "abstract": "Variational Autoencoders (VAEs) are essential for large-scale audio tasks like diffusion-based generation. However, existing open-source models often neglect auditory perceptual aspects during training, leading to weaknesses in phase accuracy and stereophonic spatial representation. To address these challenges, we propose {\\epsilon}ar-VAE, an open-source music signal reconstruction model that rethinks and optimizes the VAE training paradigm. Our contributions are threefold: (i) A K-weighting perceptual filter applied prior to loss calculation to align the objective with auditory perception. (ii) Two novel phase losses: a Correlation Loss for stereo coherence, and a Phase Loss using its derivatives--Instantaneous Frequency and Group Delay--for precision. (iii) A new spectral supervision paradigm where magnitude is supervised by all four Mid/Side/Left/Right components, while phase is supervised only by the LR components. Experiments show {\\epsilon}ar-VAE at 44.1kHz substantially outperforms leading open-source models across diverse metrics, showing particular strength in reconstructing high-frequency harmonics and the spatial characteristics.", "date": "2025-09-18", "categories": []}
{"id": "2511.04139", "title": "CantoASR: Prosody-Aware ASR-LALM Collaboration for Low-Resource Cantonese", "abstract": "Automatic speech recognition (ASR) is critical for language accessibility, yet low-resource Cantonese remains challenging due to limited annotated data, six lexical tones, tone sandhi, and accent variation. Existing ASR models, such as Whisper, often suffer from high word error rates. Large audio-language models (LALMs), in contrast, can leverage broader contextual reasoning but still require explicit tonal and prosodic acoustic cues. We introduce CantoASR, a collaborative ASR-LALM error correction framework that integrates forced alignment for acoustic feature extraction, a LoRA-finetuned Whisper for improved tone discrimination, and an instruction-tuned Qwen-Audio for prosody-aware correction. Evaluations on spontaneous Cantonese data show substantial CER gains over Whisper-Large-V3. These findings suggest that integrating acoustic cues with LALM reasoning provides a scalable strategy for low-resource tonal and dialectal ASR.", "date": "2025-11-06", "categories": []}
{"id": "2506.21448", "title": "ThinkSound: Chain-of-Thought Reasoning in Multimodal Large Language Models for Audio Generation and Editing", "abstract": "While end-to-end video-to-audio generation has greatly improved, producing high-fidelity audio that authentically captures the nuances of visual content remains challenging. Like professionals in the creative industries, this generation requires sophisticated reasoning about items such as visual dynamics, acoustic environments, and temporal relationships. We present ThinkSound, a novel framework that leverages Chain-of-Thought (CoT) reasoning to enable stepwise, interactive audio generation and editing for videos. Our approach decomposes the process into three complementary stages: foundational foley generation that creates semantically coherent soundscapes, interactive object-centric refinement through precise user interactions, and targeted editing guided by natural language instructions. At each stage, a multimodal large language model generates contextually aligned CoT reasoning that guides a unified audio foundation model. Furthermore, we introduce AudioCoT, a comprehensive dataset with structured reasoning annotations that establishes connections between visual content, textual descriptions, and sound synthesis. Experiments demonstrate that ThinkSound achieves state-of-the-art performance in video-to-audio generation across both audio metrics and CoT metrics, and excels in the out-of-distribution Movie Gen Audio benchmark. The project page is available at https://ThinkSound-Project.github.io.", "date": "2025-06-26", "categories": []}
{"id": "2501.15417", "title": "AnyEnhance: A Unified Generative Model with Prompt-Guidance and Self-Critic for Voice Enhancement", "abstract": "We introduce AnyEnhance, a unified generative model for voice enhancement that processes both speech and singing voices. Based on a masked generative model, AnyEnhance is capable of handling both speech and singing voices, supporting a wide range of enhancement tasks including denoising, dereverberation, declipping, super-resolution, and target speaker extraction, all simultaneously and without fine-tuning. AnyEnhance introduces a prompt-guidance mechanism for in-context learning, which allows the model to natively accept a reference speaker's timbre. In this way, it could boost enhancement performance when a reference audio is available and enable the target speaker extraction task without altering the underlying architecture. Moreover, we also introduce a self-critic mechanism into the generative process for masked generative models, yielding higher-quality outputs through iterative self-assessment and refinement. Extensive experiments on various enhancement tasks demonstrate AnyEnhance outperforms existing methods in terms of both objective metrics and subjective listening tests. Demo audios are publicly available at https://amphionspace.github.io/anyenhance. An open-source implementation is provided at https://github.com/viewfinder-annn/anyenhance-v1-ccf-aatc.", "date": "2025-01-26", "categories": []}
{"id": "2503.02585", "title": "As Good as It KAN Get: High-Fidelity Audio Representation", "abstract": "Implicit neural representations (INR) have gained prominence for efficiently encoding multimedia data, yet their applications in audio signals remain limited. This study introduces the Kolmogorov-Arnold Network (KAN), a novel architecture using learnable activation functions, as an effective INR model for audio representation. KAN demonstrates superior perceptual performance over previous INRs, achieving the lowest Log-SpectralDistance of 1.29 and the highest Perceptual Evaluation of Speech Quality of 3.57 for 1.5 s audio. To extend KAN's utility, we propose FewSound, a hypernetwork-based architecture that enhances INR parameter updates. FewSound outperforms the state-of-the-art HyperSound, with a 33.3% improvement in MSE and 60.87% in SI-SNR. These results show KAN as a robust and adaptable audio representation with the potential for scalability and integration into various hypernetwork frameworks. The source code can be accessed at https://github.com/gmum/fewsound.git.", "date": "2025-03-04", "categories": []}
{"id": "2507.18161", "title": "Recent Trends in Distant Conversational Speech Recognition: A Review of CHiME-7 and 8 DASR Challenges", "abstract": "The CHiME-7 and 8 distant speech recognition (DASR) challenges focus on multi-channel, generalizable, joint automatic speech recognition (ASR) and diarization of conversational speech. With participation from 9 teams submitting 32 diverse systems, these challenges have contributed to state-of-the-art research in the field. This paper outlines the challenges' design, evaluation metrics, datasets, and baseline systems while analyzing key trends from participant submissions. From this analysis it emerges that: 1) Most participants use end-to-end (e2e) ASR systems, whereas hybrid systems were prevalent in previous CHiME challenges. This transition is mainly due to the availability of robust large-scale pre-trained models, which lowers the data burden for e2e-ASR. 2) Despite recent advances in neural speech separation and enhancement (SSE), all teams still heavily rely on guided source separation, suggesting that current neural SSE techniques are still unable to reliably deal with complex scenarios and different recording setups. 3) All best systems employ diarization refinement via target-speaker diarization techniques. Accurate speaker counting in the first diarization pass is thus crucial to avoid compounding errors and CHiME-8 DASR participants especially focused on this part. 4) Downstream evaluation via meeting summarization can correlate weakly with transcription quality due to the remarkable effectiveness of large-language models in handling errors. On the NOTSOFAR-1 scenario, even systems with over 50% time-constrained minimum permutation WER can perform roughly on par with the most effective ones (around 11%). 5) Despite recent progress, accurately transcribing spontaneous speech in challenging acoustic environments remains difficult, even when using computationally intensive system ensembles.", "date": "2025-07-24", "categories": []}
{"id": "2509.13927", "title": "DSpAST: Disentangled Representations for Spatial Audio Reasoning with Large Language Models", "abstract": "Reasoning about spatial audio with large language models requires a spatial audio encoder as an acoustic front-end to obtain audio embeddings for further processing. Such an encoder needs to capture all information required to detect the type of sound events, as well as the direction and distance of their corresponding sources. Accomplishing this with a single audio encoder is demanding as the information required for each of these tasks is mostly independent of each other. As a result, the performance obtained with a single encoder is often worse than when using task-specific audio encoders. In this work, we present DSpAST, a novel audio encoder based on SpatialAST that learns disentangled representations of spatial audio while having only 0.2% additional parameters. Experiments on SpatialSoundQA with the spatial audio reasoning system BAT demonstrate that DSpAST significantly outperforms SpatialAST.", "date": "2025-09-17", "categories": []}
{"id": "2511.00252", "title": "Merlin L48 Spectrogram Dataset", "abstract": "In the single-positive multi-label (SPML) setting, each image in a dataset is labeled with the presence of a single class, while the true presence of other classes remains unknown. The challenge is to narrow the performance gap between this partially-labeled setting and fully-supervised learning, which often requires a significant annotation budget. Prior SPML methods were developed and benchmarked on synthetic datasets created by randomly sampling single positive labels from fully-annotated datasets like Pascal VOC, COCO, NUS-WIDE, and CUB200. However, this synthetic approach does not reflect real-world scenarios and fails to capture the fine-grained complexities that can lead to difficult misclassifications. In this work, we introduce the L48 dataset, a fine-grained, real-world multi-label dataset derived from recordings of bird sounds. L48 provides a natural SPML setting with single-positive annotations on a challenging, fine-grained domain, as well as two extended settings in which domain priors give access to additional negative labels. We benchmark existing SPML methods on L48 and observe significant performance differences compared to synthetic datasets and analyze method weaknesses, underscoring the need for more realistic and difficult benchmarks.", "date": "2025-10-31", "categories": []}
{"id": "2511.01773", "title": "ADNAC: Audio Denoiser using Neural Audio Codec", "abstract": "Audio denoising is critical in signal processing, enhancing intelligibility and fidelity for applications like restoring musical recordings. This paper presents a proof-of-concept for adapting a state-of-the-art neural audio codec, the Descript Audio Codec (DAC), for music denoising. This work overcomes the limitations of traditional architectures like U-Nets by training the model on a large-scale, custom-synthesized dataset built from diverse sources. Training is guided by a multi objective loss function that combines time-domain, spectral, and signal-level fidelity metrics. Ultimately, this paper aims to present a PoC for high-fidelity, generative audio restoration.", "date": "2025-11-03", "categories": []}
{"id": "2510.26825", "title": "Audio-Visual Speech Enhancement In Complex Scenarios With Separation And Dereverberation Joint Modeling", "abstract": "Audio-visual speech enhancement (AVSE) is a task that uses visual auxiliary information to extract a target speaker's speech from mixed audio. In real-world scenarios, there often exist complex acoustic environments, accompanied by various interfering sounds and reverberation. Most previous methods struggle to cope with such complex conditions, resulting in poor perceptual quality of the extracted speech. In this paper, we propose an effective AVSE system that performs well in complex acoustic environments. Specifically, we design a \"separation before dereverberation\" pipeline that can be extended to other AVSE networks. The 4th COGMHEAR Audio-Visual Speech Enhancement Challenge (AVSEC) aims to explore new approaches to speech processing in multimodal complex environments. We validated the performance of our system in AVSEC-4: we achieved excellent results in the three objective metrics on the competition leaderboard, and ultimately secured first place in the human subjective listening test.", "date": "2025-10-29", "categories": []}
{"id": "2510.26838", "title": "Multi-Representation Attention Framework for Underwater Bioacoustic Denoising and Recognition", "abstract": "Automated monitoring of marine mammals in the St. Lawrence Estuary faces extreme challenges: calls span low-frequency moans to ultrasonic clicks, often overlap, and are embedded in variable anthropogenic and environmental noise. We introduce a multi-step, attention-guided framework that first segments spectrograms to generate soft masks of biologically relevant energy and then fuses these masks with the raw inputs for multi-band, denoised classification. Image and mask embeddings are integrated via mid-level fusion, enabling the model to focus on salient spectrogram regions while preserving global context. Using real-world recordings from the Saguenay St. Lawrence Marine Park Research Station in Canada, we demonstrate that segmentation-driven attention and mid-level fusion improve signal discrimination, reduce false positive detections, and produce reliable representations for operational marine mammal monitoring across diverse environmental conditions and signal-to-noise ratios. Beyond in-distribution evaluation, we further assess the generalization of Mask-Guided Classification (MGC) under distributional shifts by testing on spectrograms generated with alternative acoustic transformations. While high-capacity baseline models lose accuracy in this Out-of-distribution (OOD) setting, MGC maintains stable performance, with even simple fusion mechanisms (gated, concat) achieving comparable results across distributions. This robustness highlights the capacity of MGC to learn transferable representations rather than overfitting to a specific transformation, thereby reinforcing its suitability for large-scale, real-world biodiversity monitoring. We show that in all experimental settings, the MGC framework consistently outperforms baseline architectures, yielding substantial gains in accuracy on both in-distribution and OOD data.", "date": "2025-10-29", "categories": []}
{"id": "2510.23802", "title": "Learning Interpretable Features in Audio Latent Spaces via Sparse Autoencoders", "abstract": "While sparse autoencoders (SAEs) successfully extract interpretable features from language models, applying them to audio generation faces unique challenges: audio's dense nature requires compression that obscures semantic meaning, and automatic feature characterization remains limited. We propose a framework for interpreting audio generative models by mapping their latent representations to human-interpretable acoustic concepts. We train SAEs on audio autoencoder latents, then learn linear mappings from SAE features to discretized acoustic properties (pitch, amplitude, and timbre). This enables both controllable manipulation and analysis of the AI music generation process, revealing how acoustic properties emerge during synthesis. We validate our approach on continuous (DiffRhythm-VAE) and discrete (EnCodec, WavTokenizer) audio latent spaces, and analyze DiffRhythm, a state-of-the-art text-to-music model, to demonstrate how pitch, timbre, and loudness evolve throughout generation. While our work is only done on audio modality, our framework can be extended to interpretable analysis of visual latent space generation models.", "date": "2025-10-27", "categories": []}
{"id": "2510.25566", "title": "PitchFlower: A flow-based neural audio codec with pitch controllability", "abstract": "We present PitchFlower, a flow-based neural audio codec with explicit pitch controllability. Our approach enforces disentanglement through a simple perturbation: during training, F0 contours are flattened and randomly shifted, while the true F0 is provided as conditioning. A vector-quantization bottleneck prevents pitch recovery, and a flow-based decoder generates high quality audio. Experiments show that PitchFlower achieves more accurate pitch control than WORLD at much higher audio quality, and outperforms SiFiGAN in controllability while maintaining comparable quality. Beyond pitch, this framework provides a simple and extensible path toward disentangling other speech attributes.", "date": "2025-10-29", "categories": []}
{"id": "2502.04465", "title": "FocalCodec: Low-Bitrate Speech Coding via Focal Modulation Networks", "abstract": "Large language models have revolutionized natural language processing through self-supervised pretraining on massive datasets. Inspired by this success, researchers have explored adapting these methods to speech by discretizing continuous audio into tokens using neural audio codecs. However, existing approaches face limitations, including high bitrates, the loss of either semantic or acoustic information, and the reliance on multi-codebook designs when trying to capture both, which increases architectural complexity for downstream tasks. To address these challenges, we introduce FocalCodec, an efficient low-bitrate codec based on focal modulation that utilizes a single binary codebook to compress speech between 0.16 and 0.65 kbps. FocalCodec delivers competitive performance in speech resynthesis and voice conversion at lower bitrates than the current state-of-the-art, while effectively handling multilingual speech and noisy environments. Evaluation on downstream tasks shows that FocalCodec successfully preserves sufficient semantic and acoustic information, while also being well-suited for generative modeling. Demo samples and code are available at https://lucadellalib.github.io/focalcodec-web/.", "date": "2025-02-06", "categories": []}
{"id": "2507.16343", "title": "Detect Any Sound: Open-Vocabulary Sound Event Detection with Multi-Modal Queries", "abstract": "Most existing sound event detection~(SED) algorithms operate under a closed-set assumption, restricting their detection capabilities to predefined classes. While recent efforts have explored language-driven zero-shot SED by exploiting audio-language models, their performance is still far from satisfactory due to the lack of fine-grained alignment and cross-modal feature fusion. In this work, we propose the Detect Any Sound Model (DASM), a query-based framework for open-vocabulary SED guided by multi-modal queries. DASM formulates SED as a frame-level retrieval task, where audio features are matched against query vectors derived from text or audio prompts. To support this formulation, DASM introduces a dual-stream decoder that explicitly decouples event recognition and temporal localization: a cross-modality event decoder performs query-feature fusion and determines the presence of sound events at the clip-level, while a context network models temporal dependencies for frame-level localization. Additionally, an inference-time attention masking strategy is proposed to leverage semantic relations between base and novel classes, substantially enhancing generalization to novel classes. Experiments on the AudioSet Strong dataset demonstrate that DASM effectively balances localization accuracy with generalization to novel classes, outperforming CLAP-based methods in open-vocabulary setting (+ 7.8 PSDS) and the baseline in the closed-set setting (+ 6.9 PSDS). Furthermore, in cross-dataset zero-shot evaluation on DESED, DASM achieves a PSDS1 score of 42.2, even exceeding the supervised CRNN baseline. The project page is available at https://cai525.github.io/Transformer4SED/demo_page/DASM/.", "date": "2025-07-22", "categories": []}
{"id": "2508.01488", "title": "PESTO: Real-Time Pitch Estimation with Self-supervised Transposition-equivariant Objective", "abstract": "In this paper, we introduce PESTO, a self-supervised learning approach for single-pitch estimation using a Siamese architecture. Our model processes individual frames of a Variable-$Q$ Transform (VQT) and predicts pitch distributions. The neural network is designed to be equivariant to translations, notably thanks to a Toeplitz fully-connected layer. In addition, we construct pitch-shifted pairs by translating and cropping the VQT frames and train our model with a novel class-based transposition-equivariant objective, eliminating the need for annotated data. Thanks to this architecture and training objective, our model achieves remarkable performances while being very lightweight ($130$k parameters). Evaluations on music and speech datasets (MIR-1K, MDB-stem-synth, and PTDB) demonstrate that PESTO not only outperforms self-supervised baselines but also competes with supervised methods, exhibiting superior cross-dataset generalization. Finally, we enhance PESTO's practical utility by developing a streamable VQT implementation using cached convolutions. Combined with our model's low latency (less than 10 ms) and minimal parameter count, this makes PESTO particularly suitable for real-time applications.", "date": "2025-08-02", "categories": []}
{"id": "2508.01493", "title": "Translation-Equivariant Self-Supervised Learning for Pitch Estimation with Optimal Transport", "abstract": "In this paper, we propose an Optimal Transport objective for learning one-dimensional translation-equivariant systems and demonstrate its applicability to single pitch estimation. Our method provides a theoretically grounded, more numerically stable, and simpler alternative for training state-of-the-art self-supervised pitch estimators.", "date": "2025-08-02", "categories": []}
{"id": "2509.10074", "title": "Prototypical Contrastive Learning For Improved Few-Shot Audio Classification", "abstract": "Few-shot learning has emerged as a powerful paradigm for training models with limited labeled data, addressing challenges in scenarios where large-scale annotation is impractical. While extensive research has been conducted in the image domain, few-shot learning in audio classification remains relatively underexplored. In this work, we investigate the effect of integrating supervised contrastive loss into prototypical few shot training for audio classification. In detail, we demonstrate that angular loss further improves the performance compared to the standard contrastive loss. Our method leverages SpecAugment followed by a self-attention mechanism to encapsulate diverse information of augmented input versions into one unified embedding. We evaluate our approach on MetaAudio, a benchmark including five datasets with predefined splits, standardized preprocessing, and a comprehensive set of few-shot learning models for comparison. The proposed approach achieves state-of-the-art performance in a 5-way, 5-shot setting.", "date": "2025-09-12", "categories": []}
{"id": "2510.11507", "title": "Automatic Music Sample Identification with Multi-Track Contrastive Learning", "abstract": "Sampling, the technique of reusing pieces of existing audio tracks to create new music content, is a very common practice in modern music production. In this paper, we tackle the challenging task of automatic sample identification, that is, detecting such sampled content and retrieving the material from which it originates. To do so, we adopt a self-supervised learning approach that leverages a multi-track dataset to create positive pairs of artificial mixes, and design a novel contrastive learning objective. We show that such method significantly outperforms previous state-of-the-art baselines, that is robust to various genres, and that scales well when increasing the number of noise songs in the reference database. In addition, we extensively analyze the contribution of the different components of our training pipeline and highlight, in particular, the need for high-quality separated stems for this task.", "date": "2025-10-13", "categories": []}
{"id": "2510.22795", "title": "SAO-Instruct: Free-form Audio Editing using Natural Language Instructions", "abstract": "Generative models have made significant progress in synthesizing high-fidelity audio from short textual descriptions. However, editing existing audio using natural language has remained largely underexplored. Current approaches either require the complete description of the edited audio or are constrained to predefined edit instructions that lack flexibility. In this work, we introduce SAO-Instruct, a model based on Stable Audio Open capable of editing audio clips using any free-form natural language instruction. To train our model, we create a dataset of audio editing triplets (input audio, edit instruction, output audio) using Prompt-to-Prompt, DDPM inversion, and a manual editing pipeline. Although partially trained on synthetic data, our model generalizes well to real in-the-wild audio clips and unseen edit instructions. We demonstrate that SAO-Instruct achieves competitive performance on objective metrics and outperforms other audio editing approaches in a subjective listening study. To encourage future research, we release our code and model weights.", "date": "2025-10-26", "categories": []}
{"id": "2504.01690", "title": "Token Pruning in Audio Transformers: Optimizing Performance and Decoding Patch Importance", "abstract": "Vision Transformers (ViTs) have achieved state-of-the-art performance across various computer vision tasks, but their high computational cost remains a challenge. Token pruning has been proposed to reduce this cost by selectively removing less important tokens. While effective in vision tasks by discarding non-object regions, applying this technique to audio tasks presents unique challenges, as distinguishing relevant from irrelevant regions in time-frequency representations is less straightforward. In this study, for the first time, we applied token pruning to ViT-based audio classification models using Mel-spectrograms and analyzed the trade-offs between model performance and computational cost: TopK token pruning can reduce MAC operations of AudioMAE and AST by 30-40%, with less than a 1% drop in accuracy. Our analysis reveals that while high-intensity or high-variation tokens contribute significantly to model accuracy, low-intensity or low variation tokens also remain important when token pruning is applied; pruning solely based on the intensity or variation of signals in a patch leads to a noticeable drop in accuracy. We support our claim by measuring high correlation between attention scores and these statistical features and by showing retained tokens consistently receive distinct attention compared to pruned ones. We also show that AudioMAE retains more low-intensity tokens than AST. This can be explained by AudioMAE's self-supervised reconstruction objective, which encourages attention to all patches, whereas AST's supervised training focuses on label-relevant tokens.", "date": "2025-04-02", "categories": []}
{"id": "2510.20441", "title": "UniSE: A Unified Framework for Decoder-only Autoregressive LM-based Speech Enhancement", "abstract": "The development of neural audio codecs (NACs) has largely promoted applications of language models (LMs) to speech processing and understanding. However, there lacks the verification on the effectiveness of autoregressive (AR) LMbased models in unifying different sub-tasks of speech enhancement (SE). In this work, we propose UniSE, a unified decoder-only LM-based framework to handle different SE tasks including speech restoration, target speaker extraction and speech separation. It takes input speech features as conditions and generates discrete tokens of the target speech using AR modeling, which facilitates a compatibility between distinct learning patterns of multiple tasks. Experiments on several benchmarks indicate the proposed UniSE can achieve competitive performance compared to discriminative and generative baselines, showing the capacity of LMs in unifying SE tasks. The demo page is available here: https://github.com/hyyan2k/UniSE.", "date": "2025-10-23", "categories": []}
{"id": "2510.20602", "title": "Resounding Acoustic Fields with Reciprocity", "abstract": "Achieving immersive auditory experiences in virtual environments requires flexible sound modeling that supports dynamic source positions. In this paper, we introduce a task called resounding, which aims to estimate room impulse responses at arbitrary emitter location from a sparse set of measured emitter positions, analogous to the relighting problem in vision. We leverage the reciprocity property and introduce Versa, a physics-inspired approach to facilitating acoustic field learning. Our method creates physically valid samples with dense virtual emitter positions by exchanging emitter and listener poses. We also identify challenges in deploying reciprocity due to emitter/listener gain patterns and propose a self-supervised learning approach to address them. Results show that Versa substantially improve the performance of acoustic field learning on both simulated and real-world datasets across different metrics. Perceptual user studies show that Versa can greatly improve the immersive spatial sound experience. Code, dataset and demo videos are available on the project website: https://waves.seas.upenn.edu/projects/versa.", "date": "2025-10-23", "categories": []}
{"id": "2510.19055", "title": "The MUSE Benchmark: Probing Music Perception and Auditory Relational Reasoning in Audio LLMS", "abstract": "Multimodal Large Language Models (MLLMs) have demonstrated capabilities in audio understanding, but current evaluations may obscure fundamental weaknesses in relational reasoning. We introduce the Music Understanding and Structural Evaluation (MUSE) Benchmark, an open-source resource with 10 tasks designed to probe fundamental music perception skills. We evaluate four SOTA models (Gemini Pro and Flash, Qwen2.5-Omni, and Audio-Flamingo 3) against a large human baseline (N=200). Our results reveal a wide variance in SOTA capabilities and a persistent gap with human experts. While Gemini Pro succeeds on basic perception, Qwen and Audio Flamingo 3 perform at or near chance, exposing severe perceptual deficits. Furthermore, we find Chain-of-Thought (CoT) prompting provides inconsistent, often detrimental results. Our work provides a critical tool for evaluating invariant musical representations and driving development of more robust AI systems.", "date": "2025-10-21", "categories": []}
{"id": "2505.23305", "title": "MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction", "abstract": "We present MGE-LDM, a unified latent diffusion framework for simultaneous music generation, source imputation, and query-driven source separation. Unlike prior approaches constrained to fixed instrument classes, MGE-LDM learns a joint distribution over full mixtures, submixtures, and individual stems within a single compact latent diffusion model. At inference, MGE-LDM enables (1) complete mixture generation, (2) partial generation (i.e., source imputation), and (3) text-conditioned extraction of arbitrary sources. By formulating both separation and imputation as conditional inpainting tasks in the latent space, our approach supports flexible, class-agnostic manipulation of arbitrary instrument sources. Notably, MGE-LDM can be trained jointly across heterogeneous multi-track datasets (e.g., Slakh2100, MUSDB18, MoisesDB) without relying on predefined instrument categories. Audio samples are available at our project page: https://yoongi43.github.io/MGELDM_Samples/.", "date": "2025-05-29", "categories": []}
{"id": "2507.20408", "title": "A Multi-Stage Hybrid CNN-Transformer Network for Automated Pediatric Lung Sound Classification", "abstract": "Automated analysis of lung sound auscultation is essential for monitoring respiratory health, especially in regions facing a shortage of skilled healthcare workers. While respiratory sound classification has been widely studied in adults, its ap plication in pediatric populations, particularly in children aged <6 years, remains an underexplored area. The developmental changes in pediatric lungs considerably alter the acoustic proper ties of respiratory sounds, necessitating specialized classification approaches tailored to this age group. To address this, we propose a multistage hybrid CNN-Transformer framework that combines CNN-extracted features with an attention-based architecture to classify pediatric respiratory diseases using scalogram images from both full recordings and individual breath events. Our model achieved an overall score of 0.9039 in binary event classifi cation and 0.8448 in multiclass event classification by employing class-wise focal loss to address data imbalance. At the recording level, the model attained scores of 0.720 for ternary and 0.571 for multiclass classification. These scores outperform the previous best models by 3.81% and 5.94%, respectively. This approach offers a promising solution for scalable pediatric respiratory disease diagnosis, especially in resource-limited settings.", "date": "2025-07-27", "categories": []}
{"id": "2505.11315", "title": "Improving Inference-Time Optimisation for Vocal Effects Style Transfer with a Gaussian Prior", "abstract": "Style Transfer with Inference-Time Optimisation (ST-ITO) is a recent approach for transferring the applied effects of a reference audio to an audio track. It optimises the effect parameters to minimise the distance between the style embeddings of the processed audio and the reference. However, this method treats all possible configurations equally and relies solely on the embedding space, which can result in unrealistic configurations or biased outcomes. We address this pitfall by introducing a Gaussian prior derived from the DiffVox vocal preset dataset over the parameter space. The resulting optimisation is equivalent to maximum-a-posteriori estimation. Evaluations on vocal effects transfer on the MedleyDB dataset show significant improvements across metrics compared to baselines, including a blind audio effects estimator, nearest-neighbour approaches, and uncalibrated ST-ITO. The proposed calibration reduces the parameter mean squared error by up to 33% and more closely matches the reference style. Subjective evaluations with 16 participants confirm the superiority of our method in limited data regimes. This work demonstrates how incorporating prior knowledge at inference time enhances audio effects transfer, paving the way for more effective and realistic audio processing systems.", "date": "2025-05-16", "categories": []}
{"id": "2510.14249", "title": "Do Joint Language-Audio Embeddings Encode Perceptual Timbre Semantics?", "abstract": "Understanding and modeling the relationship between language and sound is critical for applications such as music information retrieval,text-guided music generation, and audio captioning. Central to these tasks is the use of joint language-audio embedding spaces, which map textual descriptions and auditory content into a shared embedding space. While multimodal embedding models such as MS-CLAP, LAION-CLAP, and MuQ-MuLan have shown strong performance in aligning language and audio, their correspondence to human perception of timbre, a multifaceted attribute encompassing qualities such as brightness, roughness, and warmth, remains underexplored. In this paper, we evaluate the above three joint language-audio embedding models on their ability to capture perceptual dimensions of timbre. Our findings show that LAION-CLAP consistently provides the most reliable alignment with human-perceived timbre semantics across both instrumental sounds and audio effects.", "date": "2025-10-16", "categories": []}
{"id": "2510.12819", "title": "Beyond Discrete Categories: Multi-Task Valence-Arousal Modeling for Pet Vocalization Analysis", "abstract": "Traditional pet emotion recognition from vocalizations, based on discrete classification, struggles with ambiguity and capturing intensity variations. We propose a continuous Valence-Arousal (VA) model that represents emotions in a two-dimensional space. Our method uses an automatic VA label generation algorithm, enabling large-scale annotation of 42,553 pet vocalization samples. A multi-task learning framework jointly trains VA regression with auxiliary tasks (emotion, body size, gender) to enhance prediction by improving feature learning. Our Audio Transformer model achieves a validation Valence Pearson correlation of r = 0.9024 and an Arousal r = 0.7155, effectively resolving confusion between discrete categories like \"territorial\" and \"happy.\" This work introduces the first continuous VA framework for pet vocalization analysis, offering a more expressive representation for human-pet interaction, veterinary diagnostics, and behavioral training. The approach shows strong potential for deployment in consumer products like AI pet emotion translators.", "date": "2025-10-09", "categories": []}
{"id": "2510.12851", "title": "Adaptive vector steering: A training-free, layer-wise intervention for hallucination mitigation in large audio and multimodal models", "abstract": "Large Audio-Language Models and Multi-Modal Large Language Models have demonstrated strong capabilities in tasks such as Audio Question Answering (AQA), Audio Captioning, and Automatic Speech Recognition (ASR). However, there is growing evidence that these models can hallucinate about the content of the audio. To address this issue, we probe the models' internal states and propose Adaptive Vector Steering (AVS), a method that better grounds generation in audio content. We also identify a strong correlation between output correctness and internal representations. Experiments show consistent performance gains across two models and two benchmarks. On the Audio Hallucination QA dataset, our method boosts the F1-score of Gemma from 0.550 to 0.619 and Qwen from 0.626 to 0.632. Furthermore, our method increases the accuracy of Qwen on MMAU from 0.548 to 0.592, marking an 8% relative increase. To the best of our knowledge, this is the first work to apply vector steering to mitigate hallucination in audio.", "date": "2025-10-14", "categories": []}
{"id": "2510.13344", "title": "UniMoE-Audio: Unified Speech and Music Generation with Dynamic-Capacity MoE", "abstract": "Recent advances in unified multimodal models indicate a clear trend towards comprehensive content generation. However, the auditory domain remains a significant challenge, with music and speech often developed in isolation, hindering progress towards universal audio synthesis. This separation stems from inherent task conflicts and severe data imbalances, which impede the development of a truly unified audio generation model. To address this challenge, we propose UniMoE-Audio, a unified speech and music generation model within a novel Dynamic-Capacity Mixture-of-Experts (MoE) framework. Architecturally, UniMoE-Audio introduces a Top-P routing strategy for dynamic expert number allocation, and a hybrid expert design comprising routed experts for domain-specific knowledge, shared experts for domain-agnostic features, and null experts for adaptive computation skipping. To tackle data imbalance, we introduce a three-stage training curriculum: 1) Independent Specialist Training leverages original datasets to instill domain-specific knowledge into each \"proto-expert\" without interference; 2) MoE Integration and Warmup incorporates these specialists into the UniMoE-Audio architecture, warming up the gate module and shared expert using a subset of balanced dataset; and 3) Synergistic Joint Training trains the entire model end-to-end on the fully balanced dataset, fostering enhanced cross-domain synergy. Extensive experiments show that UniMoE-Audio not only achieves state-of-the-art performance on major speech and music generation benchmarks, but also demonstrates superior synergistic learning, mitigating the performance degradation typically seen in naive joint training. Our findings highlight the substantial potential of specialized MoE architecture and curated training strategies in advancing the field of universal audio generation. Homepage: https://mukioxun.github.io/Uni-MoE-site/home.html", "date": "2025-10-15", "categories": []}
{"id": "2509.15666", "title": "TISDiSS: A Training-Time and Inference-Time Scalable Framework for Discriminative Source Separation", "abstract": "Source separation is a fundamental task in speech, music, and audio processing, and it also provides cleaner and larger data for training generative models. However, improving separation performance in practice often depends on increasingly large networks, inflating training and deployment costs. Motivated by recent advances in inference-time scaling for generative modeling, we propose Training-Time and Inference-Time Scalable Discriminative Source Separation (TISDiSS), a unified framework that integrates early-split multi-loss supervision, shared-parameter design, and dynamic inference repetitions. TISDiSS enables flexible speed-performance trade-offs by adjusting inference depth without retraining additional models. We further provide systematic analyses of architectural and training choices and show that training with more inference repetitions improves shallow-inference performance, benefiting low-latency applications. Experiments on standard speech separation benchmarks demonstrate state-of-the-art performance with a reduced parameter count, establishing TISDiSS as a scalable and practical framework for adaptive source separation. Code is available at https://github.com/WingSingFung/TISDiSS.", "date": "2025-09-19", "categories": []}
{"id": "2510.12000", "title": "UALM: Unified Audio Language Model for Understanding, Generation and Reasoning", "abstract": "Recent advances in the audio language modeling (ALM) domain tackle audio understanding and text-to-audio generation as separate tasks. Very few studies attempt to unify these tasks -- an essential step toward advanced multimodal reasoning. This paper introduces U}nified Audio Language Model (UALM), which aims to unify audio understanding, text-to-audio generation, and multimodal reasoning in a single model. To achieve this goal, we first present UALM-Gen, a text-to-audio language model that directly predicts audio tokens and is comparable to state-of-the-art diffusion-based models. We then demonstrate, using proper data blending, training recipes, and inference techniques, that our single UALM model matches the quality of state-of-the-art specialized models in audio understanding, text-to-audio generation, and text reasoning. Furthermore, we present UALM-Reason, a multimodal reasoning model that utilizes both text and audio in the intermediate thinking steps to facilitate complex generation tasks. To our knowledge, this is the first demonstration in audio research of cross-modal generative reasoning, with its effectiveness confirmed by subjective evaluations.", "date": "2025-10-13", "categories": []}
{"id": "2510.05295", "title": "AUREXA-SE: Audio-Visual Unified Representation Exchange Architecture with Cross-Attention and Squeezeformer for Speech Enhancement", "abstract": "In this paper, we propose AUREXA-SE (Audio-Visual Unified Representation Exchange Architecture with Cross-Attention and Squeezeformer for Speech Enhancement), a progressive bimodal framework tailored for audio-visual speech enhancement (AVSE). AUREXA-SE jointly leverages raw audio waveforms and visual cues by employing a U-Net-based 1D convolutional encoder for audio and a Swin Transformer V2 for efficient and expressive visual feature extraction. Central to the architecture is a novel bidirectional cross-attention mechanism, which facilitates deep contextual fusion between modalities, enabling rich and complementary representation learning. To capture temporal dependencies within the fused embeddings, a stack of lightweight Squeezeformer blocks combining convolutional and attention modules is introduced. The enhanced embeddings are then decoded via a U-Net-style decoder for direct waveform reconstruction, ensuring perceptually consistent and intelligible speech output. Experimental evaluations demonstrate the effectiveness of AUREXA-SE, achieving significant performance improvements over noisy baselines, with STOI of 0.516, PESQ of 1.323, and SI-SDR of -4.322 dB. The source code of AUREXA-SE is available at https://github.com/mtanveer1/AVSEC-4-Challenge-2025.", "date": "2025-10-06", "categories": []}
{"id": "2510.09926", "title": "Phase-Aware Deep Learning with Complex-Valued CNNs for Audio Signal Applications", "abstract": "This study explores the design and application of Complex-Valued Convolutional Neural Networks (CVCNNs) in audio signal processing, with a focus on preserving and utilizing phase information often neglected in real-valued networks. We begin by presenting the foundational theoretical concepts of CVCNNs, including complex convolutions, pooling layers, Wirtinger-based differentiation, and various complex-valued activation functions. These are complemented by critical adaptations of training techniques, including complex batch normalization and weight initialization schemes, to ensure stability in training dynamics. Empirical evaluations are conducted across three stages. First, CVCNNs are benchmarked on standard image datasets, where they demonstrate competitive performance with real-valued CNNs, even under synthetic complex perturbations. Although our focus is audio signal processing, we first evaluate CVCNNs on image datasets to establish baseline performance and validate training stability before applying them to audio tasks. In the second experiment, we focus on audio classification using Mel-Frequency Cepstral Coefficients (MFCCs). CVCNNs trained on real-valued MFCCs slightly outperform real CNNs, while preserving phase in input workflows highlights challenges in exploiting phase without architectural modifications. Finally, a third experiment introduces GNNs to model phase information via edge weighting, where the inclusion of phase yields measurable gains in both binary and multi-class genre classification. These results underscore the expressive capacity of complex-valued architectures and confirm phase as a meaningful and exploitable feature in audio processing applications. While current methods show promise, especially with activations like cardioid, future advances in phase-aware design will be essential to leverage the potential of complex representations in neural networks.", "date": "2025-10-10", "categories": []}
{"id": "2510.10719", "title": "SS-DPPN: A self-supervised dual-path foundation model for the generalizable cardiac audio representation", "abstract": "The automated analysis of phonocardiograms is vital for the early diagnosis of cardiovascular disease, yet supervised deep learning is often constrained by the scarcity of expert-annotated data. In this paper, we propose the Self-Supervised Dual-Path Prototypical Network (SS-DPPN), a foundation model for cardiac audio representation and classification from unlabeled data. The framework introduces a dual-path contrastive learning based architecture that simultaneously processes 1D waveforms and 2D spectrograms using a novel hybrid loss. For the downstream task, a metric-learning approach using a Prototypical Network was used that enhances sensitivity and produces well-calibrated and trustworthy predictions. SS-DPPN achieves state-of-the-art performance on four cardiac audio benchmarks. The framework demonstrates exceptional data efficiency with a fully supervised model on three-fold reduction in labeled data. Finally, the learned representations generalize successfully across lung sound classification and heart rate estimation. Our experiments and findings validate SS-DPPN as a robust, reliable, and scalable foundation model for physiological signals.", "date": "2025-10-12", "categories": []}
{"id": "2510.10948", "title": "Unify Variables in Neural Scaling Laws for General Audio Representations via Embedding Effective Rank", "abstract": "Scaling laws have profoundly shaped our understanding of model performance in computer vision and natural language processing, yet their application to general audio representation learning remains underexplored. A key challenge lies in the multifactorial nature of general audio representation-representation quality is jointly influenced by variables such as audio length, embedding dimensionality, model depth, model architecture, data volume, etc., many of which are difficult to isolate or express analytically. In this work, we present a systematic study of scaling laws for general audio representations by utilizing embedding effective rank (RankMe) as a unifying metric that encapsulates the impact of diverse variables on representation quality. RankMe enables a label-free, information-theoretic quantification of audio embeddings, allowing us to examine scaling behaviors across a wide hyper-parameter space, including model size, training data volume, computational budget, architectural configurations, etc. Our empirical findings reveal a consistent power-law relationship between RankMe and representation quality, suggesting that embedding effective rank serves as a reliable proxy for assessing and predicting model performance in audio representation learning. This work not only validates the applicability of classical scaling principles to the general audio domain but also offers a theoretically grounded and empirically robust framework for guiding future model scaling strategies in audio foundation models.", "date": "2025-10-13", "categories": []}
{"id": "2510.11454", "title": "Audio-Maestro: Enhancing Large Audio-Language Models with Tool-Augmented Reasoning", "abstract": "Recent advancements in large multimodal models (LMMs) have shown strong capabilities in audio understanding. However, most systems rely solely on end-to-end reasoning, limiting interpretability and accuracy for tasks that require structured knowledge or specialized signal analysis. In this work, we present Audio-Maestro -- a tool-augmented audio reasoning framework that enables audio-language models to autonomously call external tools and integrate their timestamped outputs into the reasoning process. This design allows the model to analyze, transform, and interpret audio signals through specialized tools rather than relying solely on end-to-end inference. Experiments show that Audio-Maestro consistently improves general audio reasoning performance: Gemini-2.5-flash's average accuracy on MMAU-Test rises from 67.4% to 72.1%, DeSTA-2.5 from 58.3% to 62.8%, and GPT-4o from 60.8% to 63.9%. To our knowledge, Audio-Maestro is the first framework to integrate structured tool output into the large audio language model reasoning process.", "date": "2025-10-13", "categories": []}
{"id": "2508.02521", "title": "Towards Reliable Audio Deepfake Attribution and Model Recognition: A Multi-Level Autoencoder-Based Framework", "abstract": "The proliferation of audio deepfakes poses a growing threat to trust in digital communications. While detection methods have advanced, attributing audio deepfakes to their source models remains an underexplored yet crucial challenge. In this paper we introduce LAVA (Layered Architecture for Voice Attribution), a hierarchical framework for audio deepfake detection and model recognition that leverages attention-enhanced latent representations extracted by a convolutional autoencoder trained solely on fake audio. Two specialized classifiers operate on these features: Audio Deepfake Attribution (ADA), which identifies the generation technology, and Audio Deepfake Model Recognition (ADMR), which recognize the specific generative model instance. To improve robustness under open-set conditions, we incorporate confidence-based rejection thresholds. Experiments on ASVspoof2021, FakeOrReal, and CodecFake show strong performance: the ADA classifier achieves F1-scores over 95% across all datasets, and the ADMR module reaches 96.31% macro F1 across six classes. Additional tests on unseen attacks from ASVpoof2019 LA and error propagation analysis confirm LAVA's robustness and reliability. The framework advances the field by introducing a supervised approach to deepfake attribution and model recognition under open-set conditions, validated on public benchmarks and accompanied by publicly released models and code. Models and code are available at https://www.github.com/adipiz99/lava-framework.", "date": "2025-08-04", "categories": []}
{"id": "2510.09307", "title": "Target speaker anonymization in multi-speaker recordings", "abstract": "Most of the existing speaker anonymization research has focused on single-speaker audio, leading to the development of techniques and evaluation metrics optimized for such condition. This study addresses the significant challenge of speaker anonymization within multi-speaker conversational audio, specifically when only a single target speaker needs to be anonymized. This scenario is highly relevant in contexts like call centers, where customer privacy necessitates anonymizing only the customer's voice in interactions with operators. Conventional anonymization methods are often not suitable for this task. Moreover, current evaluation methodology does not allow us to accurately assess privacy protection and utility in this complex multi-speaker scenario. This work aims to bridge these gaps by exploring effective strategies for targeted speaker anonymization in conversational audio, highlighting potential problems in their development and proposing corresponding improved evaluation methodologies.", "date": "2025-10-10", "categories": []}
{"id": "2506.22783", "title": "PhonemeFake: Redefining Deepfake Realism with Language-Driven Segmental Manipulation and Adaptive Bilevel Detection", "abstract": "Deepfake (DF) attacks pose a growing threat as generative models become increasingly advanced. However, our study reveals that existing DF datasets fail to deceive human perception, unlike real DF attacks that influence public discourse. It highlights the need for more realistic DF attack vectors. We introduce PhonemeFake (PF), a DF attack that manipulates critical speech segments using language reasoning, significantly reducing human perception by up to 42% and benchmark accuracies by up to 94%. We release an easy-to-use PF dataset on HuggingFace and open-source bilevel DF segment detection model that adaptively prioritizes compute on manipulated regions. Our extensive experiments across three known DF datasets reveal that our detection model reduces EER by 91% while achieving up to 90% speed-up, with minimal compute overhead and precise localization beyond existing models as a scalable solution.", "date": "2025-06-28", "categories": []}
{"id": "2510.07293", "title": "AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs", "abstract": "Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.", "date": "2025-10-08", "categories": []}
{"id": "2506.20995", "title": "Step-by-Step Video-to-Audio Synthesis via Negative Audio Guidance", "abstract": "We propose a step-by-step video-to-audio (V2A) generation method for finer controllability over the generation process and more realistic audio synthesis. Inspired by traditional Foley workflows, our approach aims to comprehensively capture all sound events induced by a video through the incremental generation of missing sound events. To avoid the need for costly multi-reference video-audio datasets, each generation step is formulated as a negatively guided V2A process that discourages duplication of existing sounds. The guidance model is trained by finetuning a pre-trained V2A model on audio pairs from adjacent segments of the same video, allowing training with standard single-reference audiovisual datasets that are easily accessible. Objective and subjective evaluations demonstrate that our method enhances the separability of generated sounds at each step and improves the overall quality of the final composite audio, outperforming existing baselines.", "date": "2025-06-26", "categories": []}
{"id": "2508.20983", "title": "Multilingual Dataset Integration Strategies for Robust Audio Deepfake Detection: A SAFE Challenge System", "abstract": "The SAFE Challenge evaluates synthetic speech detection across three tasks: unmodified audio, processed audio with compression artifacts, and laundered audio designed to evade detection. We systematically explore self-supervised learning (SSL) front-ends, training data compositions, and audio length configurations for robust deepfake detection. Our AASIST-based approach incorporates WavLM large frontend with RawBoost augmentation, trained on a multilingual dataset of 256,600 samples spanning 9 languages and over 70 TTS systems from CodecFake, MLAAD v5, SpoofCeleb, Famous Figures, and MAILABS. Through extensive experimentation with different SSL front-ends, three training data versions, and two audio lengths, we achieved second place in both Task 1 (unmodified audio detection) and Task 3 (laundered audio detection), demonstrating strong generalization and robustness.", "date": "2025-08-28", "categories": []}
{"id": "2509.18424", "title": "Scattering Transformer: A Training-Free Transformer Architecture for Heart Murmur Detection", "abstract": "In an attempt to address the need for skilled clinicians in heart sound interpretation, recent research efforts on automating cardiac auscultation have explored deep learning approaches. The majority of these approaches have been based on supervised learning that is always challenged in occasions where training data is limited. More recently, there has been a growing interest in potentials of pre-trained self-supervised audio foundation models for biomedical end tasks. Despite exhibiting promising results, these foundational models are typically computationally intensive. Within the context of automatic cardiac auscultation, this study explores a lightweight alternative to these general-purpose audio foundation models by introducing the Scattering Transformer, a novel, training-free transformer architecture for heart murmur detection. The proposed method leverages standard wavelet scattering networks by introducing contextual dependencies in a transformer-like architecture without any backpropagation. We evaluate our approach on the public CirCor DigiScope dataset, directly comparing it against leading general-purpose foundational models. The Scattering Transformer achieves a Weighted Accuracy(WAR) of 0.786 and an Unweighted Average Recall(UAR) of 0.697, demonstrating performance highly competitive with contemporary state of the art methods. This study establishes the Scattering Transformer as a viable and promising alternative in resource-constrained setups.", "date": "2025-09-22", "categories": []}
{"id": "2510.05542", "title": "Sci-Phi: A Large Language Model Spatial Audio Descriptor", "abstract": "Acoustic scene perception involves describing the type of sounds, their timing, their direction and distance, as well as their loudness and reverberation. While audio language models excel in sound recognition, single-channel input fundamentally limits spatial understanding. This work presents Sci-Phi, a spatial audio large language model with dual spatial and spectral encoders that estimates a complete parameter set for all sound sources and the surrounding environment. Learning from over 4,000 hours of synthetic first-order Ambisonics recordings including metadata, Sci-Phi enumerates and describes up to four directional sound sources in one pass, alongside non-directional background sounds and room characteristics. We evaluate the model with a permutation-invariant protocol and 15 metrics covering content, location, timing, loudness, and reverberation, and analyze its robustness across source counts, signal-to-noise ratios, reverberation levels, and challenging mixtures of acoustically, spatially, or temporally similar sources. Notably, Sci-Phi generalizes to real room impulse responses with only minor performance degradation. Overall, this work establishes the first audio LLM capable of full spatial-scene description, with strong potential for real-world deployment. Demo: https://sci-phi-audio.github.io/demo", "date": "2025-10-07", "categories": []}
{"id": "2510.05696", "title": "Sparse deepfake detection promotes better disentanglement", "abstract": "Due to the rapid progress of speech synthesis, deepfake detection has become a major concern in the speech processing community. Because it is a critical task, systems must not only be efficient and robust, but also provide interpretable explanations. Among the different approaches for explainability, we focus on the interpretation of latent representations. In such paper, we focus on the last layer of embeddings of AASIST, a deepfake detection architecture. We use a TopK activation inspired by SAEs on this layer to obtain sparse representations which are used in the decision process. We demonstrate that sparse deepfake detection can improve detection performance, with an EER of 23.36% on ASVSpoof5 test set, with 95% of sparsity. We then show that these representations provide better disentanglement, using completeness and modularity metrics based on mutual information. Notably, some attacks are directly encoded in the latent space.", "date": "2025-10-07", "categories": []}
{"id": "2510.05828", "title": "StereoSync: Spatially-Aware Stereo Audio Generation from Video", "abstract": "Although audio generation has been widely studied over recent years, video-aligned audio generation still remains a relatively unexplored frontier. To address this gap, we introduce StereoSync, a novel and efficient model designed to generate audio that is both temporally synchronized with a reference video and spatially aligned with its visual context. Moreover, StereoSync also achieves efficiency by leveraging pretrained foundation models, reducing the need for extensive training while maintaining high-quality synthesis. Unlike existing methods that primarily focus on temporal synchronization, StereoSync introduces a significant advancement by incorporating spatial awareness into video-aligned audio generation. Indeed, given an input video, our approach extracts spatial cues from depth maps and bounding boxes, using them as cross-attention conditioning in a diffusion-based audio generation model. Such an approach allows StereoSync to go beyond simple synchronization, producing stereo audio that dynamically adapts to the spatial structure and movement of a video scene. We evaluate StereoSync on Walking The Maps, a curated dataset comprising videos from video games that feature animated characters walking through diverse environments. Experimental results demonstrate the ability of StereoSync to achieve both temporal and spatial alignment, advancing the state of the art in video-to-audio generation and resulting in a significantly more immersive and realistic audio experience.", "date": "2025-10-07", "categories": []}
{"id": "2510.06204", "title": "Modulation Discovery with Differentiable Digital Signal Processing", "abstract": "Modulations are a critical part of sound design and music production, enabling the creation of complex and evolving audio. Modern synthesizers provide envelopes, low frequency oscillators (LFOs), and more parameter automation tools that allow users to modulate the output with ease. However, determining the modulation signals used to create a sound is difficult, and existing sound-matching / parameter estimation systems are often uninterpretable black boxes or predict high-dimensional framewise parameter values without considering the shape, structure, and routing of the underlying modulation curves. We propose a neural sound-matching approach that leverages modulation extraction, constrained control signal parameterizations, and differentiable digital signal processing (DDSP) to discover the modulations present in a sound. We demonstrate the effectiveness of our approach on highly modulated synthetic and real audio samples, its applicability to different DDSP synth architectures, and investigate the trade-off it incurs between interpretability and sound-matching accuracy. We make our code and audio samples available and provide the trained DDSP synths in a VST plugin.", "date": "2025-10-07", "categories": []}
{"id": "2510.04339", "title": "Pitch-Conditioned Instrument Sound Synthesis From an Interactive Timbre Latent Space", "abstract": "This paper presents a novel approach to neural instrument sound synthesis using a two-stage semi-supervised learning framework capable of generating pitch-accurate, high-quality music samples from an expressive timbre latent space. Existing approaches that achieve sufficient quality for music production often rely on high-dimensional latent representations that are difficult to navigate and provide unintuitive user experiences. We address this limitation through a two-stage training paradigm: first, we train a pitch-timbre disentangled 2D representation of audio samples using a Variational Autoencoder; second, we use this representation as conditioning input for a Transformer-based generative model. The learned 2D latent space serves as an intuitive interface for navigating and exploring the sound landscape. We demonstrate that the proposed method effectively learns a disentangled timbre space, enabling expressive and controllable audio generation with reliable pitch conditioning. Experimental results show the model's ability to capture subtle variations in timbre while maintaining a high degree of pitch accuracy. The usability of our method is demonstrated in an interactive web application, highlighting its potential as a step towards future music production environments that are both intuitive and creatively empowering: https://pgesam.faresschulz.com", "date": "2025-10-05", "categories": []}
{"id": "2510.04738", "title": "Speak, Edit, Repeat: High-Fidelity Voice Editing and Zero-Shot TTS with Cross-Attentive Mamba", "abstract": "We introduce MAVE (Mamba with Cross-Attention for Voice Editing and Synthesis), a novel autoregressive architecture for text-conditioned voice editing and high-fidelity text-to-speech (TTS) synthesis, built on a cross-attentive Mamba backbone. MAVE achieves state-of-the-art performance in speech editing and very competitive results in zero-shot TTS, while not being explicitly trained on the latter task, outperforming leading autoregressive and diffusion models on diverse, real-world audio. By integrating Mamba for efficient audio sequence modeling with cross-attention for precise text-acoustic alignment, MAVE enables context-aware voice editing with exceptional naturalness and speaker consistency. In pairwise human evaluations on a random 40-sample subset of the RealEdit benchmark (400 judgments), 57.2% of listeners rated MAVE - edited speech as perceptually equal to the original, while 24.8% prefered the original and 18.0% MAVE - demonstrating that in the majority of cases edits are indistinguishable from the source. MAVE compares favorably with VoiceCraft and FluentSpeech both on pairwise comparisons and standalone mean opinion score (MOS) evaluations. For zero-shot TTS, MAVE exceeds VoiceCraft in both speaker similarity and naturalness, without requiring multiple inference runs or post-processing. Remarkably, these quality gains come with a significantly lower memory cost and approximately the same latency: MAVE requires ~6x less memory than VoiceCraft during inference on utterances from the RealEdit database (mean duration: 6.21s, A100, FP16, batch size 1). Our results demonstrate that MAVE establishes a new standard for flexible, high-fidelity voice editing and synthesis through the synergistic integration of structured state-space modeling and cross-modal attention.", "date": "2025-10-06", "categories": []}
{"id": "2508.20088", "title": "AudioStory: Generating Long-Form Narrative Audio with Large Language Models", "abstract": "Recent advances in text-to-audio (TTA) generation excel at synthesizing short audio clips but struggle with long-form narrative audio, which requires temporal coherence and compositional reasoning. To address this gap, we propose AudioStory, a unified framework that integrates large language models (LLMs) with TTA systems to generate structured, long-form audio narratives. AudioStory possesses strong instruction-following reasoning generation capabilities. It employs LLMs to decompose complex narrative queries into temporally ordered sub-tasks with contextual cues, enabling coherent scene transitions and emotional tone consistency. AudioStory has two appealing features: (1) Decoupled bridging mechanism: AudioStory disentangles LLM-diffuser collaboration into two specialized components, i.e., a bridging query for intra-event semantic alignment and a residual query for cross-event coherence preservation. (2) End-to-end training: By unifying instruction comprehension and audio generation within a single end-to-end framework, AudioStory eliminates the need for modular training pipelines while enhancing synergy between components. Furthermore, we establish a benchmark AudioStory-10K, encompassing diverse domains such as animated soundscapes and natural sound narratives. Extensive experiments show the superiority of AudioStory on both single-audio generation and narrative audio generation, surpassing prior TTA baselines in both instruction-following ability and audio fidelity. Our code is available at https://github.com/TencentARC/AudioStory", "date": "2025-08-27", "categories": []}
{"id": "2510.02320", "title": "WEE-Therapy: A Mixture of Weak Encoders Framework for Psychological Counseling Dialogue Analysis", "abstract": "The advancement of computational psychology requires AI tools capable of deeply understanding counseling dialogues. Existing audio language models (AudioLLMs) often rely on single speech encoders pre-trained on general data, struggling to capture domain-specific features like complex emotions and professional techniques. To address this, we propose WEE-Therapy, a multi-task AudioLLM incorporating a Weak Encoder Ensemble (WEE) mechanism. This supplements a powerful base encoder with a pool of lightweight, specialized encoders. A novel dual-routing strategy combines stable, data-independent domain knowledge with dynamic, data-dependent expert selection. Evaluated on emotion recognition, technique classification, risk detection, and summarization, WEE-Therapy achieves significant performance gains across all tasks with minimal parameter overhead, demonstrating strong potential for AI-assisted clinical analysis.", "date": "2025-09-24", "categories": []}
{"id": "2510.02401", "title": "Linear RNNs for autoregressive generation of long music samples", "abstract": "Directly learning to generate audio waveforms in an autoregressive manner is a challenging task, due to the length of the raw sequences and the existence of important structure on many different timescales. Traditional approaches based on recurrent neural networks, as well as causal convolutions and self-attention, have only had limited success on this task. However, recent work has shown that deep state space models, also referred to as linear RNNs, can be highly efficient in this context. In this work, we push the boundaries of linear RNNs applied to raw audio modeling, investigating the effects of different architectural choices and using context-parallelism to enable training on sequences up to one minute (1M tokens) in length. We present a model, HarmonicRNN, which attains state of the art log-likelihoods and perceptual metrics on small-scale datasets.", "date": "2025-10-01", "categories": []}
{"id": "2510.02916", "title": "SALSA-V: Shortcut-Augmented Long-form Synchronized Audio from Videos", "abstract": "We propose SALSA-V, a multimodal video-to-audio generation model capable of synthesizing highly synchronized, high-fidelity long-form audio from silent video content. Our approach introduces a masked diffusion objective, enabling audio-conditioned generation and the seamless synthesis of audio sequences of unconstrained length. Additionally, by integrating a shortcut loss into our training process, we achieve rapid generation of high-quality audio samples in as few as eight sampling steps, paving the way for near-real-time applications without requiring dedicated fine-tuning or retraining. We demonstrate that SALSA-V significantly outperforms existing state-of-the-art methods in both audiovisual alignment and synchronization with video content in quantitative evaluation and a human listening study. Furthermore, our use of random masking during training enables our model to match spectral characteristics of reference audio samples, broadening its applicability to professional audio synthesis tasks such as Foley generation and sound design.", "date": "2025-10-03", "categories": []}
{"id": "2510.00030", "title": "Temporal-Aware Iterative Speech Model for Dementia Detection", "abstract": "Deep learning systems often struggle with processing long sequences, where computational complexity can become a bottleneck. Current methods for automated dementia detection using speech frequently rely on static, time-agnostic features or aggregated linguistic content, lacking the flexibility to model the subtle, progressive deterioration inherent in speech production. These approaches often miss the dynamic temporal patterns that are critical early indicators of cognitive decline. In this paper, we introduce TAI-Speech, a Temporal Aware Iterative framework that dynamically models spontaneous speech for dementia detection. The flexibility of our method is demonstrated through two key innovations: 1) Optical Flow-inspired Iterative Refinement: By treating spectrograms as sequential frames, this component uses a convolutional GRU to capture the fine-grained, frame-to-frame evolution of acoustic features. 2) Cross-Attention Based Prosodic Alignment: This component dynamically aligns spectral features with prosodic patterns, such as pitch and pauses, to create a richer representation of speech production deficits linked to functional decline (IADL). TAI-Speech adaptively models the temporal evolution of each utterance, enhancing the detection of cognitive markers. Experimental results on the DementiaBank dataset show that TAI-Speech achieves a strong AUC of 0.839 and 80.6\\% accuracy, outperforming text-based baselines without relying on ASR. Our work provides a more flexible and robust solution for automated cognitive assessment, operating directly on the dynamics of raw audio.", "date": "2025-09-26", "categories": []}
{"id": "2509.26140", "title": "OWL: Geometry-Aware Spatial Reasoning for Audio Large Language Models", "abstract": "Spatial reasoning is fundamental to auditory perception, yet current audio large language models (ALLMs) largely rely on unstructured binaural cues and single step inference. This limits both perceptual accuracy in direction and distance estimation and the capacity for interpretable reasoning. Recent work such as BAT demonstrates spatial QA with binaural audio, but its reliance on coarse categorical labels (left, right, up, down) and the absence of explicit geometric supervision constrain resolution and robustness. We introduce the $\\textbf{Spatial-Acoustic Geometry Encoder (SAGE}$), a geometry-aware audio encoder that aligns binaural acoustic features with 3D spatial structure using panoramic depth images and room-impulse responses at training time, while requiring only audio at inference. Building on this representation, we present $\\textbf{OWL}$, an ALLM that integrates $\\textbf{SAGE}$ with a spatially grounded chain-of-thought to rationalize over direction-of-arrivals (DoA) and distance estimates. Through curriculum learning from perceptual QA to multi-step reasoning, $\\textbf{OWL}$ supports o'clock-level azimuth and DoA estimation. To enable large-scale training and evaluation, we construct and release $\\textbf{BiDepth}$, a dataset of over one million QA pairs combining binaural audio with panoramic depth images and room impulse responses across both in-room and out-of-room scenarios. Across two benchmark datasets, our new $\\textbf{BiDepth}$ and the public SpatialSoundQA, $\\textbf{OWL}$ reduces mean DoA error by $\\textbf{11$^{\\circ}$}$ through $\\textbf{SAGE}$ and improves spatial reasoning QA accuracy by up to $\\textbf{25}$\\% over BAT.", "date": "2025-09-30", "categories": []}
{"id": "2509.26291", "title": "Representation-Based Data Quality Audits for Audio", "abstract": "Data quality issues such as off-topic samples, near duplicates, and label errors often limit the performance of audio-based systems. This paper addresses these issues by adapting SelfClean, a representation-to-rank data auditing framework, from the image to the audio domain. This approach leverages self-supervised audio representations to identify common data quality issues, creating ranked review lists that surface distinct issues within a single, unified process. The method is benchmarked on the ESC-50, GTZAN, and a proprietary industrial dataset, using both synthetic and naturally occurring corruptions. The results demonstrate that this framework achieves state-of-the-art ranking performance, often outperforming issue-specific baselines and enabling significant annotation savings by efficiently guiding human review.", "date": "2025-09-30", "categories": []}
{"id": "2509.26329", "title": "TAU: A Benchmark for Cultural Sound Understanding Beyond Semantics", "abstract": "Large audio-language models are advancing rapidly, yet most evaluations emphasize speech or globally sourced sounds, overlooking culturally distinctive cues. This gap raises a critical question: can current models generalize to localized, non-semantic audio that communities instantly recognize but outsiders do not? To address this, we present TAU (Taiwan Audio Understanding), a benchmark of everyday Taiwanese \"soundmarks.\" TAU is built through a pipeline combining curated sources, human editing, and LLM-assisted question generation, producing 702 clips and 1,794 multiple-choice items that cannot be solved by transcripts alone. Experiments show that state-of-the-art LALMs, including Gemini 2.5 and Qwen2-Audio, perform far below local humans. TAU demonstrates the need for localized benchmarks to reveal cultural blind spots, guide more equitable multimodal evaluation, and ensure models serve communities beyond the global mainstream.", "date": "2025-09-30", "categories": []}
{"id": "2509.26564", "title": "Parametric Neural Amp Modeling with Active Learning", "abstract": "We introduce Panama, an active learning framework to train parametric guitar amp models end-to-end using a combination of an LSTM model and a WaveNet-like architecture. With \\model, one can create a virtual amp by recording samples that are determined through an ensemble-based active learning strategy to minimize the amount of datapoints needed (i.e., amp knob settings). Our strategy uses gradient-based optimization to maximize the disagreement among ensemble models, in order to identify the most informative datapoints. MUSHRA listening tests reveal that, with 75 datapoints, our models are able to match the perceptual quality of NAM, the leading open-source non-parametric amp modeler.", "date": "2025-09-30", "categories": []}
{"id": "2509.26580", "title": "Source Separation for A Cappella Music", "abstract": "In this work, we study the task of multi-singer separation in a cappella music, where the number of active singers varies across mixtures. To address this, we use a power set-based data augmentation strategy that expands limited multi-singer datasets into exponentially more training samples. To separate singers, we introduce SepACap, an adaptation of SepReformer, a state-of-the-art speaker separation model architecture. We adapt the model with periodic activations and a composite loss function that remains effective when stems are silent, enabling robust detection and separation. Experiments on the JaCappella dataset demonstrate that our approach achieves state-of-the-art performance in both full-ensemble and subset singer separation scenarios, outperforming spectrogram-based baselines while generalizing to realistic mixtures with varying numbers of singers.", "date": "2025-09-30", "categories": []}
{"id": "2509.26604", "title": "Video Object Segmentation-Aware Audio Generation", "abstract": "Existing multimodal audio generation models often lack precise user control, which limits their applicability in professional Foley workflows. In particular, these models focus on the entire video and do not provide precise methods for prioritizing a specific object within a scene, generating unnecessary background sounds, or focusing on the wrong objects. To address this gap, we introduce the novel task of video object segmentation-aware audio generation, which explicitly conditions sound synthesis on object-level segmentation maps. We present SAGANet, a new multimodal generative model that enables controllable audio generation by leveraging visual segmentation masks along with video and textual cues. Our model provides users with fine-grained and visually localized control over audio generation. To support this task and further research on segmentation-aware Foley, we propose Segmented Music Solos, a benchmark dataset of musical instrument performance videos with segmentation information. Our method demonstrates substantial improvements over current state-of-the-art methods and sets a new standard for controllable, high-fidelity Foley synthesis. Code, samples, and Segmented Music Solos are available at https://saganet.notion.site", "date": "2025-09-30", "categories": []}
{"id": "2410.15017", "title": "DM-Codec: Distilling Multimodal Representations for Speech Tokenization", "abstract": "Recent advancements in speech-language models have yielded significant improvements in speech tokenization and synthesis. However, effectively mapping the complex, multidimensional attributes of speech into discrete tokens remains challenging. This process demands acoustic, semantic, and contextual information for precise speech representations. Existing speech representations generally fall into two categories: acoustic tokens from audio codecs and semantic tokens from speech self-supervised learning models. Although recent efforts have unified acoustic and semantic tokens for improved performance, they overlook the crucial role of contextual representation in comprehensive speech modeling. Our empirical investigations reveal that the absence of contextual representations results in elevated Word Error Rate (WER) and Word Information Lost (WIL) scores in speech transcriptions. To address these limitations, we propose two novel distillation approaches: (1) a language model (LM)-guided distillation method that incorporates contextual information, and (2) a combined LM and self-supervised speech model (SM)-guided distillation technique that effectively distills multimodal representations (acoustic, semantic, and contextual) into a comprehensive speech tokenizer, termed DM-Codec. The DM-Codec architecture adopts a streamlined encoder-decoder framework with a Residual Vector Quantizer (RVQ) and incorporates the LM and SM during the training process. Experiments show DM-Codec significantly outperforms state-of-the-art speech tokenization models, reducing WER by up to 13.46%, WIL by 9.82%, and improving speech quality by 5.84% and intelligibility by 1.85% on the LibriSpeech benchmark dataset. Code, samples, and checkpoints are available at https://github.com/mubtasimahasan/DM-Codec.", "date": "2024-10-19", "categories": []}
{"id": "2411.02449", "title": "Chronic Obstructive Pulmonary Disease Prediction Using Deep Convolutional Network", "abstract": "Artificial intelligence and deep learning are increasingly applied in the clinical domain, particularly for early and accurate disease detection using medical imaging and sound. Due to limited trained personnel, there is a growing demand for automated tools to support clinicians in managing rising patient loads. Respiratory diseases such as cancer and diabetes remain major global health concerns requiring timely diagnosis and intervention. Auscultation of lung sounds, combined with chest X-rays, is an established diagnostic method for respiratory illness. This study presents a Deep Convolutional Neural Network (CNN)-based approach for the analysis of respiratory sound data to detect Chronic Obstructive Pulmonary Disease (COPD). Acoustic features extracted with the Librosa library, including Mel-Frequency Cepstral Coefficients (MFCCs), Mel-Spectrogram, Chroma, Chroma (Constant Q), and Chroma CENS, were used in training. The system also classifies disease severity as mild, moderate, or severe. Evaluation on the ICBHI database achieved 96% accuracy using 10-fold cross-validation and 90% accuracy without cross-validation. The proposed network outperforms existing methods, demonstrating potential as a practical tool for clinical deployment.", "date": "2024-11-03", "categories": []}
{"id": "2505.19203", "title": "EnvSDD: Benchmarking Environmental Sound Deepfake Detection", "abstract": "Audio generation systems now create very realistic soundscapes that can enhance media production, but also pose potential risks. Several studies have examined deepfakes in speech or singing voice. However, environmental sounds have different characteristics, which may make methods for detecting speech and singing deepfakes less effective for real-world sounds. In addition, existing datasets for environmental sound deepfake detection are limited in scale and audio types. To address this gap, we introduce EnvSDD, the first large-scale curated dataset designed for this task, consisting of 45.25 hours of real and 316.74 hours of fake audio. The test set includes diverse conditions to evaluate the generalizability, such as unseen generation models and unseen datasets. We also propose an audio deepfake detection system, based on a pre-trained audio foundation model. Results on EnvSDD show that our proposed system outperforms the state-of-the-art systems from speech and singing domains.", "date": "2025-05-25", "categories": []}
{"id": "2506.10274", "title": "Discrete Audio Tokens: More Than a Survey!", "abstract": "Discrete audio tokens are compact representations that aim to preserve perceptual quality, phonetic content, and speaker characteristics while enabling efficient storage and inference, as well as competitive performance across diverse downstream tasks. They provide a practical alternative to continuous features, enabling the integration of speech and audio into modern large language models (LLMs). As interest in token-based audio processing grows, various tokenization methods have emerged, and several surveys have reviewed the latest progress in the field. However, existing studies often focus on specific domains or tasks and lack a unified comparison across various benchmarks. This paper presents a systematic review and benchmark of discrete audio tokenizers, covering three domains: speech, music, and general audio. We propose a taxonomy of tokenization approaches based on encoder-decoder, quantization techniques, training paradigm, streamability, and application domains. We evaluate tokenizers on multiple benchmarks for reconstruction, downstream performance, and acoustic language modeling, and analyze trade-offs through controlled ablation studies. Our findings highlight key limitations, practical considerations, and open challenges, providing insight and guidance for future research in this rapidly evolving area. For more information, including our main results and tokenizer database, please refer to our website: https://poonehmousavi.github.io/dates-website/.", "date": "2025-06-12", "categories": []}
{"id": "2509.22838", "title": "Text-Independent Speaker Identification Using Audio Looping With Margin Based Loss Functions", "abstract": "Speaker identification has become a crucial component in various applications, including security systems, virtual assistants, and personalized user experiences. In this paper, we investigate the effectiveness of CosFace Loss and ArcFace Loss for text-independent speaker identification using a Convolutional Neural Network architecture based on the VGG16 model, modified to accommodate mel spectrogram inputs of variable sizes generated from the Voxceleb1 dataset. Our approach involves implementing both loss functions to analyze their effects on model accuracy and robustness, where the Softmax loss function was employed as a comparative baseline. Additionally, we examine how the sizes of mel spectrograms and their varying time lengths influence model performance. The experimental results demonstrate superior identification accuracy compared to traditional Softmax loss methods. Furthermore, we discuss the implications of these findings for future research.", "date": "2025-09-26", "categories": []}
{"id": "2509.23618", "title": "Generalizable Speech Deepfake Detection via Information Bottleneck Enhanced Adversarial Alignment", "abstract": "Neural speech synthesis techniques have enabled highly realistic speech deepfakes, posing major security risks. Speech deepfake detection is challenging due to distribution shifts across spoofing methods and variability in speakers, channels, and recording conditions. We explore learning shared discriminative features as a path to robust detection and propose Information Bottleneck enhanced Confidence-Aware Adversarial Network (IB-CAAN). Confidence-guided adversarial alignment adaptively suppresses attack-specific artifacts without erasing discriminative cues, while the information bottleneck removes nuisance variability to preserve transferable features. Experiments on ASVspoof 2019/2021, ASVspoof 5, and In-the-Wild demonstrate that IB-CAAN consistently outperforms baseline and achieves state-of-the-art performance on many benchmarks.", "date": "2025-09-28", "categories": []}
{"id": "2509.24039", "title": "End-to-end Topographic Auditory Models Replicate Signatures of Human Auditory Cortex", "abstract": "The human auditory cortex is topographically organized. Neurons with similar response properties are spatially clustered, forming smooth maps for acoustic features such as frequency in early auditory areas, and modular regions selective for music and speech in higher-order cortex. Yet, evaluations for current computational models of auditory perception do not measure whether such topographic structure is present in a candidate model. Here, we show that cortical topography is not present in the previous best-performing models at predicting human auditory fMRI responses. To encourage the emergence of topographic organization, we adapt a cortical wiring-constraint loss originally designed for visual perception. The new class of topographic auditory models, TopoAudio, are trained to classify speech, and environmental sounds from cochleagram inputs, with an added constraint that nearby units on a 2D cortical sheet develop similar tuning. Despite these additional constraints, TopoAudio achieves high accuracy on benchmark tasks comparable to the unconstrained non-topographic baseline models. Further, TopoAudio predicts the fMRI responses in the brain as well as standard models, but unlike standard models, TopoAudio develops smooth, topographic maps for tonotopy and amplitude modulation (common properties of early auditory representation, as well as clustered response modules for music and speech (higher-order selectivity observed in the human auditory cortex). TopoAudio is the first end-to-end biologically grounded auditory model to exhibit emergent topography, and our results emphasize that a wiring-length constraint can serve as a general-purpose regularization tool to achieve biologically aligned representations.", "date": "2025-09-28", "categories": []}
{"id": "2509.24404", "title": "From Sound to Setting: AI-Based Equalizer Parameter Prediction for Piano Tone Replication", "abstract": "This project presents an AI-based system for tone replication in music production, focusing on predicting EQ parameter settings directly from audio features. Unlike traditional audio-to-audio methods, our approach outputs interpretable parameter values (e.g., EQ band gains) that musicians can further adjust in their workflow. Using a dataset of piano recordings with systematically varied EQ settings, we evaluate both regression and neural network models. The neural network achieves a mean squared error of 0.0216 on multi-band tasks. The system enables practical, flexible, and automated tone matching for music producers and lays the foundation for extensions to more complex audio effects.", "date": "2025-09-29", "categories": []}
{"id": "2408.13068", "title": "On Class Separability Pitfalls In Audio-Text Contrastive Zero-Shot Learning", "abstract": "Recent advances in audio-text cross-modal contrastive learning have shown its potential towards zero-shot learning. One possibility for this is by projecting item embeddings from pre-trained backbone neural networks into a cross-modal space in which item similarity can be calculated in either domain. This process relies on a strong unimodal pre-training of the backbone networks, and on a data-intensive training task for the projectors. These two processes can be biased by unintentional data leakage, which can arise from using supervised learning in pre-training or from inadvertently training the cross-modal projection using labels from the zero-shot learning evaluation. In this study, we show that a significant part of the measured zero-shot learning accuracy is due to strengths inherited from the audio and text backbones, that is, they are not learned in the cross-modal domain and are not transferred from one modality to another.", "date": "2024-08-23", "categories": []}
{"id": "2505.13577", "title": "VocalAgent: Large Language Models for Vocal Health Diagnostics with Safety-Aware Evaluation", "abstract": "Vocal health plays a crucial role in peoples' lives, significantly impacting their communicative abilities and interactions. However, despite the global prevalence of voice disorders, many lack access to convenient diagnosis and treatment. This paper introduces VocalAgent, an audio large language model (LLM) to address these challenges through vocal health diagnosis. We leverage Qwen-Audio-Chat fine-tuned on three datasets collected in-situ from hospital patients, and present a multifaceted evaluation framework encompassing a safety assessment to mitigate diagnostic biases, cross-lingual performance analysis, and modality ablation studies. VocalAgent demonstrates superior accuracy on voice disorder classification compared to state-of-the-art baselines. Its LLM-based method offers a scalable solution for broader adoption of health diagnostics, while underscoring the importance of ethical and technical validation.", "date": "2025-05-19", "categories": []}
{"id": "2509.21625", "title": "Guiding Audio Editing with Audio Language Model", "abstract": "Audio editing plays a central role in VR/AR immersion, virtual conferencing, sound design, and other interactive media. However, recent generative audio editing models depend on template-like instruction formats and are restricted to mono-channel audio. These models fail to deal with declarative audio editing, where the user declares what the desired outcome should be, while leaving the details of editing operations to the system. We introduce SmartDJ, a novel framework for stereo audio editing that combines the reasoning capability of audio language models with the generative power of latent diffusion. Given a high-level instruction, SmartDJ decomposes it into a sequence of atomic edit operations, such as adding, removing, or spatially relocating events. These operations are then executed by a diffusion model trained to manipulate stereo audio. To support this, we design a data synthesis pipeline that produces paired examples of high-level instructions, atomic edit operations, and audios before and after each edit operation. Experiments demonstrate that SmartDJ achieves superior perceptual quality, spatial realism, and semantic alignment compared to prior audio editing methods. Demos are available at https://zitonglan.github.io/project/smartdj/smartdj.html.", "date": "2025-09-25", "categories": []}
{"id": "2509.21676", "title": "HuLA: Prosody-Aware Anti-Spoofing with Multi-Task Learning for Expressive and Emotional Synthetic Speech", "abstract": "Current anti-spoofing systems remain vulnerable to expressive and emotional synthetic speech, since they rarely leverage prosody as a discriminative cue. Prosody is central to human expressiveness and emotion, and humans instinctively use prosodic cues such as F0 patterns and voiced/unvoiced structure to distinguish natural from synthetic speech. In this paper, we propose HuLA, a two-stage prosody-aware multi-task learning framework for spoof detection. In Stage 1, a self-supervised learning (SSL) backbone is trained on real speech with auxiliary tasks of F0 prediction and voiced/unvoiced classification, enhancing its ability to capture natural prosodic variation similar to human perceptual learning. In Stage 2, the model is jointly optimized for spoof detection and prosody tasks on both real and synthetic data, leveraging prosodic awareness to detect mismatches between natural and expressive synthetic speech. Experiments show that HuLA consistently outperforms strong baselines on challenging out-of-domain dataset, including expressive, emotional, and cross-lingual attacks. These results demonstrate that explicit prosodic supervision, combined with SSL embeddings, substantially improves robustness against advanced synthetic speech attacks.", "date": "2025-09-25", "categories": []}
{"id": "2509.21749", "title": "Thinking with Sound: Audio Chain-of-Thought Enables Multimodal Reasoning in Large Audio-Language Models", "abstract": "Recent Large Audio-Language Models (LALMs) have shown strong performance on various audio understanding tasks such as speech translation and Audio Q\\&A. However, they exhibit significant limitations on challenging audio reasoning tasks in complex acoustic scenarios. These situations would greatly benefit from the use of acoustic tools like noise suppression, source separation, and precise temporal alignment, but current LALMs lack access to such tools. To address this limitation, we introduce Thinking-with-Sound (TwS), a framework that equips LALMs with Audio CoT by combining linguistic reasoning with on-the-fly audio-domain analysis. Unlike existing approaches that treat audio as static input, TwS enables models to actively think with audio signals, performing numerical analysis and digital manipulation through multimodal reasoning. To evaluate this approach, we construct MELD-Hard1k, a new robustness benchmark created by introducing various acoustic perturbations. Experiments reveal that state-of-the-art LALMs suffer dramatic performance degradation on MELD-Hard1k, with accuracy dropping by more than $50\\%$ compared to clean audio. TwS achieves substantial improvements in robustness, demonstrating both effectiveness and scalability: small models gain $24.73\\%$ absolute accuracy, with improvements scaling consistently up to $36.61\\%$ for larger models. Our findings demonstrate that Audio CoT can significantly enhance robustness without retraining, opening new directions for developing more robust audio understanding systems.", "date": "2025-09-26", "categories": []}
{"id": "2509.22063", "title": "High-Quality Sound Separation Across Diverse Categories via Visually-Guided Generative Modeling", "abstract": "We propose DAVIS, a Diffusion-based Audio-VIsual Separation framework that solves the audio-visual sound source separation task through generative learning. Existing methods typically frame sound separation as a mask-based regression problem, achieving significant progress. However, they face limitations in capturing the complex data distribution required for high-quality separation of sounds from diverse categories. In contrast, DAVIS circumvents these issues by leveraging potent generative modeling paradigms, specifically Denoising Diffusion Probabilistic Models (DDPM) and the more recent Flow Matching (FM), integrated within a specialized Separation U-Net architecture. Our framework operates by synthesizing the desired separated sound spectrograms directly from a noise distribution, conditioned concurrently on the mixed audio input and associated visual information. The inherent nature of its generative objective makes DAVIS particularly adept at producing high-quality sound separations for diverse sound categories. We present comparative evaluations of DAVIS, encompassing both its DDPM and Flow Matching variants, against leading methods on the standard AVE and MUSIC datasets. The results affirm that both variants surpass existing approaches in separation quality, highlighting the efficacy of our generative framework for tackling the audio-visual source separation task.", "date": "2025-09-26", "categories": []}
{"id": "2509.18412", "title": "Identifying birdsong syllables without labelled data", "abstract": "Identifying sequences of syllables within birdsongs is key to tackling a wide array of challenges, including bird individual identification and better understanding of animal communication and sensory-motor learning. Recently, machine learning approaches have demonstrated great potential to alleviate the need for experts to label long audio recordings by hand. However, they still typically rely on the availability of labelled data for model training, restricting applicability to a few species and datasets. In this work, we build the first fully unsupervised algorithm to decompose birdsong recordings into sequences of syllables. We first detect syllable events, then cluster them to extract templates -- syllable representations -- before performing matching pursuit to decompose the recording as a sequence of syllables. We evaluate our automatic annotations against human labels on a dataset of Bengalese finch songs and find that our unsupervised method achieves high performance. We also demonstrate that our approach can distinguish individual birds within a species through their unique vocal signatures, for both Bengalese finches and another species, the great tit.", "date": "2025-09-22", "categories": []}
{"id": "2509.20736", "title": "The Impact of Audio Watermarking on Audio Anti-Spoofing Countermeasures", "abstract": "This paper presents the first study on the impact of audio watermarking on spoofing countermeasures. While anti-spoofing systems are essential for securing speech-based applications, the influence of widely used audio watermarking, originally designed for copyright protection, remains largely unexplored. We construct watermark-augmented training and evaluation datasets, named the Watermark-Spoofing dataset, by applying diverse handcrafted and neural watermarking methods to existing anti-spoofing datasets. Experiments show that watermarking consistently degrades anti-spoofing performance, with higher watermark density correlating with higher Equal Error Rates (EERs). To mitigate this, we propose the Knowledge-Preserving Watermark Learning (KPWL) framework, enabling models to adapt to watermark-induced shifts while preserving their original-domain spoofing detection capability. These findings reveal audio watermarking as a previously overlooked domain shift and establish the first benchmark for developing watermark-resilient anti-spoofing systems. All related protocols are publicly available at https://github.com/Alphawarheads/Watermark_Spoofing.git", "date": "2025-09-25", "categories": []}
{"id": "2509.20741", "title": "Real-Time System for Audio-Visual Target Speech Enhancement", "abstract": "We present a live demonstration for RAVEN, a real-time audio-visual speech enhancement system designed to run entirely on a CPU. In single-channel, audio-only settings, speech enhancement is traditionally approached as the task of extracting clean speech from environmental noise. More recent work has explored the use of visual cues, such as lip movements, to improve robustness, particularly in the presence of interfering speakers. However, to our knowledge, no prior work has demonstrated an interactive system for real-time audio-visual speech enhancement operating on CPU hardware. RAVEN fills this gap by using pretrained visual embeddings from an audio-visual speech recognition model to encode lip movement information. The system generalizes across environmental noise, interfering speakers, transient sounds, and even singing voices. In this demonstration, attendees will be able to experience live audio-visual target speech enhancement using a microphone and webcam setup, with clean speech playback through headphones.", "date": "2025-09-25", "categories": []}
{"id": "2504.04751", "title": "Unsupervised Estimation of Nonlinear Audio Effects: Comparing Diffusion-Based and Adversarial approaches", "abstract": "Accurately estimating nonlinear audio effects without access to paired input-output signals remains a challenging problem. This work studies unsupervised probabilistic approaches for solving this task. We introduce a method, novel for this application, based on diffusion generative models for blind system identification, enabling the estimation of unknown nonlinear effects using black- and gray-box models. This study compares this method with a previously proposed adversarial approach, analyzing the performance of both methods under different parameterizations of the effect operator and varying lengths of available effected recordings. Through experiments on guitar distortion effects, we show that the diffusion-based approach provides more stable results and is less sensitive to data availability, while the adversarial approach is superior at estimating more pronounced distortion effects. Our findings contribute to the robust unsupervised blind estimation of audio effects, demonstrating the potential of diffusion models for system identification in music technology.", "date": "2025-04-07", "categories": []}
{"id": "2509.19676", "title": "Thinking While Listening: Simple Test Time Scaling For Audio Classification", "abstract": "We propose a framework that enables neural models to \"think while listening\" to everyday sounds, thereby enhancing audio classification performance. Motivated by recent advances in the reasoning capabilities of large language models, we address two central questions: (i) how can thinking be incorporated into existing audio classification pipelines to enable reasoning in the category space and improve performance, and (ii) can a new architecture be designed from the ground up to support both thinking and test-time scaling? We demonstrate that in both settings, our models exhibit improved classification accuracy. Leveraging test-time scaling, we observe consistent gains as the number of sampled traces increases. Furthermore, we evaluate two open-source reasoning models, GPT-OSS-20B and Qwen3-14B, showing that while such models are capable of zero-shot reasoning, a lightweight approach--retraining only the embedding matrix of a frozen, smaller model like GPT-2--can surpass the performance of billion-parameter text-based reasoning models.", "date": "2025-09-24", "categories": []}
{"id": "2509.18579", "title": "Teaching Audio Models to Reason: A Unified Framework for Source- and Layer-wise Distillation", "abstract": "While large audio language models excel at tasks like ASR and emotion recognition, they still struggle with complex reasoning due to the modality gap between audio and text as well as the lack of structured intermediate supervision. To address this, we propose a unified knowledge distillation framework to transfer reasoning capabilities from a high-capacity textual teacher model to a student audio models while preserving its acoustic competence. Our method introduces two key dimensions: source-wise distillation, which leverages both textual and acoustic teachers to provide complementary modality-specific supervision; and layer-wise distillation, which aligns teacher signals with appropriate student layers to improve transfer efficiency. This dual-dimensional strategy enables fine-grained control over the distillation process, effectively bridging the gap between symbolic reasoning and speech representations. Experimental results show significant improvements in audio reasoning performance, demonstrating the effectiveness of our framework as a reasoning transfer solution for audio modeling.", "date": "2025-09-23", "categories": []}
{"id": "2509.18603", "title": "SynSonic: Augmenting Sound Event Detection through Text-to-Audio Diffusion ControlNet and Effective Sample Filtering", "abstract": "Data synthesis and augmentation are essential for Sound Event Detection (SED) due to the scarcity of temporally labeled data. While augmentation methods like SpecAugment and Mix-up can enhance model performance, they remain constrained by the diversity of existing samples. Recent generative models offer new opportunities, yet their direct application to SED is challenging due to the lack of precise temporal annotations and the risk of introducing noise through unreliable filtering. To address these challenges and enable generative-based augmentation for SED, we propose SynSonic, a data augmentation method tailored for this task. SynSonic leverages text-to-audio diffusion models guided by an energy-envelope ControlNet to generate temporally coherent sound events. A joint score filtering strategy with dual classifiers ensures sample quality, and we explore its practical integration into training pipelines. Experimental results show that SynSonic improves Polyphonic Sound Detection Scores (PSDS1 and PSDS2), enhancing both temporal localization and sound class discrimination.", "date": "2025-09-23", "categories": []}
{"id": "2509.18606", "title": "FlexSED: Towards Open-Vocabulary Sound Event Detection", "abstract": "Despite recent progress in large-scale sound event detection (SED) systems capable of handling hundreds of sound classes, existing multi-class classification frameworks remain fundamentally limited. They cannot process free-text sound queries, which enable more flexible and user-friendly interaction, and they lack zero-shot capabilities and offer poor few-shot adaptability. Although text-query-based separation methods have been explored, they primarily focus on source separation and are ill-suited for SED tasks that require precise temporal localization and efficient detection across large and diverse sound vocabularies. In this paper, we propose FlexSED, an open-vocabulary sound event detection system. FlexSED builds on a pretrained audio SSL model and the CLAP text encoder, introducing an encoder-decoder composition and an adaptive fusion strategy to enable effective continuous training from pretrained weights. To ensure robust supervision, it also employs large language models (LLMs) to assist in event query selection during training, addressing challenges related to missing labels. As a result, FlexSED achieves superior performance compared to vanilla SED models on AudioSet-Strong, while demonstrating strong zero-shot and few-shot capabilities. We release the code and pretrained models to support future research and applications based on FlexSED.", "date": "2025-09-23", "categories": []}
{"id": "2509.18691", "title": "An overview of neural architectures for self-supervised audio representation learning from masked spectrograms", "abstract": "In recent years, self-supervised learning has amassed significant interest for training deep neural representations without labeled data. One such self-supervised learning approach is masked spectrogram modeling, where the objective is to learn semantically rich contextual representations by predicting removed or hidden portions of the input audio spectrogram. With the Transformer neural architecture at its core, masked spectrogram modeling has emerged as the prominent approach for learning general purpose audio representations, a.k.a. audio foundation models. Meanwhile, addressing the issues of the Transformer architecture, in particular the underlying Scaled Dot-product Attention operation, which scales quadratically with input sequence length, has led to renewed interest in recurrent sequence modeling approaches. Among them, Selective structured state space models (such as Mamba) and extended Long Short-Term Memory (xLSTM) are the two most promising approaches which have experienced widespread adoption. While the body of work on these two topics continues to grow, there is currently a lack of an adequate overview encompassing the intersection of these topics. In this paper, we present a comprehensive overview of the aforementioned research domains, covering masked spectrogram modeling and the previously mentioned neural sequence modeling architectures, Mamba and xLSTM. Further, we compare Transformers, Mamba and xLSTM based masked spectrogram models in a unified, reproducible framework on ten diverse downstream audio classification tasks, which will help interested readers to make informed decisions regarding suitability of the evaluated approaches to adjacent applications.", "date": "2025-09-23", "categories": []}
{"id": "2509.19295", "title": "Audio-Based Pedestrian Detection in the Presence of Vehicular Noise", "abstract": "Audio-based pedestrian detection is a challenging task and has, thus far, only been explored in noise-limited environments. We present a new dataset, results, and a detailed analysis of the state-of-the-art in audio-based pedestrian detection in the presence of vehicular noise. In our study, we conduct three analyses: (i) cross-dataset evaluation between noisy and noise-limited environments, (ii) an assessment of the impact of noisy data on model performance, highlighting the influence of acoustic context, and (iii) an evaluation of the model's predictive robustness on out-of-domain sounds. The new dataset is a comprehensive 1321-hour roadside dataset. It incorporates traffic-rich soundscapes. Each recording includes 16kHz audio synchronized with frame-level pedestrian annotations and 1fps video thumbnails.", "date": "2025-09-23", "categories": []}
{"id": "2309.11895", "title": "Audio Contrastive-based Fine-tuning: Decoupling Representation Learning and Classification", "abstract": "Standard fine-tuning of pre-trained audio models couples representation learning with classifier training, which can obscure the true quality of the learned representations. In this work, we advocate for a disentangled two-stage framework that separates representation refinement from downstream evaluation. First, we employ a \"contrastive-tuning\" stage to explicitly improve the geometric structure of the model's embedding space. Subsequently, we introduce a dual-probe evaluation protocol to assess the quality of these refined representations from a geometric perspective. This protocol uses a linear probe to measure global linear separability and a k-Nearest Neighbours probe to investigate the local structure of class clusters. Our experiments on a diverse set of audio classification tasks show that our framework provides a better foundation for classification, leading to improved accuracy. Our newly proposed dual-probing framework acts as a powerful analytical lens, demonstrating why contrastive learning is more effective by revealing a superior embedding space. It significantly outperforms vanilla fine-tuning, particularly on single-label datasets with a large number of classes, and also surpasses strong baselines on multi-label tasks using a Jaccard-weighted loss. Our findings demonstrate that decoupling representation refinement from classifier training is a broadly effective strategy for unlocking the full potential of pre-trained audio models. Our code will be publicly available.", "date": "2023-09-21", "categories": []}
{"id": "2503.02318", "title": "Audio-Reasoner: Improving Reasoning Capability in Large Audio Language Models", "abstract": "Recent advancements in multimodal reasoning have largely overlooked the audio modality. We introduce Audio-Reasoner, a large-scale audio language model for deep reasoning in audio tasks. We meticulously curated a large-scale and diverse multi-task audio dataset with simple annotations. Then, we leverage closed-source models to conduct secondary labeling, QA generation, along with structured COT process. These datasets together form a high-quality reasoning dataset with 1.2 million reasoning-rich samples, which we name CoTA. Following inference scaling principles, we train Audio-Reasoner on CoTA, enabling it to achieve great logical capabilities in audio reasoning. Experiments show state-of-the-art performance across key benchmarks, including MMAU-mini (+25.42%), AIR-Bench chat/foundation(+14.57%/+10.13%), and MELD (+8.01%). Our findings stress the core of structured CoT training in advancing audio reasoning.", "date": "2025-03-04", "categories": []}
{"id": "2506.12935", "title": "SoundMind: RL-Incentivized Logic Reasoning for Audio-Language Models", "abstract": "While large language models have demonstrated impressive reasoning abilities, their extension to the audio modality, particularly within large audio-language models (LALMs), remains underexplored. Addressing this gap requires a systematic approach that involves a capable base model, high-quality reasoning-oriented audio data, and effective training algorithms. In this work, we present a comprehensive solution for audio logical reasoning (ALR) tasks: we introduce SoundMind, a dataset of 6,446 audio-text annotated samples specifically curated to support complex reasoning. Building on this resource, we propose SoundMind-RL, a rule-based reinforcement learning (RL) algorithm designed to equip audio-language models with robust audio-text reasoning capabilities. By fine-tuning Qwen2.5-Omni-7B on the proposed SoundMind dataset using SoundMind-RL, we achieve strong and consistent improvements over state-of-the-art baselines on the SoundMind benchmark. This work highlights the benefit of combining high-quality, reasoning-focused datasets with specialized RL techniques, and contributes to advancing auditory intelligence in language models. The code and dataset introduced in this work are publicly available at https://github.com/xid32/SoundMind.", "date": "2025-06-15", "categories": []}
{"id": "2509.16342", "title": "Similarity-Guided Diffusion for Long-Gap Music Inpainting", "abstract": "Music inpainting aims to reconstruct missing segments of a corrupted recording. While diffusion-based generative models improve reconstruction for medium-length gaps, they often struggle to preserve musical plausibility over multi-second gaps. We introduce Similarity-Guided Diffusion Posterior Sampling (SimDPS), a hybrid method that combines diffusion-based inference with similarity search. Candidate segments are first retrieved from a corpus based on contextual similarity, then incorporated into a modified likelihood that guides the diffusion process toward contextually consistent reconstructions. Subjective evaluation on piano music inpainting with 2-s gaps shows that the proposed SimDPS method enhances perceptual plausibility compared to unguided diffusion and frequently outperforms similarity search alone when moderately similar candidates are available. These results demonstrate the potential of a hybrid similarity approach for diffusion-based audio enhancement with long gaps.", "date": "2025-09-19", "categories": []}
{"id": "2509.16649", "title": "AISTAT lab system for DCASE2025 Task6: Language-based audio retrieval", "abstract": "This report presents the AISTAT team's submission to the language-based audio retrieval task in DCASE 2025 Task 6. Our proposed system employs dual encoder architecture, where audio and text modalities are encoded separately, and their representations are aligned using contrastive learning. Drawing inspiration from methodologies of the previous year's challenge, we implemented a distillation approach and leveraged large language models (LLMs) for effective data augmentation techniques, including back-translation and LLM mix. Additionally, we incorporated clustering to introduce an auxiliary classification task for further finetuning. Our best single system achieved a mAP@16 of 46.62, while an ensemble of four systems reached a mAP@16 of 48.83 on the Clotho development test split.", "date": "2025-09-20", "categories": []}
{"id": "2509.16781", "title": "MoRoVoc: A Large Dataset for Geographical Variation Identification of the Spoken Romanian Language", "abstract": "This paper introduces MoRoVoc, the largest dataset for analyzing the regional variation of spoken Romanian. It has more than 93 hours of audio and 88,192 audio samples, balanced between the Romanian language spoken in Romania and the Republic of Moldova. We further propose a multi-target adversarial training framework for speech models that incorporates demographic attributes (i.e., age and gender of the speakers) as adversarial targets, making models discriminative for primary tasks while remaining invariant to secondary attributes. The adversarial coefficients are dynamically adjusted via meta-learning to optimize performance. Our approach yields notable gains: Wav2Vec2-Base achieves 78.21% accuracy for the variation identification of spoken Romanian using gender as an adversarial target, while Wav2Vec2-Large reaches 93.08% accuracy for gender classification when employing both dialect and age as adversarial objectives.", "date": "2025-09-20", "categories": []}
{"id": "2509.16926", "title": "Cross-Attention with Confidence Weighting for Multi-Channel Audio Alignment", "abstract": "Multi-channel audio alignment is a key requirement in bioacoustic monitoring, spatial audio systems, and acoustic localization. However, existing methods often struggle to address nonlinear clock drift and lack mechanisms for quantifying uncertainty. Traditional methods like Cross-correlation and Dynamic Time Warping assume simple drift patterns and provide no reliability measures. Meanwhile, recent deep learning models typically treat alignment as a binary classification task, overlooking inter-channel dependencies and uncertainty estimation. We introduce a method that combines cross-attention mechanisms with confidence-weighted scoring to improve multi-channel audio synchronization. We extend BEATs encoders with cross-attention layers to model temporal relationships between channels. We also develop a confidence-weighted scoring function that uses the full prediction distribution instead of binary thresholding. Our method achieved first place in the BioDCASE 2025 Task 1 challenge with 0.30 MSE average across test datasets, compared to 0.58 for the deep learning baseline. On individual datasets, we achieved 0.14 MSE on ARU data (77% reduction) and 0.45 MSE on zebra finch data (18% reduction). The framework supports probabilistic temporal alignment, moving beyond point estimates. While validated in a bioacoustic context, the approach is applicable to a broader range of multi-channel audio tasks where alignment confidence is critical. Code available on: https://github.com/Ragib-Amin-Nihal/BEATsCA", "date": "2025-09-21", "categories": []}
{"id": "2509.17219", "title": "Virtual Consistency for Audio Editing", "abstract": "Free-form, text-based audio editing remains a persistent challenge, despite progress in inversion-based neural methods. Current approaches rely on slow inversion procedures, limiting their practicality. We present a virtual-consistency based audio editing system that bypasses inversion by adapting the sampling process of diffusion models. Our pipeline is model-agnostic, requiring no fine-tuning or architectural changes, and achieves substantial speed-ups over recent neural editing baselines. Crucially, it achieves this efficiency without compromising quality, as demonstrated by quantitative benchmarks and a user study involving 16 participants.", "date": "2025-09-21", "categories": []}
{"id": "2509.17404", "title": "SongPrep: A Preprocessing Framework and End-to-end Model for Full-song Structure Parsing and Lyrics Transcription", "abstract": "Artificial Intelligence Generated Content (AIGC) is currently a popular research area. Among its various branches, song generation has attracted growing interest. Despite the abundance of available songs, effective data preparation remains a significant challenge. Converting these songs into training-ready datasets typically requires extensive manual labeling, which is both time consuming and costly. To address this issue, we propose SongPrep, an automated preprocessing pipeline designed specifically for song data. This framework streamlines key processes such as source separation, structure analysis, and lyric recognition, producing structured data that can be directly used to train song generation models. Furthermore, we introduce SongPrepE2E, an end-to-end structured lyrics recognition model based on pretrained language models. Without the need for additional source separation, SongPrepE2E is able to analyze the structure and lyrics of entire songs and provide precise timestamps. By leveraging context from the whole song alongside pretrained semantic knowledge, SongPrepE2E achieves low Diarization Error Rate (DER) and Word Error Rate (WER) on the proposed SSLD-200 dataset. Downstream tasks demonstrate that training song generation models with the data output by SongPrepE2E enables the generated songs to closely resemble those produced by humans.", "date": "2025-09-22", "categories": []}
{"id": "2509.17883", "title": "Brainprint-Modulated Target Speaker Extraction", "abstract": "Achieving robust and personalized performance in neuro-steered Target Speaker Extraction (TSE) remains a significant challenge for next-generation hearing aids. This is primarily due to two factors: the inherent non-stationarity of EEG signals across sessions, and the high inter-subject variability that limits the efficacy of generalized models. To address these issues, we propose Brainprint-Modulated Target Speaker Extraction (BM-TSE), a novel framework for personalized and high-fidelity extraction. BM-TSE first employs a spatio-temporal EEG encoder with an Adaptive Spectral Gain (ASG) module to extract stable features resilient to non-stationarity. The core of our framework is a personalized modulation mechanism, where a unified brainmap embedding is learned under the joint supervision of subject identification (SID) and auditory attention decoding (AAD) tasks. This learned brainmap, encoding both static user traits and dynamic attentional states, actively refines the audio separation process, dynamically tailoring the output to each user. Evaluations on the public KUL and Cocktail Party datasets demonstrate that BM-TSE achieves state-of-the-art performance, significantly outperforming existing methods. Our code is publicly accessible at: https://github.com/rosshan-orz/BM-TSE.", "date": "2025-09-22", "categories": []}
{"id": "2507.07318", "title": "Generating Moving 3D Soundscapes with Latent Diffusion Models", "abstract": "Spatial audio has become central to immersive applications such as VR/AR, cinema, and music. Existing generative audio models are largely limited to mono or stereo formats and cannot capture the full 3D localization cues available in first-order Ambisonics (FOA). Recent FOA models extend text-to-audio generation but remain restricted to static sources. In this work, we introduce SonicMotion, the first end-to-end latent diffusion framework capable of generating FOA audio with explicit control over moving sound sources. SonicMotion is implemented in two variations: 1) a descriptive model conditioned on natural language prompts, and 2) a parametric model conditioned on both text and spatial trajectory parameters for higher precision. To support training and evaluation, we construct a new dataset of over one million simulated FOA caption pairs that include both static and dynamic sources with annotated azimuth, elevation, and motion attributes. Experiments show that SonicMotion achieves state-of-the-art semantic alignment and perceptual quality comparable to leading text-to-audio systems, while uniquely attaining low spatial localization error.", "date": "2025-07-09", "categories": []}
{"id": "2509.12845", "title": "Improving Anomalous Sound Detection with Attribute-aware Representation from Domain-adaptive Pre-training", "abstract": "Anomalous Sound Detection (ASD) is often formulated as a machine attribute classification task, a strategy necessitated by the common scenario where only normal data is available for training. However, the exhaustive collection of machine attribute labels is laborious and impractical. To address the challenge of missing attribute labels, this paper proposes an agglomerative hierarchical clustering method for the assignment of pseudo-attribute labels using representations derived from a domain-adaptive pre-trained model, which are expected to capture machine attribute characteristics. We then apply model adaptation to this pre-trained model through supervised fine-tuning for machine attribute classification, resulting in a new state-of-the-art performance. Evaluation on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2025 Challenge dataset demonstrates that our proposed approach yields significant performance gains, ultimately outperforming our previous top-ranking system in the challenge.", "date": "2025-09-16", "categories": []}
{"id": "2509.14049", "title": "Comprehensive Evaluation of CNN-Based Audio Tagging Models on Resource-Constrained Devices", "abstract": "Convolutional Neural Networks (CNNs) have demonstrated exceptional performance in audio tagging tasks. However, deploying these models on resource-constrained devices like the Raspberry Pi poses challenges related to computational efficiency and thermal management. In this paper, a comprehensive evaluation of multiple convolutional neural network (CNN) architectures for audio tagging on the Raspberry Pi is conducted, encompassing all 1D and 2D models from the Pretrained Audio Neural Networks (PANNs) framework, a ConvNeXt-based model adapted for audio classification, as well as MobileNetV3 architectures. In addition, two PANNs-derived networks, CNN9 and CNN13, recently proposed, are also evaluated. To enhance deployment efficiency and portability across diverse hardware platforms, all models are converted to the Open Neural Network Exchange (ONNX) format. Unlike previous works that focus on a single model, our analysis encompasses a broader range of architectures and involves continuous 24-hour inference sessions to assess performance stability. Our experiments reveal that, with appropriate model selection and optimization, it is possible to maintain consistent inference latency and manage thermal behavior effectively over extended periods. These findings provide valuable insights for deploying audio tagging models in real-world edge computing scenarios.", "date": "2025-09-17", "categories": []}
{"id": "2509.15473", "title": "Breathing and Semantic Pause Detection and Exertion-Level Classification in Post-Exercise Speech", "abstract": "Post-exercise speech contains rich physiological and linguistic cues, often marked by semantic pauses, breathing pauses, and combined breathing-semantic pauses. Detecting these events enables assessment of recovery rate, lung function, and exertion-related abnormalities. However, existing works on identifying and distinguishing different types of pauses in this context are limited. In this work, building on a recently released dataset with synchronized audio and respiration signals, we provide systematic annotations of pause types. Using these annotations, we systematically conduct exploratory breathing and semantic pause detection and exertion-level classification across deep learning models (GRU, 1D CNN-LSTM, AlexNet, VGG16), acoustic features (MFCC, MFB), and layer-stratified Wav2Vec2 representations. We evaluate three setups-single feature, feature fusion, and a two-stage detection-classification cascade-under both classification and regression formulations. Results show per-type detection accuracy up to 89$\\%$ for semantic, 55$\\%$ for breathing, 86$\\%$ for combined pauses, and 73$\\%$overall, while exertion-level classification achieves 90.5$\\%$ accuracy, outperformin prior work.", "date": "2025-09-18", "categories": []}
{"id": "2509.15570", "title": "Contrastive Learning with Spectrum Information Augmentation in Abnormal Sound Detection", "abstract": "The outlier exposure method is an effective approach to address the unsupervised anomaly sound detection problem. The key focus of this method is how to make the model learn the distribution space of normal data. Based on biological perception and data analysis, it is found that anomalous audio and noise often have higher frequencies. Therefore, we propose a data augmentation method for high-frequency information in contrastive learning. This enables the model to pay more attention to the low-frequency information of the audio, which represents the normal operational mode of the machine. We evaluated the proposed method on the DCASE 2020 Task 2. The results showed that our method outperformed other contrastive learning methods used on this dataset. We also evaluated the generalizability of our method on the DCASE 2022 Task 2 dataset.", "date": "2025-09-19", "categories": []}
{"id": "2509.16195", "title": "FocalCodec-Stream: Streaming Low-Bitrate Speech Coding via Causal Distillation", "abstract": "Neural audio codecs are a fundamental component of modern generative audio pipelines. Although recent codecs achieve strong low-bitrate reconstruction and provide powerful representations for downstream tasks, most are non-streamable, limiting their use in real-time applications. We present FocalCodec-Stream, a hybrid codec based on focal modulation that compresses speech into a single binary codebook at 0.55 - 0.80 kbps with a theoretical latency of 80 ms. Our approach combines multi-stage causal distillation of WavLM with targeted architectural improvements, including a lightweight refiner module that enhances quality under latency constraints. Experiments show that FocalCodec-Stream outperforms existing streamable codecs at comparable bitrates, while preserving both semantic and acoustic information. The result is a favorable trade-off between reconstruction quality, downstream task performance, latency, and efficiency. Code and checkpoints will be released at https://github.com/lucadellalib/focalcodec.", "date": "2025-09-19", "categories": []}
{"id": "2509.12275", "title": "Omni-CLST: Error-aware Curriculum Learning with guided Selective chain-of-Thought for audio question answering", "abstract": "With the rapid progress of large audio-language models (LALMs), audio question answering (AQA) has emerged as a challenging task requiring both fine-grained audio understanding and complex reasoning. While current methods mainly rely on constructing new datasets via captioning or reasoning traces, existing high-quality AQA data remains underutilized. To address this, we propose Omni-CLST, an error-aware Curriculum Learning framework with guided Selective Chain-of-Thought. The framework efficiently leverages existing high-quality dataset through two key strategies: an error-aware curriculum that organizes samples by difficulty, and a guided thought dropout mechanism that focuses reasoning on challenging cases. Experiments show that Omni-CLST achieves 73.80% on MMAU-mini and a new state of the art of 64.30% on MMAR, demonstrating robust generalization in multimodal audio-language understanding.", "date": "2025-09-14", "categories": []}
{"id": "2509.13853", "title": "Noise Supervised Contrastive Learning and Feature-Perturbed for Anomalous Sound Detection", "abstract": "Unsupervised anomalous sound detection aims to detect unknown anomalous sounds by training a model using only normal audio data. Despite advancements in self-supervised methods, the issue of frequent false alarms when handling samples of the same type from different machines remains unresolved. This paper introduces a novel training technique called one-stage supervised contrastive learning (OS-SCL), which significantly addresses this problem by perturbing features in the embedding space and employing a one-stage noisy supervised contrastive learning approach. On the DCASE 2020 Challenge Task 2, it achieved 94.64\\% AUC, 88.42\\% pAUC, and 89.24\\% mAUC using only Log-Mel features. Additionally, a time-frequency feature named TFgram is proposed, which is extracted from raw audio. This feature effectively captures critical information for anomalous sound detection, ultimately achieving 95.71\\% AUC, 90.23\\% pAUC, and 91.23\\% mAUC. The source code is available at: \\underline{www.github.com/huangswt/OS-SCL}.", "date": "2025-09-17", "categories": []}
{"id": "2509.14379", "title": "Diffusion-Based Unsupervised Audio-Visual Speech Separation in Noisy Environments with Noise Prior", "abstract": "In this paper, we address the problem of single-microphone speech separation in the presence of ambient noise. We propose a generative unsupervised technique that directly models both clean speech and structured noise components, training exclusively on these individual signals rather than noisy mixtures. Our approach leverages an audio-visual score model that incorporates visual cues to serve as a strong generative speech prior. By explicitly modelling the noise distribution alongside the speech distribution, we enable effective decomposition through the inverse problem paradigm. We perform speech separation by sampling from the posterior distributions via a reverse diffusion process, which directly estimates and removes the modelled noise component to recover clean constituent signals. Experimental results demonstrate promising performance, highlighting the effectiveness of our direct noise modelling approach in challenging acoustic environments.", "date": "2025-09-17", "categories": []}
{"id": "2509.14666", "title": "Spatial Audio Motion Understanding and Reasoning", "abstract": "Spatial audio reasoning enables machines to interpret auditory scenes by understanding events and their spatial attributes. In this work, we focus on spatial audio understanding with an emphasis on reasoning about moving sources. First, we introduce a spatial audio encoder that processes spatial audio to detect multiple overlapping events and estimate their spatial attributes, Direction of Arrival (DoA) and source distance, at the frame level. To generalize to unseen events, we incorporate an audio grounding model that aligns audio features with semantic audio class text embeddings via a cross-attention mechanism. Second, to answer complex queries about dynamic audio scenes involving moving sources, we condition a large language model (LLM) on structured spatial attributes extracted by our model. Finally, we introduce a spatial audio motion understanding and reasoning benchmark dataset and demonstrate our framework's performance against the baseline model.", "date": "2025-09-18", "categories": []}
{"id": "2509.15140", "title": "FCPE: A Fast Context-based Pitch Estimation Model", "abstract": "Pitch estimation (PE) in monophonic audio is crucial for MIDI transcription and singing voice conversion (SVC), but existing methods suffer significant performance degradation under noise. In this paper, we propose FCPE, a fast context-based pitch estimation model that employs a Lynx-Net architecture with depth-wise separable convolutions to effectively capture mel spectrogram features while maintaining low computational cost and robust noise tolerance. Experiments show that our method achieves 96.79\\% Raw Pitch Accuracy (RPA) on the MIR-1K dataset, on par with the state-of-the-art methods. The Real-Time Factor (RTF) is 0.0062 on a single RTX 4090 GPU, which significantly outperforms existing algorithms in efficiency. Code is available at https://github.com/CNChTu/FCPE.", "date": "2025-09-18", "categories": []}
{"id": "2408.11915", "title": "Video-Foley: Two-Stage Video-To-Sound Generation via Temporal Event Condition For Foley Sound", "abstract": "Foley sound synthesis is crucial for multimedia production, enhancing user experience by synchronizing audio and video both temporally and semantically. Recent studies on automating this labor-intensive process through video-to-sound generation face significant challenges. Systems lacking explicit temporal features suffer from poor alignment and controllability, while timestamp-based models require costly and subjective human annotation. We propose Video-Foley, a video-to-sound system using Root Mean Square (RMS) as an intuitive condition with semantic timbre prompts (audio or text). RMS, a frame-level intensity envelope closely related to audio semantics, acts as a temporal event feature to guide audio generation from video. The annotation-free self-supervised learning framework consists of two stages, Video2RMS and RMS2Sound, incorporating novel ideas including RMS discretization and RMS-ControlNet with a pretrained text-to-audio model. Our extensive evaluation shows that Video-Foley achieves state-of-the-art performance in audio-visual alignment and controllability for sound timing, intensity, timbre, and nuance. Source code, model weights and demos are available on our companion website. (https://jnwnlee.github.io/video-foley-demo)", "date": "2024-08-21", "categories": []}
{"id": "2509.13390", "title": "A Domain Knowledge Informed Approach for Anomaly Detection of Electric Vehicle Interior Sounds", "abstract": "The detection of anomalies in automotive cabin sounds is critical for ensuring vehicle quality and maintaining passenger comfort. In many real-world settings, this task is more appropriately framed as an unsupervised learning problem rather than the supervised case due to the scarcity or complete absence of labeled faulty data. In such an unsupervised setting, the model is trained exclusively on healthy samples and detects anomalies as deviations from normal behavior. However, in the absence of labeled faulty samples for validation and the limited reliability of commonly used metrics, such as validation reconstruction error, effective model selection remains a significant challenge. To overcome these limitations, a domain-knowledge-informed approach for model selection is proposed, in which proxy-anomalies engineered through structured perturbations of healthy spectrograms are used in the validation set to support model selection. The proposed methodology is evaluated on a high-fidelity electric vehicle dataset comprising healthy and faulty cabin sounds across five representative fault types viz., Imbalance, Modulation, Whine, Wind, and Pulse Width Modulation. This dataset, generated using advanced sound synthesis techniques, and validated via expert jury assessments, has been made publicly available to facilitate further research. Experimental evaluations on the five fault cases demonstrate the selection of optimal models using proxy-anomalies, significantly outperform conventional model selection strategies.", "date": "2025-09-16", "categories": []}
{"id": "2509.13878", "title": "Mixture of Low-Rank Adapter Experts in Generalizable Audio Deepfake Detection", "abstract": "Foundation models such as Wav2Vec2 excel at representation learning in speech tasks, including audio deepfake detection. However, after being fine-tuned on a fixed set of bonafide and spoofed audio clips, they often fail to generalize to novel deepfake methods not represented in training. To address this, we propose a mixture-of-LoRA-experts approach that integrates multiple low-rank adapters (LoRA) into the model's attention layers. A routing mechanism selectively activates specialized experts, enhancing adaptability to evolving deepfake attacks. Experimental results show that our method outperforms standard fine-tuning in both in-domain and out-of-domain scenarios, reducing equal error rates relative to baseline models. Notably, our best MoE-LoRA model lowers the average out-of-domain EER from 8.55\\% to 6.08\\%, demonstrating its effectiveness in achieving generalizable audio deepfake detection.", "date": "2025-09-17", "categories": []}
{"id": "2504.00730", "title": "Detection of Disease on Nasal Breath Sound by New Lightweight Architecture: Using COVID-19 as An Example", "abstract": "Background. Infectious diseases, particularly COVID-19, continue to be a significant global health issue. Although many countries have reduced or stopped large-scale testing measures, the detection of such diseases remains a propriety. Objective. This study aims to develop a novel, lightweight deep neural network for efficient, accurate, and cost-effective detection of COVID-19 using a nasal breathing audio data collected via smartphones. Methodology. Nasal breathing audio from 128 patients diagnosed with the Omicron variant was collected. Mel-Frequency Cepstral Coefficients (MFCCs), a widely used feature in speech and sound analysis, were employed for extracting important characteristics from the audio signals. Additional feature selection was performed using Random Forest (RF) and Principal Component Analysis (PCA) for dimensionality reduction. A Dense-ReLU-Dropout model was trained with K-fold cross-validation (K=3), and performance metrics like accuracy, precision, recall, and F1-score were used to evaluate the model. Results. The proposed model achieved 97% accuracy in detecting COVID-19 from nasal breathing sounds, outperforming state-of-the-art methods such as those by [23] and [13]. Our Dense-ReLU-Dropout model, using RF and PCA for feature selection, achieves high accuracy with greater computational efficiency compared to existing methods that require more complex models or larger datasets. Conclusion. The findings suggest that the proposed method holds significant potential for clinical implementation, advancing smartphone-based diagnostics in infectious diseases. The Dense-ReLU-Dropout model, combined with innovative feature processing techniques, offers a promising approach for efficient and accurate COVID-19 detection, showcasing the capabilities of mobile device-based diagnostics", "date": "2025-04-01", "categories": []}
{"id": "2509.13285", "title": "Contrastive timbre representations for musical instrument and synthesizer retrieval", "abstract": "Efficiently retrieving specific instrument timbres from audio mixtures remains a challenge in digital music production. This paper introduces a contrastive learning framework for musical instrument retrieval, enabling direct querying of instrument databases using a single model for both single- and multi-instrument sounds. We propose techniques to generate realistic positive/negative pairs of sounds for virtual musical instruments, such as samplers and synthesizers, addressing limitations in common audio data augmentation methods.\n  The first experiment focuses on instrument retrieval from a dataset of 3,884 instruments, using single-instrument audio as input. Contrastive approaches are competitive with previous works based on classification pre-training. The second experiment considers multi-instrument retrieval with a mixture of instruments as audio input. In this case, the proposed contrastive framework outperforms related works, achieving 81.7\\% top-1 and 95.7\\% top-5 accuracies for three-instrument mixtures.", "date": "2025-09-16", "categories": []}
{"id": "2410.06927", "title": "Spectral and Rhythm Features for Audio Classification with Deep Convolutional Neural Networks", "abstract": "Convolutional neural networks (CNNs) are widely used in computer vision. They can be used not only for conventional digital image material to recognize patterns, but also for feature extraction from digital imagery representing spectral and rhythm features extracted from time-domain digital audio signals for the acoustic classification of sounds. Different spectral and rhythm feature representations like mel-scaled spectrograms, mel-frequency cepstral coefficients (MFCCs), cyclic tempograms, short-time Fourier transform (STFT) chromagrams, constant-Q transform (CQT) chromagrams and chroma energy normalized statistics (CENS) chromagrams are investigated in terms of the audio classification performance using a deep convolutional neural network. It can be clearly shown that the mel-scaled spectrograms and the mel-frequency cepstral coefficients (MFCCs) perform significantly better than the other spectral and rhythm features investigated in this research for audio classification tasks using deep CNNs. The experiments were carried out with the aid of the ESC-50 dataset with 2,000 labeled environmental audio recordings.", "date": "2024-10-09", "categories": []}
{"id": "2506.07233", "title": "Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding", "abstract": "Large Audio-Language Models (LALMs) can take audio and text as the inputs and answer questions about the audio. While prior LALMs have shown strong performance on standard benchmarks, there has been alarming evidence that LALMs can hallucinate what is presented in the audio. To mitigate the hallucination of LALMs, we introduce Audio-Aware Decoding (AAD), a lightweight inference-time strategy that uses contrastive decoding to compare the token prediction logits with and without the audio context. By contrastive decoding, AAD promotes the tokens whose probability increases when the audio is present. We conduct our experiment on object hallucination datasets with three LALMs and show that AAD improves the F1 score by 0.046 to 0.428. We also show that AAD can improve the accuracy on general audio QA datasets like Clotho-AQA by 5.4% to 10.3%. We conduct thorough ablation studies to understand the effectiveness of each component in AAD.", "date": "2025-06-08", "categories": []}
{"id": "2507.06070", "title": "Contrastive and Transfer Learning for Effective Audio Fingerprinting through a Real-World Evaluation Protocol", "abstract": "Recent advances in song identification leverage deep neural networks to learn compact audio fingerprints directly from raw waveforms. While these methods perform well under controlled conditions, their accuracy drops significantly in real-world scenarios where the audio is captured via mobile devices in noisy environments. In this paper, we introduce a novel evaluation protocol designed to better reflect such real-world conditions. We generate three recordings of the same audio, each with increasing levels of noise, captured using a mobile device's microphone. Our results reveal a substantial performance drop for two state-of-the-art CNN-based models under this protocol, compared to previously reported benchmarks. Additionally, we highlight the critical role of the augmentation pipeline during training with contrastive loss. By introduction low pass and high pass filters in the augmentation pipeline we significantly increase the performance of both systems in our proposed evaluation. Furthermore, we develop a transformer-based model with a tailored projection module and demonstrate that transferring knowledge from a semantically relevant domain yields a more robust solution. The transformer architecture outperforms CNN-based models across all noise levels, and query durations. In low noise conditions it achieves 47.99% for 1-sec queries, and 97% for 10-sec queries in finding the correct song, surpassing by 14%, and by 18.5% the second-best performing model, respectively, Under heavy noise levels, we achieve a detection rate 56.5% for 15-second query duration. All experiments are conducted on public large-scale dataset of over 100K songs, with queries matched against a database of 56 million vectors.", "date": "2025-07-08", "categories": []}
{"id": "2509.07756", "title": "Spectral and Rhythm Feature Performance Evaluation for Category and Class Level Audio Classification with Deep Convolutional Neural Networks", "abstract": "Next to decision tree and k-nearest neighbours algorithms deep convolutional neural networks (CNNs) are widely used to classify audio data in many domains like music, speech or environmental sounds. To train a specific CNN various spectral and rhythm features like mel-scaled spectrograms, mel-frequency cepstral coefficients (MFCC), cyclic tempograms, short-time Fourier transform (STFT) chromagrams, constant-Q transform (CQT) chromagrams and chroma energy normalized statistics (CENS) chromagrams can be used as digital image input data for the neural network. The performance of these spectral and rhythm features for audio category level as well as audio class level classification is investigated in detail with a deep CNN and the ESC-50 dataset with 2,000 labeled environmental audio recordings using an end-to-end deep learning pipeline. The evaluated metrics accuracy, precision, recall and F1 score for multiclass classification clearly show that the mel-scaled spectrograms and the mel-frequency cepstral coefficients (MFCC) perform significantly better then the other spectral and rhythm features investigated in this research for audio classification tasks using deep CNNs.", "date": "2025-09-09", "categories": []}
{"id": "2509.12003", "title": "Improving Out-of-Domain Audio Deepfake Detection via Layer Selection and Fusion of SSL-Based Countermeasures", "abstract": "Audio deepfake detection systems based on frozen pre-trained self-supervised learning (SSL) encoders show a high level of performance when combined with layer-weighted pooling methods, such as multi-head factorized attentive pooling (MHFA). However, they still struggle to generalize to out-of-domain (OOD) conditions. We tackle this problem by studying the behavior of six different pre-trained SSLs, on four different test corpora. We perform a layer-by-layer analysis to determine which layers contribute most. Next, we study the pooling head, comparing a strategy based on a single layer with automatic selection via MHFA. We observed that selecting the best layer gave very good results, while reducing system parameters by up to 80%. A wide variation in performance as a function of test corpus and SSL model is also observed, showing that the pre-training strategy of the encoder plays a role. Finally, score-level fusion of several encoders improved generalization to OOD attacks.", "date": "2025-09-15", "categories": []}
{"id": "2509.02622", "title": "IS${}^3$ : Generic Impulsive--Stationary Sound Separation in Acoustic Scenes using Deep Filtering", "abstract": "We are interested in audio systems capable of performing a differentiated processing of stationary backgrounds and isolated acoustic events within an acoustic scene, whether for applying specific processing methods to each part or for focusing solely on one while ignoring the other. Such systems have applications in real-world scenarios, including robust adaptive audio rendering systems (e.g., EQ or compression), plosive attenuation in voice mixing, noise suppression or reduction, robust acoustic event classification or even bioacoustics. To this end, we introduce IS${}^3$, a neural network designed for Impulsive--Stationary Sound Separation, that isolates impulsive acoustic events from the stationary background using a deep filtering approach, that can act as a pre-processing stage for the above-mentioned tasks. To ensure optimal training, we propose a sophisticated data generation pipeline that curates and adapts existing datasets for this task. We demonstrate that a learning-based approach, build on a relatively lightweight neural architecture and trained with well-designed and varied data, is successful in this previously unaddressed task, outperforming the Harmonic--Percussive Sound Separation masking method, adapted from music signal processing research, and wavelet filtering on objective separation metrics.", "date": "2025-09-01", "categories": []}
{"id": "2509.09550", "title": "Finite Scalar Quantization Enables Redundant and Transmission-Robust Neural Audio Compression at Low Bit-rates", "abstract": "Neural Audio Codecs (NACs) have become increasingly adopted in speech processing tasks due to their excellent rate-distortion performance and compatibility with Large Language Models (LLMs) as discrete feature representations for audio generation. While most existing codecs rely on Residual Vector Quantization (RVQ), Finite Scalar Quantization (FSQ) has recently emerged as a compelling alternative that simplifies training and natively supports single codebooks. We introduce NeuCodec, an FSQ-based NAC, and show that FSQ encodes baked-in redundancy which produces an encoding which is robust when transmitted through noisy channels. First, through an encoder distillation experiment, we show that two different encoders can learn to encode identical audio into vastly different code sequences whilst maintaining comparable reconstruction quality with the same quantizer and decoder. Second, we demonstrate that FSQ has vastly superior bit-level perturbation robustness by comparing the performance of RVQ and FSQ codecs when simulating the transmission of code sequences through a noisy channel.", "date": "2025-09-11", "categories": []}
{"id": "2509.09836", "title": "CoDiCodec: Unifying Continuous and Discrete Compressed Representations of Audio", "abstract": "Efficiently representing audio signals in a compressed latent space is critical for latent generative modelling. However, existing autoencoders often force a choice between continuous embeddings and discrete tokens. Furthermore, achieving high compression ratios while maintaining audio fidelity remains a challenge. We introduce CoDiCodec, a novel audio autoencoder that overcomes these limitations by both efficiently encoding global features via summary embeddings, and by producing both compressed continuous embeddings at ~ 11 Hz and discrete tokens at a rate of 2.38 kbps from the same trained model, offering unprecedented flexibility for different downstream generative tasks. This is achieved through Finite Scalar Quantization (FSQ) and a novel FSQ-dropout technique, and does not require additional loss terms beyond the single consistency loss used for end-to-end training. CoDiCodec supports both autoregressive decoding and a novel parallel decoding strategy, with the latter achieving superior audio quality and faster decoding. CoDiCodec outperforms existing continuous and discrete autoencoders at similar bitrates in terms of reconstruction audio quality. Our work enables a unified approach to audio compression, bridging the gap between continuous and discrete generative modelling paradigms.", "date": "2025-09-11", "categories": []}
{"id": "2509.10391", "title": "Improving Audio Event Recognition with Consistency Regularization", "abstract": "Consistency regularization (CR), which enforces agreement between model predictions on augmented views, has found recent benefits in automatic speech recognition [1]. In this paper, we propose the use of consistency regularization for audio event recognition, and demonstrate its effectiveness on AudioSet. With extensive ablation studies for both small ($\\sim$20k) and large ($\\sim$1.8M) supervised training sets, we show that CR brings consistent improvement over supervised baselines which already heavily utilize data augmentation, and CR using stronger augmentation and multiple augmentations leads to additional gain for the small training set. Furthermore, we extend the use of CR into the semi-supervised setup with 20K labeled samples and 1.8M unlabeled samples, and obtain performance improvement over our best model trained on the small set.", "date": "2025-09-12", "categories": []}
{"id": "2508.11609", "title": "Pretrained Conformers for Audio Fingerprinting and Retrieval", "abstract": "Conformers have shown great results in speech processing due to their ability to capture both local and global interactions. In this work, we utilize a self-supervised contrastive learning framework to train conformer-based encoders that are capable of generating unique embeddings for small segments of audio, generalizing well to previously unseen data. We achieve state-of-the-art results for audio retrieval tasks while using only 3 seconds of audio to generate embeddings. Our models are almost completely immune to temporal misalignments and achieve state-of-the-art results in cases of other audio distortions such as noise, reverb or extreme temporal stretching. Code and models are made publicly available and the results are easy to reproduce as we train and test using popular and freely available datasets of different sizes.", "date": "2025-08-15", "categories": []}
{"id": "2509.09204", "title": "Bona fide Cross Testing Reveals Weak Spot in Audio Deepfake Detection Systems", "abstract": "Audio deepfake detection (ADD) models are commonly evaluated using datasets that combine multiple synthesizers, with performance reported as a single Equal Error Rate (EER). However, this approach disproportionately weights synthesizers with more samples, underrepresenting others and reducing the overall reliability of EER. Additionally, most ADD datasets lack diversity in bona fide speech, often featuring a single environment and speech style (e.g., clean read speech), limiting their ability to simulate real-world conditions. To address these challenges, we propose bona fide cross-testing, a novel evaluation framework that incorporates diverse bona fide datasets and aggregates EERs for more balanced assessments. Our approach improves robustness and interpretability compared to traditional evaluation methods. We benchmark over 150 synthesizers across nine bona fide speech types and release a new dataset to facilitate further research at https://github.com/cyaaronk/audio_deepfake_eval.", "date": "2025-09-11", "categories": []}
{"id": "2509.09262", "title": "Adaptive Knowledge Distillation using a Device-Aware Teacher for Low-Complexity Acoustic Scene Classification", "abstract": "In this technical report, we describe our submission for Task 1, Low-Complexity Device-Robust Acoustic Scene Classification, of the DCASE 2025 Challenge. Our work tackles the dual challenges of strict complexity constraints and robust generalization to both seen and unseen devices, while also leveraging the new rule allowing the use of device labels at test time. Our proposed system is based on a knowledge distillation framework where an efficient CP-MobileNet student learns from a compact, specialized two-teacher ensemble. This ensemble combines a baseline PaSST teacher, trained with standard cross-entropy, and a 'generalization expert' teacher. This expert is trained using our novel Device-Aware Feature Alignment (DAFA) loss, adapted from prior work, which explicitly structures the feature space for device robustness. To capitalize on the availability of test-time device labels, the distilled student model then undergoes a final device-specific fine-tuning stage. Our proposed system achieves a final accuracy of 57.93\\% on the development set, demonstrating a significant improvement over the official baseline, particularly on unseen devices.", "date": "2025-09-11", "categories": []}
{"id": "2509.09515", "title": "Cough Classification using Few-Shot Learning", "abstract": "This paper investigates the effectiveness of few-shot learning for respiratory sound classification, focusing on coughbased detection of COVID-19, Flu, and healthy conditions. We leverage Prototypical Networks with spectrogram representations of cough sounds to address the challenge of limited labeled data. Our study evaluates whether few-shot learning can enable models to achieve performance comparable to traditional deep learning approaches while using significantly fewer training samples. Additionally, we compare multi-class and binary classification models to assess whether multi-class models can perform comparably to their binary counterparts. Experimental findings show that few-shot learning models can achieve competitive accuracy. Our model attains 74.87% accuracy in multi-class classification with only 15 support examples per class, while binary classification achieves over 70% accuracy across all class pairs. Class-wise analysis reveals Flu as the most distinguishable class, and Healthy as the most challenging. Statistical tests (paired t-test p = 0.149, Wilcoxon p = 0.125) indicate no significant performance difference between binary and multiclass models, supporting the viability of multi-class classification in this setting. These results highlight the feasibility of applying few-shot learning in medical diagnostics, particularly when large labeled datasets are unavailable.", "date": "2025-09-11", "categories": []}
{"id": "2411.04337", "title": "Neural-Enhanced Dynamic Range Compression Inversion: A Hybrid Approach for Restoring Audio Dynamics", "abstract": "Dynamic Range Compression (DRC) is a widely used audio effect that adjusts signal dynamics for applications in music production, broadcasting, and speech processing. Inverting DRC is of broad importance for restoring the original dynamics, enabling remixing, and enhancing the overall audio quality. Existing DRC inversion methods either overlook key parameters or rely on precise parameter values, which can be challenging to estimate accurately. To address this limitation, we introduce a hybrid approach that combines model-based DRC inversion with neural networks to achieve robust DRC parameter estimation and audio restoration simultaneously. Our method uses tailored neural network architectures (classification and regression), which are then integrated into a model-based inversion framework to reconstruct the original signal. Experimental evaluations on various music and speech datasets confirm the effectiveness and robustness of our approach, outperforming several state-of-the-art techniques.", "date": "2024-11-07", "categories": []}
{"id": "2509.08283", "title": "Segment Transformer: AI-Generated Music Detection via Music Structural Analysis", "abstract": "Audio and music generation systems have been remarkably developed in the music information retrieval (MIR) research field. The advancement of these technologies raises copyright concerns, as ownership and authorship of AI-generated music (AIGM) remain unclear. Also, it can be difficult to determine whether a piece was generated by AI or composed by humans clearly. To address these challenges, we aim to improve the accuracy of AIGM detection by analyzing the structural patterns of music segments. Specifically, to extract musical features from short audio clips, we integrated various pre-trained models, including self-supervised learning (SSL) models or an audio effect encoder, each within our suggested transformer-based framework. Furthermore, for long audio, we developed a segment transformer that divides music into segments and learns inter-segment relationships. We used the FakeMusicCaps and SONICS datasets, achieving high accuracy in both the short-audio and full-audio detection experiments. These findings suggest that integrating segment-level musical features into long-range temporal analysis can effectively enhance both the performance and robustness of AIGM detection systems.", "date": "2025-09-10", "categories": []}
{"id": "2506.00681", "title": "Learning to Upsample and Upmix Audio in the Latent Domain", "abstract": "Neural audio autoencoders create compact latent representations that preserve perceptually important information, serving as the foundation for both modern audio compression systems and generation approaches like next-token prediction and latent diffusion. Despite their prevalence, most audio processing operations, such as spatial and spectral up-sampling, still inefficiently operate on raw waveforms or spectral representations rather than directly on these compressed representations. We propose a framework that performs audio processing operations entirely within an autoencoder's latent space, eliminating the need to decode to raw audio formats. Our approach dramatically simplifies training by operating solely in the latent domain, with a latent L1 reconstruction term, augmented by a single latent adversarial discriminator. This contrasts sharply with raw-audio methods that typically require complex combinations of multi-scale losses and discriminators. Through experiments in bandwidth extension and mono-to-stereo up-mixing, we demonstrate computational efficiency gains of up to 100x while maintaining quality comparable to post-processing on raw audio. This work establishes a more efficient paradigm for audio processing pipelines that already incorporate autoencoders, enabling significantly faster and more resource-efficient workflows across various audio tasks.", "date": "2025-05-31", "categories": []}
{"id": "2507.07867", "title": "Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders", "abstract": "Neural audio codecs and autoencoders have emerged as versatile models for audio compression, transmission, feature-extraction, and latent-space generation. However, a key limitation is that most are trained to maximize reconstruction fidelity, often neglecting the specific latent structure necessary for optimal performance in diverse downstream applications. We propose a simple, post-hoc framework to address this by modifying the bottleneck of a pre-trained autoencoder. Our method introduces a \"Re-Bottleneck\", an inner bottleneck trained exclusively through latent space losses to instill user-defined structure. We demonstrate the framework's effectiveness in three experiments. First, we enforce an ordering on latent channels without sacrificing reconstruction quality. Second, we align latents with semantic embeddings, analyzing the impact on downstream diffusion modeling. Third, we introduce equivariance, ensuring that a filtering operation on the input waveform directly corresponds to a specific transformation in the latent space. Ultimately, our Re-Bottleneck framework offers a flexible and efficient way to tailor representations of neural audio models, enabling them to seamlessly meet the varied demands of different applications with minimal additional training.", "date": "2025-07-10", "categories": []}
{"id": "2509.07132", "title": "Adversarial Attacks on Audio Deepfake Detection: A Benchmark and Comparative Study", "abstract": "The widespread use of generative AI has shown remarkable success in producing highly realistic deepfakes, posing a serious threat to various voice biometric applications, including speaker verification, voice biometrics, audio conferencing, and criminal investigations. To counteract this, several state-of-the-art (SoTA) audio deepfake detection (ADD) methods have been proposed to identify generative AI signatures to distinguish between real and deepfake audio. However, the effectiveness of these methods is severely undermined by anti-forensic (AF) attacks that conceal generative signatures. These AF attacks span a wide range of techniques, including statistical modifications (e.g., pitch shifting, filtering, noise addition, and quantization) and optimization-based attacks (e.g., FGSM, PGD, C \\& W, and DeepFool). In this paper, we investigate the SoTA ADD methods and provide a comparative analysis to highlight their effectiveness in exposing deepfake signatures, as well as their vulnerabilities under adversarial conditions. We conducted an extensive evaluation of ADD methods on five deepfake benchmark datasets using two categories: raw and spectrogram-based approaches. This comparative analysis enables a deeper understanding of the strengths and limitations of SoTA ADD methods against diverse AF attacks. It does not only highlight vulnerabilities of ADD methods, but also informs the design of more robust and generalized detectors for real-world voice biometrics. It will further guide future research in developing adaptive defense strategies that can effectively counter evolving AF techniques.", "date": "2025-09-08", "categories": []}
{"id": "2509.07586", "title": "Exploring System Adaptations For Minimum Latency Real-Time Piano Transcription", "abstract": "Advances in neural network design and the availability of large-scale labeled datasets have driven major improvements in piano transcription. Existing approaches target either offline applications, with no restrictions on computational demands, or online transcription, with delays of 128-320 ms. However, most real-time musical applications require latencies below 30 ms. In this work, we investigate whether and how the current state-of-the-art online transcription model can be adapted for real-time piano transcription. Specifically, we eliminate all non-causal processing, and reduce computational load through shared computations across core model components and variations in model size. Additionally, we explore different pre- and postprocessing strategies, and related label encoding schemes, and discuss their suitability for real-time transcription. Evaluating the adaptions on the MAESTRO dataset, we find a drop in transcription accuracy due to strictly causal processing as well as a tradeoff between the preprocessing latency and prediction accuracy. We release our system as a baseline to support researchers in designing models towards minimum latency real-time transcription.", "date": "2025-09-09", "categories": []}
{"id": "2509.07635", "title": "Neural Proxies for Sound Synthesizers: Learning Perceptually Informed Preset Representations", "abstract": "Deep learning appears as an appealing solution for Automatic Synthesizer Programming (ASP), which aims to assist musicians and sound designers in programming sound synthesizers. However, integrating software synthesizers into training pipelines is challenging due to their potential non-differentiability. This work tackles this challenge by introducing a method to approximate arbitrary synthesizers. Specifically, we train a neural network to map synthesizer presets onto an audio embedding space derived from a pretrained model. This facilitates the definition of a neural proxy that produces compact yet effective representations, thereby enabling the integration of audio embedding loss into neural-based ASP systems for black-box synthesizers. We evaluate the representations derived by various pretrained audio models in the context of neural-based nASP and assess the effectiveness of several neural network architectures, including feedforward, recurrent, and transformer-based models, in defining neural proxies. We evaluate the proposed method using both synthetic and hand-crafted presets from three popular software synthesizers and assess its performance in a synthesizer sound matching downstream task. While the benefits of the learned representation are nuanced by resource requirements, encouraging results were obtained for all synthesizers, paving the way for future research into the application of synthesizer proxies for neural-based ASP systems.", "date": "2025-09-09", "categories": []}
{"id": "2503.00296", "title": "Synthetic data enables context-aware bioacoustic sound event detection", "abstract": "We propose a methodology for training foundation models that enhances their in-context learning capabilities within the domain of bioacoustic signal processing. We use synthetically generated training data, introducing a domain-randomization-based pipeline that constructs diverse acoustic scenes with temporally strong labels. We generate over 8.8 thousand hours of strongly-labeled audio and train a query-by-example, transformer-based model to perform few-shot bioacoustic sound event detection. Our second contribution is a public benchmark of 13 diverse few-shot bioacoustics tasks. Our model outperforms previously published methods, and improves relative to other training-free methods by $64\\%$. We demonstrate that this is due to increase in model size and data scale, as well as algorithmic improvements. We make our trained model available via an API, to provide ecologists and ethologists with a training-free tool for bioacoustic sound event detection.", "date": "2025-03-01", "categories": []}
{"id": "2503.02389", "title": "Robust detection of overlapping bioacoustic sound events", "abstract": "We propose a method for accurately detecting bioacoustic sound events that is robust to overlapping events, a common issue in domains such as ethology, ecology and conservation. While standard methods employ a frame-based, multi-label approach, we introduce an onset-based detection method which we name Voxaboxen. It takes inspiration from object detection methods in computer vision, but simultaneously takes advantage of recent advances in self-supervised audio encoders. For each time window, Voxaboxen predicts whether it contains the start of a vocalization and how long the vocalization is. It also does the same in reverse, predicting whether each window contains the end of a vocalization, and how long ago it started. The two resulting sets of bounding boxes are then fused using a graph-matching algorithm. We also release a new dataset designed to measure performance on detecting overlapping vocalizations. This consists of recordings of zebra finches annotated with temporally-strong labels and showing frequent overlaps. We test Voxaboxen on seven existing data sets and on our new data set. We compare Voxaboxen to natural baselines and existing sound event detection methods and demonstrate SotA results. Further experiments show that improvements are robust to frequent vocalization overlap.", "date": "2025-03-04", "categories": []}
{"id": "2509.06155", "title": "UniVerse-1: Unified Audio-Video Generation via Stitching of Experts", "abstract": "We introduce UniVerse-1, a unified, Veo-3-like model capable of simultaneously generating coordinated audio and video. To enhance training efficiency, we bypass training from scratch and instead employ a stitching of experts (SoE) technique. This approach deeply fuses the corresponding blocks of pre-trained video and music generation experts models, thereby fully leveraging their foundational capabilities. To ensure accurate annotations and temporal alignment for both ambient sounds and speech with video content, we developed an online annotation pipeline that processes the required training data and generates labels during training process. This strategy circumvents the performance degradation often caused by misalignment text-based annotations. Through the synergy of these techniques, our model, after being finetuned on approximately 7,600 hours of audio-video data, produces results with well-coordinated audio-visuals for ambient sounds generation and strong alignment for speech generation. To systematically evaluate our proposed method, we introduce Verse-Bench, a new benchmark dataset. In an effort to advance research in audio-video generation and to close the performance gap with state-of-the-art models such as Veo3, we make our model and code publicly available. We hope this contribution will benefit the broader research community. Project page: https://dorniwang.github.io/UniVerse-1/.", "date": "2025-09-07", "categories": []}
{"id": "2509.06389", "title": "MeanFlow-Accelerated Multimodal Video-to-Audio Synthesis via One-Step Generation", "abstract": "A key challenge in synthesizing audios from silent videos is the inherent trade-off between synthesis quality and inference efficiency in existing methods. For instance, flow matching based models rely on modeling instantaneous velocity, inherently require an iterative sampling process, leading to slow inference speeds. To address this efficiency bottleneck, we introduce a MeanFlow-accelerated model that characterizes flow fields using average velocity, enabling one-step generation and thereby significantly accelerating multimodal video-to-audio (VTA) synthesis while preserving audio quality, semantic alignment, and temporal synchronization. Furthermore, a scalar rescaling mechanism is employed to balance conditional and unconditional predictions when classifier-free guidance (CFG) is applied, effectively mitigating CFG-induced distortions in one step generation. Since the audio synthesis network is jointly trained with multimodal conditions, we further evaluate it on text-to-audio (TTA) synthesis task. Experimental results demonstrate that incorporating MeanFlow into the network significantly improves inference speed without compromising perceptual quality on both VTA and TTA synthesis tasks.", "date": "2025-09-08", "categories": []}
{"id": "2509.06598", "title": "Integrating Spatial and Semantic Embeddings for Stereo Sound Event Localization in Videos", "abstract": "In this study, we address the multimodal task of stereo sound event localization and detection with source distance estimation (3D SELD) in regular video content. 3D SELD is a complex task that combines temporal event classification with spatial localization, requiring reasoning across spatial, temporal, and semantic dimensions. The last is arguably the most challenging to model. Traditional SELD approaches typically rely on multichannel input, limiting their capacity to benefit from large-scale pre-training due to data constraints. To overcome this, we enhance a standard SELD architecture with semantic information by integrating pre-trained, contrastive language-aligned models: CLAP for audio and OWL-ViT for visual inputs. These embeddings are incorporated into a modified Conformer module tailored for multimodal fusion, which we refer to as the Cross-Modal Conformer. We perform an ablation study on the development set of the DCASE2025 Task3 Stereo SELD Dataset to assess the individual contributions of the language-aligned models and benchmark against the DCASE Task 3 baseline systems. Additionally, we detail the curation process of large synthetic audio and audio-visual datasets used for model pre-training. These datasets were further expanded through left-right channel swapping augmentation. Our approach, combining extensive pre-training, model ensembling, and visual post-processing, achieved second rank in the DCASE 2025 Challenge Task 3 (Track B), underscoring the effectiveness of our method. Future work will explore the modality-specific contributions and architectural refinements.", "date": "2025-09-08", "categories": []}
{"id": "2509.06635", "title": "The First Voice Timbre Attribute Detection Challenge", "abstract": "The first voice timbre attribute detection challenge is featured in a special session at NCMMSC 2025. It focuses on the explainability of voice timbre and compares the intensity of two speech utterances in a specified timbre descriptor dimension. The evaluation was conducted on the VCTK-RVA dataset. Participants developed their systems and submitted their outputs to the organizer, who evaluated the performance and sent feedback to them. Six teams submitted their outputs, with five providing descriptions of their methodologies.", "date": "2025-09-08", "categories": []}
{"id": "2509.05256", "title": "Recomposer: Event-roll-guided generative audio editing", "abstract": "Editing complex real-world sound scenes is difficult because individual sound sources overlap in time. Generative models can fill-in missing or corrupted details based on their strong prior understanding of the data domain. We present a system for editing individual sound events within complex scenes able to delete, insert, and enhance individual sound events based on textual edit descriptions (e.g., ``enhance Door'') and a graphical representation of the event timing derived from an ``event roll'' transcription. We present an encoder-decoder transformer working on SoundStream representations, trained on synthetic (input, desired output) audio example pairs formed by adding isolated sound events to dense, real-world backgrounds. Evaluation reveals the importance of each part of the edit descriptions -- action, class, timing. Our work demonstrates ``recomposition'' is an important and practical application.", "date": "2025-09-05", "categories": []}
{"id": "2509.01153", "title": "EZhouNet:A framework based on graph neural network and anchor interval for the respiratory sound event detection", "abstract": "Auscultation is a key method for early diagnosis of respiratory and pulmonary diseases, relying on skilled healthcare professionals. However, the process is often subjective, with variability between experts. As a result, numerous deep learning-based automatic classification methods have emerged, most of which focus on respiratory sound classification. In contrast, research on respiratory sound event detection remains limited. Existing sound event detection methods typically rely on frame-level predictions followed by post-processing to generate event-level outputs, making interval boundaries challenging to learn directly. Furthermore, many approaches can only handle fixed-length audio, limiting their applicability to variable-length respiratory sounds. Additionally, the impact of respiratory sound location information on detection performance has not been extensively explored. To address these issues, we propose a graph neural network-based framework with anchor intervals, capable of handling variable-length audio and providing more precise temporal localization for abnormal respiratory sound events. Our method improves both the flexibility and applicability of respiratory sound detection. Experiments on the SPRSound 2024 and HF Lung V1 datasets demonstrate the effectiveness of the proposed approach, and incorporating respiratory position information enhances the discrimination between abnormal sounds. The reference implementation is available at https://github.com/chumingqian/EzhouNet.", "date": "2025-09-01", "categories": []}
{"id": "2509.02349", "title": "AudioCodecBench: A Comprehensive Benchmark for Audio Codec Evaluation", "abstract": "Multimodal Large Language Models (MLLMs) have been widely applied in speech and music. This tendency has led to a focus on audio tokenization for Large Models (LMs). Unlike semantic-only text tokens, audio tokens must both capture global semantic content and preserve fine-grained acoustic details. Moreover, they provide a discrete method for speech and music that can be effectively integrated into MLLMs. However, existing research is unsuitable in the definitions of semantic tokens and acoustic tokens. In addition, the evaluation of different codecs typically concentrates on specific domains or tasks, such as reconstruction or Automatic Speech Recognition (ASR) task, which prevents fair and comprehensive comparisons. To address these problems, this paper provides suitable definitions for semantic and acoustic tokens and introduces a systematic evaluation framework. This framework allows for a comprehensive assessment of codecs' capabilities which evaluate across four dimensions: audio reconstruction metric, codebook index (ID) stability, decoder-only transformer perplexity, and performance on downstream probe tasks. Our results show the correctness of the provided suitable definitions and the correlation among reconstruction metrics, codebook ID stability, downstream probe tasks and perplexity.", "date": "2025-09-02", "categories": []}
{"id": "2509.03829", "title": "NE-PADD: Leveraging Named Entity Knowledge for Robust Partial Audio Deepfake Detection via Attention Aggregation", "abstract": "Different from traditional sentence-level audio deepfake detection (ADD), partial audio deepfake detection (PADD) requires frame-level positioning of the location of fake speech. While some progress has been made in this area, leveraging semantic information from audio, especially named entities, remains an underexplored aspect. To this end, we propose NE-PADD, a novel method for Partial Audio Deepfake Detection (PADD) that leverages named entity knowledge through two parallel branches: Speech Name Entity Recognition (SpeechNER) and PADD. The approach incorporates two attention aggregation mechanisms: Attention Fusion (AF) for combining attention weights and Attention Transfer (AT) for guiding PADD with named entity semantics using an auxiliary loss. Built on the PartialSpoof-NER dataset, experiments show our method outperforms existing baselines, proving the effectiveness of integrating named entity knowledge in PADD. The code is available at https://github.com/AI-S2-Lab/NE-PADD.", "date": "2025-09-04", "categories": []}
{"id": "2509.02859", "title": "Speech DF Arena: A Leaderboard for Speech DeepFake Detection Models", "abstract": "Parallel to the development of advanced deepfake audio generation, audio deepfake detection has also seen significant progress. However, a standardized and comprehensive benchmark is still missing. To address this, we introduce Speech DeepFake (DF) Arena, the first comprehensive benchmark for audio deepfake detection. Speech DF Arena provides a toolkit to uniformly evaluate detection systems, currently across 14 diverse datasets and attack scenarios, standardized evaluation metrics and protocols for reproducibility and transparency. It also includes a leaderboard to compare and rank the systems to help researchers and developers enhance their reliability and robustness. We include 14 evaluation sets, 12 state-of-the-art open-source and 3 proprietary detection systems. Our study presents many systems exhibiting high EER in out-of-domain scenarios, highlighting the need for extensive cross-domain evaluation. The leaderboard is hosted on Huggingface1 and a toolkit for reproducing results across the listed datasets is available on GitHub.", "date": "2025-09-02", "categories": []}
{"id": "2509.03256", "title": "Comparison of End-to-end Speech Assessment Models for the NOCASA 2025 Challenge", "abstract": "This paper presents an analysis of three end-to-end models developed for the NOCASA 2025 Challenge, aimed at automatic word-level pronunciation assessment for children learning Norwegian as a second language. Our models include an encoder-decoder Siamese architecture (E2E-R), a prefix-tuned direct classification model leveraging pretrained wav2vec2.0 representations, and a novel model integrating alignment-free goodness-of-pronunciation (GOP) features computed via CTC. We introduce a weighted ordinal cross-entropy loss tailored for optimizing metrics such as unweighted average recall and mean absolute error. Among the explored methods, our GOP-CTC-based model achieved the highest performance, substantially surpassing challenge baselines and attaining top leaderboard scores.", "date": "2025-09-03", "categories": []}
{"id": "2509.03292", "title": "Improving Perceptual Audio Aesthetic Assessment via Triplet Loss and Self-Supervised Embeddings", "abstract": "We present a system for automatic multi-axis perceptual quality prediction of generative audio, developed for Track 2 of the AudioMOS Challenge 2025. The task is to predict four Audio Aesthetic Scores--Production Quality, Production Complexity, Content Enjoyment, and Content Usefulness--for audio generated by text-to-speech (TTS), text-to-audio (TTA), and text-to-music (TTM) systems. A main challenge is the domain shift between natural training data and synthetic evaluation data. To address this, we combine BEATs, a pretrained transformer-based audio representation model, with a multi-branch long short-term memory (LSTM) predictor and use a triplet loss with buffer-based sampling to structure the embedding space by perceptual similarity. Our results show that this improves embedding discriminability and generalization, enabling domain-robust audio quality assessment without synthetic training data.", "date": "2025-09-03", "categories": []}
{"id": "2502.10822", "title": "NeuroAMP: A Novel End-to-end General Purpose Deep Neural Amplifier for Personalized Hearing Aids", "abstract": "The prevalence of hearing aids is increasing. However, optimizing the amplification processes of hearing aids remains challenging due to the complexity of integrating multiple modular components in traditional methods. To address this challenge, we present NeuroAMP, a novel deep neural network designed for end-to-end, personalized amplification in hearing aids. NeuroAMP leverages both spectral features and the listener's audiogram as inputs, and we investigate four architectures: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Convolutional Recurrent Neural Network (CRNN), and Transformer. We also introduce Denoising NeuroAMP, an extension that integrates noise reduction along with amplification capabilities for improved performance in real-world scenarios. To enhance generalization, a comprehensive data augmentation strategy was employed during training on diverse speech (TIMIT and TMHINT) and music (Cadenza Challenge MUSIC) datasets. Evaluation using the Hearing Aid Speech Perception Index (HASPI), Hearing Aid Speech Quality Index (HASQI), and Hearing Aid Audio Quality Index (HAAQI) demonstrates that the Transformer architecture within NeuroAMP achieves the best performance, with SRCC scores of 0.9927 (HASQI) and 0.9905 (HASPI) on TIMIT, and 0.9738 (HAAQI) on the Cadenza Challenge MUSIC dataset. Notably, our data augmentation strategy maintains high performance on unseen datasets (e.g., VCTK, MUSDB18-HQ). Furthermore, Denoising NeuroAMP outperforms both the conventional NAL-R+WDRC approach and a two-stage baseline on the VoiceBank+DEMAND dataset, achieving a 10% improvement in both HASPI (0.90) and HASQI (0.59) scores. These results highlight the potential of NeuroAMP and Denoising NeuroAMP to deliver notable improvements in personalized hearing aid amplification.", "date": "2025-02-15", "categories": []}
{"id": "2507.14534", "title": "Conan: A Chunkwise Online Network for Zero-Shot Adaptive Voice Conversion", "abstract": "Zero-shot online voice conversion (VC) holds significant promise for real-time communications and entertainment. However, current VC models struggle to preserve semantic fidelity under real-time constraints, deliver natural-sounding conversions, and adapt effectively to unseen speaker characteristics. To address these challenges, we introduce Conan, a chunkwise online zero-shot voice conversion model that preserves the content of the source while matching the voice timbre and styles of reference speech. Conan comprises three core components: 1) a Stream Content Extractor that leverages Emformer for low-latency streaming content encoding; 2) an Adaptive Style Encoder that extracts fine-grained stylistic features from reference speech for enhanced style adaptation; 3) a Causal Shuffle Vocoder that implements a fully causal HiFiGAN using a pixel-shuffle mechanism. Experimental evaluations demonstrate that Conan outperforms baseline models in subjective and objective metrics. Audio samples can be found at https://aaronz345.github.io/ConanDemo.", "date": "2025-07-19", "categories": []}
{"id": "2509.00186", "title": "Generalizable Audio Spoofing Detection using Non-Semantic Representations", "abstract": "Rapid advancements in generative modeling have made synthetic audio generation easy, making speech-based services vulnerable to spoofing attacks. Consequently, there is a dire need for robust countermeasures more than ever. Existing solutions for deepfake detection are often criticized for lacking generalizability and fail drastically when applied to real-world data. This study proposes a novel method for generalizable spoofing detection leveraging non-semantic universal audio representations. Extensive experiments have been performed to find suitable non-semantic features using TRILL and TRILLsson models. The results indicate that the proposed method achieves comparable performance on the in-domain test set while significantly outperforming state-of-the-art approaches on out-of-domain test sets. Notably, it demonstrates superior generalization on public-domain data, surpassing methods based on hand-crafted features, semantic embeddings, and end-to-end architectures.", "date": "2025-08-29", "categories": []}
{"id": "2509.00839", "title": "Adaptive Vehicle Speed Classification via BMCNN with Reinforcement Learning-Enhanced Acoustic Processing", "abstract": "Traffic congestion remains a pressing urban challenge, requiring intelligent transportation systems for real-time management. We present a hybrid framework that combines deep learning and reinforcement learning for acoustic vehicle speed classification. A dual-branch BMCNN processes MFCC and wavelet features to capture complementary frequency patterns. An attention-enhanced DQN adaptively selects the minimal number of audio frames and triggers early decisions once confidence thresholds are reached. Evaluations on IDMT-Traffic and our SZUR-Acoustic (Suzhou) datasets show 95.99% and 92.3% accuracy, with up to 1.63x faster average processing via early termination. Compared with A3C, DDDQN, SA2C, PPO, and TD3, the method provides a superior accuracy-efficiency trade-off and is suitable for real-time ITS deployment in heterogeneous urban environments.", "date": "2025-08-31", "categories": []}
{"id": "2509.00862", "title": "Speech Command Recognition Using LogNNet Reservoir Computing for Embedded Systems", "abstract": "This paper presents a low-resource speech-command recognizer combining energy-based voice activity detection (VAD), an optimized Mel-Frequency Cepstral Coefficients (MFCC) pipeline, and the LogNNet reservoir-computing classifier. Using four commands from the Speech Commands da-taset downsampled to 8 kHz, we evaluate four MFCC aggregation schemes and find that adaptive binning (64-dimensional feature vector) offers the best accuracy-to-compactness trade-off. The LogNNet classifier with architecture 64:33:9:4 reaches 92.04% accuracy under speaker-independent evaluation, while requiring significantly fewer parameters than conventional deep learn-ing models. Hardware implementation on Arduino Nano 33 IoT (ARM Cor-tex-M0+, 48 MHz, 32 KB RAM) validates the practical feasibility, achieving ~90% real-time recognition accuracy while consuming only 18 KB RAM (55% utilization). The complete pipeline (VAD -> MFCC -> LogNNet) thus enables reliable on-device speech-command recognition under strict memory and compute limits, making it suitable for battery-powered IoT nodes, wire-less sensor networks, and hands-free control interfaces.", "date": "2025-08-31", "categories": []}
{"id": "2509.01401", "title": "ArabEmoNet: A Lightweight Hybrid 2D CNN-BiLSTM Model with Attention for Robust Arabic Speech Emotion Recognition", "abstract": "Speech emotion recognition is vital for human-computer interaction, particularly for low-resource languages like Arabic, which face challenges due to limited data and research. We introduce ArabEmoNet, a lightweight architecture designed to overcome these limitations and deliver state-of-the-art performance. Unlike previous systems relying on discrete MFCC features and 1D convolutions, which miss nuanced spectro-temporal patterns, ArabEmoNet uses Mel spectrograms processed through 2D convolutions, preserving critical emotional cues often lost in traditional methods.\n  While recent models favor large-scale architectures with millions of parameters, ArabEmoNet achieves superior results with just 1 million parameters, 90 times smaller than HuBERT base and 74 times smaller than Whisper. This efficiency makes it ideal for resource-constrained environments. ArabEmoNet advances Arabic speech emotion recognition, offering exceptional performance and accessibility for real-world applications.", "date": "2025-09-01", "categories": []}
{"id": "2509.01588", "title": "From Discord to Harmony: Decomposed Consonance-based Training for Improved Audio Chord Estimation", "abstract": "Audio Chord Estimation (ACE) holds a pivotal role in music information research, having garnered attention for over two decades due to its relevance for music transcription and analysis. Despite notable advancements, challenges persist in the task, particularly concerning unique characteristics of harmonic content, which have resulted in existing systems' performances reaching a glass ceiling. These challenges include annotator subjectivity, where varying interpretations among annotators lead to inconsistencies, and class imbalance within chord datasets, where certain chord classes are over-represented compared to others, posing difficulties in model training and evaluation. As a first contribution, this paper presents an evaluation of inter-annotator agreement in chord annotations, using metrics that extend beyond traditional binary measures. In addition, we propose a consonance-informed distance metric that reflects the perceptual similarity between harmonic annotations. Our analysis suggests that consonance-based distance metrics more effectively capture musically meaningful agreement between annotations. Expanding on these findings, we introduce a novel ACE conformer-based model that integrates consonance concepts into the model through consonance-based label smoothing. The proposed model also addresses class imbalance by separately estimating root, bass, and all note activations, enabling the reconstruction of chord labels from decomposed outputs.", "date": "2025-09-01", "categories": []}
{"id": "2509.01762", "title": "Music Genre Classification Using Machine Learning Techniques", "abstract": "This paper presents a comparative analysis of machine learning methodologies for automatic music genre classification. We evaluate the performance of classical classifiers, including Support Vector Machines (SVM) and ensemble methods, trained on a comprehensive set of hand-crafted audio features, against a Convolutional Neural Network (CNN) operating on Mel spectrograms. The study is conducted on the widely-used GTZAN dataset. Our findings demonstrate a noteworthy result: the SVM, leveraging domain-specific feature engineering, achieves superior classification accuracy compared to the end-to-end CNN model. We attribute this outcome to the data-constrained nature of the benchmark dataset, where the strong inductive bias of engineered features provides a regularization effect that mitigates the risk of overfitting inherent in high-capacity deep learning models. This work underscores the enduring relevance of traditional feature extraction in practical audio processing tasks and provides a critical perspective on the universal applicability of deep learning, especially for moderately sized datasets.", "date": "2025-09-01", "categories": []}
{"id": "2509.02259", "title": "Speech transformer models for extracting information from baby cries", "abstract": "Transfer learning using latent representations from pre-trained speech models achieves outstanding performance in tasks where labeled data is scarce. However, their applicability to non-speech data and the specific acoustic properties encoded in these representations remain largely unexplored. In this study, we investigate both aspects. We evaluate five pre-trained speech models on eight baby cries datasets, encompassing 115 hours of audio from 960 babies. For each dataset, we assess the latent representations of each model across all available classification tasks. Our results demonstrate that the latent representations of these models can effectively classify human baby cries and encode key information related to vocal source instability and identity of the crying baby. In addition, a comparison of the architectures and training strategies of these models offers valuable insights for the design of future models tailored to similar tasks, such as emotion detection.", "date": "2025-09-02", "categories": []}
{"id": "2509.02471", "title": "ESTM: An Enhanced Dual-Branch Spectral-Temporal Mamba for Anomalous Sound Detection", "abstract": "The core challenge in industrial equipment anoma lous sound detection (ASD) lies in modeling the time-frequency coupling characteristics of acoustic features. Existing modeling methods are limited by local receptive fields, making it difficult to capture long-range temporal patterns and cross-band dynamic coupling effects in machine acoustic features. In this paper, we propose a novel framework, ESTM, which is based on a dual-path Mamba architecture with time-frequency decoupled modeling and utilizes Selective State-Space Models (SSM) for long-range sequence modeling. ESTM extracts rich feature representations from different time segments and frequency bands by fusing enhanced Mel spectrograms and raw audio features, while further improving sensitivity to anomalous patterns through the TriStat-Gating (TSG) module. Our experiments demonstrate that ESTM improves anomalous detection performance on the DCASE 2020 Task 2 dataset, further validating the effectiveness of the proposed method.", "date": "2025-09-02", "categories": []}
{"id": "2506.02078", "title": "Evaluating the Effectiveness of Pre-Trained Audio Embeddings for Classification of Parkinson's Disease Speech Data", "abstract": "Speech impairments are prevalent biomarkers for Parkinson's Disease (PD), motivating the development of diagnostic techniques using speech data for clinical applications. Although deep acoustic features have shown promise for PD classification, their effectiveness often varies due to individual speaker differences, a factor that has not been thoroughly explored in the existing literature. This study investigates the effectiveness of three pre-trained audio embeddings (OpenL3, VGGish and Wav2Vec2.0 models) for PD classification. Using the NeuroVoz dataset, OpenL3 outperforms others in diadochokinesis (DDK) and listen and repeat (LR) tasks, capturing critical acoustic features for PD detection. Only Wav2Vec2.0 shows significant gender bias, achieving more favorable results for male speakers, in DDK tasks. The misclassified cases reveal challenges with atypical speech patterns, highlighting the need for improved feature extraction and model robustness in PD detection.", "date": "2025-06-02", "categories": []}
{"id": "2508.21243", "title": "Full-Frequency Temporal Patching and Structured Masking for Enhanced Audio Classification", "abstract": "Transformers and State-Space Models (SSMs) have advanced audio classification by modeling spectrograms as sequences of patches. However, existing models such as the Audio Spectrogram Transformer (AST) and Audio Mamba (AuM) adopt square patching from computer vision, which disrupts continuous frequency patterns and produces an excessive number of patches, slowing training, and increasing computation. We propose Full-Frequency Temporal Patching (FFTP), a patching strategy that better matches the time-frequency asymmetry of spectrograms by spanning full frequency bands with localized temporal context, preserving harmonic structure, and significantly reducing patch count and computation. We also introduce SpecMask, a patch-aligned spectrogram augmentation that combines full-frequency and localized time-frequency masks under a fixed masking budget, enhancing temporal robustness while preserving spectral continuity. When applied on both AST and AuM, our patching method with SpecMask improves mAP by up to +6.76 on AudioSet-18k and accuracy by up to +8.46 on SpeechCommandsV2, while reducing computation by up to 83.26%, demonstrating both performance and efficiency gains.", "date": "2025-08-28", "categories": []}
{"id": "2508.21407", "title": "DRASP: A Dual-Resolution Attentive Statistics Pooling Framework for Automatic MOS Prediction", "abstract": "A pooling mechanism is essential for mean opinion score (MOS) prediction, facilitating the transformation of variable-length audio features into a concise fixed-size representation that effectively encodes speech quality. Existing pooling methods typically operate at a singular granularity, concentrating either on a comprehensive global perspective or a detailed frame-level analysis, which may overlook complementary perceptual insights. To address this limitation, we introduce the Dual-Resolution Attentive Statistics Pooling (DRASP) framework. DRASP integrates both coarse-grained, global statistical summaries and fine-grained, attentive analyses of perceptually significant segments. This dual-view architecture empowers our model to formulate a more thorough and robust representation, capturing both the overarching structural context and salient local details concurrently. Extensive experiments validate the effectiveness and strong generalization ability of the proposed framework. It consistently outperforms various baseline methods across diverse datasets (MusicEval and AES-Natural), MOS prediction backbones (including a CLAP-based model and AudioBox-Aesthetics), and different audio generation systems, achieving a relative improvement of 10.39% in system-level Spearman's rank correlation coefficient (SRCC) over the widely-used average pooling approach.", "date": "2025-08-29", "categories": []}
{"id": "2410.01108", "title": "Augmentation through Laundering Attacks for Audio Spoof Detection", "abstract": "Recent text-to-speech (TTS) developments have made voice cloning (VC) more realistic, affordable, and easily accessible. This has given rise to many potential abuses of this technology, including Joe Biden's New Hampshire deepfake robocall. Several methodologies have been proposed to detect such clones. However, these methodologies have been trained and evaluated on relatively clean databases. Recently, ASVspoof 5 Challenge introduced a new crowd-sourced database of diverse acoustic conditions including various spoofing attacks and codec conditions. This paper is our submission to the ASVspoof 5 Challenge and aims to investigate the performance of Audio Spoof Detection, trained using data augmentation through laundering attacks, on the ASVSpoof 5 database. The results demonstrate that our system performs worst on A18, A19, A20, A26, and A30 spoofing attacks and in the codec and compression conditions of C08, C09, and C10.", "date": "2024-10-01", "categories": []}
{"id": "2411.19770", "title": "Noro: Noise-Robust One-shot Voice Conversion with Hidden Speaker Representation Learning", "abstract": "The effectiveness of one-shot voice conversion (VC) decreases in real-world scenarios where reference speeches, which are often sourced from the internet, contain various disturbances like background noise. To address this issue, we introduce Noro, a noise-robust one-shot VC system. Noro features innovative components tailored for VC using noisy reference speeches, including a dual-branch reference encoding module and a noise-agnostic contrastive speaker loss. Experimental results demonstrate that Noro outperforms our baseline system in both clean and noisy scenarios, highlighting its efficacy for real-world applications. Additionally, we investigate the hidden speaker representation capabilities of our baseline system by repurposing its reference encoder as a speaker encoder. The results show that it is competitive with several advanced self-supervised learning models for speaker representation under the SUPERB settings, highlighting the potential for advancing speaker representation learning through one-shot VC tasks.", "date": "2024-11-29", "categories": []}
{"id": "2505.18722", "title": "Evaluating the Usefulness of Non-Diagnostic Speech Data for Developing Parkinson's Disease Classifiers", "abstract": "Speech-based Parkinson's disease (PD) detection has gained attention for its automated, cost-effective, and non-intrusive nature. As research studies usually rely on data from diagnostic-oriented speech tasks, this work explores the feasibility of diagnosing PD on the basis of speech data not originally intended for diagnostic purposes, using the Turn-Taking (TT) dataset. Our findings indicate that TT can be as useful as diagnostic-oriented PD datasets like PC-GITA. We also investigate which specific dataset characteristics impact PD classification performance. The results show that concatenating audio recordings and balancing participants' gender and status distributions can be beneficial. Cross-dataset evaluation reveals that models trained on PC-GITA generalize poorly to TT, whereas models trained on TT perform better on PC-GITA. Furthermore, we provide insights into the high variability across folds, which is mainly due to large differences in individual speaker performance.", "date": "2025-05-24", "categories": []}
{"id": "2508.20914", "title": "Learning Robust Spatial Representations from Binaural Audio through Feature Distillation", "abstract": "Recently, deep representation learning has shown strong performance in multiple audio tasks. However, its use for learning spatial representations from multichannel audio is underexplored. We investigate the use of a pretraining stage based on feature distillation to learn a robust spatial representation of binaural speech without the need for data labels. In this framework, spatial features are computed from clean binaural speech samples to form prediction labels. These clean features are then predicted from corresponding augmented speech using a neural network. After pretraining, we throw away the spatial feature predictor and use the learned encoder weights to initialize a DoA estimation model which we fine-tune for DoA estimation. Our experiments demonstrate that the pretrained models show improved performance in noisy and reverberant environments after fine-tuning for direction-of-arrival estimation, when compared to fully supervised models and classic signal processing methods.", "date": "2025-08-28", "categories": []}
{"id": "2508.20976", "title": "WoW-Bench: Evaluating Fine-Grained Acoustic Perception in Audio-Language Models via Marine Mammal Vocalizations", "abstract": "Large audio language models (LALMs) extend language understanding into the auditory domain, yet their ability to perform low-level listening, such as pitch and duration detection, remains underexplored. However, low-level listening is critical for real-world, out-of-distribution tasks where models must reason about unfamiliar sounds based on fine-grained acoustic cues. To address this gap, we introduce the World-of-Whale benchmark (WoW-Bench) to evaluate low-level auditory perception and cognition using marine mammal vocalizations. WoW-bench is composed of a Perception benchmark for categorizing novel sounds and a Cognition benchmark, inspired by Bloom's taxonomy, to assess the abilities to remember, understand, apply, and analyze sound events. For the Cognition benchmark, we additionally introduce distractor questions to evaluate whether models are truly solving problems through listening rather than relying on other heuristics. Experiments with state-of-the-art LALMs show performance far below human levels, indicating a need for stronger auditory grounding in LALMs.", "date": "2025-08-28", "categories": []}
{"id": "2505.11889", "title": "Revisiting SSL for sound event detection: complementary fusion and adaptive post-processing", "abstract": "Self-supervised learning (SSL) models offer powerful representations for sound event detection (SED), yet their synergistic potential remains underexplored. This study systematically evaluates state-of-the-art SSL models to guide optimal model selection and integration for SED. We propose a framework that combines heterogeneous SSL representations (e.g., BEATs, HuBERT, WavLM) through three fusion strategies: individual SSL embedding integration, dual-modal fusion, and full aggregation. Experiments on the DCASE 2023 Task 4 Challenge reveal that dual-modal fusion (e.g., CRNN+BEATs+WavLM) achieves complementary performance gains, while CRNN+BEATs alone delivers the best results among individual SSL models. We further introduce normalized sound event bounding boxes (nSEBBs), an adaptive post-processing method that dynamically adjusts event boundary predictions, improving PSDS1 by up to 4% for standalone SSL models. These findings highlight the compatibility and complementarity of SSL architectures, providing guidance for task-specific fusion and robust SED system design.", "date": "2025-05-17", "categories": []}
{"id": "2508.18440", "title": "SwiftF0: Fast and Accurate Monophonic Pitch Detection", "abstract": "Accurate and real-time monophonic pitch estimation in noisy conditions, particularly on resource-constrained devices, remains an open challenge in audio processing. We present \\emph{SwiftF0}, a novel, lightweight neural model that sets a new state-of-the-art for monophonic pitch estimation. Through training on diverse speech, music, and synthetic datasets with extensive data augmentation, SwiftF0 achieves robust generalization across acoustic domains while maintaining computational efficiency. SwiftF0 achieves a 91.80\\% harmonic mean (HM) at 10 dB SNR, outperforming baselines like CREPE by over 12 percentage points and degrading by only 2.3 points from clean audio. SwiftF0 requires only 95,842 parameters and runs approximately 42x faster than CREPE on CPU, making it ideal for efficient, real-time deployment. To address the critical lack of perfectly accurate ground truth pitch in speech corpora (which typically rely on algorithmic estimators or laryngograph signals), we introduce \\emph{SpeechSynth}. This synthetic speech dataset, generated by a phoneme-level TTS model, provides exact, on-demand ground-truth pitch curves, enabling more robust model training and evaluation. Furthermore, we propose a unified metric, combining six complementary performance measures for comprehensive and reliable pitch evaluation, and release an open-source pitch benchmark suite. A live demo of SwiftF0 is available at https://swift-f0.github.io/, the source code at https://github.com/lars76/swift-f0, and the benchmark framework at https://github.com/lars76/pitch-benchmark.", "date": "2025-08-25", "categories": []}
{"id": "2502.12714", "title": "Playing with Voices: Tabletop Role-Playing Game Recordings as a Diarization Challenge", "abstract": "This paper provides a proof of concept that audio of tabletop role-playing games (TTRPG) could serve as a challenge for diarization systems. TTRPGs are carried out mostly by conversation. Participants often alter their voices to indicate that they are talking as a fictional character. Audio processing systems are susceptible to voice conversion with or without technological assistance. TTRPG present a conversational phenomenon in which voice conversion is an inherent characteristic for an immersive gaming experience. This could make it more challenging for diarizers to pick the real speaker and determine that impersonating is just that. We present the creation of a small TTRPG audio dataset and compare it against the AMI and the ICSI corpus. The performance of two diarizers, pyannote.audio and wespeaker, were evaluated. We observed that TTRPGs' properties result in a higher confusion rate for both diarizers. Additionally, wespeaker strongly underestimates the number of speakers in the TTRPG audio files. We propose TTRPG audio as a promising challenge for diarization systems.", "date": "2025-02-18", "categories": []}
{"id": "2508.16790", "title": "TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling", "abstract": "Speech tokenizers serve as foundational components for speech language models, yet current designs exhibit several limitations, including: 1) dependence on multi-layer residual vector quantization structures or high frame rates, 2) reliance on auxiliary pre-trained models for semantic distillation, and 3) requirements for complex two-stage training processes. In this work, we introduce the Text-aware Diffusion Transformer Speech Codec (TaDiCodec), a novel approach designed to overcome these challenges. TaDiCodec employs end-to-end optimization for quantization and reconstruction through a diffusion autoencoder, while integrating text guidance into the diffusion decoder to enhance reconstruction quality and achieve optimal compression. TaDiCodec achieves an extremely low frame rate of 6.25 Hz and a corresponding bitrate of 0.0875 kbps with a single-layer codebook for 24 kHz speech, while maintaining superior performance on critical speech generation evaluation metrics such as Word Error Rate (WER), speaker similarity (SIM), and speech quality (UTMOS). Notably, TaDiCodec employs a single-stage, end-to-end training paradigm, and obviating the need for auxiliary pre-trained models. We also validate the compatibility of TaDiCodec in language model based zero-shot text-to-speech with both autoregressive modeling and masked generative modeling, demonstrating its effectiveness and efficiency for speech language modeling, as well as a significantly small reconstruction-generation gap. We will open source our code and model checkpoints. Audio samples are are available at https:/tadicodec.github.io/. We release code and model checkpoints at https:/github.com/HeCheng0625/Diffusion-Speech-Tokenizer.", "date": "2025-08-22", "categories": []}
{"id": "2508.16930", "title": "HunyuanVideo-Foley: Multimodal Diffusion with Representation Alignment for High-Fidelity Foley Audio Generation", "abstract": "Recent advances in video generation produce visually realistic content, yet the absence of synchronized audio severely compromises immersion. To address key challenges in video-to-audio generation, including multimodal data scarcity, modality imbalance and limited audio quality in existing methods, we propose HunyuanVideo-Foley, an end-to-end text-video-to-audio framework that synthesizes high-fidelity audio precisely aligned with visual dynamics and semantic context. Our approach incorporates three core innovations: (1) a scalable data pipeline curating 100k-hour multimodal datasets through automated annotation; (2) a representation alignment strategy using self-supervised audio features to guide latent diffusion training, efficiently improving audio quality and generation stability; (3) a novel multimodal diffusion transformer resolving modal competition, containing dual-stream audio-video fusion through joint attention, and textual semantic injection via cross-attention. Comprehensive evaluations demonstrate that HunyuanVideo-Foley achieves new state-of-the-art performance across audio fidelity, visual-semantic alignment, temporal alignment and distribution matching. The demo page is available at: https://szczesnys.github.io/hunyuanvideo-foley/.", "date": "2025-08-23", "categories": []}
{"id": "2508.16576", "title": "Benchmarking Training Paradigms, Dataset Composition, and Model Scaling for Child ASR in ESPnet", "abstract": "Despite advancements in ASR, child speech recognition remains challenging due to acoustic variability and limited annotated data. While fine-tuning adult ASR models on child speech is common, comparisons with flat-start training remain underexplored. We compare flat-start training across multiple datasets, SSL representations (WavLM, XEUS), and decoder architectures. Our results show that SSL representations are biased toward adult speech, with flat-start training on child speech mitigating these biases. We also analyze model scaling, finding consistent improvements up to 1B parameters, beyond which performance plateaus. Additionally, age-related ASR and speaker verification analysis highlights the limitations of proprietary models like Whisper, emphasizing the need for open-data models for reliable child speech research. All investigations are conducted using ESPnet, and our publicly available benchmark provides insights into training strategies for robust child speech processing.", "date": "2025-08-22", "categories": []}
{"id": "2505.11378", "title": "Machine Learning Approaches to Vocal Register Classification in Contemporary Male Pop Music", "abstract": "For singers of all experience levels, one of the most daunting challenges in learning technical repertoire is navigating placement and vocal register in and around the passagio (passage between chest voice and head voice registers). Particularly in pop music, where a single artist may use a variety of timbre's and textures to achieve a desired quality, it can be difficult to identify what vocal register within the vocal range a singer is using. This paper presents two methods for classifying vocal registers in an audio signal of male pop music through the analysis of textural features of mel-spectrogram images. Additionally, we will discuss the practical integration of these models for vocal analysis tools, and introduce a concurrently developed software called AVRA which stands for Automatic Vocal Register Analysis. Our proposed methods achieved consistent classification of vocal register through both Support Vector Machine (SVM) and Convolutional Neural Network (CNN) models, which supports the promise of more robust classification possibilities across more voice types and genres of singing.", "date": "2025-05-16", "categories": []}
{"id": "2508.15316", "title": "CUPE: Contextless Universal Phoneme Encoder for Language-Agnostic Speech Processing", "abstract": "Universal phoneme recognition typically requires analyzing long speech segments and language-specific patterns. Many speech processing tasks require pure phoneme representations free from contextual influence, which motivated our development of CUPE - a lightweight model that captures key phoneme features in just 120 milliseconds, about one phoneme's length. CUPE processes short, fixed-width windows independently and, despite fewer parameters than current approaches, achieves competitive cross-lingual performance by learning fundamental acoustic patterns common to all languages. Our extensive evaluation through supervised and self-supervised training on diverse languages, including zero-shot tests on the UCLA Phonetic Corpus, demonstrates strong cross-lingual generalization and reveals that effective universal speech processing is possible through modeling basic acoustic patterns within phoneme-length windows.", "date": "2025-08-21", "categories": []}
{"id": "2508.15334", "title": "An Enhanced Audio Feature Tailored for Anomalous Sound Detection Based on Pre-trained Models", "abstract": "Anomalous Sound Detection (ASD) aims at identifying anomalous sounds from machines and has gained extensive research interests from both academia and industry. However, the uncertainty of anomaly location and much redundant information such as noise in machine sounds hinder the improvement of ASD system performance. This paper proposes a novel audio feature of filter banks with evenly distributed intervals, ensuring equal attention to all frequency ranges in the audio, which enhances the detection of anomalies in machine sounds. Moreover, based on pre-trained models, this paper presents a parameter-free feature enhancement approach to remove redundant information in machine audio. It is believed that this parameter-free strategy facilitates the effective transfer of universal knowledge from pre-trained tasks to the ASD task during model fine-tuning. Evaluation results on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2024 Challenge dataset demonstrate significant improvements in ASD performance with our proposed methods.", "date": "2025-08-21", "categories": []}
{"id": "2502.04519", "title": "GenVC: Self-Supervised Zero-Shot Voice Conversion", "abstract": "Most current zero-shot voice conversion methods rely on externally supervised components, particularly speaker encoders, for training. To explore alternatives that eliminate this dependency, this paper introduces GenVC, a novel framework that disentangles speaker identity and linguistic content from speech signals in a self-supervised manner. GenVC leverages speech tokenizers and an autoregressive, Transformer-based language model as its backbone for speech generation. This design supports large-scale training while enhancing both source speaker privacy protection and target speaker cloning fidelity. Experimental results demonstrate that GenVC achieves notably higher speaker similarity, with naturalness on par with leading zero-shot approaches. Moreover, due to its autoregressive formulation, GenVC introduces flexibility in temporal alignment, reducing the preservation of source prosody and speaker-specific traits, and making it highly effective for voice anonymization.", "date": "2025-02-06", "categories": []}
{"id": "2507.02755", "title": "Multi-agent Auditory Scene Analysis", "abstract": "Auditory scene analysis (ASA) aims to retrieve information from the acoustic environment, by carrying out three main tasks: sound source location, separation, and classification. These tasks are traditionally executed with a linear data flow, where the sound sources are first located; then, using their location, each source is separated into its own audio stream; from each of which, information is extracted that is relevant to the application scenario (audio event detection, speaker identification, emotion classification, etc.). However, running these tasks linearly increases the overall response time, while making the last tasks (separation and classification) highly sensitive to errors of the first task (location). A considerable amount of effort and computational complexity has been employed in the state-of-the-art to develop techniques that are the least error-prone possible. However, doing so gives rise to an ASA system that is non-viable in many applications that require a small computational footprint and a low response time, such as bioacoustics, hearing-aid design, search and rescue, human-robot interaction, etc. To this effect, in this work, a multi-agent approach is proposed to carry out ASA where the tasks are run in parallel, with feedback loops between them to compensate for local errors, such as: using the quality of the separation output to correct the location error; and using the classification result to reduce the localization's sensitivity towards interferences. The result is a multi-agent auditory scene analysis (MASA) system that is robust against local errors, without a considerable increase in complexity, and with a low response time. The complete proposed MASA system is provided as a publicly available framework that uses open-source tools for sound acquisition and reproduction (JACK) and inter-agent communication (ROS2), allowing users to add their own agents.", "date": "2025-07-03", "categories": []}
{"id": "2504.12880", "title": "Can Masked Autoencoders Also Listen to Birds?", "abstract": "Masked Autoencoders (MAEs) learn rich semantic representations in audio classification through an efficient self-supervised reconstruction task. However, general-purpose models fail to generalize well when applied directly to fine-grained audio domains. Specifically, bird-sound classification requires distinguishing subtle inter-species differences and managing high intra-species acoustic variability, revealing the performance limitations of general-domain Audio-MAEs. This work demonstrates that bridging this domain gap domain gap requires full-pipeline adaptation, not just domain-specific pretraining data. We systematically revisit and adapt the pretraining recipe, fine-tuning methods, and frozen feature utilization to bird sounds using BirdSet, a large-scale bioacoustic dataset comparable to AudioSet. Our resulting Bird-MAE achieves new state-of-the-art results in BirdSet's multi-label classification benchmark. Additionally, we introduce the parameter-efficient prototypical probing, enhancing the utility of frozen MAE representations and closely approaching fine-tuning performance in low-resource settings. Bird-MAE's prototypical probes outperform linear probing by up to 37 percentage points in mean average precision and narrow the gap to fine-tuning across BirdSet downstream tasks. Bird-MAE also demonstrates robust few-shot capabilities with prototypical probing in our newly established few-shot benchmark on BirdSet, highlighting the potential of tailored self-supervised learning pipelines for fine-grained audio domains.", "date": "2025-04-17", "categories": []}
{"id": "2505.23509", "title": "Spectrotemporal Modulation: Efficient and Interpretable Feature Representation for Classifying Speech, Music, and Environmental Sounds", "abstract": "Audio DNNs have demonstrated impressive performance on various machine listening tasks; however, most of their representations are computationally costly and uninterpretable, leaving room for optimization. Here, we propose a novel approach centered on spectrotemporal modulation (STM) features, a signal processing method that mimics the neurophysiological representation in the human auditory cortex. The classification performance of our STM-based model, without any pretraining, is comparable to that of pretrained audio DNNs across diverse naturalistic speech, music, and environmental sounds, which are essential categories for both human cognition and machine perception. These results show that STM is an efficient and interpretable feature representation for audio classification, advancing the development of machine listening and unlocking exciting new possibilities for basic understanding of speech and auditory sciences, as well as developing audio BCI and cognitive computing.", "date": "2025-05-29", "categories": []}
{"id": "2507.10642", "title": "First-of-its-kind AI model for bioacoustic detection using a lightweight associative memory Hopfield neural network", "abstract": "A growing issue within conservation bioacoustics is the task of analysing the vast amount of data generated from the use of passive acoustic monitoring devices. In this paper, we present an alternative AI model which has the potential to help alleviate this problem. Our model formulation addresses the key issues encountered when using current AI models for bioacoustic analysis, namely the: limited training data available; environmental impact, particularly in energy consumption and carbon footprint of training and implementing these models; and associated hardware requirements. The model developed in this work uses associative memory via a transparent, explainable Hopfield neural network to store signals and detect similar signals which can then be used to classify species. Training is rapid ($3$\\,ms), as only one representative signal is required for each target sound within a dataset. The model is fast, taking only $5.4$\\,s to pre-process and classify all $10384$ publicly available bat recordings, on a standard Apple MacBook Air. The model is also lightweight with a small memory footprint of $144.09$\\,MB of RAM usage. Hence, the low computational demands make the model ideal for use on a variety of standard personal devices with potential for deployment in the field via edge-processing devices. It is also competitively accurate, with up to $86\\%$ precision on the dataset used to evaluate the model. In fact, we could not find a single case of disagreement between model and manual identification via expert field guides. Although a dataset of bat echolocation calls was chosen to demo this first-of-its-kind AI model, trained on only two representative calls, the model is not species specific. In conclusion, we propose an equitable AI model that has the potential to be a game changer for fast, lightweight, sustainable, transparent, explainable and accurate bioacoustic analysis.", "date": "2025-07-14", "categories": []}
{"id": "2409.09866", "title": "S2Cap: A Benchmark and a Baseline for Singing Style Captioning", "abstract": "Singing voices contain much richer information than common voices, including varied vocal and acoustic properties. However, current open-source audio-text datasets for singing voices capture only a narrow range of attributes and lack acoustic features, leading to limited utility towards downstream tasks, such as style captioning. To fill this gap, we formally define the singing style captioning task and present S2Cap, a dataset of singing voices with detailed descriptions covering diverse vocal, acoustic, and demographic characteristics. Using this dataset, we develop an efficient and straightforward baseline algorithm for singing style captioning. The dataset is available at https://zenodo.org/records/15673764.", "date": "2024-09-15", "categories": []}
{"id": "2504.07402", "title": "LauraTSE: Target Speaker Extraction using Auto-Regressive Decoder-Only Language Models", "abstract": "We propose LauraTSE, an Auto-Regressive Decoder-Only Language Model for Target Speaker Extraction built upon the LauraGPT backbone. LauraTSE employs a small-scale auto-regressive decoder-only language model that generates the initial layers of the target speech's discrete codec representations from the continuous embeddings of both the mixture and reference speech. These outputs serve as coarse-grained predictions. To refine them, a one-step encoder-only language model reconstructs the full codec representation by integrating information from both the mixture and the reference speech, adding fine-grained details. Experimental results show that our approach can achieve promising performance. Additionally, we conduct ablation studies to investigate the data scalability and the contribution of the encoder-only model.", "date": "2025-04-10", "categories": []}
{"id": "2505.09304", "title": "Adaptive Noise Resilient Keyword Spotting Using One-Shot Learning", "abstract": "Keyword spotting (KWS) is a key component of smart devices, enabling efficient and intuitive audio interaction. However, standard KWS systems deployed on embedded devices often suffer performance degradation under real-world operating conditions. Resilient KWS systems address this issue by enabling dynamic adaptation, with applications such as adding or replacing keywords, adjusting to specific users, and improving noise robustness. However, deploying resilient, standalone KWS systems with low latency on resource-constrained devices remains challenging due to limited memory and computational resources. This study proposes a low computational approach for continuous noise adaptation of pretrained neural networks used for KWS classification, requiring only 1-shot learning and one epoch. The proposed method was assessed using two pretrained models and three real-world noise sources at signal-to-noise ratios (SNRs) ranging from 24 to -3 dB. The adapted models consistently outperformed the pretrained models across all scenarios, especially at SNR $\\leq$ 18 dB, achieving accuracy improvements of 4.9% to 46.0%. These results highlight the efficacy of the proposed methodology while being lightweight enough for deployment on resource-constrained devices.", "date": "2025-05-14", "categories": []}
{"id": "2506.07294", "title": "Towards Generalized Source Tracing for Codec-Based Deepfake Speech", "abstract": "Recent attempts at source tracing for codec-based deepfake speech (CodecFake), generated by neural audio codec-based speech generation (CoSG) models, have exhibited suboptimal performance. However, how to train source tracing models using simulated CoSG data while maintaining strong performance on real CoSG-generated audio remains an open challenge. In this paper, we show that models trained solely on codec-resynthesized data tend to overfit to non-speech regions and struggle to generalize to unseen content. To mitigate these challenges, we introduce the Semantic-Acoustic Source Tracing Network (SASTNet), which jointly leverages Whisper for semantic feature encoding and Wav2vec2 with AudioMAE for acoustic feature encoding. Our proposed SASTNet achieves state-of-the-art performance on the CoSG test set of the CodecFake+ dataset, demonstrating its effectiveness for reliable source tracing.", "date": "2025-06-08", "categories": []}
{"id": "2506.18843", "title": "USAD: Universal Speech and Audio Representation via Distillation", "abstract": "Self-supervised learning (SSL) has revolutionized audio representations, yet models often remain domain-specific, focusing on either speech or non-speech tasks. In this work, we present Universal Speech and Audio Distillation (USAD), a unified approach to audio representation learning that integrates diverse audio types - speech, sound, and music - into a single model. USAD employs efficient layer-to-layer distillation from domain-specific SSL models to train a student on a comprehensive audio dataset. USAD offers competitive performance across various benchmarks and datasets, including frame and instance-level speech processing tasks, audio tagging, and sound classification, achieving near state-of-the-art results with a single encoder on SUPERB and HEAR benchmarks.", "date": "2025-06-23", "categories": []}
{"id": "2508.11818", "title": "Audio Flamingo Sound-CoT Technical Report: Improving Chain-of-Thought Reasoning in Sound Understanding", "abstract": "Chain-of-thought reasoning has demonstrated significant improvements in large language models and vision language models, yet its potential for audio language models remains largely unexplored. In this technical report, we take a preliminary step towards closing this gap. For better assessment of sound reasoning, we propose AF-Reasoning-Eval, a benchmark targeting common-sense reasoning and the ability to discriminate among closely related choices. To prepare training corpus for sound reasoning abilities, we propose automatic pipelines that transform existing audio question answering and classification data into explicit reasoning chains, yielding AF-CoT-Train with 1.24M samples. We study the effect of finetuning Audio Flamingo series on AF-CoT-Train and observe considerable improvements on several reasoning benchmarks, validating the effectiveness of chain-of-thought finetuning on advanced sound understanding.", "date": "2025-08-15", "categories": []}
{"id": "2508.12709", "title": "MATPAC++: Enhanced Masked Latent Prediction for Self-Supervised Audio Representation Learning", "abstract": "Masked latent prediction has emerged as a leading paradigm in self-supervised learning (SSL), especially for general audio and music representation learning. While recent methods have demonstrated strong performance, the role of the predictor module used at the output of such SSL systems remains mainly overlooked, despite being crucial for solving the pretext task at hand. In particular, this module should be able to deal with the ambiguity inherent in audio content, especially when it is composed of multiple sound sources. This work proposes a novel enhancement: integrating Multiple Choice Learning (MCL) to explicitly model prediction ambiguity and improve representation quality. We build on top of the recently proposed MATPAC system, improving its prediction and unsupervised classification pretext tasks with MCL. We extensively evaluate our method, MATPAC++, through both linear probing across multiple downstream tasks and fine-tuning on AudioSet, employing a unified protocol that enables rigorous and fair comparisons with state-of-the-art SSL approaches. Results show that our proposal achieves state-of-the-art when fine-tuned on AudioSet and overall state-of-the-art scores on downstream tasks. Additionally, we examine domain specialisation by training exclusively on music data, where our model achieves state-of-the-art performance with significantly improved efficiency.", "date": "2025-08-18", "categories": []}
{"id": "2508.12968", "title": "Arabic ASR on the SADA Large-Scale Arabic Speech Corpus with Transformer-Based Models", "abstract": "We explore the performance of several state-of-the-art automatic speech recognition (ASR) models on a large-scale Arabic speech dataset, the SADA (Saudi Audio Dataset for Arabic), which contains 668 hours of high-quality audio from Saudi television shows. The dataset includes multiple dialects and environments, specifically a noisy subset that makes it particularly challenging for ASR. We evaluate the performance of the models on the SADA test set, and we explore the impact of fine-tuning, language models, as well as noise and denoising on their performance. We find that the best performing model is the MMS 1B model finetuned on SADA with a 4-gram language model that achieves a WER of 40.9\\% and a CER of 17.6\\% on the SADA test clean set.", "date": "2025-08-18", "categories": []}
{"id": "2502.10838", "title": "Generalizable speech deepfake detection via meta-learned LoRA", "abstract": "Reliable detection of speech deepfakes (spoofs) must remain effective when the distribution of spoofing attacks shifts. We frame the task as domain generalization and show that inserting Low-Rank Adaptation (LoRA) adapters into every attention head of a self-supervised (SSL) backbone, then training only those adapters with Meta-Learning Domain Generalization (MLDG), yields strong zero-shot performance. The resulting model updates about 3.6 million parameters, roughly 1.1% of the 318 million updated in full fine-tuning, yet surpasses a fully fine-tuned counterpart on five of six evaluation corpora. A first-order MLDG loop encourages the adapters to focus on cues that persist across attack types, lowering the average EER from 8.84% for the fully fine-tuned model to 5.30% with our best MLDG-LoRA configuration. Our findings show that combining meta-learning with parameter-efficient adaptation offers an effective method for zero-shot, distribution-shift-aware speech deepfake detection.", "date": "2025-02-15", "categories": []}
{"id": "2504.04949", "title": "L3AC: Towards a Lightweight and Lossless Audio Codec", "abstract": "Neural audio codecs have recently gained traction for their ability to compress high-fidelity audio and provide discrete tokens for generative modeling. However, leading approaches often rely on resource-intensive models and complex multi-quantizer architectures, limiting their practicality in real-world applications. In this work, we introduce L3AC, a lightweight neural audio codec that addresses these challenges by leveraging a single quantizer and a highly efficient architecture. To enhance reconstruction fidelity while minimizing model complexity, L3AC explores streamlined convolutional networks and local Transformer modules, alongside TConv--a novel structure designed to capture acoustic variations across multiple temporal scales. Despite its compact design, extensive experiments across diverse datasets demonstrate that L3AC matches or exceeds the reconstruction quality of leading codecs while reducing computational overhead by an order of magnitude. The single-quantizer design further enhances its adaptability for downstream tasks. The source code is publicly available at https://github.com/zhai-lw/L3AC.", "date": "2025-04-07", "categories": []}
{"id": "2506.11049", "title": "15,500 Seconds: Lean UAV Classification Using EfficientNet and Lightweight Fine-Tuning", "abstract": "As unmanned aerial vehicles (UAVs) become increasingly prevalent in both consumer and defense applications, the need for reliable, modality-specific classification systems grows in urgency. This paper addresses the challenge of data scarcity in UAV audio classification by expanding on prior work through the integration of pre-trained deep learning models, parameter-efficient fine-tuning (PEFT) strategies, and targeted data augmentation techniques. Using a custom dataset of 3,100 UAV audio clips (15,500 seconds) spanning 31 distinct drone types, we evaluate the performance of transformer-based and convolutional neural network (CNN) architectures under various fine-tuning configurations. Experiments were conducted with five-fold cross-validation, assessing accuracy, training efficiency, and robustness. Results show that full fine-tuning of the EfficientNet-B0 model with three augmentations achieved the highest validation accuracy (95.95), outperforming both the custom CNN and transformer-based models like AST. These findings suggest that combining lightweight architectures with PEFT and well-chosen augmentations provides an effective strategy for UAV audio classification on limited datasets. Future work will extend this framework to multimodal UAV classification using visual and radar telemetry.", "date": "2025-05-21", "categories": []}
{"id": "2508.10332", "title": "Layer-Wise Analysis of Self-Supervised Representations for Age and Gender Classification in Children's Speech", "abstract": "Children's speech presents challenges for age and gender classification due to high variability in pitch, articulation, and developmental traits. While self-supervised learning (SSL) models perform well on adult speech tasks, their ability to encode speaker traits in children remains underexplored. This paper presents a detailed layer-wise analysis of four Wav2Vec2 variants using the PFSTAR and CMU Kids datasets. Results show that early layers (1-7) capture speaker-specific cues more effectively than deeper layers, which increasingly focus on linguistic information. Applying PCA further improves classification, reducing redundancy and highlighting the most informative components. The Wav2Vec2-large-lv60 model achieves 97.14% (age) and 98.20% (gender) on CMU Kids; base-100h and large-lv60 models reach 86.05% and 95.00% on PFSTAR. These results reveal how speaker traits are structured across SSL model depth and support more targeted, adaptive strategies for child-aware speech interfaces.", "date": "2025-08-14", "categories": []}
{"id": "2508.10436", "title": "Alternating Approach-Putt Models for Multi-Stage Speech Enhancement", "abstract": "Speech enhancement using artificial neural networks aims to remove noise from noisy speech signals while preserving the speech content. However, speech enhancement networks often introduce distortions to the speech signal, referred to as artifacts, which can degrade audio quality. In this work, we propose a post-processing neural network designed to mitigate artifacts introduced by speech enhancement models. Inspired by the analogy of making a `Putt' after an `Approach' in golf, we name our model PuttNet. We demonstrate that alternating between a speech enhancement model and the proposed Putt model leads to improved speech quality, as measured by perceptual quality scores (PESQ), objective intelligibility (STOI), and background noise intrusiveness (CBAK) scores. Furthermore, we illustrate with graphical analysis why this alternating Approach outperforms repeated application of either model alone.", "date": "2025-08-14", "categories": []}
{"id": "2508.10559", "title": "Fake Speech Wild: Detecting Deepfake Speech on Social Media Platform", "abstract": "The rapid advancement of speech generation technology has led to the widespread proliferation of deepfake speech across social media platforms. While deepfake audio countermeasures (CMs) achieve promising results on public datasets, their performance degrades significantly in cross-domain scenarios. To advance CMs for real-world deepfake detection, we first propose the Fake Speech Wild (FSW) dataset, which includes 254 hours of real and deepfake audio from four different media platforms, focusing on social media. As CMs, we establish a benchmark using public datasets and advanced selfsupervised learning (SSL)-based CMs to evaluate current CMs in real-world scenarios. We also assess the effectiveness of data augmentation strategies in enhancing CM robustness for detecting deepfake speech on social media. Finally, by augmenting public datasets and incorporating the FSW training set, we significantly advanced real-world deepfake audio detection performance, achieving an average equal error rate (EER) of 3.54% across all evaluation sets.", "date": "2025-08-14", "categories": []}
{"id": "2501.11311", "title": "A2SB: Audio-to-Audio Schrodinger Bridges", "abstract": "Real-world audio is often degraded by numerous factors. This work presents an audio restoration model tailored for high-res music at 44.1kHz. Our model, Audio-to-Audio Schr\\\"odinger Bridges (A2SB), is capable of both bandwidth extension (predicting high-frequency components) and inpainting (re-generating missing segments). Critically, A2SB is end-to-end requiring no vocoder to predict waveform outputs, able to restore hour-long audio inputs, and trained on permissively licensed music data. A2SB is capable of achieving state-of-the-art band-width extension and inpainting quality on several out-of-distribution music test sets.", "date": "2025-01-20", "categories": []}
{"id": "2504.20678", "title": "Non-native Children's Automatic Speech Assessment Challenge (NOCASA)", "abstract": "This paper presents the \"Non-native Children's Automatic Speech Assessment\" (NOCASA) - a data competition part of the IEEE MLSP 2025 conference. NOCASA challenges participants to develop new systems that can assess single-word pronunciations of young second language (L2) learners as part of a gamified pronunciation training app. To achieve this, several issues must be addressed, most notably the limited nature of available training data and the highly unbalanced distribution among the pronunciation level categories. To expedite the development, we provide a pseudo-anonymized training data (TeflonNorL2), containing 10,334 recordings from 44 speakers attempting to pronounce 205 distinct Norwegian words, human-rated on a 1 to 5 scale (number of stars that should be given in the game). In addition to the data, two already trained systems are released as official baselines: an SVM classifier trained on the ComParE_16 acoustic feature set and a multi-task wav2vec 2.0 model. The latter achieves the best performance on the challenge test set, with an unweighted average recall (UAR) of 36.37%.", "date": "2025-04-29", "categories": []}
{"id": "2508.09294", "title": "Fake-Mamba: Real-Time Speech Deepfake Detection Using Bidirectional Mamba as Self-Attention's Alternative", "abstract": "Advances in speech synthesis intensify security threats, motivating real-time deepfake detection research. We investigate whether bidirectional Mamba can serve as a competitive alternative to Self-Attention in detecting synthetic speech. Our solution, Fake-Mamba, integrates an XLSR front-end with bidirectional Mamba to capture both local and global artifacts. Our core innovation introduces three efficient encoders: TransBiMamba, ConBiMamba, and PN-BiMamba. Leveraging XLSR's rich linguistic representations, PN-BiMamba can effectively capture the subtle cues of synthetic speech. Evaluated on ASVspoof 21 LA, 21 DF, and In-The-Wild benchmarks, Fake-Mamba achieves 0.97%, 1.74%, and 5.85% EER, respectively, representing substantial relative gains over SOTA models XLSR-Conformer and XLSR-Mamba. The framework maintains real-time inference across utterance lengths, demonstrating strong generalization and practical viability. The code is available at https://github.com/xuanxixi/Fake-Mamba.", "date": "2025-08-12", "categories": []}
{"id": "2508.09389", "title": "ProMode: A Speech Prosody Model Conditioned on Acoustic and Textual Inputs", "abstract": "Prosody conveys rich emotional and semantic information of the speech signal as well as individual idiosyncrasies. We propose a stand-alone model that maps text-to-prosodic features such as F0 and energy and can be used in downstream tasks such as TTS. The ProMode encoder takes as input acoustic features and time-aligned textual content, both are partially masked, and obtains a fixed-length latent prosodic embedding. The decoder predicts acoustics in the masked region using both the encoded prosody input and unmasked textual content. Trained on the GigaSpeech dataset, we compare our method with state-of-the-art style encoders. For F0 and energy predictions, we show consistent improvements for our model at different levels of granularity. We also integrate these predicted prosodic features into a TTS system and conduct perceptual tests, which show higher prosody preference compared to the baselines, demonstrating the model's potential in tasks where prosody modeling is important.", "date": "2025-08-12", "categories": []}
{"id": "2410.23279", "title": "Learning Marmoset Vocal Patterns with a Masked Autoencoder for Robust Call Segmentation, Classification, and Caller Identification", "abstract": "The marmoset, a highly vocal primate, is a key model for studying social-communicative behavior. Unlike human speech, marmoset vocalizations are less structured, highly variable, and recorded in noisy, low-resource conditions. Learning marmoset communication requires joint call segmentation, classification, and caller identification -- challenging domain tasks. Previous CNNs handle local patterns but struggle with long-range temporal structure. We applied Transformers using self-attention for global dependencies. However, Transformers show overfitting and instability on small, noisy annotated datasets. To address this, we pretrain Transformers with MAE -- a self-supervised method reconstructing masked segments from hundreds of hours of unannotated marmoset recordings. The pretraining improved stability and generalization. Results show MAE-pretrained Transformers outperform CNNs, demonstrating modern self-supervised architectures effectively model low-resource non-human vocal communication.", "date": "2024-10-30", "categories": []}
{"id": "2411.15447", "title": "Gotta Hear Them All: Towards Sound Source Aware Audio Generation", "abstract": "Audio synthesis has broad applications in multimedia. Recent advancements have made it possible to generate relevant audios from inputs describing an audio scene, such as images or texts. However, the immersiveness and expressiveness of the generation are limited. One possible problem is that existing methods solely rely on the global scene and overlook details of local sounding objects (i.e., sound sources). To address this issue, we propose a Sound Source-Aware Audio (SS2A) generator. SS2A is able to locally perceive multimodal sound sources from a scene with visual detection and cross-modality translation. It then contrastively learns a Cross-Modal Sound Source (CMSS) Manifold to semantically disambiguate each source. Finally, we attentively mix their CMSS semantics into a rich audio representation, from which a pretrained audio generator outputs the sound. To model the CMSS manifold, we curate a novel single-sound-source visual-audio dataset VGGS3 from VGGSound. We also design a Sound Source Matching Score to clearly measure localized audio relevance. With the effectiveness of explicit sound source modeling, SS2A achieves state-of-the-art performance in extensive image-to-audio tasks. We also qualitatively demonstrate SS2A's ability to achieve intuitive synthesis control by compositing vision, text, and audio conditions. Furthermore, we show that our sound source modeling can achieve competitive video-to-audio performance with a straightforward temporal aggregation mechanism.", "date": "2024-11-23", "categories": []}
{"id": "2508.08892", "title": "Sound Signal Synthesis with Auxiliary Classifier GAN, COVID-19 cough as an example", "abstract": "One of the fastest-growing domains in AI is healthcare. Given its importance, it has been the interest of many researchers to deploy ML models into the ever-demanding healthcare domain to aid doctors and increase accessibility. Delivering reliable models, however, demands a sizable amount of data, and the recent COVID-19 pandemic served as a reminder of the rampant and scary nature of healthcare that makes training models difficult. To alleviate such scarcity, many published works attempted to synthesize radiological cough data to train better COVID-19 detection models on the respective radiological data. To accommodate the time sensitivity expected during a pandemic, this work focuses on detecting COVID-19 through coughs using synthetic data to improve the accuracy of the classifier. The work begins by training a CNN on a balanced subset of the Coughvid dataset, establishing a baseline classification test accuracy of 72%. The paper demonstrates how an Auxiliary Classification GAN (ACGAN) may be trained to conditionally generate novel synthetic Mel Spectrograms of both healthy and COVID-19 coughs. These coughs are used to augment the training dataset of the CNN classifier, allowing it to reach a new test accuracy of 75%. The work highlights the expected messiness and inconsistency in training and offers insights into detecting and handling such shortcomings.", "date": "2025-08-12", "categories": []}
{"id": "2508.08912", "title": "Munsit at NADI 2025 Shared Task 2: Pushing the Boundaries of Multidialectal Arabic ASR with Weakly Supervised Pretraining and Continual Supervised Fine-tuning", "abstract": "Automatic speech recognition (ASR) plays a vital role in enabling natural human-machine interaction across applications such as virtual assistants, industrial automation, customer support, and real-time transcription. However, developing accurate ASR systems for low-resource languages like Arabic remains a significant challenge due to limited labeled data and the linguistic complexity introduced by diverse dialects. In this work, we present a scalable training pipeline that combines weakly supervised learning with supervised fine-tuning to develop a robust Arabic ASR model. In the first stage, we pretrain the model on 15,000 hours of weakly labeled speech covering both Modern Standard Arabic (MSA) and various Dialectal Arabic (DA) variants. In the subsequent stage, we perform continual supervised fine-tuning using a mixture of filtered weakly labeled data and a small, high-quality annotated dataset. Our approach achieves state-of-the-art results, ranking first in the multi-dialectal Arabic ASR challenge. These findings highlight the effectiveness of weak supervision paired with fine-tuning in overcoming data scarcity and delivering high-quality ASR for low-resource, dialect-rich languages.", "date": "2025-08-12", "categories": []}
{"id": "2508.08957", "title": "QAMRO: Quality-aware Adaptive Margin Ranking Optimization for Human-aligned Assessment of Audio Generation Systems", "abstract": "Evaluating audio generation systems, including text-to-music (TTM), text-to-speech (TTS), and text-to-audio (TTA), remains challenging due to the subjective and multi-dimensional nature of human perception. Existing methods treat mean opinion score (MOS) prediction as a regression problem, but standard regression losses overlook the relativity of perceptual judgments. To address this limitation, we introduce QAMRO, a novel Quality-aware Adaptive Margin Ranking Optimization framework that seamlessly integrates regression objectives from different perspectives, aiming to highlight perceptual differences and prioritize accurate ratings. Our framework leverages pre-trained audio-text models such as CLAP and Audiobox-Aesthetics, and is trained exclusively on the official AudioMOS Challenge 2025 dataset. It demonstrates superior alignment with human evaluations across all dimensions, significantly outperforming robust baseline models.", "date": "2025-08-12", "categories": []}
{"id": "2507.16845", "title": "Enhancing Lung Disease Diagnosis via Semi-Supervised Machine Learning", "abstract": "Lung diseases, including lung cancer and COPD, are significant health concerns globally. Traditional diagnostic methods can be costly, time-consuming, and invasive. This study investigates the use of semi supervised learning methods for lung sound signal detection using a model combination of MFCC+CNN. By introducing semi supervised learning modules such as Mix Match, Co-Refinement, and Co Refurbishing, we aim to enhance the detection performance while reducing dependence on manual annotations. With the add-on semi-supervised modules, the accuracy rate of the MFCC+CNN model is 92.9%, an increase of 3.8% to the baseline model. The research contributes to the field of lung disease sound detection by addressing challenges such as individual differences, feature insufficient labeled data.", "date": "2025-07-20", "categories": []}
{"id": "2508.07561", "title": "A Small-footprint Acoustic Echo Cancellation Solution for Mobile Full-Duplex Speech Interactions", "abstract": "In full-duplex speech interaction systems, effective Acoustic Echo Cancellation (AEC) is crucial for recovering echo-contaminated speech. This paper presents a neural network-based AEC solution to address challenges in mobile scenarios with varying hardware, nonlinear distortions and long latency. We first incorporate diverse data augmentation strategies to enhance the model's robustness across various environments. Moreover, progressive learning is employed to incrementally improve AEC effectiveness, resulting in a considerable improvement in speech quality. To further optimize AEC's downstream applications, we introduce a novel post-processing strategy employing tailored parameters designed specifically for tasks such as Voice Activity Detection (VAD) and Automatic Speech Recognition (ASR), thus enhancing their overall efficacy. Finally, our method employs a small-footprint model with streaming inference, enabling seamless deployment on mobile devices. Empirical results demonstrate effectiveness of the proposed method in Echo Return Loss Enhancement and Perceptual Evaluation of Speech Quality, alongside significant improvements in both VAD and ASR results.", "date": "2025-08-11", "categories": []}
{"id": "2508.07587", "title": "Voice Pathology Detection Using Phonation", "abstract": "Voice disorders significantly affect communication and quality of life, requiring an early and accurate diagnosis. Traditional methods like laryngoscopy are invasive, subjective, and often inaccessible. This research proposes a noninvasive, machine learning-based framework for detecting voice pathologies using phonation data.\n  Phonation data from the Saarbr\\\"ucken Voice Database are analyzed using acoustic features such as Mel Frequency Cepstral Coefficients (MFCCs), chroma features, and Mel spectrograms. Recurrent Neural Networks (RNNs), including LSTM and attention mechanisms, classify samples into normal and pathological categories. Data augmentation techniques, including pitch shifting and Gaussian noise addition, enhance model generalizability, while preprocessing ensures signal quality. Scale-based features, such as H\\\"older and Hurst exponents, further capture signal irregularities and long-term dependencies.\n  The proposed framework offers a noninvasive, automated diagnostic tool for early detection of voice pathologies, supporting AI-driven healthcare, and improving patient outcomes.", "date": "2025-08-11", "categories": []}
{"id": "2508.07829", "title": "Auditory Intelligence: Understanding the World Through Sound", "abstract": "Recent progress in auditory intelligence has yielded high-performing systems for sound event detection (SED), acoustic scene classification (ASC), automated audio captioning (AAC), and audio question answering (AQA). Yet these tasks remain largely constrained to surface-level recognition-capturing what happened but not why, what it implies, or how it unfolds in context. I propose a conceptual reframing of auditory intelligence as a layered, situated process that encompasses perception, reasoning, and interaction. To instantiate this view, I introduce four cognitively inspired task paradigms-ASPIRE, SODA, AUX, and AUGMENT-those structure auditory understanding across time-frequency pattern captioning, hierarchical event/scene description, causal explanation, and goal-driven interpretation, respectively. Together, these paradigms provide a roadmap toward more generalizable, explainable, and human-aligned auditory intelligence, and are intended to catalyze a broader discussion of what it means for machines to understand sound.", "date": "2025-08-11", "categories": []}
{"id": "2508.07973", "title": "Joint Transcription of Acoustic Guitar Strumming Directions and Chords", "abstract": "Automatic transcription of guitar strumming is an underrepresented and challenging task in Music Information Retrieval (MIR), particularly for extracting both strumming directions and chord progressions from audio signals. While existing methods show promise, their effectiveness is often hindered by limited datasets. In this work, we extend a multimodal approach to guitar strumming transcription by introducing a novel dataset and a deep learning-based transcription model. We collect 90 min of real-world guitar recordings using an ESP32 smartwatch motion sensor and a structured recording protocol, complemented by a synthetic dataset of 4h of labeled strumming audio. A Convolutional Recurrent Neural Network (CRNN) model is trained to detect strumming events, classify their direction, and identify the corresponding chords using only microphone audio. Our evaluation demonstrates significant improvements over baseline onset detection algorithms, with a hybrid method combining synthetic and real-world data achieving the highest accuracy for both strumming action detection and chord classification. These results highlight the potential of deep learning for robust guitar strumming transcription and open new avenues for automatic rhythm guitar analysis.", "date": "2025-08-11", "categories": []}
{"id": "2504.21772", "title": "Solving Copyright Infringement on Short Video Platforms: Novel Datasets and an Audio Restoration Deep Learning Pipeline", "abstract": "Short video platforms like YouTube Shorts and TikTok face significant copyright compliance challenges, as infringers frequently embed arbitrary background music (BGM) to obscure original soundtracks (OST) and evade content originality detection. To tackle this issue, we propose a novel pipeline that integrates Music Source Separation (MSS) and cross-modal video-music retrieval (CMVMR). Our approach effectively separates arbitrary BGM from the original OST, enabling the restoration of authentic video audio tracks. To support this work, we introduce two domain-specific datasets: OASD-20K for audio separation and OSVAR-160 for pipeline evaluation. OASD-20K contains 20,000 audio clips featuring mixed BGM and OST pairs, while OSVAR-160 is a unique benchmark dataset comprising 1,121 video and mixed-audio pairs, specifically designed for short video restoration tasks. Experimental results demonstrate that our pipeline not only removes arbitrary BGM with high accuracy but also restores OSTs, ensuring content integrity. This approach provides an ethical and scalable solution to copyright challenges in user-generated content on short video platforms.", "date": "2025-04-30", "categories": []}
{"id": "2508.05835", "title": "NanoCodec: Towards High-Quality Ultra Fast Speech LLM Inference", "abstract": "Large Language Models (LLMs) have significantly advanced audio processing by leveraging audio codecs to discretize audio into tokens, enabling the application of language modeling techniques to speech data. However, existing audio codecs often operate at high frame rates, leading to slow training and inference, particularly for autoregressive models. To address this, there is growing interest in low frame-rate audio codecs, which reduce the number of autoregressive steps required to generate one second of audio. In this paper, we conduct ablation studies to examine the impact of frame rate, bitrate, and causality on codec reconstruction quality. Based on our findings, we introduce NanoCodec, a state-of-the-art audio codec that achieves high-quality compression at just 12.5 frames per second (FPS). NanoCodec outperforms related works across various bitrate ranges, establishing a new benchmark for low-latency and efficient Speech LLM training and inference.", "date": "2025-08-07", "categories": []}
{"id": "2508.05878", "title": "Training chord recognition models on artificially generated audio", "abstract": "One of the challenging problems in Music Information Retrieval is the acquisition of enough non-copyrighted audio recordings for model training and evaluation. This study compares two Transformer-based neural network models for chord sequence recognition in audio recordings and examines the effectiveness of using an artificially generated dataset for this purpose. The models are trained on various combinations of Artificial Audio Multitracks (AAM), Schubert's Winterreise Dataset, and the McGill Billboard Dataset and evaluated with three metrics: Root, MajMin and Chord Content Metric (CCM). The experiments prove that even though there are certainly differences in complexity and structure between artificially generated and human-composed music, the former can be useful in certain scenarios. Specifically, AAM can enrich a smaller training dataset of music composed by a human or can even be used as a standalone training set for a model that predicts chord sequences in pop music, if no other data is available.", "date": "2025-08-07", "categories": []}
{"id": "2508.06321", "title": "EmoAugNet: A Signal-Augmented Hybrid CNN-LSTM Framework for Speech Emotion Recognition", "abstract": "Recognizing emotional signals in speech has a significant impact on enhancing the effectiveness of human-computer interaction (HCI). This study introduces EmoAugNet, a hybrid deep learning framework, that incorporates Long Short-Term Memory (LSTM) layers with one-dimensional Convolutional Neural Networks (1D-CNN) to enable reliable Speech Emotion Recognition (SER). The quality and variety of the features that are taken from speech signals have a significant impact on how well SER systems perform. A comprehensive speech data augmentation strategy was used to combine both traditional methods, such as noise addition, pitch shifting, and time stretching, with a novel combination-based augmentation pipeline to enhance generalization and reduce overfitting. Each audio sample was transformed into a high-dimensional feature vector using root mean square energy (RMSE), Mel-frequency Cepstral Coefficient (MFCC), and zero-crossing rate (ZCR). Our model with ReLU activation has a weighted accuracy of 95.78\\% and unweighted accuracy of 92.52\\% on the IEMOCAP dataset and, with ELU activation, has a weighted accuracy of 96.75\\% and unweighted accuracy of 91.28\\%. On the RAVDESS dataset, we get a weighted accuracy of 94.53\\% and 94.98\\% unweighted accuracy for ReLU activation and 93.72\\% weighted accuracy and 94.64\\% unweighted accuracy for ELU activation. These results highlight EmoAugNet's effectiveness in improving the robustness and performance of SER systems through integated data augmentation and hybrid modeling.", "date": "2025-08-06", "categories": []}
{"id": "2508.00733", "title": "AudioGen-Omni: A Unified Multimodal Diffusion Transformer for Video-Synchronized Audio, Speech, and Song Generation", "abstract": "We present AudioGen-Omni - a unified approach based on multimodal diffusion transformers (MMDit), capable of generating high-fidelity audio, speech, and song coherently synchronized with the input video. AudioGen-Omni introduces a novel joint training paradigm that seamlessly integrates large-scale video-text-audio corpora, enabling a model capable of generating semantically rich, acoustically diverse audio conditioned on multimodal inputs and adaptable to a wide range of audio generation tasks. AudioGen-Omni employs a unified lyrics-transcription encoder that encodes graphemes and phonemes from both song and spoken inputs into dense frame-level representations. Dense frame-level representations are fused using an AdaLN-based joint attention mechanism enhanced with phase-aligned anisotropic positional infusion (PAAPI), wherein RoPE is selectively applied to temporally structured modalities to ensure precise and robust cross-modal alignment. By unfreezing all modalities and masking missing inputs, AudioGen-Omni mitigates the semantic constraints of text-frozen paradigms, enabling effective cross-modal conditioning. This joint training approach enhances audio quality, semantic alignment, and lip-sync accuracy, while also achieving state-of-the-art results on Text-to-Audio/Speech/Song tasks. With an inference time of 1.91 seconds for 8 seconds of audio, it offers substantial improvements in both efficiency and generality.", "date": "2025-08-01", "categories": []}
{"id": "2508.05306", "title": "Estimating Musical Surprisal from Audio in Autoregressive Diffusion Model Noise Spaces", "abstract": "Recently, the information content (IC) of predictions from a Generative Infinite-Vocabulary Transformer (GIVT) has been used to model musical expectancy and surprisal in audio. We investigate the effectiveness of such modelling using IC calculated with autoregressive diffusion models (ADMs). We empirically show that IC estimates of models based on two different diffusion ordinary differential equations (ODEs) describe diverse data better, in terms of negative log-likelihood, than a GIVT. We evaluate diffusion model IC's effectiveness in capturing surprisal aspects by examining two tasks: (1) capturing monophonic pitch surprisal, and (2) detecting segment boundaries in multi-track audio. In both tasks, the diffusion models match or exceed the performance of a GIVT. We hypothesize that the surprisal estimated at different diffusion process noise levels corresponds to the surprisal of music and audio features present at different audio granularities. Testing our hypothesis, we find that, for appropriate noise levels, the studied musical surprisal tasks' results improve. Code is provided on github.com/SonyCSLParis/audioic.", "date": "2025-08-07", "categories": []}
{"id": "2405.04124", "title": "Comparative Study of State-based Neural Networks for Virtual Analog Audio Effects Modeling", "abstract": "Artificial neural networks are a promising technique for virtual analog modeling, having shown particular success in emulating distortion circuits. Despite their potential, enhancements are needed to enable effect parameters to influence the network's response and to achieve a low-latency output. While hybrid solutions, which incorporate both analytical and black-box techniques, offer certain advantages, black-box approaches, such as neural networks, can be preferable in contexts where rapid deployment, simplicity, or adaptability are required, and where understanding the internal mechanisms of the system is less critical. In this article, we explore the application of recent machine learning advancements for virtual analog modeling. We compare State-Space models and Linear Recurrent Units against the more common LSTM networks, with a variety of audio effects. We evaluate the performance and limitations of these models using multiple metrics, providing insights for future research and development. Our metrics aim to assess the models' ability to accurately replicate the signal's energy and frequency contents, with a particular focus on transients. The Feature-wise Linear Modulation method is employed to incorporate effect parameters that influence the network's response, enabling dynamic adaptability based on specified conditions. Experimental results suggest that LSTM networks offer an advantage in emulating distortions and equalizers, although performance differences are sometimes subtle yet statistically significant. On the other hand, encoder-decoder configurations of Long Short-Term Memory networks and State-Space models excel in modeling saturation and compression, effectively managing the dynamic aspects inherent in these effects. However, no models effectively emulate the low-pass filter, and Linear Recurrent Units show inconsistent performance across various audio effects.", "date": "2024-05-07", "categories": []}
{"id": "2501.02786", "title": "CCStereo: Audio-Visual Contextual and Contrastive Learning for Binaural Audio Generation", "abstract": "Binaural audio generation (BAG) aims to convert monaural audio to stereo audio using visual prompts, requiring a deep understanding of spatial and semantic information. However, current models risk overfitting to room environments and lose fine-grained spatial details. In this paper, we propose a new audio-visual binaural generation model incorporating an audio-visual conditional normalisation layer that dynamically aligns the mean and variance of the target difference audio features using visual context, along with a new contrastive learning method to enhance spatial sensitivity by mining negative samples from shuffled visual features. We also introduce a cost-efficient way to utilise test-time augmentation in video data to enhance performance. Our approach achieves state-of-the-art generation accuracy on the FAIR-Play and MUSIC-Stereo benchmarks.", "date": "2025-01-06", "categories": []}
{"id": "2508.03764", "title": "CoughViT: A Self-Supervised Vision Transformer for Cough Audio Representation Learning", "abstract": "Physicians routinely assess respiratory sounds during the diagnostic process, providing insight into the condition of a patient's airways. In recent years, AI-based diagnostic systems operating on respiratory sounds, have demonstrated success in respiratory disease detection. These systems represent a crucial advancement in early and accessible diagnosis which is essential for timely treatment. However, label and data scarcity remain key challenges, especially for conditions beyond COVID-19, limiting diagnostic performance and reliable evaluation. In this paper, we propose CoughViT, a novel pre-training framework for learning general-purpose cough sound representations, to enhance diagnostic performance in tasks with limited data. To address label scarcity, we employ masked data modelling to train a feature encoder in a self-supervised learning manner. We evaluate our approach against other pre-training strategies on three diagnostically important cough classification tasks. Experimental results show that our representations match or exceed current state-of-the-art supervised audio representations in enhancing performance on downstream tasks.", "date": "2025-08-04", "categories": []}
{"id": "2507.01339", "title": "User-guided Generative Source Separation", "abstract": "Music source separation (MSS) aims to extract individual instrument sources from their mixture. While most existing methods focus on the widely adopted four-stem separation setup (vocals, bass, drums, and other instruments), this approach lacks the flexibility needed for real-world applications. To address this, we propose GuideSep, a diffusion-based MSS model capable of instrument-agnostic separation beyond the four-stem setup. GuideSep is conditioned on multiple inputs: a waveform mimicry condition, which can be easily provided by humming or playing the target melody, and mel-spectrogram domain masks, which offer additional guidance for separation. Unlike prior approaches that relied on fixed class labels or sound queries, our conditioning scheme, coupled with the generative approach, provides greater flexibility and applicability. Additionally, we design a mask-prediction baseline using the same model architecture to systematically compare predictive and generative approaches. Our objective and subjective evaluations demonstrate that GuideSep achieves high-quality separation while enabling more versatile instrument extraction, highlighting the potential of user participation in the diffusion-based generative process for MSS. Our code and demo page are available at https://yutongwen.github.io/GuideSep/", "date": "2025-07-02", "categories": []}
{"id": "2507.12701", "title": "Task-Specific Audio Coding for Machines: Machine-Learned Latent Features Are Codes for That Machine", "abstract": "Neural audio codecs, leveraging quantization algorithms, have significantly impacted various speech/audio tasks. While high-fidelity reconstruction is paramount for human perception, audio coding for machines (ACoM) prioritizes efficient compression and downstream task performance, disregarding perceptual nuances. This work introduces an efficient ACoM method that can compress and quantize any chosen intermediate feature representation of an already trained speech/audio downstream model. Our approach employs task-specific loss guidance alongside residual vector quantization (RVQ) losses, providing ultra-low bitrates (i.e., less than 200 bps) with a minimal loss of the downstream model performance. The resulting tokenizer is adaptable to various bitrates and model sizes for flexible deployment. Evaluated on automatic speech recognition and audio classification, our method demonstrates its efficacy and potential for broader task and architectural applicability through appropriate regularization.", "date": "2025-07-17", "categories": []}
{"id": "2507.21202", "title": "Combolutional Neural Networks", "abstract": "Selecting appropriate inductive biases is an essential step in the design of machine learning models, especially when working with audio, where even short clips may contain millions of samples. To this end, we propose the combolutional layer: a learned-delay IIR comb filter and fused envelope detector, which extracts harmonic features in the time domain. We demonstrate the efficacy of the combolutional layer on three information retrieval tasks, evaluate its computational cost relative to other audio frontends, and provide efficient implementations for training. We find that the combolutional layer is an effective replacement for convolutional layers in audio tasks where precise harmonic analysis is important, e.g., piano transcription, speaker classification, and key detection. Additionally, the combolutional layer has several other key benefits over existing frontends, namely: low parameter count, efficient CPU inference, strictly real-valued computations, and improved interpretability.", "date": "2025-07-28", "categories": []}
{"id": "2508.02801", "title": "Adaptive Knowledge Distillation for Device-Directed Speech Detection", "abstract": "Device-directed speech detection (DDSD) is a binary classification task that separates the user's queries to a voice assistant (VA) from background speech or side conversations. This is important for achieving naturalistic user experience. To this end, we propose knowledge distillation (KD) to enhance DDSD accuracy while ensuring efficient deployment. Specifically, we introduce a novel adaptive KD method that transfers knowledge from general representations of an ASR large pre-trained acoustic encoder (teacher). We apply task-specific adapters, on top of the (frozen) teacher encoder, trained jointly with the student model on DDSD. We demonstrate that the proposed adaptive KD outperforms the student model without distillation in the keyword and keyword-free (follow-up) invocations, with an improvement of +26% and +19% in terms of Equal Error Rate, respectively. We also show that this approach generalizes across the transformer and conformer-based model architectures.", "date": "2025-08-04", "categories": []}
{"id": "2508.02905", "title": "How Would It Sound? Material-Controlled Multimodal Acoustic Profile Generation for Indoor Scenes", "abstract": "How would the sound in a studio change with a carpeted floor and acoustic tiles on the walls? We introduce the task of material-controlled acoustic profile generation, where, given an indoor scene with specific audio-visual characteristics, the goal is to generate a target acoustic profile based on a user-defined material configuration at inference time. We address this task with a novel encoder-decoder approach that encodes the scene's key properties from an audio-visual observation and generates the target Room Impulse Response (RIR) conditioned on the material specifications provided by the user. Our model enables the generation of diverse RIRs based on various material configurations defined dynamically at inference time. To support this task, we create a new benchmark, the Acoustic Wonderland Dataset, designed for developing and evaluating material-aware RIR prediction methods under diverse and challenging settings. Our results demonstrate that the proposed model effectively encodes material information and generates high-fidelity RIRs, outperforming several baselines and state-of-the-art methods.", "date": "2025-08-04", "categories": []}
{"id": "2409.06690", "title": "Benchmarking Sub-Genre Classification For Mainstage Dance Music", "abstract": "Music classification, a cornerstone of music information retrieval, supports a wide array of applications. To address the lack of comprehensive datasets and effective methods for sub-genre classification in mainstage dance music, we introduce a novel benchmark featuring a new dataset and baseline. Our dataset expands the scope of sub-genres to reflect the diversity of recent mainstage live sets performed by leading DJs at global music festivals, capturing the vibrant and rapidly evolving electronic dance music (EDM) scene that engages millions of fans worldwide. We employ a continuous soft labeling approach to accommodate tracks blending multiple sub-genres, preserving their inherent complexity. Experiments demonstrate that even state-of-the-art multimodal large language models (MLLMs) struggle with this task, while our specialized baseline models achieve high accuracy. This benchmark supports applications such as music recommendation, DJ set curation, and interactive multimedia systems, with video demos provided. Our code and data are all open-sourced at https://github.com/Gariscat/housex-v2.git.", "date": "2024-09-10", "categories": []}
{"id": "2409.15672", "title": "Language-based Audio Moment Retrieval", "abstract": "In this paper, we propose and design a new task called audio moment retrieval (AMR). Unlike conventional language-based audio retrieval tasks that search for short audio clips from an audio database, AMR aims to predict relevant moments in untrimmed long audio based on a text query. Given the lack of prior work in AMR, we first build a dedicated dataset, Clotho-Moment, consisting of large-scale simulated audio recordings with moment annotations. We then propose a DETR-based model, named Audio Moment DETR (AM-DETR), as a fundamental framework for AMR tasks. This model captures temporal dependencies within audio features, inspired by similar video moment retrieval tasks, thus surpassing conventional clip-level audio retrieval methods. Additionally, we provide manually annotated datasets to properly measure the effectiveness and robustness of our methods on real data. Experimental results show that AM-DETR, trained with Clotho-Moment, outperforms a baseline model that applies a clip-level audio retrieval method with a sliding window on all metrics, particularly improving Recall1@0.7 by 9.00 points. Our datasets and code are publicly available in https://h-munakata.github.io/Language-based-Audio-Moment-Retrieval.", "date": "2024-09-24", "categories": []}
{"id": "2506.11811", "title": "Abstract Sound Fusion with Unconditional Inversion Models", "abstract": "An abstract sound is defined as a sound that does not disclose identifiable real-world sound events to a listener. Sound fusion aims to synthesize an original sound and a reference sound to generate a novel sound that exhibits auditory features beyond mere additive superposition of the sound constituents. To achieve this fusion, we employ inversion techniques that preserve essential features of the original sample while enabling controllable synthesis. We propose novel SDE and ODE inversion models based on DPMSolver++ samplers that reverse the sampling process by configuring model outputs as constants, eliminating circular dependencies incurred by noise prediction terms. Our inversion approach requires no prompt conditioning while maintaining flexible guidance during sampling.", "date": "2025-06-13", "categories": []}
{"id": "2507.07043", "title": "Advances in Intelligent Hearing Aids: Deep Learning Approaches to Selective Noise Cancellation", "abstract": "The integration of artificial intelligence into hearing assistance marks a paradigm shift from traditional amplification-based systems to intelligent, context-aware audio processing. This systematic literature review evaluates advances in AI-driven selective noise cancellation (SNC) for hearing aids, highlighting technological evolution, implementation challenges, and future research directions. We synthesize findings across deep learning architectures, hardware deployment strategies, clinical validation studies, and user-centric design. The review traces progress from early machine learning models to state-of-the-art deep networks, including Convolutional Recurrent Networks for real-time inference and Transformer-based architectures for high-accuracy separation. Key findings include significant gains over traditional methods, with recent models achieving up to 18.3 dB SI-SDR improvement on noisy-reverberant benchmarks, alongside sub-10 ms real-time implementations and promising clinical outcomes. Yet, challenges remain in bridging lab-grade models with real-world deployment - particularly around power constraints, environmental variability, and personalization. Identified research gaps include hardware-software co-design, standardized evaluation protocols, and regulatory considerations for AI-enhanced hearing devices. Future work must prioritize lightweight models, continual learning, contextual-based classification and clinical translation to realize transformative hearing solutions for millions globally.", "date": "2025-06-25", "categories": []}
{"id": "2507.21448", "title": "Real-Time Audio-Visual Speech Enhancement Using Pre-trained Visual Representations", "abstract": "Speech enhancement in audio-only settings remains challenging, particularly in the presence of interfering speakers. This paper presents a simple yet effective real-time audio-visual speech enhancement (AVSE) system, RAVEN, which isolates and enhances the on-screen target speaker while suppressing interfering speakers and background noise. We investigate how visual embeddings learned from audio-visual speech recognition (AVSR) and active speaker detection (ASD) contribute to AVSE across different SNR conditions and numbers of interfering speakers. Our results show concatenating embeddings from AVSR and ASD models provides the greatest improvement in low-SNR, multi-speaker environments, while AVSR embeddings alone perform best in noise-only scenarios. In addition, we develop a real-time streaming system that operates on a computer CPU and we provide a video demonstration and code repository. To our knowledge, this is the first open-source implementation of a real-time AVSE system.", "date": "2025-07-29", "categories": []}
{"id": "2508.01172", "title": "GeHirNet: A Gender-Aware Hierarchical Model for Voice Pathology Classification", "abstract": "AI-based voice analysis shows promise for disease diagnostics, but existing classifiers often fail to accurately identify specific pathologies because of gender-related acoustic variations and the scarcity of data for rare diseases. We propose a novel two-stage framework that first identifies gender-specific pathological patterns using ResNet-50 on Mel spectrograms, then performs gender-conditioned disease classification. We address class imbalance through multi-scale resampling and time warping augmentation. Evaluated on a merged dataset from four public repositories, our two-stage architecture with time warping achieves state-of-the-art performance (97.63\\% accuracy, 95.25\\% MCC), with a 5\\% MCC improvement over single-stage baseline. This work advances voice pathology classification while reducing gender bias through hierarchical modeling of vocal characteristics.", "date": "2025-08-02", "categories": []}
{"id": "2508.01178", "title": "Advancing the Foundation Model for Music Understanding", "abstract": "The field of Music Information Retrieval (MIR) is fragmented, with specialized models excelling at isolated tasks. In this work, we challenge this paradigm by introducing a unified foundation model named MuFun for holistic music understanding. Our model features a novel architecture that jointly processes instrumental and lyrical content, and is trained on a large-scale dataset covering diverse tasks such as genre classification, music tagging, and question answering. To facilitate robust evaluation, we also propose a new benchmark for multi-faceted music understanding called MuCUE (Music Comprehensive Understanding Evaluation). Experiments show our model significantly outperforms existing audio large language models across the MuCUE tasks, demonstrating its state-of-the-art effectiveness and generalization ability.", "date": "2025-08-02", "categories": []}
{"id": "2508.01691", "title": "Voxlect: A Speech Foundation Model Benchmark for Modeling Dialects and Regional Languages Around the Globe", "abstract": "We present Voxlect, a novel benchmark for modeling dialects and regional languages worldwide using speech foundation models. Specifically, we report comprehensive benchmark evaluations on dialects and regional language varieties in English, Arabic, Mandarin and Cantonese, Tibetan, Indic languages, Thai, Spanish, French, German, Brazilian Portuguese, and Italian. Our study used over 2 million training utterances from 30 publicly available speech corpora that are provided with dialectal information. We evaluate the performance of several widely used speech foundation models in classifying speech dialects. We assess the robustness of the dialectal models under noisy conditions and present an error analysis that highlights modeling results aligned with geographic continuity. In addition to benchmarking dialect classification, we demonstrate several downstream applications enabled by Voxlect. Specifically, we show that Voxlect can be applied to augment existing speech recognition datasets with dialect information, enabling a more detailed analysis of ASR performance across dialectal variations. Voxlect is also used as a tool to evaluate the performance of speech generation systems. Voxlect is publicly available with the license of the RAIL family at: https://github.com/tiantiaf0627/voxlect.", "date": "2025-08-03", "categories": []}
{"id": "2508.02210", "title": "WhiSQA: Non-Intrusive Speech Quality Prediction Using Whisper Encoder Features", "abstract": "There has been significant research effort developing neural-network-based predictors of SQ in recent years. While a primary objective has been to develop non-intrusive, i.e.~reference-free, metrics to assess the performance of SE systems, recent work has also investigated the direct inference of neural SQ predictors within the loss function of downstream speech tasks. To aid in the training of SQ predictors, several large datasets of audio with corresponding human labels of quality have been created. Recent work in this area has shown that speech representations derived from large unsupervised or semi-supervised foundational speech models are useful input feature representations for neural SQ prediction. In this work, a novel and robust SQ predictor is proposed based on feature representations extracted from an ASR model, found to be a powerful input feature for the SQ prediction task. The proposed system achieves higher correlation with human MOS ratings than recent approaches on all NISQA test sets and shows significantly better domain adaption compared to the commonly used DNSMOS metric.", "date": "2025-08-04", "categories": []}
{"id": "2508.02391", "title": "Inference-time Scaling for Diffusion-based Audio Super-resolution", "abstract": "Diffusion models have demonstrated remarkable success in generative tasks, including audio super-resolution (SR). In many applications like movie post-production and album mastering, substantial computational budgets are available for achieving superior audio quality. However, while existing diffusion approaches typically increase sampling steps to improve quality, the performance remains fundamentally limited by the stochastic nature of the sampling process, leading to high-variance and quality-limited outputs. Here, rather than simply increasing the number of sampling steps, we propose a different paradigm through inference-time scaling for SR, which explores multiple solution trajectories during the sampling process. Different task-specific verifiers are developed, and two search algorithms, including the random search and zero-order search for SR, are introduced. By actively guiding the exploration of the high-dimensional solution space through verifier-algorithm combinations, we enable more robust and higher-quality outputs. Through extensive validation across diverse audio domains (speech, music, sound effects) and frequency ranges, we demonstrate consistent performance gains, achieving improvements of up to 9.70% in aesthetics, 5.88% in speaker similarity, 15.20% in word error rate, and 46.98% in spectral distance for speech SR from 4kHz to 24kHz, showcasing the effectiveness of our approach. Audio samples are available at: https://racerk.github.io/tt-scale-audiosr/.", "date": "2025-08-04", "categories": []}
{"id": "2508.02643", "title": "CAK: Emergent Audio Effects from Minimal Deep Learning", "abstract": "We demonstrate that a single 3x3 convolutional kernel can produce emergent audio effects when trained on 200 samples from a personalized corpus. We achieve this through two key techniques: (1) Conditioning Aware Kernels (CAK), where output = input + (learned_pattern x control), with a soft-gate mechanism supporting identity preservation at zero control; and (2) AuGAN (Audit GAN), which reframes adversarial training from \"is this real?\" to \"did you apply the requested value?\" Rather than learning to generate or detect forgeries, our networks cooperate to verify control application, discovering unique transformations. The learned kernel exhibits a diagonal structure creating frequency-dependent temporal shifts that are capable of producing musical effects based on input characteristics. Our results show the potential of adversarial training to discover audio transformations from minimal data, enabling new approaches to effect design.", "date": "2025-08-04", "categories": []}
{"id": "2507.23091", "title": "Moravec's Paradox: Towards an Auditory Turing Test", "abstract": "This research work demonstrates that current AI systems fail catastrophically on auditory tasks that humans perform effortlessly. Drawing inspiration from Moravec's paradox (i.e., tasks simple for humans often prove difficult for machines, and vice versa), we introduce an auditory Turing test comprising 917 challenges across seven categories: overlapping speech, speech in noise, temporal distortion, spatial audio, coffee-shop noise, phone distortion, and perceptual illusions. Our evaluation of state-of-the-art audio models including GPT-4's audio capabilities and OpenAI's Whisper reveals a striking failure rate exceeding 93%, with even the best-performing model achieving only 6.9% accuracy on tasks that humans solved at 7.5 times higher success (52%). These results expose focusing failures in how AI systems process complex auditory scenes, particularly in selective attention, noise robustness, and contextual adaptation. Our benchmark not only quantifies the human-machine auditory gap but also provides insights into why these failures occur, suggesting that current architectures lack fundamental mechanisms for human-like auditory scene analysis. The traditional design of audio CAPTCHAs highlights common filters that humans evolved but machines fail to select in multimodal language models. This work establishes a diagnostic framework for measuring progress toward human-level machine listening and highlights the need for novel approaches integrating selective attention, physics-based audio understanding, and context-aware perception into multimodal AI systems.", "date": "2025-07-30", "categories": []}
{"id": "2502.20040", "title": "CleanMel: Mel-Spectrogram Enhancement for Improving Both Speech Quality and ASR", "abstract": "In this work, we propose CleanMel, a single-channel Mel-spectrogram denoising and dereverberation network for improving both speech quality and automatic speech recognition (ASR) performance. The proposed network takes as input the noisy and reverberant microphone recording and predicts the corresponding clean Mel-spectrogram. The enhanced Mel-spectrogram can be either transformed to the speech waveform with a neural vocoder or directly used for ASR. The proposed network is composed of interleaved cross-band and narrow-band processing in the Mel-frequency domain, for learning the full-band spectral pattern and the narrow-band properties of signals, respectively. Compared to linear-frequency domain or time-domain speech enhancement, the key advantage of Mel-spectrogram enhancement is that Mel-frequency presents speech in a more compact way and thus is easier to learn, which will benefit both speech quality and ASR. Experimental results on five English and one Chinese datasets demonstrate a significant improvement in both speech quality and ASR performance achieved by the proposed model.Code and audio examples of our model are available online.", "date": "2025-02-27", "categories": []}
{"id": "2503.11026", "title": "MAVFlow: Preserving Paralinguistic Elements with Conditional Flow Matching for Zero-Shot AV2AV Multilingual Translation", "abstract": "Despite recent advances in text-to-speech (TTS) models, audio-visual-to-audio-visual (AV2AV) translation still faces a critical challenge: maintaining speaker consistency between the original and translated vocal and facial features. To address this issue, we propose a conditional flow matching (CFM) zero-shot audio-visual renderer that utilizes strong dual guidance from both audio and visual modalities. By leveraging multimodal guidance with CFM, our model robustly preserves speaker-specific characteristics and enhances zero-shot AV2AV translation abilities. For the audio modality, we enhance the CFM process by integrating robust speaker embeddings with x-vectors, which serve to bolster speaker consistency. Additionally, we convey emotional nuances to the face rendering module. The guidance provided by both audio and visual cues remains independent of semantic or linguistic content, allowing our renderer to effectively handle zero-shot translation tasks for monolingual speakers in different languages. We empirically demonstrate that the inclusion of high-quality mel-spectrograms conditioned on facial information not only enhances the quality of the synthesized speech but also positively influences facial generation, leading to overall performance improvements in LSE and FID score. Our code is available at https://github.com/Peter-SungwooCho/MAVFlow.", "date": "2025-03-14", "categories": []}
{"id": "2507.22208", "title": "Quantum-Inspired Audio Unlearning: Towards Privacy-Preserving Voice Biometrics", "abstract": "The widespread adoption of voice-enabled authentication and audio biometric systems have significantly increased privacy vulnerabilities associated with sensitive speech data. Compliance with privacy regulations such as GDPR's right to be forgotten and India's DPDP Act necessitates targeted and efficient erasure of individual-specific voice signatures from already-trained biometric models. Existing unlearning methods designed for visual data inadequately handle the sequential, temporal, and high-dimensional nature of audio signals, leading to ineffective or incomplete speaker and accent erasure. To address this, we introduce QPAudioEraser, a quantum-inspired audio unlearning framework. Our our-phase approach involves: (1) weight initialization using destructive interference to nullify target features, (2) superposition-based label transformations that obscure class identity, (3) an uncertainty-maximizing quantum loss function, and (4) entanglement-inspired mixing of correlated weights to retain model knowledge. Comprehensive evaluations with ResNet18, ViT, and CNN architectures across AudioMNIST, Speech Commands, LibriSpeech, and Speech Accent Archive datasets validate QPAudioEraser's superior performance. The framework achieves complete erasure of target data (0% Forget Accuracy) while incurring minimal impact on model utility, with a performance degradation on retained data as low as 0.05%. QPAudioEraser consistently surpasses conventional baselines across single-class, multi-class, sequential, and accent-level erasure scenarios, establishing the proposed approach as a robust privacy-preserving solution.", "date": "2025-07-29", "categories": []}
{"id": "2507.08128", "title": "Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models", "abstract": "We present Audio Flamingo 3 (AF3), a fully open state-of-the-art (SOTA) large audio-language model that advances reasoning and understanding across speech, sound, and music. AF3 introduces: (i) AF-Whisper, a unified audio encoder trained using a novel strategy for joint representation learning across all 3 modalities of speech, sound, and music; (ii) flexible, on-demand thinking, allowing the model to do chain-of-thought-type reasoning before answering; (iii) multi-turn, multi-audio chat; (iv) long audio understanding and reasoning (including speech) up to 10 minutes; and (v) voice-to-voice interaction. To enable these capabilities, we propose several large-scale training datasets curated using novel strategies, including AudioSkills-XL, LongAudio-XL, AF-Think, and AF-Chat, and train AF3 with a novel five-stage curriculum-based training strategy. Trained on only open-source audio data, AF3 achieves new SOTA results on over 20+ (long) audio understanding and reasoning benchmarks, surpassing both open-weight and closed-source models trained on much larger datasets.", "date": "2025-07-10", "categories": []}
{"id": "2507.21331", "title": "A Deep Learning Automatic Speech Recognition Model for Shona Language", "abstract": "This study presented the development of a deep learning-based Automatic Speech Recognition system for Shona, a low-resource language characterized by unique tonal and grammatical complexities. The research aimed to address the challenges posed by limited training data, lack of labelled data, and the intricate tonal nuances present in Shona speech, with the objective of achieving significant improvements in recognition accuracy compared to traditional statistical models. The research first explored the feasibility of using deep learning to develop an accurate ASR system for Shona. Second, it investigated the specific challenges involved in designing and implementing deep learning architectures for Shona speech recognition and proposed strategies to mitigate these challenges. Lastly, it compared the performance of the deep learning-based model with existing statistical models in terms of accuracy. The developed ASR system utilized a hybrid architecture consisting of a Convolutional Neural Network for acoustic modelling and a Long Short-Term Memory network for language modelling. To overcome the scarcity of data, data augmentation techniques and transfer learning were employed. Attention mechanisms were also incorporated to accommodate the tonal nature of Shona speech. The resulting ASR system achieved impressive results, with a Word Error Rate of 29%, Phoneme Error Rate of 12%, and an overall accuracy of 74%. These metrics indicated the potential of deep learning to enhance ASR accuracy for under-resourced languages like Shona. This study contributed to the advancement of ASR technology for under-resourced languages like Shona, ultimately fostering improved accessibility and communication for Shona speakers worldwide.", "date": "2025-07-28", "categories": []}
{"id": "2211.15406", "title": "Automated Detection of Dolphin Whistles with Convolutional Networks and Transfer Learning", "abstract": "Effective conservation of maritime environments and wildlife management of endangered species require the implementation of efficient, accurate and scalable solutions for environmental monitoring. Ecoacoustics offers the advantages of non-invasive, long-duration sampling of environmental sounds and has the potential to become the reference tool for biodiversity surveying. However, the analysis and interpretation of acoustic data is a time-consuming process that often requires a great amount of human supervision. This issue might be tackled by exploiting modern techniques for automatic audio signal analysis, which have recently achieved impressive performance thanks to the advances in deep learning research. In this paper we show that convolutional neural networks can indeed significantly outperform traditional automatic methods in a challenging detection task: identification of dolphin whistles from underwater audio recordings. The proposed system can detect signals even in the presence of ambient noise, at the same time consistently reducing the likelihood of producing false positives and false negatives. Our results further support the adoption of artificial intelligence technology to improve the automatic monitoring of marine ecosystems.", "date": "2022-11-28", "categories": []}
{"id": "2412.05167", "title": "Benchmarking Open-ended Audio Dialogue Understanding for Large Audio-Language Models", "abstract": "Large Audio-Language Models (LALMs), such as GPT-4o, have recently unlocked audio dialogue capabilities, enabling direct spoken exchanges with humans. The potential of LALMs broadens their applicability across a wide range of practical scenarios supported by audio dialogues. However, given these advancements, a comprehensive benchmark to evaluate the performance of LALMs in the open-ended audio dialogue understanding remains absent currently. To address this gap, we propose an Audio Dialogue Understanding Benchmark (ADU-Bench), which consists of 4 benchmark datasets. They assess the open-ended audio dialogue ability for LALMs in 3 general scenarios, 12 skills, 9 multilingual languages, and 4 categories of ambiguity handling. Notably, we firstly propose the evaluation of ambiguity handling in audio dialogues that expresses different intentions beyond the same literal meaning of sentences, e.g., \"Really!?\" with different intonations. In summary, ADU-Bench includes over 20,000 open-ended audio dialogues for the assessment of LALMs. Through extensive experiments on 16 LALMs, our analysis reveals that existing LALMs struggle with mathematical symbols and formulas, understanding human behavior such as roleplay, comprehending multiple languages, and handling audio dialogue ambiguities from different phonetic elements, such as intonations, pause positions, and homophones. The benchmark is available at https://adu-bench.github.io/.", "date": "2024-12-06", "categories": []}
{"id": "2505.20788", "title": "Enhancing Wearable Tap Water Audio Detection through Subclass Annotation in the HD-Epic Dataset", "abstract": "Wearable human activity recognition has been shown to benefit from the inclusion of acoustic data, as the sounds around a person often contain valuable context. However, due to privacy concerns, it is usually not ethically feasible to record and save microphone data from the device, since the audio could, for instance, also contain private conversations. Rather, the data should be processed locally, which in turn requires processing power and consumes energy on the wearable device. One special use case of contextual information that can be utilized to augment special tasks in human activity recognition is water flow detection, which can, e.g., be used to aid wearable hand washing detection. We created a new label called tap water for the recently released HD-Epic data set, creating 717 hand-labeled annotations of tap water flow, based on existing annotations of the water class. We analyzed the relation of tap water and water in the dataset and additionally trained and evaluated two lightweight classifiers to evaluate the newly added label class, showing that the new class can be learned more easily.", "date": "2025-05-27", "categories": []}
{"id": "2506.06732", "title": "Neural Spectral Band Generation for Audio Coding", "abstract": "Spectral band replication (SBR) enables bit-efficient coding by generating high-frequency bands from the low-frequency ones. However, it only utilizes coarse spectral features upon a subband-wise signal replication, limiting adaptability to diverse acoustic signals. In this paper, we explore the efficacy of a deep neural network (DNN)-based generative approach for coding the high-frequency bands, which we call neural spectral band generation (n-SBG). Specifically, we propose a DNN-based encoder-decoder structure to extract and quantize the side information related to the high-frequency components and generate the components given both the side information and the decoded core-band signals. The whole coding pipeline is optimized with generative adversarial criteria to enable the generation of perceptually plausible sound. From experiments using AAC as the core codec, we show that the proposed method achieves a better perceptual quality than HE-AAC-v1 with much less side information.", "date": "2025-06-07", "categories": []}
{"id": "2507.20036", "title": "Improving Audio Classification by Transitioning from Zero- to Few-Shot", "abstract": "State-of-the-art audio classification often employs a zero-shot approach, which involves comparing audio embeddings with embeddings from text describing the respective audio class. These embeddings are usually generated by neural networks trained through contrastive learning to align audio and text representations. Identifying the optimal text description for an audio class is challenging, particularly when the class comprises a wide variety of sounds. This paper examines few-shot methods designed to improve classification accuracy beyond the zero-shot approach. Specifically, audio embeddings are grouped by class and processed to replace the inherently noisy text embeddings. Our results demonstrate that few-shot classification typically outperforms the zero-shot baseline.", "date": "2025-07-26", "categories": []}
{"id": "2507.20052", "title": "Improving Deep Learning-based Respiratory Sound Analysis with Frequency Selection and Attention Mechanism", "abstract": "Accurate classification of respiratory sounds requires deep learning models that effectively capture fine-grained acoustic features and long-range temporal dependencies. Convolutional Neural Networks (CNNs) are well-suited for extracting local time-frequency patterns but are limited in modeling global context. In contrast, transformer-based models can capture long-range dependencies, albeit with higher computational demands. To address these limitations, we propose a compact CNN-Temporal Self-Attention (CNN-TSA) network that integrates lightweight self-attention into an efficient CNN backbone. Central to our approach is a Frequency Band Selection (FBS) module that suppresses noisy and non-informative frequency regions, substantially improving accuracy and reducing FLOPs by up to 50%. We also introduce age-specific models to enhance robustness across diverse patient groups. Evaluated on the SPRSound-2022/2023 and ICBHI-2017 lung sound datasets, CNN-TSA with FBS sets new benchmarks on SPRSound and achieves state-of-the-art performance on ICBHI, all with a significantly smaller computational footprint. Furthermore, integrating FBS into an existing transformer baseline yields a new record on ICBHI, confirming FBS as an effective drop-in enhancement. These results demonstrate that our framework enables reliable, real-time respiratory sound analysis suitable for deployment in resource-constrained settings.", "date": "2025-07-26", "categories": []}
{"id": "2507.20666", "title": "MIMII-Agent: Leveraging LLMs with Function Calling for Relative Evaluation of Anomalous Sound Detection", "abstract": "This paper proposes a method for generating machine-type-specific anomalies to evaluate the relative performance of unsupervised anomalous sound detection (UASD) systems across different machine types, even in the absence of real anomaly sound data. Conventional keyword-based data augmentation methods often produce unrealistic sounds due to their reliance on manually defined labels, limiting scalability as machine types and anomaly patterns diversify. Advanced audio generative models, such as MIMII-Gen, show promise but typically depend on anomalous training data, making them less effective when diverse anomalous examples are unavailable. To address these limitations, we propose a novel synthesis approach leveraging large language models (LLMs) to interpret textual descriptions of faults and automatically select audio transformation functions, converting normal machine sounds into diverse and plausible anomalous sounds. We validate this approach by evaluating a UASD system trained only on normal sounds from five machine types, using both real and synthetic anomaly data. Experimental results reveal consistent trends in relative detection difficulty across machine types between synthetic and real anomalies. This finding supports our hypothesis and highlights the effectiveness of the proposed LLM-based synthesis approach for relative evaluation of UASD systems.", "date": "2025-07-28", "categories": []}
{"id": "2407.15423", "title": "Integrating IP Broadcasting with Audio Tags: Workflow and Challenges", "abstract": "The broadcasting industry has adopted IP technologies, revolutionising both live and pre-recorded content production, from news gathering to live music events. IP broadcasting allows for the transport of audio and video signals in an easily configurable way, aligning with modern networking techniques. This shift towards an IP workflow allows for much greater flexibility, not only in routing signals but with the integration of tools using standard web development techniques. One possible tool could include the use of live audio tagging, which has a number of uses in the production of content. These could include adding sound effects to automated closed captioning or identifying unwanted sound events within a scene. In this paper, we describe the process of containerising an audio tagging model into a microservice, a small segregated code module that can be integrated into a multitude of different network setups. The goal is to develop a modular, accessible, and flexible tool capable of seamless deployment into broadcasting workflows of all sizes, from small productions to large corporations. Challenges surrounding latency of the selected audio tagging model and its effect on the usefulness of the end product are discussed.", "date": "2024-07-22", "categories": []}
{"id": "2410.21478", "title": "Knowledge Distillation for Real-Time Classification of Early Media in Voice Communications", "abstract": "This paper investigates the industrial setting of real-time classification of early media exchanged during the initialization phase of voice calls. We explore the application of state-of-the-art audio tagging models and highlight some limitations when applied to the classification of early media. While most existing approaches leverage convolutional neural networks, we propose a novel approach for low-resource requirements based on gradient-boosted trees. Our approach not only demonstrates a substantial improvement in runtime performance, but also exhibits a comparable accuracy. We show that leveraging knowledge distillation and class aggregation techniques to train a simpler and smaller model accelerates the classification of early media in voice calls. We provide a detailed analysis of the results on a proprietary and publicly available dataset, regarding accuracy and runtime performance. We additionally report a case study of the achieved performance improvements at a regional data center in India.", "date": "2024-10-28", "categories": []}
{"id": "2507.18897", "title": "HH-Codec: High Compression High-fidelity Discrete Neural Codec for Spoken Language Modeling", "abstract": "Discrete speech tokenization is a fundamental component in speech codecs. However, in large-scale speech-to-speech systems, the complexity of parallel streams from multiple quantizers and the computational cost of high-time-dimensional codecs pose significant challenges. In this paper, we introduce HH-Codec, a neural codec that achieves extreme compression at 24 tokens per second for 24 kHz audio while relying on single-quantizer inference. Our approach involves a carefully designed Vector Quantization space for Spoken Language Modeling, optimizing compression efficiency while minimizing information loss. Building on this, we propose an asymmetric encoder-decoder architecture (Audio-VQ-Mel-Audio) that leverages dual supervision and progressive training to enhance reconstruction stability and fidelity. HH-Codec achieves state-of-the-art performance in speech reconstruction with an ultra-low bandwidth of 0.3 kbps. We further evaluate its effectiveness in codebook utilization and generative model adaptation, with extensive ablations validating the necessity of each module. HH-Codec is available at https://github.com/opendilab/HH-Codec.", "date": "2025-07-25", "categories": []}
{"id": "2507.19202", "title": "Latent Granular Resynthesis using Neural Audio Codecs", "abstract": "We introduce a novel technique for creative audio resynthesis that operates by reworking the concept of granular synthesis at the latent vector level. Our approach creates a \"granular codebook\" by encoding a source audio corpus into latent vector segments, then matches each latent grain of a target audio signal to its closest counterpart in the codebook. The resulting hybrid sequence is decoded to produce audio that preserves the target's temporal structure while adopting the source's timbral characteristics. This technique requires no model training, works with diverse audio materials, and naturally avoids the discontinuities typical of traditional concatenative synthesis through the codec's implicit interpolation during decoding. We include supplementary material at https://github.com/naotokui/latentgranular/ , as well as a proof-of-concept implementation to allow users to experiment with their own sounds at https://huggingface.co/spaces/naotokui/latentgranular .", "date": "2025-07-25", "categories": []}
{"id": "2505.23782", "title": "4,500 Seconds: Small Data Training Approaches for Deep UAV Audio Classification", "abstract": "Unmanned aerial vehicle (UAV) usage is expected to surge in the coming decade, raising the need for heightened security measures to prevent airspace violations and security threats. This study investigates deep learning approaches to UAV classification focusing on the key issue of data scarcity. To investigate this we opted to train the models using a total of 4,500 seconds of audio samples, evenly distributed across a 9-class dataset. We leveraged parameter efficient fine-tuning (PEFT) and data augmentations to mitigate the data scarcity. This paper implements and compares the use of convolutional neural networks (CNNs) and attention-based transformers. Our results show that, CNNs outperform transformers by 1-2\\% accuracy, while still being more computationally efficient. These early findings, however, point to potential in using transformers models; suggesting that with more data and further optimizations they could outperform CNNs. Future works aims to upscale the dataset to better understand the trade-offs between these approaches.", "date": "2025-05-21", "categories": []}
{"id": "2507.18334", "title": "Improving Bird Classification with Primary Color Additives", "abstract": "We address the problem of classifying bird species using their song recordings, a challenging task due to environmental noise, overlapping vocalizations, and missing labels. Existing models struggle with low-SNR or multi-species recordings. We hypothesize that birds can be classified by visualizing their pitch pattern, speed, and repetition, collectively called motifs. Deep learning models applied to spectrogram images help, but similar motifs across species cause confusion. To mitigate this, we embed frequency information into spectrograms using primary color additives. This enhances species distinction and improves classification accuracy. Our experiments show that the proposed approach achieves statistically significant gains over models without colorization and surpasses the BirdCLEF 2024 winner, improving F1 by 7.3%, ROC-AUC by 6.2%, and CMAP by 6.6%. These results demonstrate the effectiveness of incorporating frequency information via colorization.", "date": "2025-07-24", "categories": []}
{"id": "2507.16832", "title": "Does Language Matter for Early Detection of Parkinson's Disease from Speech?", "abstract": "Using speech samples as a biomarker is a promising avenue for detecting and monitoring the progression of Parkinson's disease (PD), but there is considerable disagreement in the literature about how best to collect and analyze such data. Early research in detecting PD from speech used a sustained vowel phonation (SVP) task, while some recent research has explored recordings of more cognitively demanding tasks. To assess the role of language in PD detection, we tested pretrained models with varying data types and pretraining objectives and found that (1) text-only models match the performance of vocal-feature models, (2) multilingual Whisper outperforms self-supervised models whereas monolingual Whisper does worse, and (3) AudioSet pretraining improves performance on SVP but not spontaneous speech. These findings together highlight the critical role of language for the early detection of Parkinson's disease.", "date": "2025-07-14", "categories": []}
{"id": "2507.17297", "title": "On Temporal Guidance and Iterative Refinement in Audio Source Separation", "abstract": "Spatial semantic segmentation of sound scenes (S5) involves the accurate identification of active sound classes and the precise separation of their sources from complex acoustic mixtures. Conventional systems rely on a two-stage pipeline - audio tagging followed by label-conditioned source separation - but are often constrained by the absence of fine-grained temporal information critical for effective separation. In this work, we address this limitation by introducing a novel approach for S5 that enhances the synergy between the event detection and source separation stages. Our key contributions are threefold. First, we fine-tune a pre-trained Transformer to detect active sound classes. Second, we utilize a separate instance of this fine-tuned Transformer to perform sound event detection (SED), providing the separation module with detailed, time-varying guidance. Third, we implement an iterative refinement mechanism that progressively enhances separation quality by recursively reusing the separator's output from previous iterations. These advancements lead to significant improvements in both audio tagging and source separation performance, as demonstrated by our system's second-place finish in Task 4 of the DCASE Challenge 2025. Our implementation and model checkpoints are available in our GitHub repository: https://github.com/theMoro/dcase25task4 .", "date": "2025-07-23", "categories": []}
{"id": "2507.14215", "title": "Developing an AI-Guided Assistant Device for the Deaf and Hearing Impaired", "abstract": "This study aims to develop a deep learning system for an accessibility device for the deaf or hearing impaired. The device will accurately localize and identify sound sources in real time. This study will fill an important gap in current research by leveraging machine learning techniques to target the underprivileged community. The system includes three main components. 1. JerryNet: A custom designed CNN architecture that determines the direction of arrival (DoA) for nine possible directions. 2. Audio Classification: This model is based on fine-tuning the Contrastive Language-Audio Pretraining (CLAP) model to identify the exact sound classes only based on audio. 3. Multimodal integration model: This is an accurate sound localization model that combines audio, visual, and text data to locate the exact sound sources in the images. The part consists of two modules, one object detection using Yolov9 to generate all the bounding boxes of the objects, and an audio visual localization model to identify the optimal bounding box using complete Intersection over Union (CIoU). The hardware consists of a four-microphone rectangular formation and a camera mounted on glasses with a wristband for displaying necessary information like direction. On a custom collected data set, JerryNet achieved a precision of 91. 1% for the sound direction, outperforming all the baseline models. The CLAP model achieved 98.5% and 95% accuracy on custom and AudioSet datasets, respectively. The audio-visual localization model within component 3 yielded a cIoU of 0.892 and an AUC of 0.658, surpassing other similar models. There are many future potentials to this study, paving the way to creating a new generation of accessibility devices.", "date": "2025-07-16", "categories": []}
{"id": "2507.15523", "title": "An Investigation of Test-time Adaptation for Audio Classification under Background Noise", "abstract": "Domain shift is a prominent problem in Deep Learning, causing a model pre-trained on a source dataset to suffer significant performance degradation on test datasets. This research aims to address the issue of audio classification under domain shift caused by background noise using Test-Time Adaptation (TTA), a technique that adapts a pre-trained model during testing using only unlabelled test data before making predictions. We adopt two common TTA methods, TTT and TENT, and a state-of-the-art method CoNMix, and investigate their respective performance on two popular audio classification datasets, AudioMNIST (AM) and SpeechCommands V1 (SC), against different types of background noise and noise severity levels. The experimental results reveal that our proposed modified version of CoNMix produced the highest classification accuracy under domain shift (5.31% error rate under 10 dB exercise bike background noise and 12.75% error rate under 3 dB running tap background noise for AM) compared to TTT and TENT. The literature search provided no evidence of similar works, thereby motivating the work reported here as the first study to leverage TTA techniques for audio classification under domain shift.", "date": "2025-07-21", "categories": []}
{"id": "2302.09409", "title": "LOCUS: LOcalization with Channel Uncertainty and Sporadic Energy", "abstract": "Accurate sound source localization (SSL), such as direction-of-arrival (DoA) estimation, relies on consistent multichannel data. However, batteryless systems often suffer from missing data due to the stochastic nature of energy harvesting, degrading localization performance. We propose LOCUS, a deep learning framework that recovers corrupted features in such settings. LOCUS integrates three modules: (1) Information-Weighted Focus (InFo) to identify corrupted regions, (2) Latent Feature Synthesizer (LaFS) to reconstruct missing features, and (3) Guided Replacement (GRep) to restore data without altering valid inputs. LOCUS significantly improves DoA accuracy under missing-channel conditions, achieving up to 36.91% error reduction on DCASE and LargeSet, and 25.87-59.46% gains in real-world deployments. We release a 50-hour multichannel dataset to support future research on localization under energy constraints. Our code and data are available at: https://bashlab.github.io/locus_project/", "date": "2023-02-18", "categories": []}
{"id": "2405.18386", "title": "Instruct-MusicGen: Unlocking Text-to-Music Editing for Music Language Models via Instruction Tuning", "abstract": "Recent advances in text-to-music editing, which employ text queries to modify music (e.g.\\ by changing its style or adjusting instrumental components), present unique challenges and opportunities for AI-assisted music creation. Previous approaches in this domain have been constrained by the necessity to train specific editing models from scratch, which is both resource-intensive and inefficient; other research uses large language models to predict edited music, resulting in imprecise audio reconstruction. To Combine the strengths and address these limitations, we introduce Instruct-MusicGen, a novel approach that finetunes a pretrained MusicGen model to efficiently follow editing instructions such as adding, removing, or separating stems. Our approach involves a modification of the original MusicGen architecture by incorporating a text fusion module and an audio fusion module, which allow the model to process instruction texts and audio inputs concurrently and yield the desired edited music. Remarkably, Instruct-MusicGen only introduces 8% new parameters to the original MusicGen model and only trains for 5K steps, yet it achieves superior performance across all tasks compared to existing baselines, and demonstrates performance comparable to the models trained for specific tasks. This advancement not only enhances the efficiency of text-to-music editing but also broadens the applicability of music language models in dynamic music production environments.", "date": "2024-05-28", "categories": []}
{"id": "2505.16195", "title": "SpecMaskFoley: Steering Pretrained Spectral Masked Generative Transformer Toward Synchronized Video-to-audio Synthesis via ControlNet", "abstract": "Foley synthesis aims to synthesize high-quality audio that is both semantically and temporally aligned with video frames. Given its broad application in creative industries, the task has gained increasing attention in the research community. To avoid the non-trivial task of training audio generative models from scratch, adapting pretrained audio generative models for video-synchronized foley synthesis presents an attractive direction. ControlNet, a method for adding fine-grained controls to pretrained generative models, has been applied to foley synthesis, but its use has been limited to handcrafted human-readable temporal conditions. In contrast, from-scratch models achieved success by leveraging high-dimensional deep features extracted using pretrained video encoders. We have observed a performance gap between ControlNet-based and from-scratch foley models. To narrow this gap, we propose SpecMaskFoley, a method that steers the pretrained SpecMaskGIT model toward video-synchronized foley synthesis via ControlNet. To unlock the potential of a single ControlNet branch, we resolve the discrepancy between the temporal video features and the time-frequency nature of the pretrained SpecMaskGIT via a frequency-aware temporal feature aligner, eliminating the need for complicated conditioning mechanisms widely used in prior arts. Evaluations on a common foley synthesis benchmark demonstrate that SpecMaskFoley could even outperform strong from-scratch baselines, substantially advancing the development of ControlNet-based foley synthesis models. Demo page: https://zzaudio.github.io/SpecMaskFoley_Demo/", "date": "2025-05-22", "categories": []}
{"id": "2507.13727", "title": "Adversarial Training Improves Generalization Under Distribution Shifts in Bioacoustics", "abstract": "Adversarial training is a promising strategy for enhancing model robustness against adversarial attacks. However, its impact on generalization under substantial data distribution shifts in audio classification remains largely unexplored. To address this gap, this work investigates how different adversarial training strategies improve generalization performance and adversarial robustness in audio classification. The study focuses on two model architectures: a conventional convolutional neural network (ConvNeXt) and an inherently interpretable prototype-based model (AudioProtoPNet). The approach is evaluated using a challenging bird sound classification benchmark. This benchmark is characterized by pronounced distribution shifts between training and test data due to varying environmental conditions and recording methods, a common real-world challenge. The investigation explores two adversarial training strategies: one based on output-space attacks that maximize the classification loss function, and another based on embedding-space attacks designed to maximize embedding dissimilarity. These attack types are also used for robustness evaluation. Additionally, for AudioProtoPNet, the study assesses the stability of its learned prototypes under targeted embedding-space attacks. Results show that adversarial training, particularly using output-space attacks, improves clean test data performance by an average of 10.5% relative and simultaneously strengthens the adversarial robustness of the models. These findings, although derived from the bird sound domain, suggest that adversarial training holds potential to enhance robustness against both strong distribution shifts and adversarial attacks in challenging audio classification settings.", "date": "2025-07-18", "categories": []}
{"id": "2507.13170", "title": "SHIELD: A Secure and Highly Enhanced Integrated Learning for Robust Deepfake Detection against Adversarial Attacks", "abstract": "Audio plays a crucial role in applications like speaker verification, voice-enabled smart devices, and audio conferencing. However, audio manipulations, such as deepfakes, pose significant risks by enabling the spread of misinformation. Our empirical analysis reveals that existing methods for detecting deepfake audio are often vulnerable to anti-forensic (AF) attacks, particularly those attacked using generative adversarial networks. In this article, we propose a novel collaborative learning method called SHIELD to defend against generative AF attacks. To expose AF signatures, we integrate an auxiliary generative model, called the defense (DF) generative model, which facilitates collaborative learning by combining input and output. Furthermore, we design a triplet model to capture correlations for real and AF attacked audios with real-generated and attacked-generated audios using auxiliary generative models. The proposed SHIELD strengthens the defense against generative AF attacks and achieves robust performance across various generative models. The proposed AF significantly reduces the average detection accuracy from 95.49% to 59.77% for ASVspoof2019, from 99.44% to 38.45% for In-the-Wild, and from 98.41% to 51.18% for HalfTruth for three different generative models. The proposed SHIELD mechanism is robust against AF attacks and achieves an average accuracy of 98.13%, 98.58%, and 99.57% in match, and 98.78%, 98.62%, and 98.85% in mismatch settings for the ASVspoof2019, In-the-Wild, and HalfTruth datasets, respectively.", "date": "2025-07-17", "categories": []}
{"id": "2302.00646", "title": "Epic-Sounds: A Large-scale Dataset of Actions That Sound", "abstract": "We introduce EPIC-SOUNDS, a large-scale dataset of audio annotations capturing temporal extents and class labels within the audio stream of the egocentric videos. We propose an annotation pipeline where annotators temporally label distinguishable audio segments and describe the action that could have caused this sound. We identify actions that can be discriminated purely from audio, through grouping these free-form descriptions of audio into classes. For actions that involve objects colliding, we collect human annotations of the materials of these objects (e.g. a glass object being placed on a wooden surface), which we verify from video, discarding ambiguities. Overall, EPIC-SOUNDS includes 78.4k categorised segments of audible events and actions, distributed across 44 classes as well as 39.2k non-categorised segments. We train and evaluate state-of-the-art audio recognition and detection models on our dataset, for both audio-only and audio-visual methods. We also conduct analysis on: the temporal overlap between audio events, the temporal and label correlations between audio and visual modalities, the ambiguities in annotating materials from audio-only input, the importance of audio-only labels and the limitations of current models to understand actions that sound.", "date": "2023-02-01", "categories": []}
{"id": "2507.11636", "title": "JSQA: Speech Quality Assessment with Perceptually-Inspired Contrastive Pretraining Based on JND Audio Pairs", "abstract": "Speech quality assessment (SQA) is often used to learn a mapping from a high-dimensional input space to a scalar that represents the mean opinion score (MOS) of the perceptual speech quality. Learning such a mapping is challenging for many reasons, but largely because MOS exhibits high levels of inherent variance due to perceptual and experimental-design differences. Many solutions have been proposed, but many approaches do not properly incorporate perceptual factors into their learning algorithms (beyond the MOS label), which could lead to unsatisfactory results. To this end, we propose JSQA, a two-stage framework that pretrains an audio encoder using perceptually-guided contrastive learning on just noticeable difference (JND) pairs, followed by fine-tuning for MOS prediction. We first generate pairs of audio data within JND levels, which are then used to pretrain an encoder to leverage perceptual quality similarity information and map it into an embedding space. The JND pairs come from clean LibriSpeech utterances that are mixed with background noise from CHiME-3, at different signal-to-noise ratios (SNRs). The encoder is later fine-tuned with audio samples from the NISQA dataset for MOS prediction. Experimental results suggest that perceptually-inspired contrastive pretraining significantly improves the model performance evaluated by various metrics when compared against the same network trained from scratch without pretraining. These findings suggest that incorporating perceptual factors into pretraining greatly contributes to the improvement in performance for SQA.", "date": "2025-07-15", "categories": []}
{"id": "2507.12042", "title": "Stereo Sound Event Localization and Detection with Onscreen/offscreen Classification", "abstract": "This paper presents the objective, dataset, baseline, and metrics of Task 3 of the DCASE2025 Challenge on sound event localization and detection (SELD). In previous editions, the challenge used four-channel audio formats of first-order Ambisonics (FOA) and microphone array. In contrast, this year's challenge investigates SELD with stereo audio data (termed stereo SELD). This change shifts the focus from more specialized 360{\\deg} audio and audiovisual scene analysis to more commonplace audio and media scenarios with limited field-of-view (FOV). Due to inherent angular ambiguities in stereo audio data, the task focuses on direction-of-arrival (DOA) estimation in the azimuth plane (left-right axis) along with distance estimation. The challenge remains divided into two tracks: audio-only and audiovisual, with the audiovisual track introducing a new sub-task of onscreen/offscreen event classification necessitated by the limited FOV. This challenge introduces the DCASE2025 Task3 Stereo SELD Dataset, whose stereo audio and perspective video clips are sampled and converted from the STARSS23 recordings. The baseline system is designed to process stereo audio and corresponding video frames as inputs. In addition to the typical SELD event classification and localization, it integrates onscreen/offscreen classification for the audiovisual track. The evaluation metrics have been modified to introduce an onscreen/offscreen accuracy metric, which assesses the models' ability to identify which sound sources are onscreen. In the experimental evaluation, the baseline system performs reasonably well with the stereo audio data.", "date": "2025-07-16", "categories": []}
{"id": "2507.12175", "title": "RUMAA: Repeat-Aware Unified Music Audio Analysis for Score-Performance Alignment, Transcription, and Mistake Detection", "abstract": "This study introduces RUMAA, a transformer-based framework for music performance analysis that unifies score-to-performance alignment, score-informed transcription, and mistake detection in a near end-to-end manner. Unlike prior methods addressing these tasks separately, RUMAA integrates them using pre-trained score and audio encoders and a novel tri-stream decoder capturing task interdependencies through proxy tasks. It aligns human-readable MusicXML scores with repeat symbols to full-length performance audio, overcoming traditional MIDI-based methods that rely on manually unfolded score-MIDI data with pre-specified repeat structures. RUMAA matches state-of-the-art alignment methods on non-repeated scores and outperforms them on scores with repeats in a public piano music dataset, while also delivering promising transcription and mistake detection results.", "date": "2025-07-16", "categories": []}
{"id": "2507.11096", "title": "EditGen: Harnessing Cross-Attention Control for Instruction-Based Auto-Regressive Audio Editing", "abstract": "In this study, we investigate leveraging cross-attention control for efficient audio editing within auto-regressive models. Inspired by image editing methodologies, we develop a Prompt-to-Prompt-like approach that guides edits through cross and self-attention mechanisms. Integrating a diffusion-based strategy, influenced by Auffusion, we extend the model's functionality to support refinement edits, establishing a baseline for prompt-guided audio editing. Additionally, we introduce an alternative approach by incorporating MUSICGEN, a pre-trained frozen auto-regressive model, and propose three editing mechanisms, based on Replacement, Reweighting, and Refinement of the attention scores. We employ commonly-used music-specific evaluation metrics and a human study, to gauge time-varying controllability, adherence to global text cues, and overall audio realism. The automatic and human evaluations indicate that the proposed combination of prompt-to-prompt guidance with autoregressive generation models significantly outperforms the diffusion-based baseline in terms of melody, dynamics, and tempo of the generated audio. Our code is available at https://github.com/billsioros/EditGen", "date": "2025-07-15", "categories": []}
{"id": "2507.08882", "title": "Less Stress, More Privacy: Stress Detection on Anonymized Speech of Air Traffic Controllers", "abstract": "Air traffic control (ATC) demands multi-tasking under time pressure with high consequences of an error. This can induce stress. Detecting stress is a key point in maintaining the high safety standards of ATC. However, processing ATC voice data entails privacy restrictions, e.g. the General Data Protection Regulation (GDPR) law. Anonymizing the ATC voice data is one way to comply with these restrictions. In this paper, different architectures for stress detection for anonymized ATCO speech are evaluated. Our best networks reach a stress detection accuracy of 93.6% on an anonymized version of the Speech Under Simulated and Actual Stress (SUSAS) dataset and an accuracy of 80.1% on our anonymized ATC simulation dataset. This shows that privacy does not have to be an impediment in building well-performing deep-learning-based models.", "date": "2025-07-10", "categories": []}
{"id": "2507.09310", "title": "Voice Conversion for Lombard Speaking Style with Implicit and Explicit Acoustic Feature Conditioning", "abstract": "Text-to-Speech (TTS) systems in Lombard speaking style can improve the overall intelligibility of speech, useful for hearing loss and noisy conditions. However, training those models requires a large amount of data and the Lombard effect is challenging to record due to speaker and noise variability and tiring recording conditions. Voice conversion (VC) has been shown to be a useful augmentation technique to train TTS systems in the absence of recorded data from the target speaker in the target speaking style. In this paper, we are concerned with Lombard speaking style transfer. Our goal is to convert speaker identity while preserving the acoustic attributes that define the Lombard speaking style. We compare voice conversion models with implicit and explicit acoustic feature conditioning. We observe that our proposed implicit conditioning strategy achieves an intelligibility gain comparable to the model conditioned on explicit acoustic features, while also preserving speaker similarity.", "date": "2025-07-12", "categories": []}
{"id": "2507.09323", "title": "Dynamic Inter-Class Confusion-Aware Encoder for Audio-Visual Fusion in Human Activity Recognition", "abstract": "Humans do not understand individual events in isolation; rather, they generalize concepts within classes and compare them to others. Existing audio-video pre-training paradigms only focus on the alignment of the overall audio-video modalities, without considering the reinforcement of distinguishing easily confused classes through cognitive induction and contrast during training. This paper proposes the Dynamic Inter-Class Confusion-Aware Encoder (DICCAE), an encoder that aligns audio-video representations at a fine-grained, category-level. DICCAE addresses category confusion by dynamically adjusting the confusion loss based on inter-class confusion degrees, thereby enhancing the model's ability to distinguish between similar activities. To further extend the application of DICCAE, we also introduce a novel training framework that incorporates both audio and video modalities, as well as their fusion. To mitigate the scarcity of audio-video data in the human activity recognition task, we propose a cluster-guided audio-video self-supervised pre-training strategy for DICCAE. DICCAE achieves near state-of-the-art performance on the VGGSound dataset, with a top-1 accuracy of 65.5%. We further evaluate its feature representation quality through extensive ablation studies, validating the necessity of each module.", "date": "2025-07-12", "categories": []}
{"id": "2507.09834", "title": "Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction", "abstract": "Autoregressive next-token prediction with the Transformer decoder has become a de facto standard in large language models (LLMs), achieving remarkable success in Natural Language Processing (NLP) at scale. Extending this paradigm to audio poses unique challenges due to its inherently continuous nature. We research audio generation with a causal language model (LM) without discrete tokens. We leverage token-wise diffusion to model the continuous distribution of the next continuous-valued token. Our approach delivers significant improvements over previous discrete solution, AudioGen, achieving 20% and 40% relative gains on AudioCaps in Frechet Audio Distance (FAD) and Kullback-Leibler (KL) divergence, respectively. Additionally, we propose a novel masked next-token prediction task that incorporates masked prediction into the causal LM framework. On AudioCaps, the innovation yields 41% and 33% relative FAD improvements over AudioGen Base (285M) and AudioGen Large (1B) models, respectively, and is on par with the state-of-the-art (SOTA) diffusion models. Furthermore, we achieve these results with significantly fewer parameters -- 193M for our Base and 462M for our Large models.", "date": "2025-07-14", "categories": []}
{"id": "2507.10311", "title": "Recognizing Dementia from Neuropsychological Tests with State Space Models", "abstract": "Early detection of dementia is critical for timely medical intervention and improved patient outcomes. Neuropsychological tests are widely used for cognitive assessment but have traditionally relied on manual scoring. Automatic dementia classification (ADC) systems aim to infer cognitive decline directly from speech recordings of such tests. We propose Demenba, a novel ADC framework based on state space models, which scale linearly in memory and computation with sequence length. Trained on over 1,000 hours of cognitive assessments administered to Framingham Heart Study participants, some of whom were diagnosed with dementia through adjudicated review, our method outperforms prior approaches in fine-grained dementia classification by 21\\%, while using fewer parameters. We further analyze its scaling behavior and demonstrate that our model gains additional improvement when fused with large language models, paving the way for more transparent and scalable dementia assessment tools. Code: https://anonymous.4open.science/r/Demenba-0861", "date": "2025-07-14", "categories": []}
{"id": "2507.10447", "title": "Evaluating Fake Music Detection Performance Under Audio Augmentations", "abstract": "With the rapid advancement of generative audio models, distinguishing between human-composed and generated music is becoming increasingly challenging. As a response, models for detecting fake music have been proposed. In this work, we explore the robustness of such systems under audio augmentations. To evaluate model generalization, we constructed a dataset consisting of both real and synthetic music generated using several systems. We then apply a range of audio transformations and analyze how they affect classification accuracy. We test the performance of a recent state-of-the-art musical deepfake detection model in the presence of audio augmentations. The performance of the model decreases significantly even with the introduction of light augmentations.", "date": "2025-07-07", "categories": []}
{"id": "2507.10464", "title": "AudioMAE++: learning better masked audio representations with SwiGLU FFNs", "abstract": "Masked Autoencoders (MAEs) trained on audio spectrogram patches have emerged as a prominent approach for learning self-supervised audio representations. While several recent papers have evaluated key aspects of training MAEs on audio data, the majority of these approaches still leverage vanilla transformer building blocks, whereas the transformer community has seen steady integration of newer architectural advancements. In this work, we propose AudioMAE++, a revamped audio masked autoencoder with two such enhancements, namely macaron-style transformer blocks with gated linear units. When pretrained on the AudioSet dataset, the proposed AudioMAE++ models outperform existing MAE based approaches on 10 diverse downstream tasks, demonstrating excellent performance on audio classification and speech-based benchmarks. The proposed AudioMAE++ models also demonstrate excellent scaling characteristics, outperforming directly comparable standard MAE baselines with up to 4x more parameters.", "date": "2025-07-14", "categories": []}
{"id": "2507.08530", "title": "MIDI-VALLE: Improving Expressive Piano Performance Synthesis Through Neural Codec Language Modelling", "abstract": "Generating expressive audio performances from music scores requires models to capture both instrument acoustics and human interpretation. Traditional music performance synthesis pipelines follow a two-stage approach, first generating expressive performance MIDI from a score, then synthesising the MIDI into audio. However, the synthesis models often struggle to generalise across diverse MIDI sources, musical styles, and recording environments. To address these challenges, we propose MIDI-VALLE, a neural codec language model adapted from the VALLE framework, which was originally designed for zero-shot personalised text-to-speech (TTS) synthesis. For performance MIDI-to-audio synthesis, we improve the architecture to condition on a reference audio performance and its corresponding MIDI. Unlike previous TTS-based systems that rely on piano rolls, MIDI-VALLE encodes both MIDI and audio as discrete tokens, facilitating a more consistent and robust modelling of piano performances. Furthermore, the model's generalisation ability is enhanced by training on an extensive and diverse piano performance dataset. Evaluation results show that MIDI-VALLE significantly outperforms a state-of-the-art baseline, achieving over 75% lower Frechet Audio Distance on the ATEPP and Maestro datasets. In the listening test, MIDI-VALLE received 202 votes compared to 58 for the baseline, demonstrating improved synthesis quality and generalisation across diverse performance MIDI inputs.", "date": "2025-07-11", "categories": []}
{"id": "2507.06329", "title": "MixAssist: An Audio-Language Dataset for Co-Creative AI Assistance in Music Mixing", "abstract": "While AI presents significant potential for enhancing music mixing and mastering workflows, current research predominantly emphasizes end-to-end automation or generation, often overlooking the collaborative and instructional dimensions vital for co-creative processes. This gap leaves artists, particularly amateurs seeking to develop expertise, underserved. To bridge this, we introduce MixAssist, a novel audio-language dataset capturing the situated, multi-turn dialogue between expert and amateur music producers during collaborative mixing sessions. Comprising 431 audio-grounded conversational turns derived from 7 in-depth sessions involving 12 producers, MixAssist provides a unique resource for training and evaluating audio-language models that can comprehend and respond to the complexities of real-world music production dialogues. Our evaluations, including automated LLM-as-a-judge assessments and human expert comparisons, demonstrate that fine-tuning models such as Qwen-Audio on MixAssist can yield promising results, with Qwen significantly outperforming other tested models in generating helpful, contextually relevant mixing advice. By focusing on co-creative instruction grounded in audio context, MixAssist enables the development of intelligent AI assistants designed to support and augment the creative process in music mixing.", "date": "2025-07-08", "categories": []}
{"id": "2507.07066", "title": "Latent Acoustic Mapping for Direction of Arrival Estimation: A Self-Supervised Approach", "abstract": "Acoustic mapping techniques have long been used in spatial audio processing for direction of arrival estimation (DoAE). Traditional beamforming methods for acoustic mapping, while interpretable, often rely on iterative solvers that can be computationally intensive and sensitive to acoustic variability. On the other hand, recent supervised deep learning approaches offer feedforward speed and robustness but require large labeled datasets and lack interpretability. Despite their strengths, both methods struggle to consistently generalize across diverse acoustic setups and array configurations, limiting their broader applicability. We introduce the Latent Acoustic Mapping (LAM) model, a self-supervised framework that bridges the interpretability of traditional methods with the adaptability and efficiency of deep learning methods. LAM generates high-resolution acoustic maps, adapts to varying acoustic conditions, and operates efficiently across different microphone arrays. We assess its robustness on DoAE using the LOCATA and STARSS benchmarks. LAM achieves comparable or superior localization performance to existing supervised methods. Additionally, we show that LAM's acoustic maps can serve as effective features for supervised models, further enhancing DoAE accuracy and underscoring its potential to advance adaptive, high-performance sound localization systems.", "date": "2025-07-08", "categories": []}
{"id": "2410.21849", "title": "Joint Beamforming and Speaker-Attributed ASR for Real Distant-Microphone Meeting Transcription", "abstract": "Distant-microphone meeting transcription is a challenging task. State-of-the-art end-to-end speaker-attributed automatic speech recognition (SA-ASR) architectures lack a multichannel noise and reverberation reduction front-end, which limits their performance. In this paper, we introduce a joint beamforming and SA-ASR approach for real meeting transcription. We first describe a data alignment and augmentation method to pretrain a neural beamformer on real meeting data. We then compare fixed, hybrid, and fully neural beamformers as front-ends to the SA-ASR model. Finally, we jointly optimize the fully neural beamformer and the SA-ASR model. Experiments on the real AMI corpus show that, while state-of-the-art multi-frame cross-channel attention based channel fusion fails to improve ASR performance, fine-tuning SA-ASR on the fixed beamformer's output and jointly fine-tuning SA-ASR with the neural beamformer reduce the word error rate by 8% and 9% relative, respectively.", "date": "2024-10-29", "categories": []}
{"id": "2505.11079", "title": "ALLM4ADD: Unlocking the Capabilities of Audio Large Language Models for Audio Deepfake Detection", "abstract": "Audio deepfake detection (ADD) has grown increasingly important due to the rise of high-fidelity audio generative models and their potential for misuse. Given that audio large language models (ALLMs) have made significant progress in various audio processing tasks, a heuristic question arises: \\textit{Can ALLMs be leveraged to solve ADD?}. In this paper, we first conduct a comprehensive zero-shot evaluation of ALLMs on ADD, revealing their ineffectiveness. To this end, we propose ALLM4ADD, an ALLM-driven framework for ADD. Specifically, we reformulate ADD task as an audio question answering problem, prompting the model with the question: ``Is this audio fake or real?''. We then perform supervised fine-tuning to enable the ALLM to assess the authenticity of query audio. Extensive experiments are conducted to demonstrate that our ALLM-based method can achieve superior performance in fake audio detection, particularly in data-scarce scenarios. As a pioneering study, we anticipate that this work will inspire the research community to leverage ALLMs to develop more effective ADD systems. Code is available at https://github.com/ucas-hao/qwen_audio_for_add.git", "date": "2025-05-16", "categories": []}
{"id": "2507.04554", "title": "Self-supervised learning of speech representations with Dutch archival data", "abstract": "This paper explores the use of Dutch archival television broadcast data for self-supervised learning of speech foundation models, specifically wav2vec 2.0. We first study data quality assumptions for pre-training, and show how music, noise and speaker overlap affect SSL convergence and downstream fine-tuning performance. Secondly, we explore effectively pre-processing strategies to convert the noisy broadcast dataset into a qualitative dataset for pre-training, by using Whisper and WhisperX. Thirdly, we compare mono-lingual and multi-lingual pre-training with equivalent amounts of data, and show that mono-lingual pre-training is more robust to out-of-domain data. Lastly, we achieve a state-of-the-art LARGE wav2vec 2.0 model for the Dutch language, by a continuation of pre-training a wav2vec 2.0 XLS-R model checkpoint with our 55k hour archival dataset.", "date": "2025-07-06", "categories": []}
{"id": "2507.06116", "title": "Speech Quality Assessment Model Based on Mixture of Experts: System-Level Performance Enhancement and Utterance-Level Challenge Analysis", "abstract": "Automatic speech quality assessment plays a crucial role in the development of speech synthesis systems, but existing models exhibit significant performance variations across different granularity levels of prediction tasks. This paper proposes an enhanced MOS prediction system based on self-supervised learning speech models, incorporating a Mixture of Experts (MoE) classification head and utilizing synthetic data from multiple commercial generation models for data augmentation. Our method builds upon existing self-supervised models such as wav2vec2, designing a specialized MoE architecture to address different types of speech quality assessment tasks. We also collected a large-scale synthetic speech dataset encompassing the latest text-to-speech, speech conversion, and speech enhancement systems. However, despite the adoption of the MoE architecture and expanded dataset, the model's performance improvements in sentence-level prediction tasks remain limited. Our work reveals the limitations of current methods in handling sentence-level quality assessment, provides new technical pathways for the field of automatic speech quality assessment, and also delves into the fundamental causes of performance differences across different assessment granularities.", "date": "2025-07-08", "categories": []}
{"id": "2406.06612", "title": "SEE-2-SOUND: Zero-Shot Spatial Environment-to-Spatial Sound", "abstract": "Generating combined visual and auditory sensory experiences is critical for the consumption of immersive content. Recent advances in neural generative models have enabled the creation of high-resolution content across multiple modalities such as images, text, speech, and videos. Despite these successes, there remains a significant gap in the generation of high-quality spatial audio that complements generated visual content. Furthermore, current audio generation models excel in either generating natural audio or speech or music but fall short in integrating spatial audio cues necessary for immersive experiences. In this work, we introduce SEE-2-SOUND, a zero-shot approach that decomposes the task into (1) identifying visual regions of interest; (2) locating these elements in 3D space; (3) generating mono-audio for each; and (4) integrating them into spatial audio. Using our framework, we demonstrate compelling results for generating spatial audio for high-quality videos, images, and dynamic images from the internet, as well as media generated by learned approaches.", "date": "2024-06-06", "categories": []}
{"id": "2502.03897", "title": "UniForm: A Unified Multi-Task Diffusion Transformer for Audio-Video Generation", "abstract": "With the rise of diffusion models, audio-video generation has been revolutionized. However, most existing methods rely on separate modules for each modality, with limited exploration of unified generative architectures. In addition, many are confined to a single task and small-scale datasets. To overcome these limitations, we introduce UniForm, a unified multi-task diffusion transformer that generates both audio and visual modalities in a shared latent space. By using a unified denoising network, UniForm captures the inherent correlations between sound and vision. Additionally, we propose task-specific noise schemes and task tokens, enabling the model to support multiple tasks with a single set of parameters, including video-to-audio, audio-to-video and text-to-audio-video generation. Furthermore, by leveraging large language models and a large-scale text-audio-video combined dataset, UniForm achieves greater generative diversity than prior approaches. Experiments show that UniForm achieves performance close to the state-of-the-art single-task models across three generation tasks, with generated content that is not only highly aligned with real-world data distributions but also enables more diverse and fine-grained generation.", "date": "2025-02-06", "categories": []}
{"id": "2507.02915", "title": "Audio-JEPA: Joint-Embedding Predictive Architecture for Audio Representation Learning", "abstract": "Building on the Joint-Embedding Predictive Architecture (JEPA) paradigm, a recent self-supervised learning framework that predicts latent representations of masked regions in high-level feature spaces, we propose Audio-JEPA (Audio Joint-Embedding Predictive Architecture), tailored specifically for audio data. Audio-JEPA uses a simple Vision Transformer backbone to predict latent representations of masked spectrogram patches rather than reconstructing raw audio. We pre-train on unlabeled AudioSet clips (10s, 32kHz) with random patch masking on mel-spectrograms. We evaluate on the X-ARES suite covering speech, music, and environmental sound tasks. Although our implementation is a straightforward translation of the original model to audio, the results still show comparable performance to wav2vec 2.0 and data2vec while using less than one-fifth of their training data and with no hyper-parameter tuning. All code and pretrained checkpoints will be released on GitHub.", "date": "2025-06-25", "categories": []}
{"id": "2507.03641", "title": "Improving Low-Resource Dialect Classification Using Retrieval-based Voice Conversion", "abstract": "Deep learning models for dialect identification are often limited by the scarcity of dialectal data. To address this challenge, we propose to use Retrieval-based Voice Conversion (RVC) as an effective data augmentation method for a low-resource German dialect classification task. By converting audio samples to a uniform target speaker, RVC minimizes speaker-related variability, enabling models to focus on dialect-specific linguistic and phonetic features. Our experiments demonstrate that RVC enhances classification performance when utilized as a standalone augmentation method. Furthermore, combining RVC with other augmentation methods such as frequency masking and segment removal leads to additional performance gains, highlighting its potential for improving dialect classification in low-resource scenarios.", "date": "2025-07-04", "categories": []}
{"id": "2507.04230", "title": "High-Resolution Sustain Pedal Depth Estimation from Piano Audio Across Room Acoustics", "abstract": "Piano sustain pedal detection has previously been approached as a binary on/off classification task, limiting its application in real-world piano performance scenarios where pedal depth significantly influences musical expression. This paper presents a novel approach for high-resolution estimation that predicts continuous pedal depth values. We introduce a Transformer-based architecture that not only matches state-of-the-art performance on the traditional binary classification task but also achieves high accuracy in continuous pedal depth estimation. Furthermore, by estimating continuous values, our model provides musically meaningful predictions for sustain pedal usage, whereas baseline models struggle to capture such nuanced expressions with their binary detection approach. Additionally, this paper investigates the influence of room acoustics on sustain pedal estimation using a synthetic dataset that includes varied acoustic conditions. We train our model with different combinations of room settings and test it in an unseen new environment using a \"leave-one-out\" approach. Our findings show that the two baseline models and ours are not robust to unseen room conditions. Statistical analysis further confirms that reverberation influences model predictions and introduces an overestimation bias.", "date": "2025-07-06", "categories": []}
{"id": "2507.04817", "title": "Fast-VGAN: Lightweight Voice Conversion with Explicit Control of F0 and Duration Parameters", "abstract": "Precise control over speech characteristics, such as pitch, duration, and speech rate, remains a significant challenge in the field of voice conversion. The ability to manipulate parameters like pitch and syllable rate is an important element for effective identity conversion, but can also be used independently for voice transformation, achieving goals that were historically addressed by vocoder-based methods.\n  In this work, we explore a convolutional neural network-based approach that aims to provide means for modifying fundamental frequency (F0), phoneme sequences, intensity, and speaker identity. Rather than relying on disentanglement techniques, our model is explicitly conditioned on these factors to generate mel spectrograms, which are then converted into waveforms using a universal neural vocoder. Accordingly, during inference, F0 contours, phoneme sequences, and speaker embeddings can be freely adjusted, allowing for intuitively controlled voice transformations.\n  We evaluate our approach on speaker conversion and expressive speech tasks using both perceptual and objective metrics. The results suggest that the proposed method offers substantial flexibility, while maintaining high intelligibility and speaker similarity.", "date": "2025-07-07", "categories": []}
{"id": "2507.04845", "title": "Spatial and Semantic Embedding Integration for Stereo Sound Event Localization and Detection in Regular Videos", "abstract": "This report presents our systems submitted to the audio-only and audio-visual tracks of the DCASE2025 Task 3 Challenge: Stereo Sound Event Localization and Detection (SELD) in Regular Video Content. SELD is a complex task that combines temporal event classification with spatial localization, requiring reasoning across spatial, temporal, and semantic dimensions. The last is arguably the most challenging to model. Traditional SELD architectures rely on multichannel input, which limits their ability to leverage large-scale pre-training due to data constraints. To address this, we enhance standard SELD architectures with semantic information by integrating pre-trained, contrastive language-aligned models: CLAP for audio and OWL-ViT for visual inputs. These embeddings are incorporated into a modified Conformer module tailored for multimodal fusion, which we refer to as the Cross-Modal Conformer. Additionally, we incorporate autocorrelation-based acoustic features to improve distance estimation. We pre-train our models on curated synthetic audio and audio-visual datasets and apply a left-right channel swapping augmentation to further increase the training data. Both our audio-only and audio-visual systems substantially outperform the challenge baselines on the development set, demonstrating the effectiveness of our strategy. Performance is further improved through model ensembling and a visual post-processing step based on human keypoints. Future work will investigate the contribution of each modality and explore architectural variants to further enhance results.", "date": "2025-07-07", "categories": []}
{"id": "2507.04858", "title": "Towards Human-in-the-Loop Onset Detection: A Transfer Learning Approach for Maracatu", "abstract": "We explore transfer learning strategies for musical onset detection in the Afro-Brazilian Maracatu tradition, which features complex rhythmic patterns that challenge conventional models. We adapt two Temporal Convolutional Network architectures: one pre-trained for onset detection (intra-task) and another for beat tracking (inter-task). Using only 5-second annotated snippets per instrument, we fine-tune these models through layer-wise retraining strategies for five traditional percussion instruments. Our results demonstrate significant improvements over baseline performance, with F1 scores reaching up to 0.998 in the intra-task setting and improvements of over 50 percentage points in best-case scenarios. The cross-task adaptation proves particularly effective for time-keeping instruments, where onsets naturally align with beat positions. The optimal fine-tuning configuration varies by instrument, highlighting the importance of instrument-specific adaptation strategies. This approach addresses the challenges of underrepresented musical traditions, offering an efficient human-in-the-loop methodology that minimizes annotation effort while maximizing performance. Our findings contribute to more inclusive music information retrieval tools applicable beyond Western musical contexts.", "date": "2025-07-07", "categories": []}
{"id": "2507.02109", "title": "Parametric Neural Amp Modeling with Active Learning", "abstract": "We introduce PANAMA, an active learning framework for the training of end-to-end parametric guitar amp models using a WaveNet-like architecture. With \\model, one can create a virtual amp by recording samples that are determined by an active learning strategy to use a minimum amount of datapoints (i.e., amp knob settings). We show that gradient-based optimization algorithms can be used to determine the optimal datapoints to sample, and that the approach helps under a constrained number of samples.", "date": "2025-07-02", "categories": []}
{"id": "2507.02407", "title": "Benchmarking Akan ASR Models Across Domain-Specific Datasets: A Comparative Evaluation of Performance, Scalability, and Adaptability", "abstract": "Most existing automatic speech recognition (ASR) research evaluate models using in-domain datasets. However, they seldom evaluate how they generalize across diverse speech contexts. This study addresses this gap by benchmarking seven Akan ASR models built on transformer architectures, such as Whisper and Wav2Vec2, using four Akan speech corpora to determine their performance. These datasets encompass various domains, including culturally relevant image descriptions, informal conversations, biblical scripture readings, and spontaneous financial dialogues. A comparison of the word error rate and character error rate highlighted domain dependency, with models performing optimally only within their training domains while showing marked accuracy degradation in mismatched scenarios. This study also identified distinct error behaviors between the Whisper and Wav2Vec2 architectures. Whereas fine-tuned Whisper Akan models led to more fluent but potentially misleading transcription errors, Wav2Vec2 produced more obvious yet less interpretable outputs when encountering unfamiliar inputs. This trade-off between readability and transparency in ASR errors should be considered when selecting architectures for low-resource language (LRL) applications. These findings highlight the need for targeted domain adaptation techniques, adaptive routing strategies, and multilingual training frameworks for Akan and other LRLs.", "date": "2025-07-03", "categories": []}
{"id": "2507.02666", "title": "ASDA: Audio Spectrogram Differential Attention Mechanism for Self-Supervised Representation Learning", "abstract": "In recent advancements in audio self-supervised representation learning, the standard Transformer architecture has emerged as the predominant approach, yet its attention mechanism often allocates a portion of attention weights to irrelevant information, potentially impairing the model's discriminative ability. To address this, we introduce a differential attention mechanism, which effectively mitigates ineffective attention allocation through the integration of dual-softmax operations and appropriately tuned differential coefficients. Experimental results demonstrate that our ASDA model achieves state-of-the-art (SOTA) performance across multiple benchmarks, including audio classification (49.0% mAP on AS-2M, 41.5% mAP on AS20K), keyword spotting (98.3% accuracy on SPC-2), and environmental sound classification (96.1% accuracy on ESC-50). These results highlight ASDA's effectiveness in audio tasks, paving the way for broader applications.", "date": "2025-07-03", "categories": []}
{"id": "2410.11062", "title": "CleanUMamba: A Compact Mamba Network for Speech Denoising using Channel Pruning", "abstract": "This paper presents CleanUMamba, a time-domain neural network architecture designed for real-time causal audio denoising directly applied to raw waveforms. CleanUMamba leverages a U-Net encoder-decoder structure, incorporating the Mamba state-space model in the bottleneck layer. By replacing conventional self-attention and LSTM mechanisms with Mamba, our architecture offers superior denoising performance while maintaining a constant memory footprint, enabling streaming operation. To enhance efficiency, we applied structured channel pruning, achieving an 8X reduction in model size without compromising audio quality. Our model demonstrates strong results in the Interspeech 2020 Deep Noise Suppression challenge. Specifically, CleanUMamba achieves a PESQ score of 2.42 and STOI of 95.1% with only 442K parameters and 468M MACs, matching or outperforming larger models in real-time performance. Code will be available at: https://github.com/lab-emi/CleanUMamba", "date": "2024-10-14", "categories": []}
{"id": "2503.15576", "title": "A Bird Song Detector for improving bird identification through Deep Learning: a case study from Do\\~nana", "abstract": "Passive Acoustic Monitoring is a key tool for biodiversity conservation, but the large volumes of unsupervised audio it generates present major challenges for extracting meaningful information. Deep Learning offers promising solutions. BirdNET, a widely used bird identification model, has shown success in many study systems but is limited at local scale due to biases in its training data, which focus on specific locations and target sounds rather than entire soundscapes. A key challenge in bird species identification is that many recordings either lack target species or contain overlapping vocalizations, complicating automatic identification. To address these problems, we developed a multi-stage pipeline for automatic bird vocalization identification in Do\\~nana National Park (SW Spain), a wetland of high conservation concern. We deployed AudioMoth recorders in three main habitats across nine locations and manually annotated 461 minutes of audio, resulting in 3749 labeled segments spanning 34 classes. We first applied a Bird Song Detector to isolate bird vocalizations using spectrogram-based image processing. Then, species were classified using custom models trained at the local scale. Applying the Bird Song Detector before classification improved species identification, as all models performed better when analyzing only the segments where birds were detected. Specifically, the combination of detector and fine-tuned BirdNET outperformed the baseline without detection. This approach demonstrates the effectiveness of integrating a Bird Song Detector with local classification models. These findings highlight the need to adapt general-purpose tools to specific ecological challenges. Automatically detecting bird species helps track the health of this threatened ecosystem, given birds sensitivity to environmental change, and supports conservation planning to reduce biodiversity loss.", "date": "2025-03-19", "categories": []}
{"id": "2507.01563", "title": "Real-Time Emergency Vehicle Siren Detection with Efficient CNNs on Embedded Hardware", "abstract": "We present a full-stack emergency vehicle (EV) siren detection system designed for real-time deployment on embedded hardware. The proposed approach is based on E2PANNs, a fine-tuned convolutional neural network derived from EPANNs, and optimized for binary sound event detection under urban acoustic conditions. A key contribution is the creation of curated and semantically structured datasets - AudioSet-EV, AudioSet-EV Augmented, and Unified-EV - developed using a custom AudioSet-Tools framework to overcome the low reliability of standard AudioSet annotations. The system is deployed on a Raspberry Pi 5 equipped with a high-fidelity DAC+microphone board, implementing a multithreaded inference engine with adaptive frame sizing, probability smoothing, and a decision-state machine to control false positive activations. A remote WebSocket interface provides real-time monitoring and facilitates live demonstration capabilities. Performance is evaluated using both framewise and event-based metrics across multiple configurations. Results show the system achieves low-latency detection with improved robustness under realistic audio conditions. This work demonstrates the feasibility of deploying IoS-compatible SED solutions that can form distributed acoustic monitoring networks, enabling collaborative emergency vehicle tracking across smart city infrastructures through WebSocket connectivity on low-cost edge devices.", "date": "2025-07-02", "categories": []}
{"id": "2505.14518", "title": "Teaching Audio-Aware Large Language Models What Does Not Hear: Mitigating Hallucinations through Synthesized Negative Samples", "abstract": "Recent advancements in audio-aware large language models (ALLMs) enable them to process and understand audio inputs. However, these models often hallucinate non-existent sound events, reducing their reliability in real-world applications. To address this, we propose LISTEN (Learning to Identify Sounds Through Extended Negative Samples), a contrastive-like training method that enhances ALLMs' ability to distinguish between present and absent sounds using synthesized data from the backbone LLM. Unlike prior approaches, our method requires no modification to LLM parameters and efficiently integrates audio representations via a lightweight adapter. Experiments show that LISTEN effectively mitigates hallucinations while maintaining impressive performance on existing audio question and reasoning benchmarks. At the same time, it is more efficient in both data and computation.", "date": "2025-05-20", "categories": []}
{"id": "2411.07186", "title": "NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics", "abstract": "Large language models (LLMs) prompted with text and audio have achieved state-of-the-art performance across various auditory tasks, including speech, music, and general audio, showing emergent abilities on unseen tasks. However, their potential has yet to be fully demonstrated in bioacoustics tasks, such as detecting animal vocalizations in large recordings, classifying rare and endangered species, and labeling context and behavior -- tasks that are crucial for conservation, biodiversity monitoring, and animal behavior studies. In this work, we present NatureLM-audio, the first audio-language foundation model specifically designed for bioacoustics. Our training dataset consists of carefully curated text-audio pairs spanning bioacoustics, speech, and music, designed to address the field's limited availability of annotated data. We demonstrate successful transfer of learned representations from music and speech to bioacoustics, and our model shows promising generalization to unseen taxa and tasks. We evaluate NatureLM-audio on a novel benchmark (BEANS-Zero) and it sets a new state of the art on several bioacoustics tasks, including zero-shot classification of unseen species. To advance bioacoustics research, we release our model weights, benchmark data, and open-source the code for training and benchmark data generation and model training.", "date": "2024-11-11", "categories": []}
{"id": "2506.12285", "title": "CMI-Bench: A Comprehensive Benchmark for Evaluating Music Instruction Following", "abstract": "Recent advances in audio-text large language models (LLMs) have opened new possibilities for music understanding and generation. However, existing benchmarks are limited in scope, often relying on simplified tasks or multi-choice evaluations that fail to reflect the complexity of real-world music analysis. We reinterpret a broad range of traditional MIR annotations as instruction-following formats and introduce CMI-Bench, a comprehensive music instruction following benchmark designed to evaluate audio-text LLMs on a diverse set of music information retrieval (MIR) tasks. These include genre classification, emotion regression, emotion tagging, instrument classification, pitch estimation, key detection, lyrics transcription, melody extraction, vocal technique recognition, instrument performance technique detection, music tagging, music captioning, and (down)beat tracking: reflecting core challenges in MIR research. Unlike previous benchmarks, CMI-Bench adopts standardized evaluation metrics consistent with previous state-of-the-art MIR models, ensuring direct comparability with supervised approaches. We provide an evaluation toolkit supporting all open-source audio-textual LLMs, including LTU, Qwen-audio, SALMONN, MusiLingo, etc. Experiment results reveal significant performance gaps between LLMs and supervised models, along with their culture, chronological and gender bias, highlighting the potential and limitations of current models in addressing MIR tasks. CMI-Bench establishes a unified foundation for evaluating music instruction following, driving progress in music-aware LLMs.", "date": "2025-06-14", "categories": []}
{"id": "2506.15981", "title": "Double Entendre: Robust Audio-Based AI-Generated Lyrics Detection via Multi-View Fusion", "abstract": "The rapid advancement of AI-based music generation tools is revolutionizing the music industry but also posing challenges to artists, copyright holders, and providers alike. This necessitates reliable methods for detecting such AI-generated content. However, existing detectors, relying on either audio or lyrics, face key practical limitations: audio-based detectors fail to generalize to new or unseen generators and are vulnerable to audio perturbations; lyrics-based methods require cleanly formatted and accurate lyrics, unavailable in practice. To overcome these limitations, we propose a novel, practically grounded approach: a multimodal, modular late-fusion pipeline that combines automatically transcribed sung lyrics and speech features capturing lyrics-related information within the audio. By relying on lyrical aspects directly from audio, our method enhances robustness, mitigates susceptibility to low-level artifacts, and enables practical applicability. Experiments show that our method, DE-detect, outperforms existing lyrics-based detectors while also being more robust to audio perturbations. Thus, it offers an effective, robust solution for detecting AI-generated music in real-world scenarios. Our code is available at https://github.com/deezer/robust-AI-lyrics-detection.", "date": "2025-06-19", "categories": []}
{"id": "2506.23271", "title": "Mettle: Meta-Token Learning for Memory-Efficient Audio-Visual Adaptation", "abstract": "We present \\textbf{Met}a-\\textbf{T}oken \\textbf{Le}arning (Mettle), a simple and memory-efficient method for adapting large-scale pretrained transformer models to downstream audio-visual tasks. Instead of sequentially modifying the output feature distribution of the transformer backbone, Mettle utilizes a lightweight \\textit{Layer-Centric Distillation (LCD)} module to distill in parallel the intact audio or visual features embedded by each transformer layer into compact meta-tokens. This distillation process considers both pretrained knowledge preservation and task-specific adaptation. The obtained meta-tokens can be directly applied to classification tasks, such as audio-visual event localization and audio-visual video parsing. To further support fine-grained segmentation tasks, such as audio-visual segmentation, we introduce a \\textit{Meta-Token Injection (MTI)} module, which utilizes the audio and visual meta-tokens distilled from the top transformer layer to guide feature adaptation in earlier layers. Extensive experiments on multiple audiovisual benchmarks demonstrate that our method significantly reduces memory usage and training time while maintaining parameter efficiency and competitive accuracy.", "date": "2025-06-29", "categories": []}
{"id": "2506.23437", "title": "From Large-scale Audio Tagging to Real-Time Explainable Emergency Vehicle Sirens Detection", "abstract": "Accurate recognition of Emergency Vehicle (EV) sirens is critical for the integration of intelligent transportation systems, smart city monitoring systems, and autonomous driving technologies. Modern automatic solutions are limited by the lack of large scale, curated datasets and by the computational demands of state of the art sound event detection models. This work introduces E2PANNs (Efficient Emergency Pre trained Audio Neural Networks), a lightweight Convolutional Neural Network architecture derived from the PANNs framework, specifically optimized for binary EV siren detection. Leveraging our dedicated subset of AudioSet (AudioSet EV) we fine-tune and evaluate E2PANNs across multiple reference datasets and test its viability on embedded hardware. The experimental campaign includes ablation studies, cross-domain benchmarking, and real-time inference deployment on edge device. Interpretability analyses exploiting Guided Backpropagation and ScoreCAM algorithms provide insights into the model internal representations and validate its ability to capture distinct spectrotemporal patterns associated with different types of EV sirens. Real time performance is assessed through frame wise and event based detection metrics, as well as a detailed analysis of false positive activations. Results demonstrate that E2PANNs establish a new state of the art in this research domain, with high computational efficiency, and suitability for edge-based audio monitoring and safety-critical applications.", "date": "2025-06-30", "categories": []}
{"id": "2506.23873", "title": "Emergent musical properties of a transformer under contrastive self-supervised learning", "abstract": "In music information retrieval (MIR), contrastive self-supervised learning for general-purpose representation models is effective for global tasks such as automatic tagging. However, for local tasks such as chord estimation, it is widely assumed that contrastively trained general-purpose self-supervised models are inadequate and that more sophisticated SSL is necessary; e.g., masked modeling. Our paper challenges this assumption by revealing the potential of contrastive SSL paired with a transformer in local MIR tasks. We consider a lightweight vision transformer with one-dimensional patches in the time--frequency domain (ViT-1D) and train it with simple contrastive SSL through normalized temperature-scaled cross-entropy loss (NT-Xent). Although NT-Xent operates only over the class token, we observe that, potentially thanks to weight sharing, informative musical properties emerge in ViT-1D's sequence tokens. On global tasks, the temporal average of class and sequence tokens offers a performance increase compared to the class token alone, showing useful properties in the sequence tokens. On local tasks, sequence tokens perform unexpectedly well, despite not being specifically trained for. Furthermore, high-level musical features such as onsets emerge from layer-wise attention maps and self-similarity matrices show different layers capture different musical dimensions. Our paper does not focus on improving performance but advances the musical interpretation of transformers and sheds light on some overlooked abilities of contrastive SSL paired with transformers for sequence modeling in MIR.", "date": "2025-06-30", "categories": []}
{"id": "2504.08524", "title": "USM-VC: Mitigating Timbre Leakage with Universal Semantic Mapping Residual Block for Voice Conversion", "abstract": "Voice conversion (VC) transforms source speech into a target voice by preserving the content. However, timbre information from the source speaker is inherently embedded in the content representations, causing significant timbre leakage and reducing similarity to the target speaker. To address this, we introduce a Universal Semantic Matching (USM) residual block to a content extractor. The residual block consists of two weighted branches: 1) universal semantic dictionary based Content Feature Re-expression (CFR) module, supplying timbre-free content representation. 2) skip connection to the original content layer, providing complementary fine-grained information. In the CFR module, each dictionary entry in the universal semantic dictionary represents a phoneme class, computed statistically using speech from multiple speakers, creating a stable, speaker-independent semantic set. We introduce a CFR method to obtain timbre-free content representations by expressing each content frame as a weighted linear combination of dictionary entries using corresponding phoneme posteriors as weights. Extensive experiments across various VC frameworks demonstrate that our approach effectively mitigates timbre leakage and significantly improves similarity to the target speaker.", "date": "2025-04-11", "categories": []}
{"id": "2506.22237", "title": "Fine-Tuning MIDI-to-Audio Alignment using a Neural Network on Piano Roll and CQT Representations", "abstract": "In this paper, we present a neural network approach for synchronizing audio recordings of human piano performances with their corresponding loosely aligned MIDI files. The task is addressed using a Convolutional Recurrent Neural Network (CRNN) architecture, which effectively captures spectral and temporal features by processing an unaligned piano roll and a spectrogram as inputs to estimate the aligned piano roll. To train the network, we create a dataset of piano pieces with augmented MIDI files that simulate common human timing errors. The proposed model achieves up to 20% higher alignment accuracy than the industry-standard Dynamic Time Warping (DTW) method across various tolerance windows. Furthermore, integrating DTW with the CRNN yields additional improvements, offering enhanced robustness and consistency. These findings demonstrate the potential of neural networks in advancing state-of-the-art MIDI-to-audio alignment.", "date": "2025-06-27", "categories": []}
{"id": "2506.19014", "title": "IndieFake Dataset: A Benchmark Dataset for Audio Deepfake Detection", "abstract": "Advancements in audio deepfake technology offers benefits like AI assistants, better accessibility for speech impairments, and enhanced entertainment. However, it also poses significant risks to security, privacy, and trust in digital communications. Detecting and mitigating these threats requires comprehensive datasets. Existing datasets lack diverse ethnic accents, making them inadequate for many real-world scenarios. Consequently, models trained on these datasets struggle to detect audio deepfakes in diverse linguistic and cultural contexts such as in South-Asian countries. Ironically, there is a stark lack of South-Asian speaker samples in the existing datasets despite constituting a quarter of the worlds population. This work introduces the IndieFake Dataset (IFD), featuring 27.17 hours of bonafide and deepfake audio from 50 English speaking Indian speakers. IFD offers balanced data distribution and includes speaker-level characterization, absent in datasets like ASVspoof21 (DF). We evaluated various baselines on IFD against existing ASVspoof21 (DF) and In-The-Wild (ITW) datasets. IFD outperforms ASVspoof21 (DF) and proves to be more challenging compared to benchmark ITW dataset. The complete dataset, along with documentation and sample reference clips, is publicly accessible for research use on project website.", "date": "2025-06-23", "categories": []}
{"id": "2506.21167", "title": "A Hierarchical Deep Learning Approach for Minority Instrument Detection", "abstract": "Identifying instrument activities within audio excerpts is vital in music information retrieval, with significant implications for music cataloging and discovery. Prior deep learning endeavors in musical instrument recognition have predominantly emphasized instrument classes with ample data availability. Recent studies have demonstrated the applicability of hierarchical classification in detecting instrument activities in orchestral music, even with limited fine-grained annotations at the instrument level. Based on the Hornbostel-Sachs classification, such a hierarchical classification system is evaluated using the MedleyDB dataset, renowned for its diversity and richness concerning various instruments and music genres. This work presents various strategies to integrate hierarchical structures into models and tests a new class of models for hierarchical music prediction. This study showcases more reliable coarse-level instrument detection by bridging the gap between detailed instrument identification and group-level recognition, paving the way for further advancements in this domain.", "date": "2025-06-26", "categories": []}
{"id": "2506.21269", "title": "Integrating Vehicle Acoustic Data for Enhanced Urban Traffic Management: A Study on Speed Classification in Suzhou", "abstract": "This study presents and publicly releases the Suzhou Urban Road Acoustic Dataset (SZUR-Acoustic Dataset), which is accompanied by comprehensive data-acquisition protocols and annotation guidelines to ensure transparency and reproducibility of the experimental workflow. To model the coupling between vehicular noise and driving speed, we propose a bimodal-feature-fusion deep convolutional neural network (BMCNN). During preprocessing, an adaptive denoising and normalization strategy is applied to suppress environmental background interference; in the network architecture, parallel branches extract Mel-frequency cepstral coefficients (MFCCs) and wavelet-packet energy features, which are subsequently fused via a cross-modal attention mechanism in the intermediate feature space to fully exploit time-frequency information. Experimental results demonstrate that BMCNN achieves a classification accuracy of 87.56% on the SZUR-Acoustic Dataset and 96.28% on the public IDMT-Traffic dataset. Ablation studies and robustness tests on the Suzhou dataset further validate the contributions of each module to performance improvement and overfitting mitigation. The proposed acoustics-based speed classification method can be integrated into smart-city traffic management systems for real-time noise monitoring and speed estimation, thereby optimizing traffic flow control, reducing roadside noise pollution, and supporting sustainable urban planning.", "date": "2025-06-26", "categories": []}
{"id": "2506.02858", "title": "DGMO: Training-Free Audio Source Separation through Diffusion-Guided Mask Optimization", "abstract": "Language-queried Audio Source Separation (LASS) enables open-vocabulary sound separation via natural language queries. While existing methods rely on task-specific training, we explore whether pretrained diffusion models, originally designed for audio generation, can inherently perform separation without further training. In this study, we introduce a training-free framework leveraging generative priors for zero-shot LASS. Analyzing naive adaptations, we identify key limitations arising from modality-specific challenges. To address these issues, we propose Diffusion-Guided Mask Optimization (DGMO), a test-time optimization framework that refines spectrogram masks for precise, input-aligned separation. Our approach effectively repurposes pretrained diffusion models for source separation, achieving competitive performance without task-specific supervision. This work expands the application of diffusion models beyond generation, establishing a new paradigm for zero-shot audio separation. The code is available at: https://wltschmrz.github.io/DGMO/", "date": "2025-06-03", "categories": []}
{"id": "2506.20609", "title": "Deciphering GunType Hierarchy through Acoustic Analysis of Gunshot Recordings", "abstract": "The escalating rates of gun-related violence and mass shootings represent a significant threat to public safety. Timely and accurate information for law enforcement agencies is crucial in mitigating these incidents. Current commercial gunshot detection systems, while effective, often come with prohibitive costs. This research explores a cost-effective alternative by leveraging acoustic analysis of gunshot recordings, potentially obtainable from ubiquitous devices like cell phones, to not only detect gunshots but also classify the type of firearm used. This paper details a study on deciphering gun type hierarchies using a curated dataset of 3459 recordings. We investigate the fundamental acoustic characteristics of gunshots, including muzzle blasts and shockwaves, which vary based on firearm type, ammunition, and shooting direction. We propose and evaluate machine learning frameworks, including Support Vector Machines (SVMs) as a baseline and a more advanced Convolutional Neural Network (CNN) architecture for joint gunshot detection and gun type classification. Results indicate that our deep learning approach achieves a mean average precision (mAP) of 0.58 on clean labeled data, outperforming the SVM baseline (mAP 0.39). Challenges related to data quality, environmental noise, and the generalization capabilities when using noisy web-sourced data (mAP 0.35) are also discussed. The long-term vision is to develop a highly accurate, real-time system deployable on common recording devices, significantly reducing detection costs and providing critical intelligence to first responders.", "date": "2025-06-25", "categories": []}
{"id": "2207.05506", "title": "Label-Efficient Self-Supervised Speaker Verification With Information Maximization and Contrastive Learning", "abstract": "State-of-the-art speaker verification systems are inherently dependent on some kind of human supervision as they are trained on massive amounts of labeled data. However, manually annotating utterances is slow, expensive and not scalable to the amount of data available today. In this study, we explore self-supervised learning for speaker verification by learning representations directly from raw audio. The objective is to produce robust speaker embeddings that have small intra-speaker and large inter-speaker variance. Our approach is based on recent information maximization learning frameworks and an intensive data augmentation pre-processing step. We evaluate the ability of these methods to work without contrastive samples before showing that they achieve better performance when combined with a contrastive loss. Furthermore, we conduct experiments to show that our method reaches competitive results compared to existing techniques and can get better performances compared to a supervised baseline when fine-tuned with a small portion of labeled data.", "date": "2022-07-12", "categories": []}
{"id": "2506.17694", "title": "SSAVSV: Towards Unified Model for Self-Supervised Audio-Visual Speaker Verification", "abstract": "Conventional audio-visual methods for speaker verification rely on large amounts of labeled data and separate modality-specific architectures, which is computationally expensive, limiting their scalability. To address these problems, we propose a self-supervised learning framework based on contrastive learning with asymmetric masking and masked data modeling to obtain robust audiovisual feature representations. In particular, we employ a unified framework for self-supervised audiovisual speaker verification using a single shared backbone for audio and visual inputs, leveraging the versatility of vision transformers. The proposed unified framework can handle audio, visual, or audiovisual inputs using a single shared vision transformer backbone during training and testing while being computationally efficient and robust to missing modalities. Extensive experiments demonstrate that our method achieves competitive performance without labeled data while reducing computational costs compared to traditional approaches.", "date": "2025-06-21", "categories": []}
{"id": "2506.18729", "title": "MuseControlLite: Multifunctional Music Generation with Lightweight Conditioners", "abstract": "We propose MuseControlLite, a lightweight mechanism designed to fine-tune text-to-music generation models for precise conditioning using various time-varying musical attributes and reference audio signals. The key finding is that positional embeddings, which have been seldom used by text-to-music generation models in the conditioner for text conditions, are critical when the condition of interest is a function of time. Using melody control as an example, our experiments show that simply adding rotary positional embeddings to the decoupled cross-attention layers increases control accuracy from 56.6% to 61.1%, while requiring 6.75 times fewer trainable parameters than state-of-the-art fine-tuning mechanisms, using the same pre-trained diffusion Transformer model of Stable Audio Open. We evaluate various forms of musical attribute control, audio inpainting, and audio outpainting, demonstrating improved controllability over MusicGen-Large and Stable Audio Open ControlNet at a significantly lower fine-tuning cost, with only 85M trainble parameters. Source code, model checkpoints, and demo examples are available at: https://musecontrollite.github.io/web/.", "date": "2025-06-23", "categories": []}
{"id": "2506.19774", "title": "Kling-Foley: Multimodal Diffusion Transformer for High-Quality Video-to-Audio Generation", "abstract": "We propose Kling-Foley, a large-scale multimodal Video-to-Audio generation model that synthesizes high-quality audio synchronized with video content. In Kling-Foley, we introduce multimodal diffusion transformers to model the interactions between video, audio, and text modalities, and combine it with a visual semantic representation module and an audio-visual synchronization module to enhance alignment capabilities. Specifically, these modules align video conditions with latent audio elements at the frame level, thereby improving semantic alignment and audio-visual synchronization. Together with text conditions, this integrated approach enables precise generation of video-matching sound effects. In addition, we propose a universal latent audio codec that can achieve high-quality modeling in various scenarios such as sound effects, speech, singing, and music. We employ a stereo rendering method that imbues synthesized audio with a spatial presence. At the same time, in order to make up for the incomplete types and annotations of the open-source benchmark, we also open-source an industrial-level benchmark Kling-Audio-Eval. Our experiments show that Kling-Foley trained with the flow matching objective achieves new audio-visual SOTA performance among public models in terms of distribution matching, semantic alignment, temporal alignment and audio quality.", "date": "2025-06-24", "categories": []}
{"id": "2409.10048", "title": "Audio-Driven Reinforcement Learning for Head-Orientation in Naturalistic Environments", "abstract": "Although deep reinforcement learning (DRL) approaches in audio signal processing have seen substantial progress in recent years, audio-driven DRL for tasks such as navigation, gaze control and head-orientation control in the context of human-robot interaction have received little attention. Here, we propose an audio-driven DRL framework in which we utilise deep Q-learning to develop an autonomous agent that orients towards a talker in the acoustic environment based on stereo speech recordings. Our results show that the agent learned to perform the task at a near perfect level when trained on speech segments in anechoic environments (that is, without reverberation). The presence of reverberation in naturalistic acoustic environments affected the agent's performance, although the agent still substantially outperformed a baseline, randomly acting agent. Finally, we quantified the degree of generalization of the proposed DRL approach across naturalistic acoustic environments. Our experiments revealed that policies learned by agents trained on medium or high reverb environments generalized to low reverb environments, but policies learned by agents trained on anechoic or low reverb environments did not generalize to medium or high reverb environments. Taken together, this study demonstrates the potential of audio-driven DRL for tasks such as head-orientation control and highlights the need for training strategies that enable robust generalization across environments for real-world audio-driven DRL applications.", "date": "2024-09-16", "categories": []}
{"id": "2505.09382", "title": "The Voice Timbre Attribute Detection 2025 Challenge Evaluation Plan", "abstract": "Voice timbre refers to the unique quality or character of a person's voice that distinguishes it from others as perceived by human hearing. The Voice Timbre Attribute Detection (VtaD) 2025 challenge focuses on explaining the voice timbre attribute in a comparative manner. In this challenge, the human impression of voice timbre is verbalized with a set of sensory descriptors, including bright, coarse, soft, magnetic, and so on. The timbre is explained from the comparison between two voices in their intensity within a specific descriptor dimension. The VtaD 2025 challenge starts in May and culminates in a special proposal at the NCMMSC2025 conference in October 2025 in Zhenjiang, China.", "date": "2025-05-14", "categories": []}
{"id": "2505.09661", "title": "Introducing voice timbre attribute detection", "abstract": "This paper focuses on explaining the timbre conveyed by speech signals and introduces a task termed voice timbre attribute detection (vTAD). In this task, voice timbre is explained with a set of sensory attributes describing its human perception. A pair of speech utterances is processed, and their intensity is compared in a designated timbre descriptor. Moreover, a framework is proposed, which is built upon the speaker embeddings extracted from the speech utterances. The investigation is conducted on the VCTK-RVA dataset. Experimental examinations on the ECAPA-TDNN and FACodec speaker encoders demonstrated that: 1) the ECAPA-TDNN speaker encoder was more capable in the seen scenario, where the testing speakers were included in the training set; 2) the FACodec speaker encoder was superior in the unseen scenario, where the testing speakers were not part of the training, indicating enhanced generalization capability. The VCTK-RVA dataset and open-source code are available on the website https://github.com/vTAD2025-Challenge/vTAD.", "date": "2025-05-14", "categories": []}
{"id": "2506.17351", "title": "Zero-Shot Cognitive Impairment Detection from Speech Using AudioLLM", "abstract": "Cognitive impairment (CI) is of growing public health concern, and early detection is vital for effective intervention. Speech has gained attention as a non-invasive and easily collectible biomarker for assessing cognitive decline. Traditional CI detection methods typically rely on supervised models trained on acoustic and linguistic features extracted from speech, which often require manual annotation and may not generalise well across datasets and languages. In this work, we propose the first zero-shot speech-based CI detection method using the Qwen2- Audio AudioLLM, a model capable of processing both audio and text inputs. By designing prompt-based instructions, we guide the model in classifying speech samples as indicative of normal cognition or cognitive impairment. We evaluate our approach on two datasets: one in English and another multilingual, spanning different cognitive assessment tasks. Our results show that the zero-shot AudioLLM approach achieves performance comparable to supervised methods and exhibits promising generalizability and consistency across languages, tasks, and datasets.", "date": "2025-06-20", "categories": []}
{"id": "2506.17409", "title": "Adaptive Control Attention Network for Underwater Acoustic Localization and Domain Adaptation", "abstract": "Localizing acoustic sound sources in the ocean is a challenging task due to the complex and dynamic nature of the environment. Factors such as high background noise, irregular underwater geometries, and varying acoustic properties make accurate localization difficult. To address these obstacles, we propose a multi-branch network architecture designed to accurately predict the distance between a moving acoustic source and a receiver, tested on real-world underwater signal arrays. The network leverages Convolutional Neural Networks (CNNs) for robust spatial feature extraction and integrates Conformers with self-attention mechanism to effectively capture temporal dependencies. Log-mel spectrogram and generalized cross-correlation with phase transform (GCC-PHAT) features are employed as input representations. To further enhance the model performance, we introduce an Adaptive Gain Control (AGC) layer, that adaptively adjusts the amplitude of input features, ensuring consistent energy levels across varying ranges, signal strengths, and noise conditions. We assess the model's generalization capability by training it in one domain and testing it in a different domain, using only a limited amount of data from the test domain for fine-tuning. Our proposed method outperforms state-of-the-art (SOTA) approaches in similar settings, establishing new benchmarks for underwater sound localization.", "date": "2025-06-20", "categories": []}
{"id": "2506.17818", "title": "CultureMERT: Continual Pre-Training for Cross-Cultural Music Representation Learning", "abstract": "Recent advances in music foundation models have improved audio representation learning, yet their effectiveness across diverse musical traditions remains limited. We introduce CultureMERT-95M, a multi-culturally adapted foundation model developed to enhance cross-cultural music representation learning and understanding. To achieve this, we propose a two-stage continual pre-training strategy that integrates learning rate re-warming and re-decaying, enabling stable adaptation even with limited computational resources. Training on a 650-hour multi-cultural data mix, comprising Greek, Turkish, and Indian music traditions, results in an average improvement of 4.9% in ROC-AUC and AP across diverse non-Western music auto-tagging tasks, surpassing prior state-of-the-art, with minimal forgetting on Western-centric benchmarks. We further investigate task arithmetic, an alternative approach to multi-cultural adaptation that merges single-culture adapted models in the weight space. Task arithmetic performs on par with our multi-culturally trained model on non-Western auto-tagging tasks and shows no regression on Western datasets. Cross-cultural evaluation reveals that single-culture models transfer with varying effectiveness across musical traditions, whereas the multi-culturally adapted model achieves the best overall performance. To support research on world music representation learning, we publicly release CultureMERT-95M and CultureMERT-TA-95M, fostering the development of more culturally aware music foundation models.", "date": "2025-06-21", "categories": []}
{"id": "2506.18035", "title": "Splitformer: An improved early-exit architecture for automatic speech recognition on edge devices", "abstract": "The ability to dynamically adjust the computational load of neural models during inference in a resource aware manner is crucial for on-device processing scenarios, characterised by limited and time-varying computational resources. Early-exit architectures represent an elegant and effective solution, since they can process the input with a subset of their layers, exiting at intermediate branches (the upmost layers are hence removed from the model).\n  From a different perspective, for automatic speech recognition applications there are memory-efficient neural architectures that apply variable frame rate analysis, through downsampling/upsampling operations in the middle layers, reducing the overall number of operations and improving significantly the performance on well established benchmarks. One example is the Zipformer. However, these architectures lack the modularity necessary to inject early-exit branches.\n  With the aim of improving the performance in early-exit models, we propose introducing parallel layers in the architecture that process downsampled versions of their inputs. % in conjunction with standard processing layers. We show that in this way the speech recognition performance on standard benchmarks significantly improve, at the cost of a small increase in the overall number of model parameters but without affecting the inference time.", "date": "2025-06-22", "categories": []}
{"id": "2506.14684", "title": "Refining music sample identification with a self-supervised graph neural network", "abstract": "Automatic sample identification (ASID), the detection and identification of portions of audio recordings that have been reused in new musical works, is an essential but challenging task in the field of audio query-based retrieval. While a related task, audio fingerprinting, has made significant progress in accurately retrieving musical content under \"real world\" (noisy, reverberant) conditions, ASID systems struggle to identify samples that have undergone musical modifications. Thus, a system robust to common music production transformations such as time-stretching, pitch-shifting, effects processing, and underlying or overlaying music is an important open challenge.\n  In this work, we propose a lightweight and scalable encoding architecture employing a Graph Neural Network within a contrastive learning framework. Our model uses only 9% of the trainable parameters compared to the current state-of-the-art system while achieving comparable performance, reaching a mean average precision (mAP) of 44.2%.\n  To enhance retrieval quality, we introduce a two-stage approach consisting of an initial coarse similarity search for candidate selection, followed by a cross-attention classifier that rejects irrelevant matches and refines the ranking of retrieved candidates - an essential capability absent in prior models. In addition, because queries in real-world applications are often short in duration, we benchmark our system for short queries using new fine-grained annotations for the Sample100 dataset, which we publish as part of this work.", "date": "2025-06-17", "categories": []}
{"id": "2506.17055", "title": "Universal Music Representations? Evaluating Foundation Models on World Music Corpora", "abstract": "Foundation models have revolutionized music information retrieval, but questions remain about their ability to generalize across diverse musical traditions. This paper presents a comprehensive evaluation of five state-of-the-art audio foundation models across six musical corpora spanning Western popular, Greek, Turkish, and Indian classical traditions. We employ three complementary methodologies to investigate these models' cross-cultural capabilities: probing to assess inherent representations, targeted supervised fine-tuning of 1-2 layers, and multi-label few-shot learning for low-resource scenarios. Our analysis shows varying cross-cultural generalization, with larger models typically outperforming on non-Western music, though results decline for culturally distant traditions. Notably, our approaches achieve state-of-the-art performance on five out of six evaluated datasets, demonstrating the effectiveness of foundation models for world music understanding. We also find that our targeted fine-tuning approach does not consistently outperform probing across all settings, suggesting foundation models already encode substantial musical knowledge. Our evaluation framework and benchmarking results contribute to understanding how far current models are from achieving universal music representations while establishing metrics for future progress.", "date": "2025-06-20", "categories": []}
{"id": "2506.15456", "title": "Factorized RVQ-GAN For Disentangled Speech Tokenization", "abstract": "We propose Hierarchical Audio Codec (HAC), a unified neural speech codec that factorizes its bottleneck into three linguistic levels-acoustic, phonetic, and lexical-within a single model. HAC leverages two knowledge distillation objectives: one from a pre-trained speech encoder (HuBERT) for phoneme-level structure, and another from a text-based encoder (LaBSE) for lexical cues. Experiments on English and multilingual data show that HAC's factorized bottleneck yields disentangled token sets: one aligns with phonemes, while another captures word-level semantics. Quantitative evaluations confirm that HAC tokens preserve naturalness and provide interpretable linguistic information, outperforming single-level baselines in both disentanglement and reconstruction quality. These findings underscore HAC's potential as a unified discrete speech representation, bridging acoustic detail and lexical meaning for downstream speech generation and understanding tasks.", "date": "2025-06-18", "categories": []}
{"id": "2506.13970", "title": "Making deep neural networks work for medical audio: representation, compression and domain adaptation", "abstract": "This thesis addresses the technical challenges of applying machine learning to understand and interpret medical audio signals. The sounds of our lungs, heart, and voice convey vital information about our health. Yet, in contemporary medicine, these sounds are primarily analyzed through auditory interpretation by experts using devices like stethoscopes. Automated analysis offers the potential to standardize the processing of medical sounds, enable screening in low-resource settings where physicians are scarce, and detect subtle patterns that may elude human perception, thereby facilitating early diagnosis and treatment.\n  Focusing on the analysis of infant cry sounds to predict medical conditions, this thesis contributes on four key fronts. First, in low-data settings, we demonstrate that large databases of adult speech can be harnessed through neural transfer learning to develop more accurate and robust models for infant cry analysis. Second, in cost-effective modeling, we introduce an end-to-end model compression approach for recurrent networks using tensor decomposition. Our method requires no post-hoc processing, achieves compression rates of several hundred-fold, and delivers accurate, portable models suitable for resource-constrained devices. Third, we propose novel domain adaptation techniques tailored for audio models and adapt existing methods from computer vision. These approaches address dataset bias and enhance generalization across domains while maintaining strong performance on the original data. Finally, to advance research in this domain, we release a unique, open-source dataset of infant cry sounds, developed in collaboration with clinicians worldwide.\n  This work lays the foundation for recognizing the infant cry as a vital sign and highlights the transformative potential of AI-driven audio monitoring in shaping the future of accessible and affordable healthcare.", "date": "2025-05-24", "categories": []}
{"id": "2409.03377", "title": "aTENNuate: Optimized Real-time Speech Enhancement with Deep SSMs on Raw Audio", "abstract": "We present aTENNuate, a simple deep state-space autoencoder configured for efficient online raw speech enhancement in an end-to-end fashion. The network's performance is primarily evaluated on raw speech denoising, with additional assessments on tasks such as super-resolution and de-quantization. We benchmark aTENNuate on the VoiceBank + DEMAND and the Microsoft DNS1 synthetic test sets. The network outperforms previous real-time denoising models in terms of PESQ score, parameter count, MACs, and latency. Even as a raw waveform processing model, the model maintains high fidelity to the clean signal with minimal audible artifacts. In addition, the model remains performant even when the noisy input is compressed down to 4000Hz and 4 bits, suggesting general speech enhancement capabilities in low-resource environments. Try it out by pip install attenuate", "date": "2024-09-05", "categories": []}
{"id": "2506.12199", "title": "ViSAGe: Video-to-Spatial Audio Generation", "abstract": "Spatial audio is essential for enhancing the immersiveness of audio-visual experiences, yet its production typically demands complex recording systems and specialized expertise. In this work, we address a novel problem of generating first-order ambisonics, a widely used spatial audio format, directly from silent videos. To support this task, we introduce YT-Ambigen, a dataset comprising 102K 5-second YouTube video clips paired with corresponding first-order ambisonics. We also propose new evaluation metrics to assess the spatial aspect of generated audio based on audio energy maps and saliency metrics. Furthermore, we present Video-to-Spatial Audio Generation (ViSAGe), an end-to-end framework that generates first-order ambisonics from silent video frames by leveraging CLIP visual features, autoregressive neural audio codec modeling with both directional and visual guidance. Experimental results demonstrate that ViSAGe produces plausible and coherent first-order ambisonics, outperforming two-stage approaches consisting of video-to-audio generation and audio spatialization. Qualitative examples further illustrate that ViSAGe generates temporally aligned high-quality spatial audio that adapts to viewpoint changes.", "date": "2025-06-13", "categories": []}
{"id": "2506.12222", "title": "SSLAM: Enhancing Self-Supervised Models with Audio Mixtures for Polyphonic Soundscapes", "abstract": "Self-supervised pre-trained audio networks have seen widespread adoption in real-world systems, particularly in multi-modal large language models. These networks are often employed in a frozen state, under the assumption that the SSL pre-training has sufficiently equipped them to handle real-world audio. However, a critical question remains: how well do these models actually perform in real-world conditions, where audio is typically polyphonic and complex, involving multiple overlapping sound sources? Current audio SSL methods are often benchmarked on datasets predominantly featuring monophonic audio, such as environmental sounds, and speech. As a result, the ability of SSL models to generalize to polyphonic audio, a common characteristic in natural scenarios, remains underexplored. This limitation raises concerns about the practical robustness of SSL models in more realistic audio settings. To address this gap, we introduce Self-Supervised Learning from Audio Mixtures (SSLAM), a novel direction in audio SSL research, designed to improve, designed to improve the model's ability to learn from polyphonic data while maintaining strong performance on monophonic data. We thoroughly evaluate SSLAM on standard audio SSL benchmark datasets which are predominantly monophonic and conduct a comprehensive comparative analysis against SOTA methods using a range of high-quality, publicly available polyphonic datasets. SSLAM not only improves model performance on polyphonic audio, but also maintains or exceeds performance on standard audio SSL benchmarks. Notably, it achieves up to a 3.9\\% improvement on the AudioSet-2M (AS-2M), reaching a mean average precision (mAP) of 50.2. For polyphonic datasets, SSLAM sets new SOTA in both linear evaluation and fine-tuning regimes with performance improvements of up to 9.1\\% (mAP).", "date": "2025-06-13", "categories": []}
{"id": "2310.03480", "title": "The ICASSP SP Cadenza Challenge: Music Demixing/Remixing for Hearing Aids", "abstract": "This paper reports on the design and results of the 2024 ICASSP SP Cadenza Challenge: Music Demixing/Remixing for Hearing Aids. The Cadenza project is working to enhance the audio quality of music for those with a hearing loss. The scenario for the challenge was listening to stereo reproduction over loudspeakers via hearing aids. The task was to: decompose pop/rock music into vocal, drums, bass and other (VDBO); rebalance the different tracks with specified gains and then remixing back to stereo. End-to-end approaches were also accepted. 17 systems were submitted by 11 teams. Causal systems performed poorer than non-causal approaches. 9 systems beat the baseline. A common approach was to fine-tuning pretrained demixing models. The best approach used an ensemble of models.", "date": "2023-10-05", "categories": []}
{"id": "2410.17834", "title": "Non-intrusive Speech Quality Assessment with Diffusion Models Trained on Clean Speech", "abstract": "Diffusion models have found great success in generating high quality, natural samples of speech, but their potential for density estimation for speech has so far remained largely unexplored. In this work, we leverage an unconditional diffusion model trained only on clean speech for the assessment of speech quality. We show that the quality of a speech utterance can be assessed by estimating the likelihood of a corresponding sample in the terminating Gaussian distribution, obtained via a deterministic noising process. The resulting method is purely unsupervised, trained only on clean speech, and therefore does not rely on annotations. Our diffusion-based approach leverages clean speech priors to assess quality based on how the input relates to the learned distribution of clean data. Our proposed log-likelihoods show promising results, correlating well with intrusive speech quality metrics and showing the best correlation with human scores in a listening experiment.", "date": "2024-10-23", "categories": []}
{"id": "2503.16862", "title": "Improving Acoustic Scene Classification with City Features", "abstract": "Acoustic scene recordings are often collected from a diverse range of cities. Most existing acoustic scene classification (ASC) approaches focus on identifying common acoustic scene patterns across cities to enhance generalization. However, the potential acoustic differences introduced by city-specific environmental and cultural factors are overlooked. In this paper, we hypothesize that the city-specific acoustic features are beneficial for the ASC task rather than being treated as noise or bias. To this end, we propose City2Scene, a novel framework that leverages city features to improve ASC. Unlike conventional approaches that may discard or suppress city information, City2Scene transfers the city-specific knowledge from pre-trained city classification models to scene classification model using knowledge distillation. We evaluate City2Scene on three datasets of DCASE Challenge Task 1, which include both scene and city labels. Experimental results demonstrate that city features provide valuable information for classifying scenes. By distilling city-specific knowledge, City2Scene effectively improves accuracy across a variety of lightweight CNN backbones, achieving competitive performance to the top-ranked solutions of DCASE Challenge in recent years.", "date": "2025-03-21", "categories": []}
{"id": "2505.17076", "title": "Impact of Frame Rates on Speech Tokenizer: A Case Study on Mandarin and English", "abstract": "The speech tokenizer plays a crucial role in recent speech tasks, generally serving as a bridge between speech signals and language models. While low-frame-rate codecs are widely employed as speech tokenizers, the impact of frame rates on speech tokens remains underexplored. In this study, we investigate how varying frame rates affect speech tokenization by examining Mandarin and English, two typologically distinct languages. We encode speech at different frame rates and evaluate the resulting semantic tokens in the speech recognition task. Our findings reveal that frame rate variations influence speech tokenization differently for each language, highlighting the interplay between frame rates, phonetic density, and language-specific acoustic features. The results provide insights into optimizing frame rate selection for speech tokenizers, with implications for automatic speech recognition, text-to-speech, and other speech-related applications.", "date": "2025-05-20", "categories": []}
{"id": "2506.08967", "title": "Step-Audio-AQAA: a Fully End-to-End Expressive Large Audio Language Model", "abstract": "Large Audio-Language Models (LALMs) have significantly advanced intelligent human-computer interaction, yet their reliance on text-based outputs limits their ability to generate natural speech responses directly, hindering seamless audio interactions. To address this, we introduce Step-Audio-AQAA, a fully end-to-end LALM designed for Audio Query-Audio Answer (AQAA) tasks. The model integrates a dual-codebook audio tokenizer for linguistic and semantic feature extraction, a 130-billion-parameter backbone LLM and a neural vocoder for high-fidelity speech synthesis. Our post-training approach employs interleaved token-output of text and audio to enhance semantic coherence and combines Direct Preference Optimization (DPO) with model merge to improve performance. Evaluations on the StepEval-Audio-360 benchmark demonstrate that Step-Audio-AQAA excels especially in speech control, outperforming the state-of-art LALMs in key areas. This work contributes a promising solution for end-to-end LALMs and highlights the critical role of token-based vocoder in enhancing overall performance for AQAA tasks.", "date": "2025-06-10", "categories": []}
{"id": "2506.11071", "title": "Embedded Acoustic Intelligence for Automotive Systems", "abstract": "Transforming sound insights into actionable streams of data, this abstract leverages findings from degree thesis research to enhance automotive system intelligence, enabling us to address road type [1].By extracting and interpreting acoustic signatures from microphones installed within the wheelbase of a car, we focus on classifying road type.Utilizing deep neural networks and feature extraction powered by pre-trained models from the Open AI ecosystem (via Hugging Face [2]), our approach enables Autonomous Driving and Advanced Driver- Assistance Systems (AD/ADAS) to anticipate road surfaces, support adaptive learning for active road noise cancellation, and generate valuable insights for urban planning. The results of this study were specifically captured to support a compelling business case for next-generation automotive systems. This forward-looking approach not only promises to redefine passenger comfort and improve vehicle safety, but also paves the way for intelligent, data-driven urban road management, making the future of mobility both achievable and sustainable.", "date": "2025-06-02", "categories": []}
{"id": "2506.11747", "title": "Enabling automatic transcription of child-centered audio recordings from real-world environments", "abstract": "Longform audio recordings obtained with microphones worn by children-also known as child-centered daylong recordings-have become a standard method for studying children's language experiences and their impact on subsequent language development. Transcripts of longform speech audio would enable rich analyses at various linguistic levels, yet the massive scale of typical longform corpora prohibits comprehensive manual annotation. At the same time, automatic speech recognition (ASR)-based transcription faces significant challenges due to the noisy, unconstrained nature of real-world audio, and no existing study has successfully applied ASR to transcribe such data. However, previous attempts have assumed that ASR must process each longform recording in its entirety. In this work, we present an approach to automatically detect those utterances in longform audio that can be reliably transcribed with modern ASR systems, allowing automatic and relatively accurate transcription of a notable proportion of all speech in typical longform data. We validate the approach on four English longform audio corpora, showing that it achieves a median word error rate (WER) of 0% and a mean WER of 18% when transcribing 13% of the total speech in the dataset. In contrast, transcribing all speech without any filtering yields a median WER of 52% and a mean WER of 51%. We also compare word log-frequencies derived from the automatic transcripts with those from manual annotations and show that the frequencies correlate at r = 0.92 (Pearson) for all transcribed words and r = 0.98 for words that appear at least five times in the automatic transcripts. Overall, the work provides a concrete step toward increasingly detailed automated linguistic analyses of child-centered longform audio.", "date": "2025-06-13", "categories": []}
{"id": "2503.11373", "title": "Exploring Performance-Complexity Trade-Offs in Sound Event Detection Models", "abstract": "We target the problem of developing new low-complexity networks for the sound event detection task. Our goal is to meticulously analyze the performance-complexity trade-off, aiming to be competitive with the large state-of-the-art models, at a fraction of the computational requirements. We find that low-complexity convolutional models previously proposed for audio tagging can be effectively adapted for event detection (which requires frame-wise prediction) by adjusting convolutional strides, removing the global pooling, and, importantly, adding a sequence model before the (now frame-wise) classification heads. Systematic experiments reveal that the best choice for the sequence model type depends on which complexity metric is most important for the given application. We also investigate the impact of enhanced training strategies such as knowledge distillation. In the end, we show that combined with an optimized training strategy, we can reach event detection performance comparable to state-of-the-art transformers while requiring only around 5% of the parameters. We release all our pre-trained models and the code for reproducing this work to support future research in low-complexity sound event detection at https://github.com/theMoro/EfficientSED.", "date": "2025-03-14", "categories": []}
{"id": "2506.10312", "title": "AC/DC: LLM-based Audio Comprehension via Dialogue Continuation", "abstract": "We propose an instruction-following audio comprehension model that leverages the dialogue continuation ability of large language models (LLMs). Instead of directly generating target captions in training data, the proposed method trains a model to produce responses as if the input caption triggered a dialogue. This dialogue continuation training mitigates the caption variation problem. Learning to continue a dialogue effectively captures the caption's meaning beyond its surface-level words. As a result, our model enables zero-shot instruction-following capability without multitask instruction tuning, even trained solely on audio captioning datasets. Experiments on AudioCaps, WavCaps, and Clotho datasets with AudioBench audio-scene question-answering tests demonstrate our model's ability to follow various unseen instructions.", "date": "2025-06-12", "categories": []}
{"id": "2402.03710", "title": "Listen, Chat, and Remix: Text-Guided Soundscape Remixing for Enhanced Auditory Experience", "abstract": "In daily life, we encounter a variety of sounds, both desirable and undesirable, with limited control over their presence and volume. Our work introduces \"Listen, Chat, and Remix\" (LCR), a novel multimodal sound remixer that controls each sound source in a mixture based on user-provided text instructions. LCR distinguishes itself with a user-friendly text interface and its unique ability to remix multiple sound sources simultaneously within a mixture, without needing to separate them. Users input open-vocabulary text prompts, which are interpreted by a large language model to create a semantic filter for remixing the sound mixture. The system then decomposes the mixture into its components, applies the semantic filter, and reassembles filtered components back to the desired output. We developed a 160-hour dataset with over 100k mixtures, including speech and various audio sources, along with text prompts for diverse remixing tasks including extraction, removal, and volume control of single or multiple sources. Our experiments demonstrate significant improvements in signal quality across all remixing tasks and robust performance in zero-shot scenarios with varying numbers and types of sound sources. An audio demo is available at: https://listenchatremix.github.io/demo.", "date": "2024-02-06", "categories": []}
{"id": "2410.16785", "title": "Annotation-Free MIDI-to-Audio Synthesis via Concatenative Synthesis and Generative Refinement", "abstract": "Recent MIDI-to-audio synthesis methods using deep neural networks have successfully generated high-quality, expressive instrumental tracks. However, these methods require MIDI annotations for supervised training, limiting the diversity of instrument timbres and expression styles in the output. We propose CoSaRef, a MIDI-to-audio synthesis method that does not require MIDI-audio paired datasets. CoSaRef first generates a synthetic audio track using concatenative synthesis based on MIDI input, then refines it with a diffusion-based deep generative model trained on datasets without MIDI annotations. This approach improves the diversity of timbres and expression styles. Additionally, it allows detailed control over timbres and expression through audio sample selection and extra MIDI design, similar to traditional functions in digital audio workstations. Experiments showed that CoSaRef could generate realistic tracks while preserving fine-grained timbre control via one-shot samples. Moreover, despite not being supervised on MIDI annotation, CoSaRef outperformed the state-of-the-art timbre-controllable method based on MIDI supervision in both objective and subjective evaluation.", "date": "2024-10-22", "categories": []}
{"id": "2506.09709", "title": "Training-Free Voice Conversion with Factorized Optimal Transport", "abstract": "This paper introduces Factorized MKL-VC, a training-free modification for kNN-VC pipeline. In contrast with original pipeline, our algorithm performs high quality any-to-any cross-lingual voice conversion with only 5 second of reference audio. MKL-VC replaces kNN regression with a factorized optimal transport map in WavLM embedding subspaces, derived from Monge-Kantorovich Linear solution. Factorization addresses non-uniform variance across dimensions, ensuring effective feature transformation. Experiments on LibriSpeech and FLEURS datasets show MKL-VC significantly improves content preservation and robustness with short reference audio, outperforming kNN-VC. MKL-VC achieves performance comparable to FACodec, especially in cross-lingual voice conversion domain.", "date": "2025-06-11", "categories": []}
{"id": "2506.09989", "title": "Hearing Hands: Generating Sounds from Physical Interactions in 3D Scenes", "abstract": "We study the problem of making 3D scene reconstructions interactive by asking the following question: can we predict the sounds of human hands physically interacting with a scene? First, we record a video of a human manipulating objects within a 3D scene using their hands. We then use these action-sound pairs to train a rectified flow model to map 3D hand trajectories to their corresponding audio. At test time, a user can query the model for other actions, parameterized as sequences of hand poses, to estimate their corresponding sounds. In our experiments, we find that our generated sounds accurately convey material properties and actions, and that they are often indistinguishable to human observers from real sounds. Project page: https://www.yimingdou.com/hearing_hands/", "date": "2025-06-11", "categories": []}
{"id": "2506.08999", "title": "Employing self-supervised learning models for cross-linguistic child speech maturity classification", "abstract": "Speech technology systems struggle with many downstream tasks for child speech due to small training corpora and the difficulties that child speech pose. We apply a novel dataset, SpeechMaturity, to state-of-the-art transformer models to address a fundamental classification task: identifying child vocalizations. Unlike previous corpora, our dataset captures maximally ecologically-valid child vocalizations across an unprecedented sample, comprising children acquiring 25+ languages in the U.S., Bolivia, Vanuatu, Papua New Guinea, Solomon Islands, and France. The dataset contains 242,004 labeled vocalizations, magnitudes larger than previous work. Models were trained to distinguish between cry, laughter, mature (consonant+vowel), and immature speech (just consonant or vowel). Models trained on the dataset outperform state-of-the-art models trained on previous datasets, achieved classification accuracy comparable to humans, and were robust across rural and urban settings.", "date": "2025-06-10", "categories": []}
{"id": "2501.13887", "title": "What Does an Audio Deepfake Detector Focus on? A Study in the Time Domain", "abstract": "Adding explanations to audio deepfake detection (ADD) models will boost their real-world application by providing insight on the decision making process. In this paper, we propose a relevancy-based explainable AI (XAI) method to analyze the predictions of transformer-based ADD models. We compare against standard Grad-CAM and SHAP-based methods, using quantitative faithfulness metrics as well as a partial spoof test, to comprehensively analyze the relative importance of different temporal regions in an audio. We consider large datasets, unlike previous works where only limited utterances are studied, and find that the XAI methods differ in their explanations. The proposed relevancy-based XAI method performs the best overall on a variety of metrics. Further investigation on the relative importance of speech/non-speech, phonetic content, and voice onsets/offsets suggest that the XAI results obtained from analyzing limited utterances don't necessarily hold when evaluated on large datasets.", "date": "2025-01-23", "categories": []}
{"id": "2506.07118", "title": "RBA-FE: A Robust Brain-Inspired Audio Feature Extractor for Depression Diagnosis", "abstract": "This article proposes a robust brain-inspired audio feature extractor (RBA-FE) model for depression diagnosis, using an improved hierarchical network architecture. Most deep learning models achieve state-of-the-art performance for image-based diagnostic tasks, ignoring the counterpart audio features. In order to tailor the noise challenge, RBA-FE leverages six acoustic features extracted from the raw audio, capturing both spatial characteristics and temporal dependencies. This hybrid attribute helps alleviate the precision limitation in audio feature extraction within other learning models like deep residual shrinkage networks. To deal with the noise issues, our model incorporates an improved spiking neuron model, called adaptive rate smooth leaky integrate-and-fire (ARSLIF). The ARSLIF model emulates the mechanism of ``retuning of cellular signal selectivity\" in the brain attention systems, which enhances the model robustness against environmental noises in audio data. Experimental results demonstrate that RBA-FE achieves state-of-the-art accuracy on the MODMA dataset, respectively with 0.8750, 0.8974, 0.8750 and 0.8750 in precision, accuracy, recall and F1 score. Extensive experiments on the AVEC2014 and DAIC-WOZ datasets both show enhancements in noise robustness. It is further indicated by comparison that the ARSLIF neuron model suggest the abnormal firing pattern within the feature extraction on depressive audio data, offering brain-inspired interpretability.", "date": "2025-06-08", "categories": []}
{"id": "2506.07199", "title": "Audio synthesizer inversion in symmetric parameter spaces with approximately equivariant flow matching", "abstract": "Many audio synthesizers can produce the same signal given different parameter configurations, meaning the inversion from sound to parameters is an inherently ill-posed problem. We show that this is largely due to intrinsic symmetries of the synthesizer, and focus in particular on permutation invariance. First, we demonstrate on a synthetic task that regressing point estimates under permutation symmetry degrades performance, even when using a permutation-invariant loss function or symmetry-breaking heuristics. Then, viewing equivalent solutions as modes of a probability distribution, we show that a conditional generative model substantially improves performance. Further, acknowledging the invariance of the implicit parameter distribution, we find that performance is further improved by using a permutation equivariant continuous normalizing flow. To accommodate intricate symmetries in real synthesizers, we also propose a relaxed equivariance strategy that adaptively discovers relevant symmetries from data. Applying our method to Surge XT, a full-featured open source synthesizer used in real world audio production, we find our method outperforms regression and generative baselines across audio reconstruction metrics.", "date": "2025-06-08", "categories": []}
{"id": "2411.14842", "title": "Who Can Withstand Chat-Audio Attacks? An Evaluation Benchmark for Large Audio-Language Models", "abstract": "Adversarial audio attacks pose a significant threat to the growing use of large audio-language models (LALMs) in voice-based human-machine interactions. While existing research focused on model-specific adversarial methods, real-world applications demand a more generalizable and universal approach to audio adversarial attacks. In this paper, we introduce the Chat-Audio Attacks (CAA) benchmark including four distinct types of audio attacks, which aims to explore the vulnerabilities of LALMs to these audio attacks in conversational scenarios. To evaluate the robustness of LALMs, we propose three evaluation strategies: Standard Evaluation, utilizing traditional metrics to quantify model performance under attacks; GPT-4o-Based Evaluation, which simulates real-world conversational complexities; and Human Evaluation, offering insights into user perception and trust. We evaluate six state-of-the-art LALMs with voice interaction capabilities, including Gemini-1.5-Pro, GPT-4o, and others, using three distinct evaluation methods on the CAA benchmark. Our comprehensive analysis reveals the impact of four types of audio attacks on the performance of these models, demonstrating that GPT-4o exhibits the highest level of resilience. Our data can be accessed via the following link: \\href{https://github.com/crystraldo/CAA}{CAA}.", "date": "2024-11-22", "categories": []}
{"id": "2502.18952", "title": "DualSpec: Text-to-spatial-audio Generation via Dual-Spectrogram Guided Diffusion Model", "abstract": "Text-to-audio (TTA), which generates audio signals from textual descriptions, has received huge attention in recent years. However, recent works focused on text to monaural audio only. As we know, spatial audio provides more immersive auditory experience than monaural audio, e.g. in virtual reality. To address this issue, we propose a text-to-spatial-audio (TTSA) generation framework named DualSpec. Specifically, it first trains variational autoencoders (VAEs) for extracting the latent acoustic representations from sound event audio. Then, given text that describes sound events and event directions, the proposed method uses the encoder of a pretrained large language model to transform the text into text features. Finally, it trains a diffusion model from the latent acoustic representations and text features for the spatial audio generation. In the inference stage, only the text description is needed to generate spatial audio. Particularly, to improve the synthesis quality and azimuth accuracy of the spatial sound events simultaneously, we propose to use two kinds of acoustic features. One is the Mel spectrograms which is good for improving the synthesis quality, and the other is the short-time Fourier transform spectrograms which is good at improving the azimuth accuracy. We provide a pipeline of constructing spatial audio dataset with text prompts, for the training of the VAEs and diffusion model. We also introduce new spatial-aware evaluation metrics to quantify the azimuth errors of the generated spatial audio recordings. Experimental results demonstrate that the proposed method can generate spatial audio with high directional and event consistency.", "date": "2025-02-26", "categories": []}
{"id": "2506.05593", "title": "Improving Neural Diarization through Speaker Attribute Attractors and Local Dependency Modeling", "abstract": "In recent years, end-to-end approaches have made notable progress in addressing the challenge of speaker diarization, which involves segmenting and identifying speakers in multi-talker recordings. One such approach, Encoder-Decoder Attractors (EDA), has been proposed to handle variable speaker counts as well as better guide the network during training. In this study, we extend the attractor paradigm by moving beyond direct speaker modeling and instead focus on representing more detailed `speaker attributes' through a multi-stage process of intermediate representations. Additionally, we enhance the architecture by replacing transformers with conformers, a convolution-augmented transformer, to model local dependencies. Experiments demonstrate improved diarization performance on the CALLHOME dataset.", "date": "2025-06-05", "categories": []}
{"id": "2504.10746", "title": "Hearing Anywhere in Any Environment", "abstract": "In mixed reality applications, a realistic acoustic experience in spatial environments is as crucial as the visual experience for achieving true immersion. Despite recent advances in neural approaches for Room Impulse Response (RIR) estimation, most existing methods are limited to the single environment on which they are trained, lacking the ability to generalize to new rooms with different geometries and surface materials. We aim to develop a unified model capable of reconstructing the spatial acoustic experience of any environment with minimum additional measurements. To this end, we present xRIR, a framework for cross-room RIR prediction. The core of our generalizable approach lies in combining a geometric feature extractor, which captures spatial context from panorama depth images, with a RIR encoder that extracts detailed acoustic features from only a few reference RIR samples. To evaluate our method, we introduce ACOUSTICROOMS, a new dataset featuring high-fidelity simulation of over 300,000 RIRs from 260 rooms. Experiments show that our method strongly outperforms a series of baselines. Furthermore, we successfully perform sim-to-real transfer by evaluating our model on four real-world environments, demonstrating the generalizability of our approach and the realism of our dataset.", "date": "2025-04-14", "categories": []}
{"id": "2506.04376", "title": "Domain Adaptation Method and Modality Gap Impact in Audio-Text Models for Prototypical Sound Classification", "abstract": "Audio-text models are widely used in zero-shot environmental sound classification as they alleviate the need for annotated data. However, we show that their performance severely drops in the presence of background sound sources. Our analysis reveals that this degradation is primarily driven by SNR levels of background soundscapes, and independent of background type. To address this, we propose a novel method that quantifies and integrates the contribution of background sources into the classification process, improving performance without requiring model retraining. Our domain adaptation technique enhances accuracy across various backgrounds and SNR conditions. Moreover, we analyze the modality gap between audio and text embeddings, showing that narrowing this gap improves classification performance. The method generalizes effectively across state-of-the-art prototypical approaches, showcasing its scalability and robustness for diverse environments.", "date": "2025-06-04", "categories": []}
{"id": "2506.04635", "title": "ViCocktail: Automated Multi-Modal Data Collection for Vietnamese Audio-Visual Speech Recognition", "abstract": "Audio-Visual Speech Recognition (AVSR) has gained significant attention recently due to its robustness against noise, which often challenges conventional speech recognition systems that rely solely on audio features. Despite this advantage, AVSR models remain limited by the scarcity of extensive datasets, especially for most languages beyond English. Automated data collection offers a promising solution. This work presents a practical approach to generate AVSR datasets from raw video, refining existing techniques for improved efficiency and accessibility. We demonstrate its broad applicability by developing a baseline AVSR model for Vietnamese. Experiments show the automatically collected dataset enables a strong baseline, achieving competitive performance with robust ASR in clean conditions and significantly outperforming them in noisy environments like cocktail parties. This efficient method provides a pathway to expand AVSR to more languages, particularly under-resourced ones.", "date": "2025-06-05", "categories": []}
{"id": "2506.04852", "title": "Improving AI-generated music with user-guided training", "abstract": "AI music generation has advanced rapidly, with models like diffusion and autoregressive algorithms enabling high-fidelity outputs. These tools can alter styles, mix instruments, or isolate them. Since sound can be visualized as spectrograms, image-generation algorithms can be applied to generate novel music. However, these algorithms are typically trained on fixed datasets, which makes it challenging for them to interpret and respond to user input accurately. This is especially problematic because music is highly subjective and requires a level of personalization that image generation does not provide. In this work, we propose a human-computation approach to gradually improve the performance of these algorithms based on user interactions. The human-computation element involves aggregating and selecting user ratings to use as the loss function for fine-tuning the model. We employ a genetic algorithm that incorporates user feedback to enhance the baseline performance of a model initially trained on a fixed dataset. The effectiveness of this approach is measured by the average increase in user ratings with each iteration. In the pilot test, the first iteration showed an average rating increase of 0.2 compared to the baseline. The second iteration further improved upon this, achieving an additional increase of 0.39 over the first iteration.", "date": "2025-06-05", "categories": []}
{"id": "2502.12031", "title": "Masked Latent Prediction and Classification for Self-Supervised Audio Representation Learning", "abstract": "Recently, self-supervised learning methods based on masked latent prediction have proven to encode input data into powerful representations. However, during training, the learned latent space can be further transformed to extract higher-level information that could be more suited for downstream classification tasks. Therefore, we propose a new method: MAsked latenT Prediction And Classification (MATPAC), which is trained with two pretext tasks solved jointly. As in previous work, the first pretext task is a masked latent prediction task, ensuring a robust input representation in the latent space. The second one is unsupervised classification, which utilises the latent representations of the first pretext task to match probability distributions between a teacher and a student. We validate the MATPAC method by comparing it to other state-of-the-art proposals and conducting ablations studies. MATPAC reaches state-of-the-art self-supervised learning results on reference audio classification datasets such as OpenMIC, GTZAN, ESC-50 and US8K and outperforms comparable supervised methods results for musical auto-tagging on Magna-tag-a-tune.", "date": "2025-02-17", "categories": []}
{"id": "2505.14470", "title": "PAST: Phonetic-Acoustic Speech Tokenizer", "abstract": "We present PAST, a novel end-to-end framework that jointly models phonetic information alongside signal reconstruction, eliminating the need for external pretrained models. Unlike previous approaches that rely on pretrained self-supervised models, PAST employs supervised phonetic data, directly integrating domain knowledge into the tokenization process via auxiliary tasks. Additionally, we introduce a streamable, causal variant of PAST, enabling real-time speech applications. Results demonstrate that PAST surpasses existing evaluated baseline tokenizers across common evaluation metrics, including phonetic representation and speech reconstruction. Notably, PAST also achieves superior performance when serving as a speech representation for speech language models, further highlighting its effectiveness as a foundation for spoken language generation. To foster further research, we release the full implementation. For code, model checkpoints, and samples see: https://pages.cs.huji.ac.il/adiyoss-lab/PAST", "date": "2025-05-20", "categories": []}
{"id": "2506.04214", "title": "Sounding that Object: Interactive Object-Aware Image to Audio Generation", "abstract": "Generating accurate sounds for complex audio-visual scenes is challenging, especially in the presence of multiple objects and sound sources. In this paper, we propose an {\\em interactive object-aware audio generation} model that grounds sound generation in user-selected visual objects within images. Our method integrates object-centric learning into a conditional latent diffusion model, which learns to associate image regions with their corresponding sounds through multi-modal attention. At test time, our model employs image segmentation to allow users to interactively generate sounds at the {\\em object} level. We theoretically validate that our attention mechanism functionally approximates test-time segmentation masks, ensuring the generated audio aligns with selected objects. Quantitative and qualitative evaluations show that our model outperforms baselines, achieving better alignment between objects and their associated sounds. Project page: https://tinglok.netlify.app/files/avobject/", "date": "2025-06-04", "categories": []}
{"id": "2503.07352", "title": "Score-informed Music Source Separation: Improving Synthetic-to-real Generalization in Classical Music", "abstract": "Music source separation is the task of separating a mixture of instruments into constituent tracks. Music source separation models are typically trained using only audio data, although additional information can be used to improve the model's separation capability. In this paper, we propose two ways of using musical scores to aid music source separation: a score-informed model where the score is concatenated with the magnitude spectrogram of the audio mixture as the input of the model, and a model where we use only the score to calculate the separation mask. We train our models on synthetic data in the SynthSOD dataset and evaluate our methods on the URMP and Aalto anechoic orchestra datasets, comprised of real recordings. The score-informed model improves separation results compared to a baseline approach, but struggles to generalize from synthetic to real data, whereas the score-only model shows a clear improvement in synthetic-to-real generalization.", "date": "2025-03-10", "categories": []}
{"id": "2504.14906", "title": "OmniAudio: Generating Spatial Audio from 360-Degree Video", "abstract": "Traditional video-to-audio generation techniques primarily focus on perspective video and non-spatial audio, often missing the spatial cues necessary for accurately representing sound sources in 3D environments. To address this limitation, we introduce a novel task, 360V2SA, to generate spatial audio from 360-degree videos, specifically producing First-order Ambisonics (FOA) audio - a standard format for representing 3D spatial audio that captures sound directionality and enables realistic 3D audio reproduction. We first create Sphere360, a novel dataset tailored for this task that is curated from real-world data. We also design an efficient semi-automated pipeline for collecting and cleaning paired video-audio data. To generate spatial audio from 360-degree video, we propose a novel framework OmniAudio, which leverages self-supervised pre-training using both spatial audio data (in FOA format) and large-scale non-spatial data. Furthermore, OmniAudio features a dual-branch framework that utilizes both panoramic and perspective video inputs to capture comprehensive local and global information from 360-degree videos. Experimental results demonstrate that OmniAudio achieves state-of-the-art performance across both objective and subjective metrics on Sphere360. Code and datasets are available at https://github.com/liuhuadai/OmniAudio. The project website is available at https://OmniAudio-360V2SA.github.io.", "date": "2025-04-21", "categories": []}
{"id": "2506.02085", "title": "Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion", "abstract": "Audio deepfakes are acquiring an unprecedented level of realism with advanced AI. While current research focuses on discerning real speech from spoofed speech, tracing the source system is equally crucial. This work proposes a novel audio source tracing system combining deep metric multi-class N-pair loss with Real Emphasis and Fake Dispersion framework, a Conformer classification network, and ensemble score-embedding fusion. The N-pair loss improves discriminative ability, while Real Emphasis and Fake Dispersion enhance robustness by focusing on differentiating real and fake speech patterns. The Conformer network captures both global and local dependencies in the audio signal, crucial for source tracing. The proposed ensemble score-embedding fusion shows an optimal trade-off between in-domain and out-of-domain source tracing scenarios. We evaluate our method using Frechet Distance and standard metrics, demonstrating superior performance in source tracing over the baseline system.", "date": "2025-06-02", "categories": []}
{"id": "2506.02091", "title": "Comparison of spectrogram scaling in multi-label Music Genre Recognition", "abstract": "As the accessibility and ease-of-use of digital audio workstations increases, so does the quantity of music available to the average listener; additionally, differences between genres are not always well defined and can be abstract, with widely varying combinations of genres across individual records. In this article, multiple preprocessing methods and approaches to model training are described and compared, accounting for the eclectic nature of today's albums. A custom, manually labeled dataset of more than 18000 entries has been used to perform the experiments.", "date": "2025-06-02", "categories": []}
{"id": "2506.02414", "title": "StarVC: A Unified Auto-Regressive Framework for Joint Text and Speech Generation in Voice Conversion", "abstract": "Voice Conversion (VC) modifies speech to match a target speaker while preserving linguistic content. Traditional methods usually extract speaker information directly from speech while neglecting the explicit utilization of linguistic content. Since VC fundamentally involves disentangling speaker identity from linguistic content, leveraging structured semantic features could enhance conversion performance. However, previous attempts to incorporate semantic features into VC have shown limited effectiveness, motivating the integration of explicit text modeling. We propose StarVC, a unified autoregressive VC framework that first predicts text tokens before synthesizing acoustic features. The experiments demonstrate that StarVC outperforms conventional VC methods in preserving both linguistic content (i.e., WER and CER) and speaker characteristics (i.e., SECS and MOS). Audio demo can be found at: https://thuhcsi.github.io/StarVC/.", "date": "2025-06-03", "categories": []}
{"id": "2506.02590", "title": "Synthetic Speech Source Tracing using Metric Learning", "abstract": "This paper addresses source tracing in synthetic speech-identifying generative systems behind manipulated audio via speaker recognition-inspired pipelines. While prior work focuses on spoofing detection, source tracing lacks robust solutions. We evaluate two approaches: classification-based and metric-learning. We tested our methods on the MLAADv5 benchmark using ResNet and self-supervised learning (SSL) backbones. The results show that ResNet achieves competitive performance with the metric learning approach, matching and even exceeding SSL-based systems. Our work demonstrates ResNet's viability for source tracing while underscoring the need to optimize SSL representations for this task. Our work bridges speaker recognition methodologies with audio forensic challenges, offering new directions for combating synthetic media manipulation.", "date": "2025-06-03", "categories": []}
{"id": "2506.02610", "title": "Speaker Diarization with Overlapping Community Detection Using Graph Attention Networks and Label Propagation Algorithm", "abstract": "In speaker diarization, traditional clustering-based methods remain widely used in real-world applications. However, these methods struggle with the complex distribution of speaker embeddings and overlapping speech segments. To address these limitations, we propose an Overlapping Community Detection method based on Graph Attention networks and the Label Propagation Algorithm (OCDGALP). The proposed framework comprises two key components: (1) a graph attention network that refines speaker embeddings and node connections by aggregating information from neighboring nodes, and (2) a label propagation algorithm that assigns multiple community labels to each node, enabling simultaneous clustering and overlapping community detection. Experimental results show that the proposed method significantly reduces the Diarization Error Rate (DER), achieving a state-of-the-art 15.94% DER on the DIHARD-III dataset without oracle Voice Activity Detection (VAD), and an impressive 11.07% with oracle VAD.", "date": "2025-06-03", "categories": []}
{"id": "2409.09401", "title": "Towards Diverse and Efficient Audio Captioning via Diffusion Models", "abstract": "We introduce Diffusion-based Audio Captioning (DAC), a non-autoregressive diffusion model tailored for diverse and efficient audio captioning. Although existing captioning models relying on language backbones have achieved remarkable success in various captioning tasks, their insufficient performance in terms of generation speed and diversity impede progress in audio understanding and multimedia applications. Our diffusion-based framework offers unique advantages stemming from its inherent stochasticity and holistic context modeling in captioning. Through rigorous evaluation, we demonstrate that DAC not only achieves SOTA performance levels compared to existing benchmarks in the caption quality, but also significantly outperforms them in terms of generation speed and diversity. The success of DAC illustrates that text generation can also be seamlessly integrated with audio and visual generation tasks using a diffusion backbone, paving the way for a unified, audio-related generative model across different modalities.", "date": "2024-09-14", "categories": []}
{"id": "2505.01199", "title": "CaReAQA: A Cardiac and Respiratory Audio Question Answering Model for Open-Ended Diagnostic Reasoning", "abstract": "Medical audio signals, such as heart and lung sounds, play a crucial role in clinical diagnosis. However, analyzing these signals remains challenging: traditional methods rely on handcrafted features or supervised deep learning models that demand extensive labeled datasets, limiting their scalability and applicability. To address these issues, we propose CaReAQA, an audio-language model that integrates a foundation audio model with the reasoning capabilities of large language models, enabling clinically relevant, open-ended diagnostic responses. Alongside CaReAQA, we introduce CaReSound, a benchmark dataset of annotated medical audio recordings enriched with metadata and paired question-answer examples, intended to drive progress in diagnostic reasoning research. Evaluation results show that CaReAQA achieves 86.2% accuracy on open-ended diagnostic reasoning tasks, outperforming baseline models. It also generalizes well to closed-ended classification tasks, achieving an average accuracy of 56.9% on unseen datasets. Our findings show how audio-language integration and reasoning advances medical diagnostics, enabling efficient AI systems for clinical decision support.", "date": "2025-05-02", "categories": []}
{"id": "2505.14862", "title": "Replay Attacks Against Audio Deepfake Detection", "abstract": "We show how replay attacks undermine audio deepfake detection: By playing and re-recording deepfake audio through various speakers and microphones, we make spoofed samples appear authentic to the detection model. To study this phenomenon in more detail, we introduce ReplayDF, a dataset of recordings derived from M-AILABS and MLAAD, featuring 109 speaker-microphone combinations across six languages and four TTS models. It includes diverse acoustic conditions, some highly challenging for detection. Our analysis of six open-source detection models across five datasets reveals significant vulnerability, with the top-performing W2V2-AASIST model's Equal Error Rate (EER) surging from 4.7% to 18.2%. Even with adaptive Room Impulse Response (RIR) retraining, performance remains compromised with an 11.0% EER. We release ReplayDF for non-commercial research use.", "date": "2025-05-20", "categories": []}
{"id": "2505.17446", "title": "Exploring the Effect of Segmentation and Vocabulary Size on Speech Tokenization for Speech Language Models", "abstract": "The purpose of speech tokenization is to transform a speech signal into a sequence of discrete representations, serving as the foundation for speech language models (SLMs). While speech tokenization has many options, their effect on the performance of SLMs remains unclear. This paper investigates two key aspects of speech tokenization: the segmentation width and the cluster size of discrete units. First, we segment speech signals into fixed/variable widths and pooled representations. We then train K-means models in multiple cluster sizes. Through the evaluation on zero-shot spoken language understanding benchmarks, we find the positive effect of moderately coarse segmentation and bigger cluster size. Notably, among the best-performing models, the most efficient one achieves a 50% reduction in training data and a 70% decrease in training runtime. Our analysis highlights the importance of combining multiple tokens to enhance fine-grained spoken language understanding.", "date": "2025-05-23", "categories": []}
{"id": "2506.00273", "title": "SoundSculpt: Direction and Semantics Driven Ambisonic Target Sound Extraction", "abstract": "This paper introduces SoundSculpt, a neural network designed to extract target sound fields from ambisonic recordings. SoundSculpt employs an ambisonic-in-ambisonic-out architecture and is conditioned on both spatial information (e.g., target direction obtained by pointing at an immersive video) and semantic embeddings (e.g., derived from image segmentation and captioning). Trained and evaluated on synthetic and real ambisonic mixtures, SoundSculpt demonstrates superior performance compared to various signal processing baselines. Our results further reveal that while spatial conditioning alone can be effective, the combination of spatial and semantic information is beneficial in scenarios where there are secondary sound sources spatially close to the target. Additionally, we compare two different semantic embeddings derived from a text description of the target sound using text encoders.", "date": "2025-05-30", "categories": []}
{"id": "2506.00343", "title": "The iNaturalist Sounds Dataset", "abstract": "We present the iNaturalist Sounds Dataset (iNatSounds), a collection of 230,000 audio files capturing sounds from over 5,500 species, contributed by more than 27,000 recordists worldwide. The dataset encompasses sounds from birds, mammals, insects, reptiles, and amphibians, with audio and species labels derived from observations submitted to iNaturalist, a global citizen science platform. Each recording in the dataset varies in length and includes a single species annotation. We benchmark multiple backbone architectures, comparing multiclass classification objectives with multilabel objectives. Despite weak labeling, we demonstrate that iNatSounds serves as a useful pretraining resource by benchmarking it on strongly labeled downstream evaluation datasets. The dataset is available as a single, freely accessible archive, promoting accessibility and research in this important domain. We envision models trained on this data powering next-generation public engagement applications, and assisting biologists, ecologists, and land use managers in processing large audio collections, thereby contributing to the understanding of species compositions in diverse soundscapes.", "date": "2025-05-31", "categories": []}
{"id": "2506.00385", "title": "MagiCodec: Simple Masked Gaussian-Injected Codec for High-Fidelity Reconstruction and Generation", "abstract": "Neural audio codecs have made significant strides in efficiently mapping raw audio waveforms into discrete token representations, which are foundational for contemporary audio generative models. However, most existing codecs are optimized primarily for reconstruction quality, often at the expense of the downstream modelability of the encoded tokens. Motivated by the need to overcome this bottleneck, we introduce $\\textbf{MagiCodec}$, a novel single-layer, streaming Transformer-based audio codec. MagiCodec is designed with a multistage training pipeline that incorporates Gaussian noise injection and latent regularization, explicitly targeting the enhancement of semantic expressiveness in the generated codes while preserving high reconstruction fidelity. We analytically derive the effect of noise injection in the frequency domain, demonstrating its efficacy in attenuating high-frequency components and fostering robust tokenization. Extensive experimental evaluations show that MagiCodec surpasses state-of-the-art codecs in both reconstruction quality and downstream tasks. Notably, the tokens produced by MagiCodec exhibit Zipf-like distributions, as observed in natural languages, thereby improving compatibility with language-model-based generative architectures. The code and pre-trained models are available at https://github.com/Ereboas/MagiCodec.", "date": "2025-05-31", "categories": []}
{"id": "2506.00800", "title": "CLAP-ART: Automated Audio Captioning with Semantic-rich Audio Representation Tokenizer", "abstract": "Automated Audio Captioning (AAC) aims to describe the semantic contexts of general sounds, including acoustic events and scenes, by leveraging effective acoustic features. To enhance performance, an AAC method, EnCLAP, employed discrete tokens from EnCodec as an effective input for fine-tuning a language model BART. However, EnCodec is designed to reconstruct waveforms rather than capture the semantic contexts of general sounds, which AAC should describe. To address this issue, we propose CLAP-ART, an AAC method that utilizes ``semantic-rich and discrete'' tokens as input. CLAP-ART computes semantic-rich discrete tokens from pre-trained audio representations through vector quantization. We experimentally confirmed that CLAP-ART outperforms baseline EnCLAP on two AAC benchmarks, indicating that semantic-rich discrete tokens derived from semantically rich AR are beneficial for AAC.", "date": "2025-06-01", "categories": []}
{"id": "2506.00927", "title": "In-the-wild Audio Spatialization with Flexible Text-guided Localization", "abstract": "To enhance immersive experiences, binaural audio offers spatial awareness of sounding objects in AR, VR, and embodied AI applications. While existing audio spatialization methods can generally map any available monaural audio to binaural audio signals, they often lack the flexible and interactive control needed in complex multi-object user-interactive environments. To address this, we propose a Text-guided Audio Spatialization (TAS) framework that utilizes flexible text prompts and evaluates our model from unified generation and comprehension perspectives. Due to the limited availability of premium and large-scale stereo data, we construct the SpatialTAS dataset, which encompasses 376,000 simulated binaural audio samples to facilitate the training of our model. Our model learns binaural differences guided by 3D spatial location and relative position prompts, augmented by flipped-channel audio. It outperforms existing methods on both simulated and real-recorded datasets, demonstrating superior generalization and accuracy. Besides, we develop an assessment model based on Llama-3.1-8B, which evaluates the spatial semantic coherence between our generated binaural audio and text prompts through a spatial reasoning task. Results demonstrate that text prompts provide flexible and interactive control to generate binaural audio with excellent quality and semantic consistency in spatial locations. Dataset is available at \\href{https://github.com/Alice01010101/TASU}", "date": "2025-06-01", "categories": []}
{"id": "2506.01023", "title": "A Two-Stage Hierarchical Deep Filtering Framework for Real-Time Speech Enhancement", "abstract": "This paper proposes a model that integrates sub-band processing and deep filtering to fully exploit information from the target time-frequency (TF) bin and its surrounding TF bins for single-channel speech enhancement. The sub-band module captures surrounding frequency bin information at the input, while the deep filtering module applies filtering at the output to both the target TF bin and its surrounding TF bins. To further improve the model performance, we decouple deep filtering into temporal and frequency components and introduce a two-stage framework, reducing the complexity of filter coefficient prediction at each stage. Additionally, we propose the TAConv module to strengthen convolutional feature extraction. Experimental results demonstrate that the proposed hierarchical deep filtering network (HDF-Net) effectively utilizes surrounding TF bin information and outperforms other advanced systems while using fewer resources.", "date": "2025-06-01", "categories": []}
{"id": "2506.01111", "title": "FusionAudio-1.2M: Towards Fine-grained Audio Captioning with Multimodal Contextual Fusion", "abstract": "High-quality, large-scale audio captioning is crucial for advancing audio understanding, yet current automated methods often generate captions that lack fine-grained detail and contextual accuracy, primarily due to their reliance on limited unimodal or superficial multimodal information. Drawing inspiration from human auditory perception, which adeptly integrates cross-modal cues and performs sophisticated auditory scene analysis, we introduce a novel two-stage automated pipeline. This pipeline first employs specialized pretrained models to extract diverse contextual cues (e.g., speech, music, general sounds, and visual information from associated video). A large language model (LLM) then synthesizes these rich, multimodal inputs to generate detailed and context-aware audio captions. Key contributions of this work include: (1) the proposed scalable method for fine-grained audio caption generation; (2) FusionAudio, a new large-scale dataset comprising 1.2 million such detailed captions, combined with 6 million QA pairs; and (3) enhanced audio models developed using FusionAudio, specifically a CLAP-based audio encoder with superior audio-text alignment and instruction following. This paper paves the way for more nuanced and accurate automated understanding of complex audio environments. Code and data can be found in https://github.com/satsuki2486441738/FusionAudio.", "date": "2025-06-01", "categories": []}
{"id": "2506.01365", "title": "Attention Is Not Always the Answer: Optimizing Voice Activity Detection with Simple Feature Fusion", "abstract": "Voice Activity Detection (VAD) plays a key role in speech processing, often utilizing hand-crafted or neural features. This study examines the effectiveness of Mel-Frequency Cepstral Coefficients (MFCCs) and pre-trained model (PTM) features, including wav2vec 2.0, HuBERT, WavLM, UniSpeech, MMS, and Whisper. We propose FusionVAD, a unified framework that combines both feature types using three fusion strategies: concatenation, addition, and cross-attention (CA). Experimental results reveal that simple fusion techniques, particularly addition, outperform CA in both accuracy and efficiency. Fusion-based models consistently surpass single-feature models, highlighting the complementary nature of MFCCs and PTM features. Notably, our best-performing fusion model exceeds the state-of-the-art Pyannote across multiple datasets, achieving an absolute average improvement of 2.04%. These results confirm that simple feature fusion enhances VAD robustness while maintaining computational efficiency.", "date": "2025-06-02", "categories": []}
{"id": "2312.10741", "title": "StyleSinger: Style Transfer for Out-of-Domain Singing Voice Synthesis", "abstract": "Style transfer for out-of-domain (OOD) singing voice synthesis (SVS) focuses on generating high-quality singing voices with unseen styles (such as timbre, emotion, pronunciation, and articulation skills) derived from reference singing voice samples. However, the endeavor to model the intricate nuances of singing voice styles is an arduous task, as singing voices possess a remarkable degree of expressiveness. Moreover, existing SVS methods encounter a decline in the quality of synthesized singing voices in OOD scenarios, as they rest upon the assumption that the target vocal attributes are discernible during the training phase. To overcome these challenges, we propose StyleSinger, the first singing voice synthesis model for zero-shot style transfer of out-of-domain reference singing voice samples. StyleSinger incorporates two critical approaches for enhanced effectiveness: 1) the Residual Style Adaptor (RSA) which employs a residual quantization module to capture diverse style characteristics in singing voices, and 2) the Uncertainty Modeling Layer Normalization (UMLN) to perturb the style attributes within the content representation during the training phase and thus improve the model generalization. Our extensive evaluations in zero-shot style transfer undeniably establish that StyleSinger outperforms baseline models in both audio quality and similarity to the reference singing voice samples. Access to singing voice samples can be found at https://aaronz345.github.io/StyleSingerDemo/.", "date": "2023-12-17", "categories": []}
{"id": "2505.21568", "title": "VoiceMark: Zero-Shot Voice Cloning-Resistant Watermarking Approach Leveraging Speaker-Specific Latents", "abstract": "Voice cloning (VC)-resistant watermarking is an emerging technique for tracing and preventing unauthorized cloning. Existing methods effectively trace traditional VC models by training them on watermarked audio but fail in zero-shot VC scenarios, where models synthesize audio from an audio prompt without training. To address this, we propose VoiceMark, the first zero-shot VC-resistant watermarking method that leverages speaker-specific latents as the watermark carrier, allowing the watermark to transfer through the zero-shot VC process into the synthesized audio. Additionally, we introduce VC-simulated augmentations and VAD-based loss to enhance robustness against distortions. Experiments on multiple zero-shot VC models demonstrate that VoiceMark achieves over 95% accuracy in watermark detection after zero-shot VC synthesis, significantly outperforming existing methods, which only reach around 50%. See our code and demos at: https://huggingface.co/spaces/haiyunli/VoiceMark", "date": "2025-05-27", "categories": []}
{"id": "2505.23781", "title": "Unified AI for Accurate Audio Anomaly Detection", "abstract": "This paper presents a unified AI framework for high-accuracy audio anomaly detection by integrating advanced noise reduction, feature extraction, and machine learning modeling techniques. The approach combines spectral subtraction and adaptive filtering to enhance audio quality, followed by feature extraction using traditional methods like MFCCs and deep embeddings from pre-trained models such as OpenL3. The modeling pipeline incorporates classical models (SVM, Random Forest), deep learning architectures (CNNs), and ensemble methods to boost robustness and accuracy. Evaluated on benchmark datasets including TORGO and LibriSpeech, the proposed framework demonstrates superior performance in precision, recall, and classification of slurred vs. normal speech. This work addresses challenges in noisy environments and real-time applications and provides a scalable solution for audio-based anomaly detection.", "date": "2025-05-20", "categories": []}
{"id": "2505.23784", "title": "Learning Normal Patterns in Musical Loops", "abstract": "This paper introduces an unsupervised framework for detecting audio patterns in musical samples (loops) through anomaly detection techniques, addressing challenges in music information retrieval (MIR). Existing methods are often constrained by reliance on handcrafted features, domain-specific limitations, or dependence on iterative user interaction. We address these limitations through an architecture combining deep feature extraction with unsupervised anomaly detection. Our approach leverages a pre-trained Hierarchical Token-semantic Audio Transformer (HTS-AT), paired with a Feature Fusion Mechanism (FFM), to generate representations from variable-length audio loops. These embeddings are processed using one-class Deep Support Vector Data Description (Deep SVDD), which learns normative audio patterns by mapping them to a compact latent hypersphere. Evaluations on curated bass and guitar datasets compare standard and residual autoencoder variants against baselines like Isolation Forest (IF) and and principle component analysis (PCA) methods. Results show our Deep SVDD models, especially the residual autoencoder variant, deliver improved anomaly separation, particularly for larger variations. This research contributes a flexible, fully unsupervised solution for processing diverse audio samples, overcoming previous structural and input limitations while enabling effective pattern identification through distance-based latent space scoring.", "date": "2025-05-22", "categories": []}
{"id": "2505.23834", "title": "Patient-Aware Feature Alignment for Robust Lung Sound Classification:Cohesion-Separation and Global Alignment Losses", "abstract": "Lung sound classification is vital for early diagnosis of respiratory diseases. However, biomedical signals often exhibit inter-patient variability even among patients with the same symptoms, requiring a learning approach that considers individual differences. We propose a Patient-Aware Feature Alignment (PAFA) framework with two novel losses, Patient Cohesion-Separation Loss (PCSL) and Global Patient Alignment Loss (GPAL). PCSL clusters features of the same patient while separating those from other patients to capture patient variability, whereas GPAL draws each patient's centroid toward a global center, preventing feature space fragmentation. Our method achieves outstanding results on the ICBHI dataset with a score of 64.84\\% for four-class and 72.08\\% for two-class classification. These findings highlight PAFA's ability to capture individualized patterns and demonstrate performance gains in distinct patient clusters, offering broader applications for patient-centered healthcare.", "date": "2025-05-28", "categories": []}
{"id": "2505.23964", "title": "Acoustic Classification of Maritime Vessels using Learnable Filterbanks", "abstract": "Reliably monitoring and recognizing maritime vessels based on acoustic signatures is complicated by the variability of different recording scenarios. A robust classification framework must be able to generalize across diverse acoustic environments and variable source-sensor distances. To this end, we present a deep learning model with robust performance across different recording scenarios. Using a trainable spectral front-end and temporal feature encoder to learn a Gabor filterbank, the model can dynamically emphasize different frequency components. Trained on the VTUAD hydrophone recordings from the Strait of Georgia, our model, CATFISH, achieves a state-of-the-art 96.63 % percent test accuracy across varying source-sensor distances, surpassing the previous benchmark by over 12 percentage points. We present the model, justify our architectural choices, analyze the learned Gabor filters, and perform ablation studies on sensor data fusion and attention-based pooling.", "date": "2025-05-29", "categories": []}
{"id": "2505.24336", "title": "When Humans Growl and Birds Speak: High-Fidelity Voice Conversion from Human to Animal and Designed Sounds", "abstract": "Human to non-human voice conversion (H2NH-VC) transforms human speech into animal or designed vocalizations. Unlike prior studies focused on dog-sounds and 16 or 22.05kHz audio transformation, this work addresses a broader range of non-speech sounds, including natural sounds (lion-roars, birdsongs) and designed voice (synthetic growls). To accomodate generation of diverse non-speech sounds and 44.1kHz high-quality audio transformation, we introduce a preprocessing pipeline and an improved CVAE-based H2NH-VC model, both optimized for human and non-human voices. Experimental results showed that the proposed method outperformed baselines in quality, naturalness, and similarity MOS, achieving effective voice conversion across diverse non-human timbres. Demo samples are available at https://nc-ai.github.io/speech/publications/nonhuman-vc/", "date": "2025-05-30", "categories": []}
{"id": "2505.24498", "title": "Efficient Neural and Numerical Methods for High-Quality Online Speech Spectrogram Inversion via Gradient Theorem", "abstract": "Recent work in online speech spectrogram inversion effectively combines Deep Learning with the Gradient Theorem to predict phase derivatives directly from magnitudes. Then, phases are estimated from their derivatives via least squares, resulting in a high quality reconstruction. In this work, we introduce three innovations that drastically reduce computational cost, while maintaining high quality: Firstly, we introduce a novel neural network architecture with just 8k parameters, 30 times smaller than previous state of the art. Secondly, increasing latency by 1 hop size allows us to further halve the cost of the neural inference step. Thirdly, we we observe that the least squares problem features a tridiagonal matrix and propose a linear-complexity solver for the least squares step that leverages tridiagonality and positive-semidefiniteness, achieving a speedup of several orders of magnitude. We release samples online.", "date": "2025-05-30", "categories": []}
{"id": "2505.21527", "title": "VietASR: Achieving Industry-level Vietnamese ASR with 50-hour labeled data and Large-Scale Speech Pretraining", "abstract": "Automatic speech recognition (ASR) has made remarkable progress but heavily relies on large-scale labeled data, which is scarce for low-resource languages like Vietnamese. While existing systems such as Whisper, USM, and MMS achieve promising performance, their efficacy remains inadequate in terms of training costs, latency, and accessibility. To address these issues, we propose VietASR, a novel ASR training pipeline that leverages vast amounts of unlabeled data and a small set of labeled data. Through multi-iteration ASR-biased self-supervised learning on a large-scale unlabeled dataset, VietASR offers a cost-effective and practical solution for enhancing ASR performance. Experiments demonstrate that pre-training on 70,000-hour unlabeled data and fine-tuning on merely 50-hour labeled data yield a lightweight but powerful ASR model. It outperforms Whisper Large-v3 and commercial ASR systems on real-world data. Our code and models will be open-sourced to facilitate research in low-resource ASR.", "date": "2025-05-23", "categories": []}
{"id": "2505.22865", "title": "BinauralFlow: A Causal and Streamable Approach for High-Quality Binaural Speech Synthesis with Flow Matching Models", "abstract": "Binaural rendering aims to synthesize binaural audio that mimics natural hearing based on a mono audio and the locations of the speaker and listener. Although many methods have been proposed to solve this problem, they struggle with rendering quality and streamable inference. Synthesizing high-quality binaural audio that is indistinguishable from real-world recordings requires precise modeling of binaural cues, room reverb, and ambient sounds. Additionally, real-world applications demand streaming inference. To address these challenges, we propose a flow matching based streaming binaural speech synthesis framework called BinauralFlow. We consider binaural rendering to be a generation problem rather than a regression problem and design a conditional flow matching model to render high-quality audio. Moreover, we design a causal U-Net architecture that estimates the current audio frame solely based on past information to tailor generative models for streaming inference. Finally, we introduce a continuous inference pipeline incorporating streaming STFT/ISTFT operations, a buffer bank, a midpoint solver, and an early skip schedule to improve rendering continuity and speed. Quantitative and qualitative evaluations demonstrate the superiority of our method over SOTA approaches. A perceptual study further reveals that our model is nearly indistinguishable from real-world recordings, with a $42\\%$ confusion rate.", "date": "2025-05-28", "categories": []}
{"id": "2505.23132", "title": "Patient Domain Supervised Contrastive Learning for Lung Sound Classification Using Mobile Phone", "abstract": "Auscultation is crucial for diagnosing lung diseases. The COVID-19 pandemic has revealed the limitations of traditional, in-person lung sound assessments. To overcome these issues, advancements in digital stethoscopes and artificial intelligence (AI) have led to the development of new diagnostic methods. In this context, our study aims to use smartphone microphones to record and analyze lung sounds. We faced two major challenges: the difference in audio style between electronic stethoscopes and smartphone microphones, and the variability among patients. To address these challenges, we developed a method called Patient Domain Supervised Contrastive Learning (PD-SCL). By integrating this method with the Audio Spectrogram Transformer (AST) model, we significantly improved its performance by 2.4\\% compared to the original AST model. This progress demonstrates that smartphones can effectively diagnose lung sounds, addressing inconsistencies in patient data and showing potential for broad use beyond traditional clinical settings. Our research contributes to making lung disease detection more accessible in the post-COVID-19 world.", "date": "2025-05-29", "categories": []}
{"id": "2505.23619", "title": "Few-Shot Speech Deepfake Detection Adaptation with Gaussian Processes", "abstract": "Recent advancements in Text-to-Speech (TTS) models, particularly in voice cloning, have intensified the demand for adaptable and efficient deepfake detection methods. As TTS systems continue to evolve, detection models must be able to efficiently adapt to previously unseen generation models with minimal data. This paper introduces ADD-GP, a few-shot adaptive framework based on a Gaussian Process (GP) classifier for Audio Deepfake Detection (ADD). We show how the combination of a powerful deep embedding model with the Gaussian processes flexibility can achieve strong performance and adaptability. Additionally, we show this approach can also be used for personalized detection, with greater robustness to new TTS models and one-shot adaptability. To support our evaluation, a benchmark dataset is constructed for this task using new state-of-the-art voice cloning models.", "date": "2025-05-29", "categories": []}
{"id": "2505.23625", "title": "ZeroSep: Separate Anything in Audio with Zero Training", "abstract": "Audio source separation is fundamental for machines to understand complex acoustic environments and underpins numerous audio applications. Current supervised deep learning approaches, while powerful, are limited by the need for extensive, task-specific labeled data and struggle to generalize to the immense variability and open-set nature of real-world acoustic scenes. Inspired by the success of generative foundation models, we investigate whether pre-trained text-guided audio diffusion models can overcome these limitations. We make a surprising discovery: zero-shot source separation can be achieved purely through a pre-trained text-guided audio diffusion model under the right configuration. Our method, named ZeroSep, works by inverting the mixed audio into the diffusion model's latent space and then using text conditioning to guide the denoising process to recover individual sources. Without any task-specific training or fine-tuning, ZeroSep repurposes the generative diffusion model for a discriminative separation task and inherently supports open-set scenarios through its rich textual priors. ZeroSep is compatible with a variety of pre-trained text-guided audio diffusion backbones and delivers strong separation performance on multiple separation benchmarks, surpassing even supervised methods.", "date": "2025-05-29", "categories": []}
{"id": "2412.08356", "title": "Zero-Shot Mono-to-Binaural Speech Synthesis", "abstract": "We present ZeroBAS, a neural method to synthesize binaural audio from monaural audio recordings and positional information without training on any binaural data. To our knowledge, this is the first published zero-shot neural approach to mono-to-binaural audio synthesis. Specifically, we show that a parameter-free geometric time warping and amplitude scaling based on source location suffices to get an initial binaural synthesis that can be refined by iteratively applying a pretrained denoising vocoder. Furthermore, we find this leads to generalization across room conditions, which we measure by introducing a new dataset, TUT Mono-to-Binaural, to evaluate state-of-the-art monaural-to-binaural synthesis methods on unseen conditions. Our zero-shot method is perceptually on-par with the performance of supervised methods on the standard mono-to-binaural dataset, and even surpasses them on our out-of-distribution TUT Mono-to-Binaural dataset. Our results highlight the potential of pretrained generative audio models and zero-shot learning to unlock robust binaural audio synthesis.", "date": "2024-12-11", "categories": []}
{"id": "2505.19663", "title": "A Comprehensive Real-World Assessment of Audio Watermarking Algorithms: Will They Survive Neural Codecs?", "abstract": "We introduce the Robust Audio Watermarking Benchmark (RAW-Bench), a benchmark for evaluating deep learning-based audio watermarking methods with standardized and systematic comparisons. To simulate real-world usage, we introduce a comprehensive audio attack pipeline with various distortions such as compression, background noise, and reverberation, along with a diverse test dataset including speech, environmental sounds, and music recordings. Evaluating four existing watermarking methods on RAW-bench reveals two main insights: (i) neural compression techniques pose the most significant challenge, even when algorithms are trained with such compressions; and (ii) training with audio attacks generally improves robustness, although it is insufficient in some cases. Furthermore, we find that specific distortions, such as polarity inversion, time stretching, or reverb, seriously affect certain methods. The evaluation framework is accessible at github.com/SonyResearch/raw_bench.", "date": "2025-05-26", "categories": []}
{"id": "2505.21809", "title": "Voice Quality Dimensions as Interpretable Primitives for Speaking Style for Atypical Speech and Affect", "abstract": "Perceptual voice quality dimensions describe key characteristics of atypical speech and other speech modulations. Here we develop and evaluate voice quality models for seven voice and speech dimensions (intelligibility, imprecise consonants, harsh voice, naturalness, monoloudness, monopitch, and breathiness). Probes were trained on the public Speech Accessibility (SAP) project dataset with 11,184 samples from 434 speakers, using embeddings from frozen pre-trained models as features. We found that our probes had both strong performance and strong generalization across speech elicitation categories in the SAP dataset. We further validated zero-shot performance on additional datasets, encompassing unseen languages and tasks: Italian atypical speech, English atypical speech, and affective speech. The strong zero-shot performance and the interpretability of results across an array of evaluations suggests the utility of using voice quality dimensions in speaking style-related tasks.", "date": "2025-05-27", "categories": []}
{"id": "2505.21827", "title": "Music Source Restoration", "abstract": "We introduce Music Source Restoration (MSR), a novel task addressing the gap between idealized source separation and real-world music production. Current Music Source Separation (MSS) approaches assume mixtures are simple sums of sources, ignoring signal degradations employed during music production like equalization, compression, and reverb. MSR models mixtures as degraded sums of individually degraded sources, with the goal of recovering original, undegraded signals. Due to the lack of data for MSR, we present RawStems, a dataset annotation of 578 songs with unprocessed source signals organized into 8 primary and 17 secondary instrument groups, totaling 354.13 hours. To the best of our knowledge, RawStems is the first dataset that contains unprocessed music stems with hierarchical categories. We consider spectral filtering, dynamic range compression, harmonic distortion, reverb and lossy codec as possible degradations, and establish U-Former as a baseline method, demonstrating the feasibility of MSR on our dataset. We release the RawStems dataset annotations, degradation simulation pipeline, training code and pre-trained models to be publicly available.", "date": "2025-05-27", "categories": []}
{"id": "2505.22027", "title": "Improving Respiratory Sound Classification with Architecture-Agnostic Knowledge Distillation from Ensembles", "abstract": "Respiratory sound datasets are limited in size and quality, making high performance difficult to achieve. Ensemble models help but inevitably increase compute cost at inference time. Soft label training distills knowledge efficiently with extra cost only at training. In this study, we explore soft labels for respiratory sound classification as an architecture-agnostic approach to distill an ensemble of teacher models into a student model. We examine different variations of our approach and find that even a single teacher, identical to the student, considerably improves performance beyond its own capability, with optimal gains achieved using only a few teachers. We achieve the new state-of-the-art Score of 64.39 on ICHBI, surpassing the previous best by 0.85 and improving average Scores across architectures by more than 1.16. Our results highlight the effectiveness of knowledge distillation with soft labels for respiratory sound classification, regardless of size or architecture.", "date": "2025-05-28", "categories": []}
{"id": "2411.01135", "title": "Music Foundation Model as Generic Booster for Music Downstream Tasks", "abstract": "We demonstrate the efficacy of using intermediate representations from a single foundation model to enhance various music downstream tasks. We introduce SoniDo, a music foundation model (MFM) designed to extract hierarchical features from target music samples. By leveraging hierarchical intermediate features, SoniDo constrains the information granularity, leading to improved performance across various downstream tasks including both understanding and generative tasks. We specifically evaluated this approach on representative tasks such as music tagging, music transcription, music source separation, and music mixing. Our results reveal that the features extracted from foundation models provide valuable enhancements in training downstream task models. This highlights the capability of using features extracted from music foundation models as a booster for downstream tasks. Our approach not only benefits existing task-specific models but also supports music downstream tasks constrained by data scarcity. This paves the way for more effective and accessible music processing solutions.", "date": "2024-11-02", "categories": []}
{"id": "2412.10792", "title": "Audio-based Anomaly Detection in Industrial Machines Using Deep One-Class Support Vector Data Description", "abstract": "The frequent breakdowns and malfunctions of industrial equipment have driven increasing interest in utilizing cost-effective and easy-to-deploy sensors, such as microphones, for effective condition monitoring of machinery. Microphones offer a low-cost alternative to widely used condition monitoring sensors with their high bandwidth and capability to detect subtle anomalies that other sensors might have less sensitivity. In this study, we investigate malfunctioning industrial machines to evaluate and compare anomaly detection performance across different machine types and fault conditions. Log-Mel spectrograms of machinery sound are used as input, and the performance is evaluated using the area under the curve (AUC) score for two different methods: baseline dense autoencoder (AE) and one-class deep Support Vector Data Description (deep SVDD) with different subspace dimensions. Our results over the MIMII sound dataset demonstrate that the deep SVDD method with a subspace dimension of 2 provides superior anomaly detection performance, achieving average AUC scores of 0.84, 0.80, and 0.69 for 6 dB, 0 dB, and -6 dB signal-to-noise ratios (SNRs), respectively, compared to 0.82, 0.72, and 0.64 for the baseline model. Moreover, deep SVDD requires 7.4 times fewer trainable parameters than the baseline dense AE, emphasizing its advantage in both effectiveness and computational efficiency.", "date": "2024-12-14", "categories": []}
{"id": "2501.18470", "title": "Resampling Filter Design for Multirate Neural Audio Effect Processing", "abstract": "Neural networks have become ubiquitous in audio effects modelling, especially for guitar amplifiers and distortion pedals. One limitation of such models is that the sample rate of the training data is implicitly encoded in the model weights and therefore not readily adjustable at inference. Recent work explored modifications to recurrent neural network architecture to approximate a sample rate independent system, enabling audio processing at a rate that differs from the original training rate. This method works well for integer oversampling and can reduce aliasing caused by nonlinear activation functions. For small fractional changes in sample rate, fractional delay filters can be used to approximate sample rate independence, but in some cases this method fails entirely. Here, we explore the use of real-time signal resampling at the input and output of the neural network as an alternative solution. We investigate several resampling filter designs and show that a two-stage design consisting of a half-band IIR filter cascaded with a Kaiser window FIR filter can give similar or better results to the previously proposed model adjustment method with many fewer filtering operations per sample and less than one millisecond of latency at typical audio rates. Furthermore, we investigate interpolation and decimation filters for the task of integer oversampling and show that cascaded half-band IIR and FIR designs can be used in conjunction with the model adjustment method to reduce aliasing in a range of distortion effect models.", "date": "2025-01-30", "categories": []}
{"id": "2505.10879", "title": "Multi-Stage Speaker Diarization for Noisy Classrooms", "abstract": "Speaker diarization, the process of identifying \"who spoke when\" in audio recordings, is essential for understanding classroom dynamics. However, classroom settings present distinct challenges, including poor recording quality, high levels of background noise, overlapping speech, and the difficulty of accurately capturing children's voices. This study investigates the effectiveness of multi-stage diarization models using Nvidia's NeMo diarization pipeline. We assess the impact of denoising on diarization accuracy and compare various voice activity detection (VAD) models, including self-supervised transformer-based frame-wise VAD models. We also explore a hybrid VAD approach that integrates Automatic Speech Recognition (ASR) word-level timestamps with frame-level VAD predictions. We conduct experiments using two datasets from English speaking classrooms to separate teacher vs. student speech and to separate all speakers. Our results show that denoising significantly improves the Diarization Error Rate (DER) by reducing the rate of missed speech. Additionally, training on both denoised and noisy datasets leads to substantial performance gains in noisy conditions. The hybrid VAD model leads to further improvements in speech detection, achieving a DER as low as 17% in teacher-student experiments and 45% in all-speaker experiments. However, we also identified trade-offs between voice activity detection and speaker confusion. Overall, our study highlights the effectiveness of multi-stage diarization models and integrating ASR-based information for enhancing speaker diarization in noisy classroom environments.", "date": "2025-05-16", "categories": []}
{"id": "2505.20635", "title": "Plug-and-Play Co-Occurring Face Attention for Robust Audio-Visual Speaker Extraction", "abstract": "Audio-visual speaker extraction isolates a target speaker's speech from a mixture speech signal conditioned on a visual cue, typically using the target speaker's face recording. However, in real-world scenarios, other co-occurring faces are often present on-screen, providing valuable speaker activity cues in the scene. In this work, we introduce a plug-and-play inter-speaker attention module to process these flexible numbers of co-occurring faces, allowing for more accurate speaker extraction in complex multi-person environments. We integrate our module into two prominent models: the AV-DPRNN and the state-of-the-art AV-TFGridNet. Extensive experiments on diverse datasets, including the highly overlapped VoxCeleb2 and sparsely overlapped MISP, demonstrate that our approach consistently outperforms baselines. Furthermore, cross-dataset evaluations on LRS2 and LRS3 confirm the robustness and generalizability of our method.", "date": "2025-05-27", "categories": []}
{"id": "2501.13375", "title": "Bridging The Multi-Modality Gaps of Audio, Visual and Linguistic for Speech Enhancement", "abstract": "Speech enhancement (SE) aims to improve the quality and intelligibility of speech in noisy environments. Recent studies have shown that incorporating visual cues in audio signal processing can enhance SE performance. Given that human speech communication naturally involves audio, visual, and linguistic modalities, it is reasonable to expect additional improvements by integrating linguistic information. However, effectively bridging these modality gaps, particularly during knowledge transfer remains a significant challenge. In this paper, we propose a novel multi-modal learning framework, termed DLAV-SE, which leverages a diffusion-based model integrating audio, visual, and linguistic information for audio-visual speech enhancement (AVSE). Within this framework, the linguistic modality is modeled using a pretrained language model (PLM), which transfers linguistic knowledge to the audio-visual domain through a cross-modal knowledge transfer (CMKT) mechanism during training. After training, the PLM is no longer required at inference, as its knowledge is embedded into the AVSE model through the CMKT process. We conduct a series of SE experiments to evaluate the effectiveness of our approach. Results show that the proposed DLAV-SE system significantly improves speech quality and reduces generative artifacts, such as phonetic confusion, compared to state-of-the-art (SOTA) methods. Furthermore, visualization analyses confirm that the CMKT method enhances the generation quality of the AVSE outputs. These findings highlight both the promise of diffusion-based methods for advancing AVSE and the value of incorporating linguistic information to further improve system performance.", "date": "2025-01-23", "categories": []}
{"id": "2502.01185", "title": "Deep Active Speech Cancellation with Mamba-Masking Network", "abstract": "We present a novel deep learning network for Active Speech Cancellation (ASC), advancing beyond Active Noise Cancellation (ANC) methods by effectively canceling both noise and speech signals. The proposed Mamba-Masking architecture introduces a masking mechanism that directly interacts with the encoded reference signal, enabling adaptive and precisely aligned anti-signal generation-even under rapidly changing, high-frequency conditions, as commonly found in speech. Complementing this, a multi-band segmentation strategy further improves phase alignment across frequency bands. Additionally, we introduce an optimization-driven loss function that provides near-optimal supervisory signals for anti-signal generation. Experimental results demonstrate substantial performance gains, achieving up to 7.2dB improvement in ANC scenarios and 6.2dB in ASC, significantly outperforming existing methods.", "date": "2025-02-03", "categories": []}
{"id": "2505.18517", "title": "LiSTEN: Learning Soft Token Embeddings for Neural Audio LLMs", "abstract": "Foundation models based on large language models (LLMs) have shown great success in handling various tasks and modalities. However, adapting these models for general-purpose audio-language tasks is challenging due to differences in acoustic environments and task variations. In this work, we introduce LiSTEN Learning Soft Token Embeddings for Neural Audio LLMs), a framework for adapting LLMs to speech and audio tasks. LiSTEN uses a dynamic prompt selection strategy with learnable key-value pairs, allowing the model to balance general and task-specific knowledge while avoiding overfitting in a multitask setting. Our approach reduces dependence on large-scale ASR or captioning datasets, achieves competitive performance with fewer trainable parameters, and simplifies training by using a single-stage process. Additionally, LiSTEN enhances interpretability by analyzing the diversity and overlap of selected prompts across different tasks.", "date": "2025-05-24", "categories": []}
{"id": "2505.19534", "title": "Training-Free Multi-Step Audio Source Separation", "abstract": "Audio source separation aims to separate a mixture into target sources. Previous audio source separation systems usually conduct one-step inference, which does not fully explore the separation ability of models. In this work, we reveal that pretrained one-step audio source separation models can be leveraged for multi-step separation without additional training. We propose a simple yet effective inference method that iteratively applies separation by optimally blending the input mixture with the previous step's separation result. At each step, we determine the optimal blending ratio by maximizing a metric. We prove that our method always yield improvement over one-step inference, provide error bounds based on model smoothness and metric robustness, and provide theoretical analysis connecting our method to denoising along linear interpolation paths between noise and clean distributions, a property we link to denoising diffusion bridge models. Our approach effectively delivers improved separation performance as a \"free lunch\" from existing models. Our empirical results demonstrate that our multi-step separation approach consistently outperforms one-step inference across both speech enhancement and music source separation tasks, and can achieve scaling performance similar to training a larger model, using more data, or in some cases employing a multi-step training objective. These improvements appear not only on the optimization metric during multi-step inference, but also extend to nearly all non-optimized metrics (with one exception). We also discuss limitations of our approach and directions for future research.", "date": "2025-05-26", "categories": []}
{"id": "2505.19951", "title": "Novel Loss-Enhanced Universal Adversarial Patches for Sustainable Speaker Privacy", "abstract": "Deep learning voice models are commonly used nowadays, but the safety processing of personal data, such as human identity and speech content, remains suspicious. To prevent malicious user identification, speaker anonymization methods were proposed. Current methods, particularly based on universal adversarial patch (UAP) applications, have drawbacks such as significant degradation of audio quality, decreased speech recognition quality, low transferability across different voice biometrics models, and performance dependence on the input audio length. To mitigate these drawbacks, in this work, we introduce and leverage the novel Exponential Total Variance (TV) loss function and provide experimental evidence that it positively affects UAP strength and imperceptibility. Moreover, we present a novel scalable UAP insertion procedure and demonstrate its uniformly high performance for various audio lengths.", "date": "2025-05-26", "categories": []}
{"id": "2505.20050", "title": "MVP: Multi-source Voice Pathology detection", "abstract": "Voice disorders significantly impact patient quality of life, yet non-invasive automated diagnosis remains under-explored due to both the scarcity of pathological voice data, and the variability in recording sources. This work introduces MVP (Multi-source Voice Pathology detection), a novel approach that leverages transformers operating directly on raw voice signals. We explore three fusion strategies to combine sentence reading and sustained vowel recordings: waveform concatenation, intermediate feature fusion, and decision-level combination. Empirical validation across the German, Portuguese, and Italian languages shows that intermediate feature fusion using transformers best captures the complementary characteristics of both recording types. Our approach achieves up to +13% AUC improvement over single-source methods.", "date": "2025-05-26", "categories": []}
{"id": "2412.03633", "title": "NBM: an Open Dataset for the Acoustic Monitoring of Nocturnal Migratory Birds in Europe", "abstract": "The persisting threats on migratory bird populations highlight the urgent need for effective monitoring techniques that could assist in their conservation. Among these, passive acoustic monitoring is an essential tool, particularly for nocturnal migratory species that are difficult to track otherwise. This work presents the Nocturnal Bird Migration (NBM) dataset, a collection of 13,359 annotated vocalizations from 117 species of the Western Palearctic. The dataset includes precise time and frequency annotations, gathered by dozens of bird enthusiasts across France, enabling novel downstream acoustic analysis. In particular, we prove the utility of this database by training an original two-stage deep object detection model tailored for the processing of audio data. While allowing the precise localization of bird calls in spectrograms, this model shows competitive accuracy on the 45 main species of the dataset with state-of-the-art systems trained on much larger audio collections. These results highlight the interest of fostering similar open-science initiatives to acquire costly but valuable fine-grained annotations of audio files. All data and code are made openly available.", "date": "2024-12-04", "categories": []}
{"id": "2505.17513", "title": "What You Read Isn't What You Hear: Linguistic Sensitivity in Deepfake Speech Detection", "abstract": "Recent advances in text-to-speech technologies have enabled realistic voice generation, fueling audio-based deepfake attacks such as fraud and impersonation. While audio anti-spoofing systems are critical for detecting such threats, prior work has predominantly focused on acoustic-level perturbations, leaving the impact of linguistic variation largely unexplored. In this paper, we investigate the linguistic sensitivity of both open-source and commercial anti-spoofing detectors by introducing transcript-level adversarial attacks. Our extensive evaluation reveals that even minor linguistic perturbations can significantly degrade detection accuracy: attack success rates surpass 60% on several open-source detector-voice pairs, and notably one commercial detection accuracy drops from 100% on synthetic audio to just 32%. Through a comprehensive feature attribution analysis, we identify that both linguistic complexity and model-level audio embedding similarity contribute strongly to detector vulnerability. We further demonstrate the real-world risk via a case study replicating the Brad Pitt audio deepfake scam, using transcript adversarial attacks to completely bypass commercial detectors. These results highlight the need to move beyond purely acoustic defenses and account for linguistic variation in the design of robust anti-spoofing systems. All source code will be publicly available.", "date": "2025-05-23", "categories": []}
{"id": "2505.17823", "title": "Source Separation of Small Classical Ensembles: Challenges and Opportunities", "abstract": "Musical (MSS) source separation of western popular music using non-causal deep learning can be very effective. In contrast, MSS for classical music is an unsolved problem. Classical ensembles are harder to separate than popular music because of issues such as the inherent greater variation in the music; the sparsity of recordings with ground truth for supervised training; and greater ambiguity between instruments. The Cadenza project has been exploring MSS for classical music. This is being done so music can be remixed to improve listening experiences for people with hearing loss. To enable the work, a new database of synthesized woodwind ensembles was created to overcome instrumental imbalances in the EnsembleSet. For the MSS, a set of ConvTasNet models was used with each model being trained to extract a string or woodwind instrument. ConvTasNet was chosen because it enabled both causal and non-causal approaches to be tested. Non-causal approaches have dominated MSS work and are useful for recorded music, but for live music or processing on hearing aids, causal signal processing is needed. The MSS performance was evaluated on the two small datasets (Bach10 and URMP) of real instrument recordings where the ground-truth is available. The performances of the causal and non-causal systems were similar. Comparing the average Signal-to-Distortion (SDR) of the synthesized validation set (6.2 dB causal; 6.9 non-causal), to the real recorded evaluation set (0.3 dB causal, 0.4 dB non-causal), shows that mismatch between synthesized and recorded data is a problem. Future work needs to either gather more real recordings that can be used for training, or to improve the realism and diversity of the synthesized recordings to reduce the mismatch...", "date": "2025-05-23", "categories": []}
{"id": "2505.16306", "title": "Layer-wise Investigation of Large-Scale Self-Supervised Music Representation Models", "abstract": "Recently, pre-trained models for music information retrieval based on self-supervised learning (SSL) are becoming popular, showing success in various downstream tasks. However, there is limited research on the specific meanings of the encoded information and their applicability. Exploring these aspects can help us better understand their capabilities and limitations, leading to more effective use in downstream tasks.\n  In this study, we analyze the advanced music representation model MusicFM and the newly emerged SSL model MuQ. We focus on three main aspects: (i) validating the advantages of SSL models across multiple downstream tasks, (ii) exploring the specialization of layer-wise information for different tasks, and (iii) comparing performance differences when selecting specific layers. Through this analysis, we reveal insights into the structure and potential applications of SSL models in music information retrieval.", "date": "2025-05-22", "categories": []}
{"id": "2407.15835", "title": "dMel: Speech Tokenization made Simple", "abstract": "Large language models have revolutionized natural language processing by leveraging self-supervised pretraining on vast textual data. Inspired by this success, researchers have investigated various compression-based speech tokenization methods to discretize continuous speech signals, enabling the application of language modeling techniques to discrete tokens. However, audio compressor introduces additional complexity and computational cost, and often fail on out-of-domain audio signals. In this work, we introduce a novel speech representation (dmel) that discretizes mel-filterbank channels into intensity bins, creating a simpler yet more effective representation compared to existing speech tokenization methods. Our approach demonstrates superior performance in preserving audio content, robustness to out-of-domain data, and offers a training-free, natural, and streamable representation. To address the high-dimensional nature of log-mel spectrograms, we propose an efficient parallel encoding and decoding method for high-dimensional tokens using an LM-style transformer architecture. This innovation enables us to develop RichTTS and RichASR, two models sharing the same architecture while achieving comparable or better results than specialized existing methods. Our results demonstrate the effectiveness of dmel in achieving high performance on both speech synthesis and recognition tasks within a unified framework, paving the way for efficient and effective joint modeling of speech and text.", "date": "2024-07-22", "categories": []}
{"id": "2410.18322", "title": "Unified Microphone Conversion: Many-to-Many Device Mapping via Feature-wise Linear Modulation", "abstract": "We present Unified Microphone Conversion, a unified generative framework designed to bolster sound event classification (SEC) systems against device variability. While our prior CycleGAN-based methods effectively simulate device characteristics, they require separate models for each device pair, limiting scalability. Our approach overcomes this constraint by conditioning the generator on frequency response data, enabling many-to-many device mappings through unpaired training. We integrate frequency-response information via Feature-wise Linear Modulation, further enhancing scalability. Additionally, incorporating synthetic frequency response differences improves the applicability of our framework for real-world application. Experimental results show that our method outperforms the state-of-the-art by 2.6% and reduces variability by 0.8% in macro-average F1 score.", "date": "2024-10-23", "categories": []}
{"id": "2505.01237", "title": "CAV-MAE Sync: Improving Contrastive Audio-Visual Mask Autoencoders via Fine-Grained Alignment", "abstract": "Recent advances in audio-visual learning have shown promising results in learning representations across modalities. However, most approaches rely on global audio representations that fail to capture fine-grained temporal correspondences with visual frames. Additionally, existing methods often struggle with conflicting optimization objectives when trying to jointly learn reconstruction and cross-modal alignment. In this work, we propose CAV-MAE Sync as a simple yet effective extension of the original CAV-MAE framework for self-supervised audio-visual learning. We address three key challenges: First, we tackle the granularity mismatch between modalities by treating audio as a temporal sequence aligned with video frames, rather than using global representations. Second, we resolve conflicting optimization goals by separating contrastive and reconstruction objectives through dedicated global tokens. Third, we improve spatial localization by introducing learnable register tokens that reduce semantic load on patch tokens. We evaluate the proposed approach on AudioSet, VGG Sound, and the ADE20K Sound dataset on zero-shot retrieval, classification and localization tasks demonstrating state-of-the-art performance and outperforming more complex architectures.", "date": "2025-05-02", "categories": []}
{"id": "2505.15406", "title": "Audio Jailbreak: An Open Comprehensive Benchmark for Jailbreaking Large Audio-Language Models", "abstract": "The rise of Large Audio Language Models (LAMs) brings both potential and risks, as their audio outputs may contain harmful or unethical content. However, current research lacks a systematic, quantitative evaluation of LAM safety especially against jailbreak attacks, which are challenging due to the temporal and semantic nature of speech. To bridge this gap, we introduce AJailBench, the first benchmark specifically designed to evaluate jailbreak vulnerabilities in LAMs. We begin by constructing AJailBench-Base, a dataset of 1,495 adversarial audio prompts spanning 10 policy-violating categories, converted from textual jailbreak attacks using realistic text to speech synthesis. Using this dataset, we evaluate several state-of-the-art LAMs and reveal that none exhibit consistent robustness across attacks. To further strengthen jailbreak testing and simulate more realistic attack conditions, we propose a method to generate dynamic adversarial variants. Our Audio Perturbation Toolkit (APT) applies targeted distortions across time, frequency, and amplitude domains. To preserve the original jailbreak intent, we enforce a semantic consistency constraint and employ Bayesian optimization to efficiently search for perturbations that are both subtle and highly effective. This results in AJailBench-APT, an extended dataset of optimized adversarial audio samples. Our findings demonstrate that even small, semantically preserved perturbations can significantly reduce the safety performance of leading LAMs, underscoring the need for more robust and semantically aware defense mechanisms.", "date": "2025-05-21", "categories": []}
{"id": "2505.15773", "title": "ToxicTone: A Mandarin Audio Dataset Annotated for Toxicity and Toxic Utterance Tonality", "abstract": "Despite extensive research on toxic speech detection in text, a critical gap remains in handling spoken Mandarin audio. The lack of annotated datasets that capture the unique prosodic cues and culturally specific expressions in Mandarin leaves spoken toxicity underexplored. To address this, we introduce ToxicTone -- the largest public dataset of its kind -- featuring detailed annotations that distinguish both forms of toxicity (e.g., profanity, bullying) and sources of toxicity (e.g., anger, sarcasm, dismissiveness). Our data, sourced from diverse real-world audio and organized into 13 topical categories, mirrors authentic communication scenarios. We also propose a multimodal detection framework that integrates acoustic, linguistic, and emotional features using state-of-the-art speech and emotion encoders. Extensive experiments show our approach outperforms text-only and baseline models, underscoring the essential role of speech-specific cues in revealing hidden toxic expressions.", "date": "2025-05-21", "categories": []}
{"id": "2505.13617", "title": "Direction-Aware Neural Acoustic Fields for Few-Shot Interpolation of Ambisonic Impulse Responses", "abstract": "The characteristics of a sound field are intrinsically linked to the geometric and spatial properties of the environment surrounding a sound source and a listener. The physics of sound propagation is captured in a time-domain signal known as a room impulse response (RIR). Prior work using neural fields (NFs) has allowed learning spatially-continuous representations of RIRs from finite RIR measurements. However, previous NF-based methods have focused on monaural omnidirectional or at most binaural listeners, which does not precisely capture the directional characteristics of a real sound field at a single point. We propose a direction-aware neural field (DANF) that more explicitly incorporates the directional information by Ambisonic-format RIRs. While DANF inherently captures spatial relations between sources and listeners, we further propose a direction-aware loss. In addition, we investigate the ability of DANF to adapt to new rooms in various ways including low-rank adaptation.", "date": "2025-05-19", "categories": []}
{"id": "2505.14517", "title": "Steering Deep Non-Linear Spatially Selective Filters for Weakly Guided Extraction of Moving Speakers in Dynamic Scenarios", "abstract": "Recent speaker extraction methods using deep non-linear spatial filtering perform exceptionally well when the target direction is known and stationary. However, spatially dynamic scenarios are considerably more challenging due to time-varying spatial features and arising ambiguities, e.g. when moving speakers cross. While in a static scenario it may be easy for a user to point to the target's direction, manually tracking a moving speaker is impractical. Instead of relying on accurate time-dependent directional cues, which we refer to as strong guidance, in this paper we propose a weakly guided extraction method solely depending on the target's initial position to cope with spatial dynamic scenarios. By incorporating our own deep tracking algorithm and developing a joint training strategy on a synthetic dataset, we demonstrate the proficiency of our approach in resolving spatial ambiguities and even outperform a mismatched, but strongly guided extraction method.", "date": "2025-05-20", "categories": []}
{"id": "2402.01591", "title": "BAT: Learning to Reason about Spatial Sounds with Large Language Models", "abstract": "Spatial sound reasoning is a fundamental human skill, enabling us to navigate and interpret our surroundings based on sound. In this paper we present BAT, which combines the spatial sound perception ability of a binaural acoustic scene analysis model with the natural language reasoning capabilities of a large language model (LLM) to replicate this innate ability. To address the lack of existing datasets of in-the-wild spatial sounds, we synthesized a binaural audio dataset using AudioSet and SoundSpaces 2.0. Next, we developed SpatialSoundQA, a spatial sound-based question-answering dataset, offering a range of QA tasks that train BAT in various aspects of spatial sound perception and reasoning. The acoustic front end encoder of BAT is a novel spatial audio encoder named Spatial Audio Spectrogram Transformer, or Spatial-AST, which by itself achieves strong performance across sound event detection, spatial localization, and distance estimation. By integrating Spatial-AST with LLaMA-2 7B model, BAT transcends standard Sound Event Localization and Detection (SELD) tasks, enabling the model to reason about the relationships between the sounds in its environment. Our experiments demonstrate BAT's superior performance on both spatial sound perception and reasoning, showcasing the immense potential of LLMs in navigating and interpreting complex spatial audio environments.", "date": "2024-02-02", "categories": []}
{"id": "2403.10380", "title": "BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics", "abstract": "Deep learning (DL) has greatly advanced audio classification, yet the field is limited by the scarcity of large-scale benchmark datasets that have propelled progress in other domains. While AudioSet is a pivotal step to bridge this gap as a universal-domain dataset, its restricted accessibility and limited range of evaluation use cases challenge its role as the sole resource. Therefore, we introduce BirdSet, a large-scale benchmark dataset for audio classification focusing on avian bioacoustics. BirdSet surpasses AudioSet with over 6,800 recording hours ($\\uparrow\\!17\\%$) from nearly 10,000 classes ($\\uparrow\\!18\\times$) for training and more than 400 hours ($\\uparrow\\!7\\times$) across eight strongly labeled evaluation datasets. It serves as a versatile resource for use cases such as multi-label classification, covariate shift or self-supervised learning. We benchmark six well-known DL models in multi-label classification across three distinct training scenarios and outline further evaluation use cases in audio classification. We host our dataset on Hugging Face for easy accessibility and offer an extensive codebase to reproduce our results.", "date": "2024-03-15", "categories": []}
{"id": "2501.09394", "title": "Quantum-Enhanced Transformers for Robust Acoustic Scene Classification in IoT Environments", "abstract": "The proliferation of Internet of Things (IoT) devices equipped with acoustic sensors necessitates robust acoustic scene classification (ASC) capabilities, even in noisy and data-limited environments. Traditional machine learning methods often struggle to generalize effectively under such conditions. To address this, we introduce Q-ASC, a novel Quantum-Inspired Acoustic Scene Classifier that leverages the power of quantum-inspired transformers. By integrating quantum concepts like superposition and entanglement, Q-ASC achieves superior feature learning and enhanced noise resilience compared to classical models. Furthermore, we introduce a Quantum Variational Autoencoder (QVAE) based data augmentation technique to mitigate the challenge of limited labeled data in IoT deployments. Extensive evaluations on the Tampere University of Technology (TUT) Acoustic Scenes 2016 benchmark dataset demonstrate that Q-ASC achieves remarkable accuracy between 68.3% and 88.5% under challenging conditions, outperforming state-of-the-art methods by over 5% in the best case. This research paves the way for deploying intelligent acoustic sensing in IoT networks, with potential applications in smart homes, industrial monitoring, and environmental surveillance, even in adverse acoustic environments.", "date": "2025-01-16", "categories": []}
{"id": "2505.12154", "title": "Learning to Highlight Audio by Watching Movies", "abstract": "Recent years have seen a significant increase in video content creation and consumption. Crafting engaging content requires the careful curation of both visual and audio elements. While visual cue curation, through techniques like optimal viewpoint selection or post-editing, has been central to media production, its natural counterpart, audio, has not undergone equivalent advancements. This often results in a disconnect between visual and acoustic saliency. To bridge this gap, we introduce a novel task: visually-guided acoustic highlighting, which aims to transform audio to deliver appropriate highlighting effects guided by the accompanying video, ultimately creating a more harmonious audio-visual experience. We propose a flexible, transformer-based multimodal framework to solve this task. To train our model, we also introduce a new dataset -- the muddy mix dataset, leveraging the meticulous audio and video crafting found in movies, which provides a form of free supervision. We develop a pseudo-data generation process to simulate poorly mixed audio, mimicking real-world scenarios through a three-step process -- separation, adjustment, and remixing. Our approach consistently outperforms several baselines in both quantitative and subjective evaluation. We also systematically study the impact of different types of contextual guidance and difficulty levels of the dataset. Our project page is here: https://wikichao.github.io/VisAH/.", "date": "2025-05-17", "categories": []}
{"id": "2505.13032", "title": "MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix", "abstract": "We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks. MMAR comprises 1,000 meticulously curated audio-question-answer triplets, collected from real-world internet videos and refined through iterative error corrections and quality checks to ensure high quality. Unlike existing benchmarks that are limited to specific domains of sound, music, or speech, MMAR extends them to a broad spectrum of real-world audio scenarios, including mixed-modality combinations of sound, music, and speech. Each question in MMAR is hierarchically categorized across four reasoning layers: Signal, Perception, Semantic, and Cultural, with additional sub-categories within each layer to reflect task diversity and complexity. To further foster research in this area, we annotate every question with a Chain-of-Thought (CoT) rationale to promote future advancements in audio reasoning. Each item in the benchmark demands multi-step deep reasoning beyond surface-level understanding. Moreover, a part of the questions requires graduate-level perceptual and domain-specific knowledge, elevating the benchmark's difficulty and depth. We evaluate MMAR using a broad set of models, including Large Audio-Language Models (LALMs), Large Audio Reasoning Models (LARMs), Omni Language Models (OLMs), Large Language Models (LLMs), and Large Reasoning Models (LRMs), with audio caption inputs. The performance of these models on MMAR highlights the benchmark's challenging nature, and our analysis further reveals critical limitations of understanding and reasoning capabilities among current models. We hope MMAR will serve as a catalyst for future advances in this important but little-explored area.", "date": "2025-05-19", "categories": []}
{"id": "2502.16936", "title": "Supervised contrastive learning from weakly-labeled audio segments for musical version matching", "abstract": "Detecting musical versions (different renditions of the same piece) is a challenging task with important applications. Because of the ground truth nature, existing approaches match musical versions at the track level (e.g., whole song). However, most applications require to match them at the segment level (e.g., 20s chunks). In addition, existing approaches resort to classification and triplet losses, disregarding more recent losses that could bring meaningful improvements. In this paper, we propose a method to learn from weakly annotated segments, together with a contrastive loss variant that outperforms well-studied alternatives. The former is based on pairwise segment distance reductions, while the latter modifies an existing loss following decoupling, hyper-parameter, and geometric considerations. With these two elements, we do not only achieve state-of-the-art results in the standard track-level evaluation, but we also obtain a breakthrough performance in a segment-level evaluation. We believe that, due to the generality of the challenges addressed here, the proposed methods may find utility in domains beyond audio or musical version matching.", "date": "2025-02-24", "categories": []}
{"id": "2505.11375", "title": "Anti-aliasing of neural distortion effects via model fine tuning", "abstract": "Neural networks have become ubiquitous with guitar distortion effects modelling in recent years. Despite their ability to yield perceptually convincing models, they are susceptible to frequency aliasing when driven by high frequency and high gain inputs. Nonlinear activation functions create both the desired harmonic distortion and unwanted aliasing distortion as the bandwidth of the signal is expanded beyond the Nyquist frequency. Here, we present a method for reducing aliasing in neural models via a teacher-student fine tuning approach, where the teacher is a pre-trained model with its weights frozen, and the student is a copy of this with learnable parameters. The student is fine-tuned against an aliasing-free dataset generated by passing sinusoids through the original model and removing non-harmonic components from the output spectra. Our results show that this method significantly suppresses aliasing for both long-short-term-memory networks (LSTM) and temporal convolutional networks (TCN). In the majority of our case studies, the reduction in aliasing was greater than that achieved by two times oversampling. One side-effect of the proposed method is that harmonic distortion components are also affected. This adverse effect was found to be model-dependent, with the LSTM models giving the best balance between anti-aliasing and preserving the perceived similarity to an analog reference device.", "date": "2025-05-16", "categories": []}
{"id": "2409.09647", "title": "Self-supervised Learning for Acoustic Few-Shot Classification", "abstract": "Labelled data are limited and self-supervised learning is one of the most important approaches for reducing labelling requirements. While it has been extensively explored in the image domain, it has so far not received the same amount of attention in the acoustic domain. Yet, reducing labelling is a key requirement for many acoustic applications. Specifically in bioacoustic, there are rarely sufficient labels for fully supervised learning available. This has led to the widespread use of acoustic recognisers that have been pre-trained on unrelated data for bioacoustic tasks. We posit that training on the actual task data and combining self-supervised pre-training with few-shot classification is a superior approach that has the ability to deliver high accuracy even when only a few labels are available. To this end, we introduce and evaluate a new architecture that combines CNN-based preprocessing with feature extraction based on state space models (SSMs). This combination is motivated by the fact that CNN-based networks alone struggle to capture temporal information effectively, which is crucial for classifying acoustic signals. SSMs, specifically S4 and Mamba, on the other hand, have been shown to have an excellent ability to capture long-range dependencies in sequence data. We pre-train this architecture using contrastive learning on the actual task data and subsequent fine-tuning with an extremely small amount of labelled data. We evaluate the performance of this proposed architecture for ($n$-shot, $n$-class) classification on standard benchmarks as well as real-world data. Our evaluation shows that it outperforms state-of-the-art architectures on the few-shot classification problem.", "date": "2024-09-15", "categories": []}
{"id": "2505.09633", "title": "Detecting Musical Deepfakes", "abstract": "The proliferation of Text-to-Music (TTM) platforms has democratized music creation, enabling users to effortlessly generate high-quality compositions. However, this innovation also presents new challenges to musicians and the broader music industry. This study investigates the detection of AI-generated songs using the FakeMusicCaps dataset by classifying audio as either deepfake or human. To simulate real-world adversarial conditions, tempo stretching and pitch shifting were applied to the dataset. Mel spectrograms were generated from the modified audio, then used to train and evaluate a convolutional neural network. In addition to presenting technical results, this work explores the ethical and societal implications of TTM platforms, arguing that carefully designed detection systems are essential to both protecting artists and unlocking the positive potential of generative AI in music.", "date": "2025-05-03", "categories": []}
{"id": "2503.11197", "title": "Reinforcement Learning Outperforms Supervised Fine-Tuning: A Case Study on Audio Question Answering", "abstract": "Recently, reinforcement learning (RL) has been shown to greatly enhance the reasoning capabilities of large language models (LLMs), and RL-based approaches have been progressively applied to visual multimodal tasks. However, the audio modality has largely been overlooked in these developments. Thus, we conduct a series of RL explorations in audio understanding and reasoning, specifically focusing on the audio question answering (AQA) task. We leverage the group relative policy optimization (GRPO) algorithm to Qwen2-Audio-7B-Instruct, and our experiments demonstrated state-of-the-art performance on the MMAU Test-mini benchmark, achieving an accuracy rate of 64.5%. The main findings in this technical report are as follows: 1) The GRPO algorithm can be effectively applied to large audio language models (LALMs), even when the model has only 8.2B parameters; 2) With only 38k post-training samples, RL significantly outperforms supervised fine-tuning (SFT), indicating that RL-based approaches can be effective without large datasets; 3) The explicit reasoning process has not shown significant benefits for AQA tasks, and how to efficiently utilize deep thinking remains an open question for further research; 4) LALMs still lag far behind humans auditory-language reasoning, suggesting that the RL-based approaches warrant further exploration. Our project is available at https://github.com/xiaomi-research/r1-aqa and https://huggingface.co/mispeech/r1-aqa.", "date": "2025-03-14", "categories": []}
{"id": "2505.09091", "title": "DPN-GAN: Inducing Periodic Activations in Generative Adversarial Networks for High-Fidelity Audio Synthesis", "abstract": "In recent years, generative adversarial networks (GANs) have made significant progress in generating audio sequences. However, these models typically rely on bandwidth-limited mel-spectrograms, which constrain the resolution of generated audio sequences, and lead to mode collapse during conditional generation. To address this issue, we propose Deformable Periodic Network based GAN (DPN-GAN), a novel GAN architecture that incorporates a kernel-based periodic ReLU activation function to induce periodic bias in audio generation. This innovative approach enhances the model's ability to capture and reproduce intricate audio patterns. In particular, our proposed model features a DPN module for multi-resolution generation utilizing deformable convolution operations, allowing for adaptive receptive fields that improve the quality and fidelity of the synthetic audio. Additionally, we enhance the discriminator network using deformable convolution to better distinguish between real and generated samples, further refining the audio quality. We trained two versions of the model: DPN-GAN small (38.67M parameters) and DPN-GAN large (124M parameters). For evaluation, we use five different datasets, covering both speech synthesis and music generation tasks, to demonstrate the efficiency of the DPN-GAN. The experimental results demonstrate that DPN-GAN delivers superior performance on both out-of-distribution and noisy data, showcasing its robustness and adaptability. Trained across various datasets, DPN-GAN outperforms state-of-the-art GAN architectures on standard evaluation metrics, and exhibits increased robustness in synthesized audio.", "date": "2025-05-14", "categories": []}
{"id": "2412.16267", "title": "A Classification Benchmark for Artificial Intelligence Detection of Laryngeal Cancer from Patient Voice", "abstract": "Cases of laryngeal cancer are predicted to rise significantly in the coming years. Current diagnostic pathways are inefficient, putting undue stress on both patients and the medical system. Artificial intelligence offers a promising solution by enabling non-invasive detection of laryngeal cancer from patient voice, which could help prioritise referrals more effectively. A major barrier in this field is the lack of reproducible methods. Our work addresses this challenge by introducing a benchmark suite comprising 36 models trained and evaluated on open-source datasets. These models classify patients with benign and malignant voice pathologies. All models are accessible in a public repository, providing a foundation for future research. We evaluate three algorithms and three audio feature sets, including both audio-only inputs and multimodal inputs incorporating demographic and symptom data. Our best model achieves a balanced accuracy of 83.7%, sensitivity of 84.0%, specificity of 83.3%, and AUROC of 91.8%.", "date": "2024-12-20", "categories": []}
{"id": "2505.03054", "title": "BLAB: Brutally Long Audio Bench", "abstract": "Developing large audio language models (LMs) capable of understanding diverse spoken interactions is essential for accommodating the multimodal nature of human communication and can increase the accessibility of language technologies across different user populations. Recent work on audio LMs has primarily evaluated their performance on short audio segments, typically under 30 seconds, with limited exploration of long-form conversational speech segments that more closely reflect natural user interactions with these models. We introduce Brutally Long Audio Bench (BLAB), a challenging long-form audio benchmark that evaluates audio LMs on localization, duration estimation, emotion, and counting tasks using audio segments averaging 51 minutes in length. BLAB consists of 833+ hours of diverse, full-length audio clips, each paired with human-annotated, text-based natural language questions and answers. Our audio data were collected from permissively licensed sources and underwent a human-assisted filtering process to ensure task compliance. We evaluate six open-source and proprietary audio LMs on BLAB and find that all of them, including advanced models such as Gemini 2.0 Pro and GPT-4o, struggle with the tasks in BLAB. Our comprehensive analysis reveals key insights into the trade-offs between task difficulty and audio duration. In general, we find that audio LMs struggle with long-form speech, with performance declining as duration increases. They perform poorly on localization, temporal reasoning, counting, and struggle to understand non-phonemic information, relying more on prompts than audio content. BLAB serves as a challenging evaluation framework to develop audio LMs with robust long-form audio understanding capabilities.", "date": "2025-05-05", "categories": []}
{"id": "2505.08215", "title": "Unveiling the Best Practices for Applying Speech Foundation Models to Speech Intelligibility Prediction for Hearing-Impaired People", "abstract": "Speech foundation models (SFMs) have demonstrated strong performance across a variety of downstream tasks, including speech intelligibility prediction for hearing-impaired people (SIP-HI). However, optimizing SFMs for SIP-HI has been insufficiently explored. In this paper, we conduct a comprehensive study to identify key design factors affecting SIP-HI performance with 5 SFMs, focusing on encoder layer selection, prediction head architecture, and ensemble configurations. Our findings show that, contrary to traditional use-all-layers methods, selecting a single encoder layer yields better results. Additionally, temporal modeling is crucial for effective prediction heads. We also demonstrate that ensembling multiple SFMs improves performance, with stronger individual models providing greater benefit. Finally, we explore the relationship between key SFM attributes and their impact on SIP-HI performance. Our study offers practical insights into effectively adapting SFMs for speech intelligibility prediction for hearing-impaired populations.", "date": "2025-05-13", "categories": []}
{"id": "2505.08681", "title": "A Mamba-based Network for Semi-supervised Singing Melody Extraction Using Confidence Binary Regularization", "abstract": "Singing melody extraction (SME) is a key task in the field of music information retrieval. However, existing methods are facing several limitations: firstly, prior models use transformers to capture the contextual dependencies, which requires quadratic computation resulting in low efficiency in the inference stage. Secondly, prior works typically rely on frequencysupervised methods to estimate the fundamental frequency (f0), which ignores that the musical performance is actually based on notes. Thirdly, transformers typically require large amounts of labeled data to achieve optimal performances, but the SME task lacks of sufficient annotated data. To address these issues, in this paper, we propose a mamba-based network, called SpectMamba, for semi-supervised singing melody extraction using confidence binary regularization. In particular, we begin by introducing vision mamba to achieve computational linear complexity. Then, we propose a novel note-f0 decoder that allows the model to better mimic the musical performance. Further, to alleviate the scarcity of the labeled data, we introduce a confidence binary regularization (CBR) module to leverage the unlabeled data by maximizing the probability of the correct classes. The proposed method is evaluated on several public datasets and the conducted experiments demonstrate the effectiveness of our proposed method.", "date": "2025-05-13", "categories": []}
{"id": "2505.06271", "title": "Tri-MTL: A Triple Multitask Learning Approach for Respiratory Disease Diagnosis", "abstract": "Auscultation remains a cornerstone of clinical practice, essential for both initial evaluation and continuous monitoring. Clinicians listen to the lung sounds and make a diagnosis by combining the patient's medical history and test results. Given this strong association, multitask learning (MTL) can offer a compelling framework to simultaneously model these relationships, integrating respiratory sound patterns with disease manifestations. While MTL has shown considerable promise in medical applications, a significant research gap remains in understanding the complex interplay between respiratory sounds, disease manifestations, and patient metadata attributes. This study investigates how integrating MTL with cutting-edge deep learning architectures can enhance both respiratory sound classification and disease diagnosis. Specifically, we extend recent findings regarding the beneficial impact of metadata on respiratory sound classification by evaluating its effectiveness within an MTL framework. Our comprehensive experiments reveal significant improvements in both lung sound classification and diagnostic performance when the stethoscope information is incorporated into the MTL architecture.", "date": "2025-05-06", "categories": []}
{"id": "2505.07609", "title": "TACOS: Temporally-aligned Audio CaptiOnS for Language-Audio Pretraining", "abstract": "Learning to associate audio with textual descriptions is valuable for a range of tasks, including pretraining, zero-shot classification, audio retrieval, audio captioning, and text-conditioned audio generation. Existing contrastive language-audio pretrained models are typically trained using global, clip-level descriptions, which provide only weak temporal supervision. We hypothesize that CLAP-like language-audio models - particularly, if they are expected to produce frame-level embeddings - can benefit from a stronger temporal supervision. To confirm our hypothesis, we curate a novel dataset of approximately 12,000 audio recordings from Freesound, each annotated with single-sentence free-text descriptions linked to a specific temporal segment in an audio recording. We use large language models to clean these annotations by removing references to non-audible events, transcribed speech, typos, and annotator language bias. We further propose a frame-wise contrastive training strategy that learns to align text descriptions with temporal regions in an audio recording and demonstrate that our model has better temporal text-audio alignment abilities compared to models trained only on global captions when evaluated on the AudioSet Strong benchmark. The dataset and our source code are available on Zenodo and GitHub, respectively.", "date": "2025-05-12", "categories": []}
{"id": "2505.07709", "title": "ISAC: An Invertible and Stable Auditory Filter Bank with Customizable Kernels for ML Integration", "abstract": "This paper introduces ISAC, an invertible and stable, perceptually-motivated filter bank that is specifically designed to be integrated into machine learning paradigms. More precisely, the center frequencies and bandwidths of the filters are chosen to follow a non-linear, auditory frequency scale, the filter kernels have user-defined maximum temporal support and may serve as learnable convolutional kernels, and there exists a corresponding filter bank such that both form a perfect reconstruction pair. ISAC provides a powerful and user-friendly audio front-end suitable for any application, including analysis-synthesis schemes.", "date": "2025-05-12", "categories": []}
{"id": "2505.06042", "title": "Learning Music Audio Representations With Limited Data", "abstract": "Large deep-learning models for music, including those focused on learning general-purpose music audio representations, are often assumed to require substantial training data to achieve high performance. If true, this would pose challenges in scenarios where audio data or annotations are scarce, such as for underrepresented music traditions, non-popular genres, and personalized music creation and listening. Understanding how these models behave in limited-data scenarios could be crucial for developing techniques to tackle them.\n  In this work, we investigate the behavior of several music audio representation models under limited-data learning regimes. We consider music models with various architectures, training paradigms, and input durations, and train them on data collections ranging from 5 to 8,000 minutes long. We evaluate the learned representations on various music information retrieval tasks and analyze their robustness to noise. We show that, under certain conditions, representations from limited-data and even random models perform comparably to ones from large-dataset models, though handcrafted features outperform all learned representations in some tasks.", "date": "2025-05-09", "categories": []}
{"id": "2505.04419", "title": "Recognizing Ornaments in Vocal Indian Art Music with Active Annotation", "abstract": "  Ornamentations, embellishments, or microtonal inflections are essential to\nmelodic expression across many musical traditions, adding depth, nuance, and\nemotional impact to performances. Recognizing ornamentations in singing voices\nis key to MIR, with potential applications in music pedagogy, singer\nidentification, genre classification, and controlled singing voice generation.\nHowever, the lack of annotated datasets and specialized modeling approaches\nremains a major obstacle for progress in this research area. In this work, we\nintroduce R\\=aga Ornamentation Detection (ROD), a novel dataset comprising\nIndian classical music recordings curated by expert musicians. The dataset is\nannotated using a custom Human-in-the-Loop tool for six vocal ornaments marked\nas event-based labels. Using this dataset, we develop an ornamentation\ndetection model based on deep time-series analysis, preserving ornament\nboundaries during the chunking of long audio recordings. We conduct experiments\nusing different train-test configurations within the ROD dataset and also\nevaluate our approach on a separate, manually annotated dataset of Indian\nclassical concert recordings. Our experimental results support the superior\nperformance of our proposed approach over the baseline CRNN.\n", "date": "2025-05-07", "categories": []}
{"id": "2309.08751", "title": "Diverse Audio Embeddings -- Bringing Features Back Outperforms CLAP!", "abstract": "  With the advent of modern AI architectures, a shift has happened towards\nend-to-end architectures. This pivot has led to neural architectures being\ntrained without domain-specific biases/knowledge, optimized according to the\ntask. We in this paper, learn audio embeddings via diverse feature\nrepresentations, in this case, domain-specific. For the case of audio\nclassification over hundreds of categories of sound, we learn robust separate\nembeddings for diverse audio properties such as pitch, timbre, and neural\nrepresentation, along with also learning it via an end-to-end architecture. We\nobserve handcrafted embeddings, e.g., pitch and timbre-based, although on their\nown, are not able to beat a fully end-to-end representation, yet adding these\ntogether with end-to-end embedding helps us, significantly improve performance.\nThis work would pave the way to bring some domain expertise with end-to-end\nmodels to learn robust, diverse representations, surpassing the performance of\njust training end-to-end models.\n", "date": "2023-09-15", "categories": []}
{"id": "2409.07016", "title": "Improving Anomalous Sound Detection via Low-Rank Adaptation Fine-Tuning\n  of Pre-Trained Audio Models", "abstract": "  Anomalous Sound Detection (ASD) has gained significant interest through the\napplication of various Artificial Intelligence (AI) technologies in industrial\nsettings. Though possessing great potential, ASD systems can hardly be readily\ndeployed in real production sites due to the generalization problem, which is\nprimarily caused by the difficulty of data collection and the complexity of\nenvironmental factors. This paper introduces a robust ASD model that leverages\naudio pre-trained models. Specifically, we fine-tune these models using machine\noperation data, employing SpecAug as a data augmentation strategy.\nAdditionally, we investigate the impact of utilizing Low-Rank Adaptation (LoRA)\ntuning instead of full fine-tuning to address the problem of limited data for\nfine-tuning. Our experiments on the DCASE2023 Task 2 dataset establish a new\nbenchmark of 77.75% on the evaluation set, with a significant improvement of\n6.48% compared with previous state-of-the-art (SOTA) models, including top-tier\ntraditional convolutional networks and speech pre-trained models, which\ndemonstrates the effectiveness of audio pre-trained models with LoRA tuning.\nAblation studies are also conducted to showcase the efficacy of the proposed\nscheme.\n", "date": "2024-09-11", "categories": []}
{"id": "2505.01880", "title": "Weakly-supervised Audio Temporal Forgery Localization via Progressive\n  Audio-language Co-learning Network", "abstract": "  Audio temporal forgery localization (ATFL) aims to find the precise forgery\nregions of the partial spoof audio that is purposefully modified. Existing ATFL\nmethods rely on training efficient networks using fine-grained annotations,\nwhich are obtained costly and challenging in real-world scenarios. To meet this\nchallenge, in this paper, we propose a progressive audio-language co-learning\nnetwork (LOCO) that adopts co-learning and self-supervision manners to prompt\nlocalization performance under weak supervision scenarios. Specifically, an\naudio-language co-learning module is first designed to capture forgery\nconsensus features by aligning semantics from temporal and global perspectives.\nIn this module, forgery-aware prompts are constructed by using utterance-level\nannotations together with learnable prompts, which can incorporate semantic\npriors into temporal content features dynamically. In addition, a forgery\nlocalization module is applied to produce forgery proposals based on fused\nforgery-class activation sequences. Finally, a progressive refinement strategy\nis introduced to generate pseudo frame-level labels and leverage supervised\nsemantic contrastive learning to amplify the semantic distinction between real\nand fake content, thereby continuously optimizing forgery-aware features.\nExtensive experiments show that the proposed LOCO achieves SOTA performance on\nthree public benchmarks.\n", "date": "2025-05-03", "categories": []}
{"id": "2505.04451", "title": "Automatic Music Transcription using Convolutional Neural Networks and\n  Constant-Q transform", "abstract": "  Automatic music transcription (AMT) is the problem of analyzing an audio\nrecording of a musical piece and detecting notes that are being played. AMT is\na challenging problem, particularly when it comes to polyphonic music. The goal\nof AMT is to produce a score representation of a music piece, by analyzing a\nsound signal containing multiple notes played simultaneously. In this work, we\ndesign a processing pipeline that can transform classical piano audio files in\n.wav format into a music score representation. The features from the audio\nsignals are extracted using the constant-Q transform, and the resulting\ncoefficients are used as an input to the convolutional neural network (CNN)\nmodel.\n", "date": "2025-05-07", "categories": []}
{"id": "2505.04621", "title": "Score Distillation Sampling for Audio: Source Separation, Synthesis, and\n  Beyond", "abstract": "  We introduce Audio-SDS, a generalization of Score Distillation Sampling (SDS)\nto text-conditioned audio diffusion models. While SDS was initially designed\nfor text-to-3D generation using image diffusion, its core idea of distilling a\npowerful generative prior into a separate parametric representation extends to\nthe audio domain. Leveraging a single pretrained model, Audio-SDS enables a\nbroad range of tasks without requiring specialized datasets. In particular, we\ndemonstrate how Audio-SDS can guide physically informed impact sound\nsimulations, calibrate FM-synthesis parameters, and perform prompt-specified\nsource separation. Our findings illustrate the versatility of\ndistillation-based methods across modalities and establish a robust foundation\nfor future work using generative priors in audio tasks.\n", "date": "2025-05-07", "categories": []}
{"id": "2406.16020", "title": "AudioBench: A Universal Benchmark for Audio Large Language Models", "abstract": "  We introduce AudioBench, a universal benchmark designed to evaluate Audio\nLarge Language Models (AudioLLMs). It encompasses 8 distinct tasks and 26\ndatasets, among which, 7 are newly proposed datasets. The evaluation targets\nthree main aspects: speech understanding, audio scene understanding, and voice\nunderstanding (paralinguistic). Despite recent advancements, there lacks a\ncomprehensive benchmark for AudioLLMs on instruction following capabilities\nconditioned on audio signals. AudioBench addresses this gap by setting up\ndatasets as well as desired evaluation metrics. Besides, we also evaluated the\ncapabilities of five popular models and found that no single model excels\nconsistently across all tasks. We outline the research outlook for AudioLLMs\nand anticipate that our open-sourced evaluation toolkit, data, and leaderboard\nwill offer a robust testbed for future model developments.\n", "date": "2024-06-23", "categories": []}
{"id": "2412.15023", "title": "FolAI: Synchronized Foley Sound Generation with Semantic and Temporal\n  Alignment", "abstract": "  Traditional sound design workflows rely on manual alignment of audio events\nto visual cues, as in Foley sound design, where everyday actions like footsteps\nor object interactions are recreated to match the on-screen motion. This\nprocess is time-consuming, difficult to scale, and lacks automation tools that\npreserve creative intent. Despite recent advances in vision-to-audio\ngeneration, producing temporally coherent and semantically controllable sound\neffects from video remains a major challenge. To address these limitations, we\nintroduce FolAI, a two-stage generative framework that decouples the when and\nthe what of sound synthesis, i.e., the temporal structure extraction and the\nsemantically guided generation, respectively. In the first stage, we estimate a\nsmooth control signal from the video that captures the motion intensity and\nrhythmic structure over time, serving as a temporal scaffold for the audio. In\nthe second stage, a diffusion-based generative model produces sound effects\nconditioned both on this temporal envelope and on high-level semantic\nembeddings, provided by the user, that define the desired auditory content\n(e.g., material or action type). This modular design enables precise control\nover both timing and timbre, streamlining repetitive tasks while preserving\ncreative flexibility in professional Foley workflows. Results on diverse visual\ncontexts, such as footstep generation and action-specific sonorization,\ndemonstrate that our model reliably produces audio that is temporally aligned\nwith visual motion, semantically consistent with user intent, and perceptually\nrealistic. These findings highlight the potential of FolAI as a controllable\nand modular solution for scalable, high-quality Foley sound synthesis in\nprofessional and interactive settings. Supplementary materials are accessible\non our dedicated demo page at https://ispamm.github.io/FolAI.\n", "date": "2024-12-19", "categories": []}
{"id": "2502.17579", "title": "VANPY: Voice Analysis Framework", "abstract": "  Voice data is increasingly being used in modern digital communications, yet\nthere is still a lack of comprehensive tools for automated voice analysis and\ncharacterization. To this end, we developed the VANPY (Voice Analysis in\nPython) framework for automated pre-processing, feature extraction, and\nclassification of voice data. The VANPY is an open-source end-to-end\ncomprehensive framework that was developed for the purpose of speaker\ncharacterization from voice data. The framework is designed with extensibility\nin mind, allowing for easy integration of new components and adaptation to\nvarious voice analysis applications. It currently incorporates over fifteen\nvoice analysis components - including music/speech separation, voice activity\ndetection, speaker embedding, vocal feature extraction, and various\nclassification models.\n  Four of the VANPY's components were developed in-house and integrated into\nthe framework to extend its speaker characterization capabilities: gender\nclassification, emotion classification, age regression, and height regression.\nThe models demonstrate robust performance across various datasets, although not\nsurpassing state-of-the-art performance.\n  As a proof of concept, we demonstrate the framework's ability to extract\nspeaker characteristics on a use-case challenge of analyzing character voices\nfrom the movie \"Pulp Fiction.\" The results illustrate the framework's\ncapability to extract multiple speaker characteristics, including gender, age,\nheight, emotion type, and emotion intensity measured across three dimensions:\narousal, dominance, and valence.\n", "date": "2025-02-17", "categories": []}
{"id": "2308.04960", "title": "Representation Learning for Audio Privacy Preservation using Source\n  Separation and Robust Adversarial Learning", "abstract": "  Privacy preservation has long been a concern in smart acoustic monitoring\nsystems, where speech can be passively recorded along with a target signal in\nthe system's operating environment. In this study, we propose the integration\nof two commonly used approaches in privacy preservation: source separation and\nadversarial representation learning. The proposed system learns the latent\nrepresentation of audio recordings such that it prevents differentiating\nbetween speech and non-speech recordings. Initially, the source separation\nnetwork filters out some of the privacy-sensitive data, and during the\nadversarial learning process, the system will learn privacy-preserving\nrepresentation on the filtered signal. We demonstrate the effectiveness of our\nproposed method by comparing our method against systems without source\nseparation, without adversarial learning, and without both. Overall, our\nresults suggest that the proposed system can significantly improve speech\nprivacy preservation compared to that of using source separation or adversarial\nlearning solely while maintaining good performance in the acoustic monitoring\ntask.\n", "date": "2023-08-09", "categories": []}
{"id": "2504.16276", "title": "An Automated Pipeline for Few-Shot Bird Call Classification: A Case\n  Study with the Tooth-Billed Pigeon", "abstract": "  This paper presents an automated one-shot bird call classification pipeline\ndesigned for rare species absent from large publicly available classifiers like\nBirdNET and Perch. While these models excel at detecting common birds with\nabundant training data, they lack options for species with only 1-3 known\nrecordings-a critical limitation for conservationists monitoring the last\nremaining individuals of endangered birds. To address this, we leverage the\nembedding space of large bird classification networks and develop a classifier\nusing cosine similarity, combined with filtering and denoising preprocessing\ntechniques, to optimize detection with minimal training data. We evaluate\nvarious embedding spaces using clustering metrics and validate our approach in\nboth a simulated scenario with Xeno-Canto recordings and a real-world test on\nthe critically endangered tooth-billed pigeon (Didunculus strigirostris), which\nhas no existing classifiers and only three confirmed recordings. The final\nmodel achieved 1.0 recall and 0.95 accuracy in detecting tooth-billed pigeon\ncalls, making it practical for use in the field. This open-source system\nprovides a practical tool for conservationists seeking to detect and monitor\nrare species on the brink of extinction.\n", "date": "2025-04-22", "categories": []}
{"id": "2409.15545", "title": "Addressing Emotion Bias in Music Emotion Recognition and Generation with\n  Frechet Audio Distance", "abstract": "  The complex nature of musical emotion introduces inherent bias in both\nrecognition and generation, particularly when relying on a single audio\nencoder, emotion classifier, or evaluation metric. In this work, we conduct a\nstudy on Music Emotion Recognition (MER) and Emotional Music Generation (EMG),\nemploying diverse audio encoders alongside Frechet Audio Distance (FAD), a\nreference-free evaluation metric. Our study begins with a benchmark evaluation\nof MER, highlighting the limitations of using a single audio encoder and the\ndisparities observed across different measurements. We then propose assessing\nMER performance using FAD derived from multiple encoders to provide a more\nobjective measure of musical emotion. Furthermore, we introduce an enhanced EMG\napproach designed to improve both the variability and prominence of generated\nmusical emotion, thereby enhancing its realism. Additionally, we investigate\nthe differences in realism between the emotions conveyed in real and synthetic\nmusic, comparing our EMG model against two baseline models. Experimental\nresults underscore the issue of emotion bias in both MER and EMG and\ndemonstrate the potential of using FAD and diverse audio encoders to evaluate\nmusical emotion more objectively and effectively.\n", "date": "2024-09-23", "categories": []}
{"id": "2409.17899", "title": "Exploring Acoustic Similarity in Emotional Speech and Music via\n  Self-Supervised Representations", "abstract": "  Emotion recognition from speech and music shares similarities due to their\nacoustic overlap, which has led to interest in transferring knowledge between\nthese domains. However, the shared acoustic cues between speech and music,\nparticularly those encoded by Self-Supervised Learning (SSL) models, remain\nlargely unexplored, given the fact that SSL models for speech and music have\nrarely been applied in cross-domain research. In this work, we revisit the\nacoustic similarity between emotion speech and music, starting with an analysis\nof the layerwise behavior of SSL models for Speech Emotion Recognition (SER)\nand Music Emotion Recognition (MER). Furthermore, we perform cross-domain\nadaptation by comparing several approaches in a two-stage fine-tuning process,\nexamining effective ways to utilize music for SER and speech for MER. Lastly,\nwe explore the acoustic similarities between emotional speech and music using\nFrechet audio distance for individual emotions, uncovering the issue of emotion\nbias in both speech and music SSL models. Our findings reveal that while speech\nand music SSL models do capture shared acoustic features, their behaviors can\nvary depending on different emotions due to their training strategies and\ndomain-specificities. Additionally, parameter-efficient fine-tuning can enhance\nSER and MER performance by leveraging knowledge from each other. This study\nprovides new insights into the acoustic similarity between emotional speech and\nmusic, and highlights the potential for cross-domain generalization to improve\nSER and MER systems.\n", "date": "2024-09-26", "categories": []}
{"id": "2504.18539", "title": "Multi-Task Corrupted Prediction for Learning Robust Audio-Visual Speech\n  Representation", "abstract": "  Audio-visual speech recognition (AVSR) incorporates auditory and visual\nmodalities to improve recognition accuracy, particularly in noisy environments\nwhere audio-only speech systems are insufficient. While previous research has\nlargely addressed audio disruptions, few studies have dealt with visual\ncorruptions, e.g., lip occlusions or blurred videos, which are also\ndetrimental. To address this real-world challenge, we propose CAV2vec, a novel\nself-supervised speech representation learning framework particularly designed\nto handle audio-visual joint corruption. CAV2vec employs a self-distillation\napproach with a corrupted prediction task, where the student model learns to\npredict clean targets, generated by the teacher model, with corrupted input\nframes. Specifically, we suggest a unimodal multi-task learning, which distills\ncross-modal knowledge and aligns the corrupted modalities, by predicting clean\naudio targets with corrupted videos, and clean video targets with corrupted\naudios. This strategy mitigates the dispersion in the representation space\ncaused by corrupted modalities, leading to more reliable and robust\naudio-visual fusion. Our experiments on robust AVSR benchmarks demonstrate that\nthe corrupted representation learning method significantly enhances recognition\naccuracy across generalized environments involving various types of corruption.\nOur code is available at https://github.com/sungnyun/cav2vec.\n", "date": "2025-01-23", "categories": []}
{"id": "2504.20923", "title": "End-to-end Audio Deepfake Detection from RAW Waveforms: a RawNet-Based\n  Approach with Cross-Dataset Evaluation", "abstract": "  Audio deepfakes represent a growing threat to digital security and trust,\nleveraging advanced generative models to produce synthetic speech that closely\nmimics real human voices. Detecting such manipulations is especially\nchallenging under open-world conditions, where spoofing methods encountered\nduring testing may differ from those seen during training. In this work, we\npropose an end-to-end deep learning framework for audio deepfake detection that\noperates directly on raw waveforms. Our model, RawNetLite, is a lightweight\nconvolutional-recurrent architecture designed to capture both spectral and\ntemporal features without handcrafted preprocessing. To enhance robustness, we\nintroduce a training strategy that combines data from multiple domains and\nadopts Focal Loss to emphasize difficult or ambiguous samples. We further\ndemonstrate that incorporating codec-based manipulations and applying\nwaveform-level audio augmentations (e.g., pitch shifting, noise, and time\nstretching) leads to significant generalization improvements under realistic\nacoustic conditions. The proposed model achieves over 99.7% F1 and 0.25% EER on\nin-domain data (FakeOrReal), and up to 83.4% F1 with 16.4% EER on a challenging\nout-of-distribution test set (AVSpoof2021 + CodecFake). These findings\nhighlight the importance of diverse training data, tailored objective functions\nand audio augmentations in building resilient and generalizable audio forgery\ndetectors. Code and pretrained models are available at\nhttps://iplab.dmi.unict.it/mfs/Deepfakes/PaperRawNet2025/.\n", "date": "2025-04-29", "categories": []}
{"id": "2504.21366", "title": "DGFNet: End-to-End Audio-Visual Source Separation Based on Dynamic\n  Gating Fusion", "abstract": "  Current Audio-Visual Source Separation methods primarily adopt two design\nstrategies. The first strategy involves fusing audio and visual features at the\nbottleneck layer of the encoder, followed by processing the fused features\nthrough the decoder. However, when there is a significant disparity between the\ntwo modalities, this approach may lead to the loss of critical information. The\nsecond strategy avoids direct fusion and instead relies on the decoder to\nhandle the interaction between audio and visual features. Nonetheless, if the\nencoder fails to integrate information across modalities adequately, the\ndecoder may be unable to effectively capture the complex relationships between\nthem. To address these issues, this paper proposes a dynamic fusion method\nbased on a gating mechanism that dynamically adjusts the modality fusion\ndegree. This approach mitigates the limitations of solely relying on the\ndecoder and facilitates efficient collaboration between audio and visual\nfeatures. Additionally, an audio attention module is introduced to enhance the\nexpressive capacity of audio features, thereby further improving model\nperformance. Experimental results demonstrate that our method achieves\nsignificant performance improvements on two benchmark datasets, validating its\neffectiveness and advantages in Audio-Visual Source Separation tasks.\n", "date": "2025-04-30", "categories": []}
{"id": "2504.15900", "title": "SARI: Structured Audio Reasoning via Curriculum-Guided Reinforcement\n  Learning", "abstract": "  Recent work shows that reinforcement learning(RL) can markedly sharpen the\nreasoning ability of large language models (LLMs) by prompting them to \"think\nbefore answering.\" Yet whether and how these gains transfer to audio-language\nreasoning remains largely unexplored. We extend the Group-Relative Policy\nOptimization (GRPO) framework from DeepSeek-R1 to a Large Audio-Language Model\n(LALM), and construct a 32k sample multiple-choice corpus. Using a two-stage\nregimen supervised fine-tuning on structured and unstructured\nchains-of-thought, followed by curriculum-guided GRPO, we systematically\ncompare implicit vs. explicit, and structured vs. free form reasoning under\nidentical architectures. Our structured audio reasoning model, SARI (Structured\nAudio Reasoning via Curriculum-Guided Reinforcement Learning), achieves a\n16.35% improvement in average accuracy over the base model\nQwen2-Audio-7B-Instruct. Furthermore, the variant built upon Qwen2.5-Omni\nreaches state-of-the-art performance of 67.08% on the MMAU test-mini benchmark.\nAblation experiments show that on the base model we use: (i) SFT warm-up is\nimportant for stable RL training, (ii) structured chains yield more robust\ngeneralization than unstructured ones, and (iii) easy-to-hard curricula\naccelerate convergence and improve final performance. These findings\ndemonstrate that explicit, structured reasoning and curriculum learning\nsubstantially enhances audio-language understanding.\n", "date": "2025-04-22", "categories": []}
{"id": "2504.18950", "title": "Speaker Retrieval in the Wild: Challenges, Effectiveness and Robustness", "abstract": "  There is a growing abundance of publicly available or company-owned\naudio/video archives, highlighting the increasing importance of efficient\naccess to desired content and information retrieval from these archives. This\npaper investigates the challenges, solutions, effectiveness, and robustness of\nspeaker retrieval systems developed \"in the wild\" which involves addressing two\nprimary challenges: extraction of task-relevant labels from limited metadata\nfor system development and evaluation, as well as the unconstrained acoustic\nconditions encountered in the archive, ranging from quiet studios to adverse\nnoisy environments. While we focus on the publicly-available BBC Rewind archive\n(spanning 1948 to 1979), our framework addresses the broader issue of speaker\nretrieval on extensive and possibly aged archives with no control over the\ncontent and acoustic conditions. Typically, these archives offer a brief and\ngeneral file description, mostly inadequate for specific applications like\nspeaker retrieval, and manual annotation of such large-scale archives is\nunfeasible. We explore various aspects of system development (e.g., speaker\ndiarisation, embedding extraction, query selection) and analyse the challenges,\npossible solutions, and their functionality. To evaluate the performance, we\nconduct systematic experiments in both clean setup and against various\ndistortions simulating real-world applications. Our findings demonstrate the\neffectiveness and robustness of the developed speaker retrieval systems,\nestablishing the versatility and scalability of the proposed framework for a\nwide range of applications beyond the BBC Rewind corpus.\n", "date": "2025-04-26", "categories": []}
{"id": "2504.20124", "title": "Pediatric Asthma Detection with Googles HeAR Model: An AI-Driven\n  Respiratory Sound Classifier", "abstract": "  Early detection of asthma in children is crucial to prevent long-term\nrespiratory complications and reduce emergency interventions. This work\npresents an AI-powered diagnostic pipeline that leverages Googles Health\nAcoustic Representations (HeAR) model to detect early signs of asthma from\npediatric respiratory sounds. The SPRSound dataset, the first open-access\ncollection of annotated respiratory sounds in children aged 1 month to 18\nyears, is used to extract 2-second audio segments labeled as wheeze, crackle,\nrhonchi, stridor, or normal. Each segment is embedded into a 512-dimensional\nrepresentation using HeAR, a foundation model pretrained on 300 million\nhealth-related audio clips, including 100 million cough sounds. Multiple\nclassifiers, including SVM, Random Forest, and MLP, are trained on these\nembeddings to distinguish between asthma-indicative and normal sounds. The\nsystem achieves over 91\\% accuracy, with strong performance on precision-recall\nmetrics for positive cases. In addition to classification, learned embeddings\nare visualized using PCA, misclassifications are analyzed through waveform\nplayback, and ROC and confusion matrix insights are provided. This method\ndemonstrates that short, low-resource pediatric recordings, when powered by\nfoundation audio models, can enable fast, noninvasive asthma screening. The\napproach is especially promising for digital diagnostics in remote or\nunderserved healthcare settings.\n", "date": "2025-04-28", "categories": []}
{"id": "2504.20625", "title": "DiffusionRIR: Room Impulse Response Interpolation using Diffusion Models", "abstract": "  Room Impulse Responses (RIRs) characterize acoustic environments and are\ncrucial in multiple audio signal processing tasks. High-quality RIR estimates\ndrive applications such as virtual microphones, sound source localization,\naugmented reality, and data augmentation. However, obtaining RIR measurements\nwith high spatial resolution is resource-intensive, making it impractical for\nlarge spaces or when dense sampling is required. This research addresses the\nchallenge of estimating RIRs at unmeasured locations within a room using\nDenoising Diffusion Probabilistic Models (DDPM). Our method leverages the\nanalogy between RIR matrices and image inpainting, transforming RIR data into a\nformat suitable for diffusion-based reconstruction.\n  Using simulated RIR data based on the image method, we demonstrate our\napproach's effectiveness on microphone arrays of different curvatures, from\nlinear to semi-circular. Our method successfully reconstructs missing RIRs,\neven in large gaps between microphones. Under these conditions, it achieves\naccurate reconstruction, significantly outperforming baseline Spline Cubic\nInterpolation in terms of Normalized Mean Square Error and Cosine Distance\nbetween actual and interpolated RIRs.\n  This research highlights the potential of using generative models for\neffective RIR interpolation, paving the way for generating additional data from\nlimited real-world measurements.\n", "date": "2025-04-29", "categories": []}
{"id": "2410.06016", "title": "Variable Bitrate Residual Vector Quantization for Audio Coding", "abstract": "  Recent state-of-the-art neural audio compression models have progressively\nadopted residual vector quantization (RVQ). Despite this success, these models\nemploy a fixed number of codebooks per frame, which can be suboptimal in terms\nof rate-distortion tradeoff, particularly in scenarios with simple input audio,\nsuch as silence. To address this limitation, we propose variable bitrate RVQ\n(VRVQ) for audio codecs, which allows for more efficient coding by adapting the\nnumber of codebooks used per frame. Furthermore, we propose a gradient\nestimation method for the non-differentiable masking operation that transforms\nfrom the importance map to the binary importance mask, improving model training\nvia a straight-through estimator. We demonstrate that the proposed training\nframework achieves superior results compared to the baseline method and shows\nfurther improvement when applied to the current state-of-the-art codec.\n", "date": "2024-10-08", "categories": []}
{"id": "2406.10719", "title": "Trading Devil: Robust backdoor attack via Stochastic investment models\n  and Bayesian approach", "abstract": "  With the growing use of voice-activated systems and speech recognition\ntechnologies, the danger of backdoor attacks on audio data has grown\nsignificantly. This research looks at a specific type of attack, known as a\nStochastic investment-based backdoor attack (MarketBack), in which adversaries\nstrategically manipulate the stylistic properties of audio to fool speech\nrecognition systems. The security and integrity of machine learning models are\nseriously threatened by backdoor attacks, in order to maintain the reliability\nof audio applications and systems, the identification of such attacks becomes\ncrucial in the context of audio data. Experimental results demonstrated that\nMarketBack is feasible to achieve an average attack success rate close to 100%\nin seven victim models when poisoning less than 1% of the training data.\n", "date": "2024-06-15", "categories": []}
{"id": "2504.18425", "title": "Kimi-Audio Technical Report", "abstract": "  We present Kimi-Audio, an open-source audio foundation model that excels in\naudio understanding, generation, and conversation. We detail the practices in\nbuilding Kimi-Audio, including model architecture, data curation, training\nrecipe, inference deployment, and evaluation. Specifically, we leverage a\n12.5Hz audio tokenizer, design a novel LLM-based architecture with continuous\nfeatures as input and discrete tokens as output, and develop a chunk-wise\nstreaming detokenizer based on flow matching. We curate a pre-training dataset\nthat consists of more than 13 million hours of audio data covering a wide range\nof modalities including speech, sound, and music, and build a pipeline to\nconstruct high-quality and diverse post-training data. Initialized from a\npre-trained LLM, Kimi-Audio is continual pre-trained on both audio and text\ndata with several carefully designed tasks, and then fine-tuned to support a\ndiverse of audio-related tasks. Extensive evaluation shows that Kimi-Audio\nachieves state-of-the-art performance on a range of audio benchmarks including\nspeech recognition, audio understanding, audio question answering, and speech\nconversation. We release the codes, model checkpoints, as well as the\nevaluation toolkits in https://github.com/MoonshotAI/Kimi-Audio.\n", "date": "2025-04-25", "categories": []}
{"id": "2504.17782", "title": "Unleashing the Power of Natural Audio Featuring Multiple Sound Sources", "abstract": "  Universal sound separation aims to extract clean audio tracks corresponding\nto distinct events from mixed audio, which is critical for artificial auditory\nperception. However, current methods heavily rely on artificially mixed audio\nfor training, which limits their ability to generalize to naturally mixed audio\ncollected in real-world environments. To overcome this limitation, we propose\nClearSep, an innovative framework that employs a data engine to decompose\ncomplex naturally mixed audio into multiple independent tracks, thereby\nallowing effective sound separation in real-world scenarios. We introduce two\nremix-based evaluation metrics to quantitatively assess separation quality and\nuse these metrics as thresholds to iteratively apply the data engine alongside\nmodel training, progressively optimizing separation performance. In addition,\nwe propose a series of training strategies tailored to these separated\nindependent tracks to make the best use of them. Extensive experiments\ndemonstrate that ClearSep achieves state-of-the-art performance across multiple\nsound separation tasks, highlighting its potential for advancing sound\nseparation in natural audio scenarios. For more examples and detailed results,\nplease visit our demo page at https://clearsep.github.io.\n", "date": "2025-04-24", "categories": []}
{"id": "2405.17615", "title": "Listenable Maps for Zero-Shot Audio Classifiers", "abstract": "  Interpreting the decisions of deep learning models, including audio\nclassifiers, is crucial for ensuring the transparency and trustworthiness of\nthis technology. In this paper, we introduce LMAC-ZS (Listenable Maps for Audio\nClassifiers in the Zero-Shot context), which, to the best of our knowledge, is\nthe first decoder-based post-hoc interpretation method for explaining the\ndecisions of zero-shot audio classifiers. The proposed method utilizes a novel\nloss function that maximizes the faithfulness to the original similarity\nbetween a given text-and-audio pair. We provide an extensive evaluation using\nthe Contrastive Language-Audio Pretraining (CLAP) model to showcase that our\ninterpreter remains faithful to the decisions in a zero-shot classification\ncontext. Moreover, we qualitatively show that our method produces meaningful\nexplanations that correlate well with different text prompts.\n", "date": "2024-05-27", "categories": []}
{"id": "2410.21897", "title": "Semi-Supervised Self-Learning Enhanced Music Emotion Recognition", "abstract": "  Music emotion recognition (MER) aims to identify the emotions conveyed in a\ngiven musical piece. However, currently, in the field of MER, the available\npublic datasets have limited sample sizes. Recently, segment-based methods for\nemotion-related tasks have been proposed, which train backbone networks on\nshorter segments instead of entire audio clips, thereby naturally augmenting\ntraining samples without requiring additional resources. Then, the predicted\nsegment-level results are aggregated to obtain the entire song prediction. The\nmost commonly used method is that the segment inherits the label of the clip\ncontaining it, but music emotion is not constant during the whole clip. Doing\nso will introduce label noise and make the training easy to overfit. To handle\nthe noisy label issue, we propose a semi-supervised self-learning (SSSL)\nmethod, which can differentiate between samples with correct and incorrect\nlabels in a self-learning manner, thus effectively utilizing the augmented\nsegment-level data. Experiments on three public emotional datasets demonstrate\nthat the proposed method can achieve better or comparable performance.\n", "date": "2024-10-29", "categories": []}
{"id": "2504.15663", "title": "FADEL: Uncertainty-aware Fake Audio Detection with Evidential Deep\n  Learning", "abstract": "  Recently, fake audio detection has gained significant attention, as\nadvancements in speech synthesis and voice conversion have increased the\nvulnerability of automatic speaker verification (ASV) systems to spoofing\nattacks. A key challenge in this task is generalizing models to detect unseen,\nout-of-distribution (OOD) attacks. Although existing approaches have shown\npromising results, they inherently suffer from overconfidence issues due to the\nusage of softmax for classification, which can produce unreliable predictions\nwhen encountering unpredictable spoofing attempts. To deal with this\nlimitation, we propose a novel framework called fake audio detection with\nevidential learning (FADEL). By modeling class probabilities with a Dirichlet\ndistribution, FADEL incorporates model uncertainty into its predictions,\nthereby leading to more robust performance in OOD scenarios. Experimental\nresults on the ASVspoof2019 Logical Access (LA) and ASVspoof2021 LA datasets\nindicate that the proposed method significantly improves the performance of\nbaseline models. Furthermore, we demonstrate the validity of uncertainty\nestimation by analyzing a strong correlation between average uncertainty and\nequal error rate (EER) across different spoofing algorithms.\n", "date": "2025-04-22", "categories": []}
{"id": "2310.04722", "title": "A Holistic Evaluation of Piano Sound Quality", "abstract": "  This paper aims to develop a holistic evaluation method for piano sound\nquality to assist in purchasing decisions. Unlike previous studies that focused\non the effect of piano performance techniques on sound quality, this study\nevaluates the inherent sound quality of different pianos. To derive quality\nevaluation systems, the study uses subjective questionnaires based on a piano\nsound quality dataset. The method selects the optimal piano classification\nmodels by comparing the fine-tuning results of different pre-training models of\nConvolutional Neural Networks (CNN). To improve the interpretability of the\nmodels, the study applies Equivalent Rectangular Bandwidth (ERB) analysis. The\nresults reveal that musically trained individuals are better able to\ndistinguish between the sound quality differences of different pianos. The best\nfine-tuned CNN pre-trained backbone achieves a high accuracy of 98.3% as the\npiano classifier. However, the dataset is limited, and the audio is sliced to\nincrease its quantity, resulting in a lack of diversity and balance, so we use\nfocal loss to reduce the impact of data imbalance. To optimize the method, the\ndataset will be expanded, or few-shot learning techniques will be employed in\nfuture research.\n", "date": "2023-10-07", "categories": []}
{"id": "2408.09269", "title": "Enhancing Audio-Language Models through Self-Supervised Post-Training\n  with Text-Audio Pairs", "abstract": "  Research on multi-modal contrastive learning strategies for audio and text\nhas rapidly gained interest. Contrastively trained Audio-Language Models\n(ALMs), such as CLAP, which establish a unified representation across audio and\nlanguage modalities, have enhanced the efficacy in various subsequent tasks by\nproviding good text aligned audio encoders and vice versa. These improvements\nare evident in areas like zero-shot audio classification and audio retrieval,\namong others. However, the ability of these models to understand natural\nlanguage and temporal relations is still a largely unexplored and open field\nfor research. In this paper, we propose to equip the multi-modal ALMs with\ntemporal understanding without loosing their inherent prior capabilities of\naudio-language tasks with a temporal instillation method TeminAL. We implement\na two-stage training scheme TeminAL A $\\&$ B, where the model first learns to\ndifferentiate between multiple sounds in TeminAL A, followed by a phase that\ninstills a sense of time, thereby enhancing its temporal understanding in\nTeminAL B. This approach results in an average performance gain of $5.28\\%$ in\ntemporal understanding on the ESC-50 dataset, while the model remains\ncompetitive in zero-shot retrieval and classification tasks on the\nAudioCap/Clotho datasets. We also note the lack of proper evaluation techniques\nfor contrastive ALMs and propose a strategy for evaluating ALMs in zero-shot\nsettings. The general-purpose zero-shot model evaluation strategy ZSTE, is used\nto evaluate various prior models. ZSTE demonstrates a general strategy to\nevaluate all ZS contrastive models. The model trained with TeminAL successfully\noutperforms current models on most downstream tasks.\n", "date": "2024-08-17", "categories": []}
{"id": "2409.06635", "title": "MoWE-Audio: Multitask AudioLLMs with Mixture of Weak Encoders", "abstract": "  The rapid advancements in large language models (LLMs) have significantly\nenhanced natural language processing capabilities, facilitating the development\nof AudioLLMs that process and understand speech and audio inputs alongside\ntext. Existing AudioLLMs typically combine a pre-trained audio encoder with a\npre-trained LLM, which are subsequently finetuned on specific audio tasks.\nHowever, the pre-trained audio encoder has constrained capacity to capture\nfeatures for new tasks and datasets. To address this, we propose to incorporate\nmixtures of `weak' encoders (MoWE) into the AudioLLM framework. MoWE\nsupplements a base encoder with a pool of relatively light weight encoders,\nselectively activated based on the audio input to enhance feature extraction\nwithout significantly increasing model size. Our empirical results demonstrate\nthat MoWE effectively improves multi-task performance, broadening the\napplicability of AudioLLMs to more diverse audio tasks.\n", "date": "2024-09-10", "categories": []}
{"id": "2504.14076", "title": "Transformation of audio embeddings into interpretable, concept-based\n  representations", "abstract": "  Advancements in audio neural networks have established state-of-the-art\nresults on downstream audio tasks. However, the black-box structure of these\nmodels makes it difficult to interpret the information encoded in their\ninternal audio representations. In this work, we explore the semantic\ninterpretability of audio embeddings extracted from these neural networks by\nleveraging CLAP, a contrastive learning model that brings audio and text into a\nshared embedding space. We implement a post-hoc method to transform CLAP\nembeddings into concept-based, sparse representations with semantic\ninterpretability. Qualitative and quantitative evaluations show that the\nconcept-based representations outperform or match the performance of original\naudio embeddings on downstream tasks while providing interpretability.\nAdditionally, we demonstrate that fine-tuning the concept-based representations\ncan further improve their performance on downstream tasks. Lastly, we publish\nthree audio-specific vocabularies for concept-based interpretability of audio\nembeddings.\n", "date": "2025-04-18", "categories": []}
{"id": "2504.13791", "title": "Collective Learning Mechanism based Optimal Transport Generative\n  Adversarial Network for Non-parallel Voice Conversion", "abstract": "  After demonstrating significant success in image synthesis, Generative\nAdversarial Network (GAN) models have likewise made significant progress in the\nfield of speech synthesis, leveraging their capacity to adapt the precise\ndistribution of target data through adversarial learning processes. Notably, in\nthe realm of State-Of-The-Art (SOTA) GAN-based Voice Conversion (VC) models,\nthere exists a substantial disparity in naturalness between real and\nGAN-generated speech samples. Furthermore, while many GAN models currently\noperate on a single generator discriminator learning approach, optimizing\ntarget data distribution is more effectively achievable through a single\ngenerator multi-discriminator learning scheme. Hence, this study introduces a\nnovel GAN model named Collective Learning Mechanism-based Optimal Transport GAN\n(CLOT-GAN) model, incorporating multiple discriminators, including the Deep\nConvolutional Neural Network (DCNN) model, Vision Transformer (ViT), and\nconformer. The objective of integrating various discriminators lies in their\nability to comprehend the formant distribution of mel-spectrograms, facilitated\nby a collective learning mechanism. Simultaneously, the inclusion of Optimal\nTransport (OT) loss aims to precisely bridge the gap between the source and\ntarget data distribution, employing the principles of OT theory. The\nexperimental validation on VCC 2018, VCTK, and CMU-Arctic datasets confirms\nthat the CLOT-GAN-VC model outperforms existing VC models in objective and\nsubjective assessments.\n", "date": "2025-04-18", "categories": []}
{"id": "2406.19388", "title": "Taming Data and Transformers for Audio Generation", "abstract": "  The scalability of ambient sound generators is hindered by data scarcity,\ninsufficient caption quality, and limited scalability in model architecture.\nThis work addresses these challenges by advancing both data and model scaling.\nFirst, we propose an efficient and scalable dataset collection pipeline\ntailored for ambient audio generation, resulting in AutoReCap-XL, the largest\nambient audio-text dataset with over 47 million clips. To provide high-quality\ntextual annotations, we propose AutoCap, a high-quality automatic audio\ncaptioning model. By adopting a Q-Former module and leveraging audio metadata,\nAutoCap substantially enhances caption quality, reaching a CIDEr score of\n$83.2$, a $3.2\\%$ improvement over previous captioning models. Finally, we\npropose GenAu, a scalable transformer-based audio generation architecture that\nwe scale up to 1.25B parameters. We demonstrate its benefits from data scaling\nwith synthetic captions as well as model size scaling. When compared to\nbaseline audio generators trained at similar size and data scale, GenAu obtains\nsignificant improvements of $4.7\\%$ in FAD score, $11.1\\%$ in IS, and $13.5\\%$\nin CLAP score. Our code, model checkpoints, and dataset are publicly available.\n", "date": "2024-06-27", "categories": []}
{"id": "2409.17285", "title": "SpoofCeleb: Speech Deepfake Detection and SASV In The Wild", "abstract": "  This paper introduces SpoofCeleb, a dataset designed for Speech Deepfake\nDetection (SDD) and Spoofing-robust Automatic Speaker Verification (SASV),\nutilizing source data from real-world conditions and spoofing attacks generated\nby Text-To-Speech (TTS) systems also trained on the same real-world data.\nRobust recognition systems require speech data recorded in varied acoustic\nenvironments with different levels of noise to be trained. However, current\ndatasets typically include clean, high-quality recordings (bona fide data) due\nto the requirements for TTS training; studio-quality or well-recorded read\nspeech is typically necessary to train TTS models. Current SDD datasets also\nhave limited usefulness for training SASV models due to insufficient speaker\ndiversity. SpoofCeleb leverages a fully automated pipeline we developed that\nprocesses the VoxCeleb1 dataset, transforming it into a suitable form for TTS\ntraining. We subsequently train 23 contemporary TTS systems. SpoofCeleb\ncomprises over 2.5 million utterances from 1,251 unique speakers, collected\nunder natural, real-world conditions. The dataset includes carefully\npartitioned training, validation, and evaluation sets with well-controlled\nexperimental protocols. We present the baseline results for both SDD and SASV\ntasks. All data, protocols, and baselines are publicly available at\nhttps://jungjee.github.io/spoofceleb.\n", "date": "2024-09-18", "categories": []}
{"id": "2504.10821", "title": "Progressive Rock Music Classification", "abstract": "  This study investigates the classification of progressive rock music, a genre\ncharacterized by complex compositions and diverse instrumentation, distinct\nfrom other musical styles. Addressing this Music Information Retrieval (MIR)\ntask, we extracted comprehensive audio features, including spectrograms,\nMel-Frequency Cepstral Coefficients (MFCCs), chromagrams, and beat positions\nfrom song snippets using the Librosa library. A winner-take-all voting strategy\nwas employed to aggregate snippet-level predictions into final song\nclassifications. We conducted a comparative analysis of various machine\nlearning techniques. Ensemble methods, encompassing Bagging (Random Forest,\nExtraTrees, Bagging Classifier) and Boosting (XGBoost, Gradient Boosting), were\nexplored, utilizing Principal Component Analysis (PCA) for dimensionality\nreduction to manage computational constraints with high-dimensional feature\nsets. Additionally, deep learning approaches were investigated, including the\ndevelopment of custom 1D Convolutional Neural Network (1D CNN) architectures\n(named \"Zuck\" and \"Satya\") featuring specific layer configurations,\nnormalization, and activation functions. Furthermore, we fine-tuned a\nstate-of-the-art Audio Spectrogram Transformer (AST) model, leveraging its\nattention-based mechanisms for audio classification. Performance evaluation on\nvalidation and test sets revealed varying effectiveness across models, with\nensemble methods like Extra Trees achieving test accuracies up to 76.38%. This\nresearch provides insights into the application and relative performance of\ndiverse machine learning paradigms for the nuanced task of progressive rock\ngenre classification.\n", "date": "2025-04-15", "categories": []}
{"id": "2504.11246", "title": "Respiratory Inhaler Sound Event Classification Using Self-Supervised\n  Learning", "abstract": "  Asthma is a chronic respiratory condition that affects millions of people\nworldwide. While this condition can be managed by administering controller\nmedications through handheld inhalers, clinical studies have shown low\nadherence to the correct inhaler usage technique. Consequently, many patients\nmay not receive the full benefit of their medication. Automated classification\nof inhaler sounds has recently been studied to assess medication adherence.\nHowever, the existing classification models were typically trained using data\nfrom specific inhaler types, and their ability to generalize to sounds from\ndifferent inhalers remains unexplored. In this study, we adapted the wav2vec\n2.0 self-supervised learning model for inhaler sound classification by\npre-training and fine-tuning this model on inhaler sounds. The proposed model\nshows a balanced accuracy of 98% on a dataset collected using a dry powder\ninhaler and smartwatch device. The results also demonstrate that re-finetuning\nthis model on minimal data from a target inhaler is a promising approach to\nadapting a generic inhaler sound classification model to a different inhaler\ndevice and audio capture hardware. This is the first study in the field to\ndemonstrate the potential of smartwatches as assistive technologies for the\npersonalized monitoring of inhaler adherence using machine learning models.\n", "date": "2025-04-15", "categories": []}
{"id": "2410.10537", "title": "Reproducible Machine Learning-based Voice Pathology Detection:\n  Introducing the Pitch Difference Feature", "abstract": "  Purpose: We introduce a novel methodology for voice pathology detection using\nthe publicly available Saarbr\\\"ucken Voice Database (SVD) and a robust feature\nset combining commonly used acoustic handcrafted features with two novel ones:\npitch difference (relative variation in fundamental frequency) and NaN feature\n(failed fundamental frequency estimation).\n  Methods: We evaluate six machine learning (ML) algorithms -- support vector\nmachine, k-nearest neighbors, naive Bayes, decision tree, random forest, and\nAdaBoost -- using grid search for feasible hyperparameters and 20480 different\nfeature subsets. Top 1000 classification models -- feature subset combinations\nfor each ML algorithm are validated with repeated stratified cross-validation.\nTo address class imbalance, we apply K-Means SMOTE to augment the training\ndata.\n  Results: Our approach achieves 85.61%, 84.69% and 85.22% unweighted average\nrecall (UAR) for females, males and combined results respectively. We\nintentionally omit accuracy as it is a highly biased metric for imbalanced\ndata.\n  Conclusion: Our study demonstrates that by following the proposed methodology\nand feature engineering, there is a potential in detection of various voice\npathologies using ML models applied to the simplest vocal task, a sustained\nutterance of the vowel /a:/. To enable easier use of our methodology and to\nsupport our claims, we provide a publicly available GitHub repository with DOI\n10.5281/zenodo.13771573. Finally, we provide a REFORMS checklist to enhance\nreadability, reproducibility and justification of our approach\n", "date": "2024-10-14", "categories": []}
{"id": "2503.11562", "title": "Designing Neural Synthesizers for Low-Latency Interaction", "abstract": "  Neural Audio Synthesis (NAS) models offer interactive musical control over\nhigh-quality, expressive audio generators. While these models can operate in\nreal-time, they often suffer from high latency, making them unsuitable for\nintimate musical interaction. The impact of architectural choices in deep\nlearning models on audio latency remains largely unexplored in the NAS\nliterature. In this work, we investigate the sources of latency and jitter\ntypically found in interactive NAS models. We then apply this analysis to the\ntask of timbre transfer using RAVE, a convolutional variational autoencoder for\naudio waveforms introduced by Caillon et al. in 2021. Finally, we present an\niterative design approach for optimizing latency. This culminates with a model\nwe call BRAVE (Bravely Realtime Audio Variational autoEncoder), which is\nlow-latency and exhibits better pitch and loudness replication while showing\ntimbre modification capabilities similar to RAVE. We implement it in a\nspecialized inference framework for low-latency, real-time inference and\npresent a proof-of-concept audio plugin compatible with audio signals from\nmusical instruments. We expect the challenges and guidelines described in this\ndocument to support NAS researchers in designing models for low-latency\ninference from the ground up, enriching the landscape of possibilities for\nmusicians.\n", "date": "2025-03-14", "categories": []}
{"id": "2502.03979", "title": "Towards Unified Music Emotion Recognition across Dimensional and\n  Categorical Models", "abstract": "  One of the most significant challenges in Music Emotion Recognition (MER)\ncomes from the fact that emotion labels can be heterogeneous across datasets\nwith regard to the emotion representation, including categorical (e.g., happy,\nsad) versus dimensional labels (e.g., valence-arousal). In this paper, we\npresent a unified multitask learning framework that combines these two types of\nlabels and is thus able to be trained on multiple datasets. This framework uses\nan effective input representation that combines musical features (i.e., key and\nchords) and MERT embeddings. Moreover, knowledge distillation is employed to\ntransfer the knowledge of teacher models trained on individual datasets to a\nstudent model, enhancing its ability to generalize across multiple tasks. To\nvalidate our proposed framework, we conducted extensive experiments on a\nvariety of datasets, including MTG-Jamendo, DEAM, PMEmo, and EmoMusic.\nAccording to our experimental results, the inclusion of musical features,\nmultitask learning, and knowledge distillation significantly enhances\nperformance. In particular, our model outperforms the state-of-the-art models,\nincluding the best-performing model from the MediaEval 2021 competition on the\nMTG-Jamendo dataset. Our work makes a significant contribution to MER by\nallowing the combination of categorical and dimensional emotion labels in one\nunified framework, thus enabling training across datasets.\n", "date": "2025-02-06", "categories": []}
{"id": "2412.11943", "title": "autrainer: A Modular and Extensible Deep Learning Toolkit for Computer\n  Audition Tasks", "abstract": "  This work introduces the key operating principles for autrainer, our new deep\nlearning training framework for computer audition tasks. autrainer is a\nPyTorch-based toolkit that allows for rapid, reproducible, and easily\nextensible training on a variety of different computer audition tasks.\nConcretely, autrainer offers low-code training and supports a wide range of\nneural networks as well as preprocessing routines. In this work, we present an\noverview of its inner workings and key capabilities.\n", "date": "2024-12-16", "categories": []}
{"id": "2412.21037", "title": "TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow\n  Matching and Clap-Ranked Preference Optimization", "abstract": "  We introduce TangoFlux, an efficient Text-to-Audio (TTA) generative model\nwith 515M parameters, capable of generating up to 30 seconds of 44.1kHz audio\nin just 3.7 seconds on a single A40 GPU. A key challenge in aligning TTA models\nlies in the difficulty of creating preference pairs, as TTA lacks structured\nmechanisms like verifiable rewards or gold-standard answers available for Large\nLanguage Models (LLMs). To address this, we propose CLAP-Ranked Preference\nOptimization (CRPO), a novel framework that iteratively generates and optimizes\npreference data to enhance TTA alignment. We demonstrate that the audio\npreference dataset generated using CRPO outperforms existing alternatives. With\nthis framework, TangoFlux achieves state-of-the-art performance across both\nobjective and subjective benchmarks. We open source all code and models to\nsupport further research in TTA generation.\n", "date": "2024-12-30", "categories": []}
{"id": "2504.07229", "title": "Visual-Aware Speech Recognition for Noisy Scenarios", "abstract": "  Humans have the ability to utilize visual cues, such as lip movements and\nvisual scenes, to enhance auditory perception, particularly in noisy\nenvironments. However, current Automatic Speech Recognition (ASR) or\nAudio-Visual Speech Recognition (AVSR) models often struggle in noisy\nscenarios. To solve this task, we propose a model that improves transcription\nby correlating noise sources to visual cues. Unlike works that rely on lip\nmotion and require the speaker's visibility, we exploit broader visual\ninformation from the environment. This allows our model to naturally filter\nspeech from noise and improve transcription, much like humans do in noisy\nscenarios. Our method re-purposes pretrained speech and visual encoders,\nlinking them with multi-headed attention. This approach enables the\ntranscription of speech and the prediction of noise labels in video inputs. We\nintroduce a scalable pipeline to develop audio-visual datasets, where visual\ncues correlate to noise in the audio. We show significant improvements over\nexisting audio-only models in noisy scenarios. Results also highlight that\nvisual cues play a vital role in improved transcription accuracy.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.07345", "title": "Quantum-Inspired Genetic Algorithm for Robust Source Separation in Smart\n  City Acoustics", "abstract": "  The cacophony of urban sounds presents a significant challenge for smart city\napplications that rely on accurate acoustic scene analysis. Effectively\nanalyzing these complex soundscapes, often characterized by overlapping sound\nsources, diverse acoustic events, and unpredictable noise levels, requires\nprecise source separation. This task becomes more complicated when only limited\ntraining data is available. This paper introduces a novel Quantum-Inspired\nGenetic Algorithm (p-QIGA) for source separation, drawing inspiration from\nquantum information theory to enhance acoustic scene analysis in smart cities.\nBy leveraging quantum superposition for efficient solution space exploration\nand entanglement to handle correlated sources, p-QIGA achieves robust\nseparation even with limited data. These quantum-inspired concepts are\nintegrated into a genetic algorithm framework to optimize source separation\nparameters. The effectiveness of our approach is demonstrated on two datasets:\nthe TAU Urban Acoustic Scenes 2020 Mobile dataset, representing typical urban\nsoundscapes, and the Silent Cities dataset, capturing quieter urban\nenvironments during the COVID-19 pandemic. Experimental results show that the\np-QIGA achieves accuracy comparable to state-of-the-art methods while\nexhibiting superior resilience to noise and limited training data, achieving up\nto 8.2 dB signal-to-distortion ratio (SDR) in noisy environments and\noutperforming baseline methods by up to 2 dB with only 10% of the training\ndata. This research highlights the potential of p-QIGA to advance acoustic\nsignal processing in smart cities, particularly for noise pollution monitoring\nand acoustic surveillance.\n", "date": "2025-04-10", "categories": []}
{"id": "2504.06710", "title": "Clustering and novel class recognition: evaluating bioacoustic deep\n  learning feature extractors", "abstract": "  In computational bioacoustics, deep learning models are composed of feature\nextractors and classifiers. The feature extractors generate vector\nrepresentations of the input sound segments, called embeddings, which can be\ninput to a classifier. While benchmarking of classification scores provides\ninsights into specific performance statistics, it is limited to species that\nare included in the models' training data. Furthermore, it makes it impossible\nto compare models trained on very different taxonomic groups. This paper aims\nto address this gap by analyzing the embeddings generated by the feature\nextractors of 15 bioacoustic models spanning a wide range of setups (model\narchitectures, training data, training paradigms). We evaluated and compared\ndifferent ways in which models structure embedding spaces through clustering\nand kNN classification, which allows us to focus our comparison on feature\nextractors independent of their classifiers. We believe that this approach lets\nus evaluate the adaptability and generalization potential of models going\nbeyond the classes they were trained on.\n", "date": "2025-04-09", "categories": []}
{"id": "2412.15322", "title": "MMAudio: Taming Multimodal Joint Training for High-Quality\n  Video-to-Audio Synthesis", "abstract": "  We propose to synthesize high-quality and synchronized audio, given video and\noptional text conditions, using a novel multimodal joint training framework\nMMAudio. In contrast to single-modality training conditioned on (limited) video\ndata only, MMAudio is jointly trained with larger-scale, readily available\ntext-audio data to learn to generate semantically aligned high-quality audio\nsamples. Additionally, we improve audio-visual synchrony with a conditional\nsynchronization module that aligns video conditions with audio latents at the\nframe level. Trained with a flow matching objective, MMAudio achieves new\nvideo-to-audio state-of-the-art among public models in terms of audio quality,\nsemantic alignment, and audio-visual synchronization, while having a low\ninference time (1.23s to generate an 8s clip) and just 157M parameters. MMAudio\nalso achieves surprisingly competitive performance in text-to-audio generation,\nshowing that joint training does not hinder single-modality performance. Code\nand demo are available at: https://hkchengrex.github.io/MMAudio\n", "date": "2024-12-19", "categories": []}
{"id": "2504.05576", "title": "SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic\n  Binding", "abstract": "  We introduce SoundVista, a method to generate the ambient sound of an\narbitrary scene at novel viewpoints. Given a pre-acquired recording of the\nscene from sparsely distributed microphones, SoundVista can synthesize the\nsound of that scene from an unseen target viewpoint. The method learns the\nunderlying acoustic transfer function that relates the signals acquired at the\ndistributed microphones to the signal at the target viewpoint, using a limited\nnumber of known recordings. Unlike existing works, our method does not require\nconstraints or prior knowledge of sound source details. Moreover, our method\nefficiently adapts to diverse room layouts, reference microphone configurations\nand unseen environments. To enable this, we introduce a visual-acoustic binding\nmodule that learns visual embeddings linked with local acoustic properties from\npanoramic RGB and depth data. We first leverage these embeddings to optimize\nthe placement of reference microphones in any given scene. During synthesis, we\nleverage multiple embeddings extracted from reference locations to get adaptive\nweights for their contribution, conditioned on target viewpoint. We benchmark\nthe task on both publicly available data and real-world settings. We\ndemonstrate significant improvements over existing methods.\n", "date": "2025-04-08", "categories": []}
{"id": "2409.17635", "title": "FlowMAC: Conditional Flow Matching for Audio Coding at Low Bit Rates", "abstract": "  This paper introduces FlowMAC, a novel neural audio codec for high-quality\ngeneral audio compression at low bit rates based on conditional flow matching\n(CFM). FlowMAC jointly learns a mel spectrogram encoder, quantizer and decoder.\nAt inference time the decoder integrates a continuous normalizing flow via an\nODE solver to generate a high-quality mel spectrogram. This is the first time\nthat a CFM-based approach is applied to general audio coding, enabling a\nscalable, simple and memory efficient training. Our subjective evaluations show\nthat FlowMAC at 3 kbps achieves similar quality as state-of-the-art GAN-based\nand DDPM-based neural audio codecs at double the bit rate. Moreover, FlowMAC\noffers a tunable inference pipeline, which permits to trade off complexity and\nquality. This enables real-time coding on CPU, while maintaining high\nperceptual quality.\n", "date": "2024-09-26", "categories": []}
{"id": "2411.05679", "title": "Tell What You Hear From What You See -- Video to Audio Generation\n  Through Text", "abstract": "  The content of visual and audio scenes is multi-faceted such that a video can\nbe paired with various audio and vice-versa. Thereby, in video-to-audio\ngeneration task, it is imperative to introduce steering approaches for\ncontrolling the generated audio. While Video-to-Audio generation is a\nwell-established generative task, existing methods lack such controllability.\nIn this work, we propose VATT, a multi-modal generative framework that takes a\nvideo and an optional text prompt as input, and generates audio and optional\ntextual description of the audio. Such a framework has two advantages: i)\nVideo-to-Audio generation process can be refined and controlled via text which\ncomplements the context of visual information, and ii) The model can suggest\nwhat audio to generate for the video by generating audio captions. VATT\nconsists of two key modules: VATT Converter, a LLM that is fine-tuned for\ninstructions and includes a projection layer that maps video features to the\nLLM vector space; and VATT Audio, a transformer that generates audio tokens\nfrom visual frames and from optional text prompt using iterative parallel\ndecoding. The audio tokens are converted to a waveform by pretrained neural\ncodec. Experiments show that when VATT is compared to existing video-to-audio\ngeneration methods in objective metrics, it achieves competitive performance\nwhen the audio caption is not provided. When the audio caption is provided as a\nprompt, VATT achieves even more refined performance (lowest KLD score of 1.41).\nFurthermore, subjective studies show that VATT Audio has been chosen as\npreferred generated audio than audio generated by existing methods. VATT\nenables controllable video-to-audio generation through text as well as\nsuggesting text prompts for videos through audio captions, unlocking novel\napplications such as text-guided video-to-audio generation and video-to-audio\ncaptioning.\n", "date": "2024-11-08", "categories": []}
{"id": "2504.04428", "title": "Formula-Supervised Sound Event Detection: Pre-Training Without Real Data", "abstract": "  In this paper, we propose a novel formula-driven supervised learning (FDSL)\nframework for pre-training an environmental sound analysis model by leveraging\nacoustic signals parametrically synthesized through formula-driven methods.\nSpecifically, we outline detailed procedures and evaluate their effectiveness\nfor sound event detection (SED). The SED task, which involves estimating the\ntypes and timings of sound events, is particularly challenged by the difficulty\nof acquiring a sufficient quantity of accurately labeled training data.\nMoreover, it is well known that manually annotated labels often contain noises\nand are significantly influenced by the subjective judgment of annotators. To\naddress these challenges, we propose a novel pre-training method that utilizes\na synthetic dataset, Formula-SED, where acoustic data are generated solely\nbased on mathematical formulas. The proposed method enables large-scale\npre-training by using the synthesis parameters applied at each time step as\nground truth labels, thereby eliminating label noise and bias. We demonstrate\nthat large-scale pre-training with Formula-SED significantly enhances model\naccuracy and accelerates training, as evidenced by our results in the DESED\ndataset used for DCASE2023 Challenge Task 4. The project page is at\nhttps://yutoshibata07.github.io/Formula-SED/\n", "date": "2025-04-06", "categories": []}
{"id": "2504.03329", "title": "Mind the Prompt: Prompting Strategies in Audio Generations for Improving\n  Sound Classification", "abstract": "  This paper investigates the design of effective prompt strategies for\ngenerating realistic datasets using Text-To-Audio (TTA) models. We also analyze\ndifferent techniques for efficiently combining these datasets to enhance their\nutility in sound classification tasks. By evaluating two sound classification\ndatasets with two TTA models, we apply a range of prompt strategies. Our\nfindings reveal that task-specific prompt strategies significantly outperform\nbasic prompt approaches in data generation. Furthermore, merging datasets\ngenerated using different TTA models proves to enhance classification\nperformance more effectively than merely increasing the training dataset size.\nOverall, our results underscore the advantages of these methods as effective\ndata augmentation techniques using synthetic data.\n", "date": "2025-04-04", "categories": []}
{"id": "2501.00398", "title": "TSPE: Task-Specific Prompt Ensemble for Improved Zero-Shot Audio\n  Classification", "abstract": "  Audio-language models (ALMs) excel in zero-shot audio classification, a task\nwhere models classify previously unseen audio clips at test time by leveraging\ndescriptive natural language prompts. We introduce TSPE (Task-Specific Prompt\nEnsemble), a simple, training-free hard prompting method that boosts ALEs'\nzero-shot performance by customizing prompts for diverse audio classification\ntasks. Rather than using generic template-based prompts like \"Sound of a car\"\nwe generate context-rich prompts, such as \"Sound of a car coming from a\ntunnel\". Specifically, we leverage label information to identify suitable sound\nattributes, such as \"loud\" and \"feeble\", and appropriate sound sources, such as\n\"tunnel\" and \"street\" and incorporate this information into the prompts used by\nAudio-Language Models (ALMs) for audio classification. Further, to enhance\naudio-text alignment, we perform prompt ensemble across TSPE-generated\ntask-specific prompts. When evaluated on 12 diverse audio classification\ndatasets, TSPE improves performance across ALMs by showing an absolute\nimprovement of 1.23-16.36% over vanilla zero-shot evaluation.\n", "date": "2024-12-31", "categories": []}
{"id": "2504.02386", "title": "VoiceCraft-Dub: Automated Video Dubbing with Neural Codec Language\n  Models", "abstract": "  We present VoiceCraft-Dub, a novel approach for automated video dubbing that\nsynthesizes high-quality speech from text and facial cues. This task has broad\napplications in filmmaking, multimedia creation, and assisting voice-impaired\nindividuals. Building on the success of Neural Codec Language Models (NCLMs)\nfor speech synthesis, our method extends their capabilities by incorporating\nvideo features, ensuring that synthesized speech is time-synchronized and\nexpressively aligned with facial movements while preserving natural prosody. To\ninject visual cues, we design adapters to align facial features with the NCLM\ntoken space and introduce audio-visual fusion layers to merge audio-visual\ninformation within the NCLM framework. Additionally, we curate CelebV-Dub, a\nnew dataset of expressive, real-world videos specifically designed for\nautomated video dubbing. Extensive experiments show that our model achieves\nhigh-quality, intelligible, and natural speech synthesis with accurate lip\nsynchronization, outperforming existing methods in human perception and\nperforming favorably in objective evaluations. We also adapt VoiceCraft-Dub for\nthe video-to-speech task, demonstrating its versatility for various\napplications.\n", "date": "2025-04-03", "categories": []}
{"id": "2410.12082", "title": "Learning to rumble: Automated elephant call classification, detection\n  and endpointing using deep architectures", "abstract": "  We consider the problem of detecting, isolating and classifying elephant\ncalls in continuously recorded audio. Such automatic call characterisation can\nassist conservation efforts and inform environmental management strategies. In\ncontrast to previous work in which call detection was performed at a segment\nlevel, we perform call detection at a frame level which implicitly also allows\ncall endpointing, the isolation of a call in a longer recording. For\nexperimentation, we employ two annotated datasets, one containing Asian and the\nother African elephant vocalisations. We evaluate several shallow and deep\nclassifier models, and show that the current best performance can be improved\nby using an audio spectrogram transformer (AST), a neural architecture which\nhas not been used for this purpose before, and which we have configured in a\nnovel sequence-to-sequence manner. We also show that using transfer learning by\npre-training leads to further improvements both in terms of computational\ncomplexity and performance. Finally, we consider sub-call classification using\nan accepted taxonomy of call types, a task which has not previously been\nconsidered. We show that also in this case the transformer architectures\nprovide the best performance. Our best classifiers achieve an average precision\n(AP) of 0.962 for framewise binary call classification, and an area under the\nreceiver operating characteristic (AUC) of 0.957 and 0.979 for call\nclassification with 5 classes and sub-call classification with 7 classes\nrespectively. All of these represent either new benchmarks (sub-call\nclassifications) or improvements on previously best systems. We conclude that a\nfully-automated elephant call detection and subcall classification system is\nwithin reach. Such a system would provide valuable information on the behaviour\nand state of elephant herds for the purposes of conservation and management.\n", "date": "2024-10-15", "categories": []}
{"id": "2411.07751", "title": "SAV-SE: Scene-aware Audio-Visual Speech Enhancement with Selective State\n  Space Model", "abstract": "  Speech enhancement plays an essential role in various applications, and the\nintegration of visual information has been demonstrated to bring substantial\nadvantages. However, the majority of current research concentrates on the\nexamination of facial and lip movements, which can be compromised or entirely\ninaccessible in scenarios where occlusions occur or when the camera view is\ndistant. Whereas contextual visual cues from the surrounding environment have\nbeen overlooked: for example, when we see a dog bark, our brain has the innate\nability to discern and filter out the barking noise. To this end, in this\npaper, we introduce a novel task, i.e. SAV-SE. To our best knowledge, this is\nthe first proposal to use rich contextual information from synchronized video\nas auxiliary cues to indicate the type of noise, which eventually improves the\nspeech enhancement performance. Specifically, we propose the VC-S$^2$E method,\nwhich incorporates the Conformer and Mamba modules for their complementary\nstrengths. Extensive experiments are conducted on public MUSIC, AVSpeech and\nAudioSet datasets, where the results demonstrate the superiority of VC-S$^2$E\nover other competitive methods. We will make the source code publicly\navailable. Project demo page: https://AVSEPage.github.io/\n", "date": "2024-11-12", "categories": []}
{"id": "2409.16644", "title": "Enabling Auditory Large Language Models for Automatic Speech Quality\n  Evaluation", "abstract": "  Speech quality assessment typically requires evaluating audio from multiple\naspects, such as mean opinion score (MOS) and speaker similarity (SIM) \\etc.,\nwhich can be challenging to cover using one small model designed for a single\ntask. In this paper, we propose leveraging recently introduced auditory large\nlanguage models (LLMs) for automatic speech quality assessment. By employing\ntask-specific prompts, auditory LLMs are finetuned to predict MOS, SIM and A/B\ntesting results, which are commonly used for evaluating text-to-speech systems.\nAdditionally, the finetuned auditory LLM is able to generate natural language\ndescriptions assessing aspects like noisiness, distortion, discontinuity, and\noverall quality, providing more interpretable outputs. Extensive experiments\nhave been performed on the NISQA, BVCC, SOMOS and VoxSim speech quality\ndatasets, using open-source auditory LLMs such as SALMONN, Qwen-Audio, and\nQwen2-Audio. For the natural language descriptions task, a commercial model\nGoogle Gemini 1.5 Pro is also evaluated. The results demonstrate that auditory\nLLMs achieve competitive performance compared to state-of-the-art task-specific\nsmall models in predicting MOS and SIM, while also delivering promising results\nin A/B testing and natural language descriptions. Our data processing scripts\nand finetuned model checkpoints can be found at\nhttps://github.com/bytedance/SALMONN.\n", "date": "2024-09-25", "categories": []}
{"id": "2504.00750", "title": "$C^2$AV-TSE: Context and Confidence-aware Audio Visual Target Speaker\n  Extraction", "abstract": "  Audio-Visual Target Speaker Extraction (AV-TSE) aims to mimic the human\nability to enhance auditory perception using visual cues. Although numerous\nmodels have been proposed recently, most of them estimate target signals by\nprimarily relying on local dependencies within acoustic features,\nunderutilizing the human-like capacity to infer unclear parts of speech through\ncontextual information. This limitation results in not only suboptimal\nperformance but also inconsistent extraction quality across the utterance, with\nsome segments exhibiting poor quality or inadequate suppression of interfering\nspeakers. To close this gap, we propose a model-agnostic strategy called the\nMask-And-Recover (MAR). It integrates both inter- and intra-modality contextual\ncorrelations to enable global inference within extraction modules.\nAdditionally, to better target challenging parts within each sample, we\nintroduce a Fine-grained Confidence Score (FCS) model to assess extraction\nquality and guide extraction modules to emphasize improvement on low-quality\nsegments. To validate the effectiveness of our proposed model-agnostic training\nparadigm, six popular AV-TSE backbones were adopted for evaluation on the\nVoxCeleb2 dataset, demonstrating consistent performance improvements across\nvarious metrics.\n", "date": "2025-04-01", "categories": []}
{"id": "2407.02264", "title": "SOAF: Scene Occlusion-aware Neural Acoustic Field", "abstract": "  This paper tackles the problem of novel view audio-visual synthesis along an\narbitrary trajectory in an indoor scene, given the audio-video recordings from\nother known trajectories of the scene. Existing methods often overlook the\neffect of room geometry, particularly wall occlusions on sound propagation,\nmaking them less accurate in multi-room environments. In this work, we propose\na new approach called Scene Occlusion-aware Acoustic Field (SOAF) for accurate\nsound generation. Our approach derives a global prior for the sound field using\ndistance-aware parametric sound-propagation modeling and then transforms it\nbased on the scene structure learned from the input video. We extract features\nfrom the local acoustic field centered at the receiver using a Fibonacci Sphere\nto generate binaural audio for novel views with a direction-aware attention\nmechanism. Extensive experiments on the real dataset RWAVS and the synthetic\ndataset SoundSpaces demonstrate that our method outperforms previous\nstate-of-the-art techniques in audio generation.\n", "date": "2024-07-02", "categories": []}
{"id": "2503.21826", "title": "Hierarchical Label Propagation: A Model-Size-Dependent Performance\n  Booster for AudioSet Tagging", "abstract": "  AudioSet is one of the most used and largest datasets in audio tagging,\ncontaining about 2 million audio samples that are manually labeled with 527\nevent categories organized into an ontology. However, the annotations contain\ninconsistencies, particularly where categories that should be labeled as\npositive according to the ontology are frequently mislabeled as negative. To\naddress this issue, we apply Hierarchical Label Propagation (HLP), which\npropagates labels up the ontology hierarchy, resulting in a mean increase in\npositive labels per audio clip from 1.98 to 2.39 and affecting 109 out of the\n527 classes. Our results demonstrate that HLP provides performance benefits\nacross various model architectures, including convolutional neural networks\n(PANN's CNN6 and ConvNeXT) and transformers (PaSST), with smaller models\nshowing more improvements. Finally, on FSD50K, another widely used dataset,\nmodels trained on AudioSet with HLP consistently outperformed those trained\nwithout HLP. Our source code will be made available on GitHub.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.22200", "title": "Enhance Generation Quality of Flow Matching V2A Model via Multi-Step\n  CoT-Like Guidance and Combined Preference Optimization", "abstract": "  Creating high-quality sound effects from videos and text prompts requires\nprecise alignment between visual and audio domains, both semantically and\ntemporally, along with step-by-step guidance for professional audio generation.\nHowever, current state-of-the-art video-guided audio generation models often\nfall short of producing high-quality audio for both general and specialized use\ncases. To address this challenge, we introduce a multi-stage, multi-modal,\nend-to-end generative framework with Chain-of-Thought-like (CoT-like) guidance\nlearning, termed Chain-of-Perform (CoP). First, we employ a transformer-based\nnetwork architecture designed to achieve CoP guidance, enabling the generation\nof both general and professional audio. Second, we implement a multi-stage\ntraining framework that follows step-by-step guidance to ensure the generation\nof high-quality sound effects. Third, we develop a CoP multi-modal dataset,\nguided by video, to support step-by-step sound effects generation. Evaluation\nresults highlight the advantages of the proposed multi-stage CoP generative\nframework compared to the state-of-the-art models on a variety of datasets,\nwith FAD 0.79 to 0.74 (+6.33%), CLIP 16.12 to 17.70 (+9.80%) on VGGSound,\nSI-SDR 1.98dB to 3.35dB (+69.19%), MOS 2.94 to 3.49(+18.71%) on PianoYT-2h, and\nSI-SDR 2.22dB to 3.21dB (+44.59%), MOS 3.07 to 3.42 (+11.40%) on Piano-10h.\n", "date": "2025-03-28", "categories": []}
{"id": "2503.22208", "title": "DeepSound-V1: Start to Think Step-by-Step in the Audio Generation from\n  Videos", "abstract": "  Currently, high-quality, synchronized audio is synthesized from video and\noptional text inputs using various multi-modal joint learning frameworks.\nHowever, the precise alignment between the visual and generated audio domains\nremains far from satisfactory. One key factor is the lack of sufficient\ntemporal and semantic alignment annotations in open-source video-audio and\ntext-audio benchmarks. Therefore, we propose a framework for audio generation\nfrom videos, leveraging the internal chain-of-thought (CoT) of a multi-modal\nlarge language model (MLLM) to enable step-by-step reasoning without requiring\nadditional annotations. Additionally, a corresponding multi-modal reasoning\ndataset is constructed to facilitate the learning of initial reasoning in audio\ngeneration. In the experiments, we demonstrate the effectiveness of the\nproposed framework in reducing misalignment (voice-over) in generated audio and\nachieving competitive performance compared to various state-of-the-art models.\nThe evaluation results show that the proposed method outperforms\nstate-of-the-art approaches across multiple metrics. Specifically, the F DP\naSST indicator is reduced by up to 10.07%, the F DP AN N s indicator by up to\n11.62%, and the F DV GG indicator by up to 38.61%. Furthermore, the IS\nindicator improves by up to 4.95%, the IB-score indicator increases by up to\n6.39%, and the DeSync indicator is reduced by up to 0.89%.\n", "date": "2025-03-28", "categories": []}
{"id": "2503.22265", "title": "DeepAudio-V1:Towards Multi-Modal Multi-Stage End-to-End Video to Speech\n  and Audio Generation", "abstract": "  Currently, high-quality, synchronized audio is synthesized using various\nmulti-modal joint learning frameworks, leveraging video and optional text\ninputs. In the video-to-audio benchmarks, video-to-audio quality, semantic\nalignment, and audio-visual synchronization are effectively achieved. However,\nin real-world scenarios, speech and audio often coexist in videos\nsimultaneously, and the end-to-end generation of synchronous speech and audio\ngiven video and text conditions are not well studied. Therefore, we propose an\nend-to-end multi-modal generation framework that simultaneously produces speech\nand audio based on video and text conditions. Furthermore, the advantages of\nvideo-to-audio (V2A) models for generating speech from videos remain unclear.\nThe proposed framework, DeepAudio, consists of a video-to-audio (V2A) module, a\ntext-to-speech (TTS) module, and a dynamic mixture of modality fusion (MoF)\nmodule. In the evaluation, the proposed end-to-end framework achieves\nstate-of-the-art performance on the video-audio benchmark, video-speech\nbenchmark, and text-speech benchmark. In detail, our framework achieves\ncomparable results in the comparison with state-of-the-art models for the\nvideo-audio and text-speech benchmarks, and surpassing state-of-the-art models\nin the video-speech benchmark, with WER 16.57% to 3.15% (+80.99%), SPK-SIM\n78.30% to 89.38% (+14.15%), EMO-SIM 66.24% to 75.56% (+14.07%), MCD 8.59 to\n7.98 (+7.10%), MCD SL 11.05 to 9.40 (+14.93%) across a variety of dubbing\nsettings.\n", "date": "2025-03-28", "categories": []}
{"id": "2503.20212", "title": "Dolphin: A Large-Scale Automatic Speech Recognition Model for Eastern\n  Languages", "abstract": "  This report introduces Dolphin, a large-scale multilingual automatic speech\nrecognition (ASR) model that extends the Whisper architecture to support a\nwider range of languages. Our approach integrates in-house proprietary and\nopen-source datasets to refine and optimize Dolphin's performance. The model is\nspecifically designed to achieve notable recognition accuracy for 40 Eastern\nlanguages across East Asia, South Asia, Southeast Asia, and the Middle East,\nwhile also supporting 22 Chinese dialects. Experimental evaluations show that\nDolphin significantly outperforms current state-of-the-art open-source models\nacross various languages. To promote reproducibility and community-driven\ninnovation, we are making our trained models and inference source code publicly\navailable.\n", "date": "2025-03-26", "categories": []}
{"id": "2408.07472", "title": "Unsupervised Blind Joint Dereverberation and Room Acoustics Estimation\n  with Diffusion Models", "abstract": "  This paper presents an unsupervised method for single-channel blind\ndereverberation and room impulse response (RIR) estimation, called BUDDy. The\nalgorithm is rooted in Bayesian posterior sampling: it combines a likelihood\nmodel enforcing fidelity to the reverberant measurement, and an anechoic speech\nprior implemented by an unconditional diffusion model. We design a parametric\nfilter representing the RIR, with exponential decay for each frequency subband.\nRoom acoustics estimation and speech dereverberation are jointly carried out,\nas the filter parameters are iteratively estimated and the speech utterance\nrefined along the reverse diffusion trajectory. In a blind scenario where the\nRIR is unknown, BUDDy successfully performs speech dereverberation in various\nacoustic scenarios, significantly outperforming other blind unsupervised\nbaselines. Unlike supervised methods, which often struggle to generalize, BUDDy\nseamlessly adapts to different acoustic conditions. This paper extends our\nprevious work by offering new experimental results and insights into the\nalgorithm's versatility. We demonstrate the robustness of our proposed method\nto new acoustic and speaker conditions, as well as its adaptability to\nhigh-resolution singing voice dereverberation, using both instrumental metrics\nand subjective listening evaluation. We study BUDDy's performance for RIR\nestimation and observe it surpasses a state-of-the-art supervised DNN-based\nestimator on mismatched acoustic conditions. Finally, we investigate the\nsensitivity of informed dereverberation methods to RIR estimation errors,\nthereby motivating the joint acoustic estimation and dereverberation design.\nAudio examples and code can be found online.\n", "date": "2024-08-14", "categories": []}
{"id": "2503.19591", "title": "Boosting the Transferability of Audio Adversarial Examples with Acoustic\n  Representation Optimization", "abstract": "  With the widespread application of automatic speech recognition (ASR)\nsystems, their vulnerability to adversarial attacks has been extensively\nstudied. However, most existing adversarial examples are generated on specific\nindividual models, resulting in a lack of transferability. In real-world\nscenarios, attackers often cannot access detailed information about the target\nmodel, making query-based attacks unfeasible. To address this challenge, we\npropose a technique called Acoustic Representation Optimization that aligns\nadversarial perturbations with low-level acoustic characteristics derived from\nspeech representation models. Rather than relying on model-specific,\nhigher-layer abstractions, our approach leverages fundamental acoustic\nrepresentations that remain consistent across diverse ASR architectures. By\nenforcing an acoustic representation loss to guide perturbations toward these\nrobust, lower-level representations, we enhance the cross-model transferability\nof adversarial examples without degrading audio quality. Our method is\nplug-and-play and can be integrated with any existing attack methods. We\nevaluate our approach on three modern ASR models, and the experimental results\ndemonstrate that our method significantly improves the transferability of\nadversarial examples generated by previous methods while preserving the audio\nquality.\n", "date": "2025-03-25", "categories": []}
{"id": "2410.04324", "title": "Where are we in audio deepfake detection? A systematic analysis over\n  generative and detection models", "abstract": "  Recent advances in Text-to-Speech (TTS) and Voice-Conversion (VC) using\ngenerative Artificial Intelligence (AI) technology have made it possible to\ngenerate high-quality and realistic human-like audio. This poses growing\nchallenges in distinguishing AI-synthesized speech from the genuine human voice\nand could raise concerns about misuse for impersonation, fraud, spreading\nmisinformation, and scams. However, existing detection methods for\nAI-synthesized audio have not kept pace and often fail to generalize across\ndiverse datasets. In this paper, we introduce SONAR, a synthetic AI-Audio\nDetection Framework and Benchmark, aiming to provide a comprehensive evaluation\nfor distinguishing cutting-edge AI-synthesized auditory content. SONAR includes\na novel evaluation dataset sourced from 9 diverse audio synthesis platforms,\nincluding leading TTS providers and state-of-the-art TTS models. It is the\nfirst framework to uniformly benchmark AI-audio detection across both\ntraditional and foundation model-based detection systems. Through extensive\nexperiments, (1) we reveal the limitations of existing detection methods and\ndemonstrate that foundation models exhibit stronger generalization\ncapabilities, likely due to their model size and the scale and quality of\npretraining data. (2) Speech foundation models demonstrate robust cross-lingual\ngeneralization capabilities, maintaining strong performance across diverse\nlanguages despite being fine-tuned solely on English speech data. This finding\nalso suggests that the primary challenges in audio deepfake detection are more\nclosely tied to the realism and quality of synthetic audio rather than\nlanguage-specific characteristics. (3) We explore the effectiveness and\nefficiency of few-shot fine-tuning in improving generalization, highlighting\nits potential for tailored applications, such as personalized detection systems\nfor specific entities or individuals.\n", "date": "2024-10-06", "categories": []}
{"id": "2503.17577", "title": "Measuring the Robustness of Audio Deepfake Detectors", "abstract": "  Deepfakes have become a universal and rapidly intensifying concern of\ngenerative AI across various media types such as images, audio, and videos.\nAmong these, audio deepfakes have been of particular concern due to the ease of\nhigh-quality voice synthesis and distribution via platforms such as social\nmedia and robocalls. Consequently, detecting audio deepfakes plays a critical\nrole in combating the growing misuse of AI-synthesized speech. However,\nreal-world scenarios often introduce various audio corruptions, such as noise,\nmodification, and compression, that may significantly impact detection\nperformance. This work systematically evaluates the robustness of 10 audio\ndeepfake detection models against 16 common corruptions, categorized into noise\nperturbation, audio modification, and compression. Using both traditional deep\nlearning models and state-of-the-art foundation models, we make four unique\nobservations. First, our findings show that while most models demonstrate\nstrong robustness to noise, they are notably more vulnerable to modifications\nand compression, especially when neural codecs are applied. Second, speech\nfoundation models generally outperform traditional models across most\nscenarios, likely due to their self-supervised learning paradigm and\nlarge-scale pre-training. Third, our results show that increasing model size\nimproves robustness, albeit with diminishing returns. Fourth, we demonstrate\nhow targeted data augmentation during training can enhance model resilience to\nunseen perturbations. A case study on political speech deepfakes highlights the\neffectiveness of foundation models in achieving high accuracy under real-world\nconditions. These findings emphasize the importance of developing more robust\ndetection frameworks to ensure reliability in practical deployment settings.\n", "date": "2025-03-21", "categories": []}
{"id": "2503.18032", "title": "Anomaly Detection and Localization for Speech Deepfakes via Feature\n  Pyramid Matching", "abstract": "  The rise of AI-driven generative models has enabled the creation of highly\nrealistic speech deepfakes - synthetic audio signals that can imitate target\nspeakers' voices - raising critical security concerns. Existing methods for\ndetecting speech deepfakes primarily rely on supervised learning, which suffers\nfrom two critical limitations: limited generalization to unseen synthesis\ntechniques and a lack of explainability. In this paper, we address these issues\nby introducing a novel interpretable one-class detection framework, which\nreframes speech deepfake detection as an anomaly detection task. Our model is\ntrained exclusively on real speech to characterize its distribution, enabling\nthe classification of out-of-distribution samples as synthetically generated.\nAdditionally, our framework produces interpretable anomaly maps during\ninference, highlighting anomalous regions across both time and frequency\ndomains. This is done through a Student-Teacher Feature Pyramid Matching\nsystem, enhanced with Discrepancy Scaling to improve generalization\ncapabilities across unseen data distributions. Extensive evaluations\ndemonstrate the superior performance of our approach compared to the considered\nbaselines, validating the effectiveness of framing speech deepfake detection as\nan anomaly detection problem.\n", "date": "2025-03-23", "categories": []}
{"id": "2503.18138", "title": "Machine learning based animal emotion classification using audio signals", "abstract": "  This paper presents the machine learning approach to the automated\nclassification of a dog's emotional state based on the processing and\nrecognition of audio signals. It offers helpful information for improving\nhuman-machine interfaces and developing more precise tools for classifying\nemotions from acoustic data. The presented model demonstrates an overall\naccuracy value above 70% for audio signals recorded for one dog.\n", "date": "2025-03-23", "categories": []}
{"id": "2502.15285", "title": "Offload Rethinking by Cloud Assistance for Efficient Environmental Sound\n  Recognition on LPWANs", "abstract": "  Learning-based environmental sound recognition has emerged as a crucial\nmethod for ultra-low-power environmental monitoring in biological research and\ncity-scale sensing systems. These systems usually operate under limited\nresources and are often powered by harvested energy in remote areas. Recent\nefforts in on-device sound recognition suffer from low accuracy due to resource\nconstraints, whereas cloud offloading strategies are hindered by high\ncommunication costs. In this work, we introduce ORCA, a novel\nresource-efficient cloud-assisted environmental sound recognition system on\nbatteryless devices operating over the Low-Power Wide-Area Networks (LPWANs),\ntargeting wide-area audio sensing applications. We propose a cloud assistance\nstrategy that remedies the low accuracy of on-device inference while minimizing\nthe communication costs for cloud offloading. By leveraging a\nself-attention-based cloud sub-spectral feature selection method to facilitate\nefficient on-device inference, ORCA resolves three key challenges for\nresource-constrained cloud offloading over LPWANs: 1) high communication costs\nand low data rates, 2) dynamic wireless channel conditions, and 3) unreliable\noffloading. We implement ORCA on an energy-harvesting batteryless\nmicrocontroller and evaluate it in a real world urban sound testbed. Our\nresults show that ORCA outperforms state-of-the-art methods by up to $80\n\\times$ in energy savings and $220 \\times$ in latency reduction while\nmaintaining comparable accuracy.\n", "date": "2025-02-21", "categories": []}
{"id": "2503.16669", "title": "Aligning Text-to-Music Evaluation with Human Preferences", "abstract": "  Despite significant recent advances in generative acoustic text-to-music\n(TTM) modeling, robust evaluation of these models lags behind, relying in\nparticular on the popular Fr\\'echet Audio Distance (FAD). In this work, we\nrigorously study the design space of reference-based divergence metrics for\nevaluating TTM models through (1) designing four synthetic meta-evaluations to\nmeasure sensitivity to particular musical desiderata, and (2) collecting and\nevaluating on MusicPrefs, the first open-source dataset of human preferences\nfor TTM systems. We find that not only is the standard FAD setup inconsistent\non both synthetic and human preference data, but that nearly all existing\nmetrics fail to effectively capture desiderata, and are only weakly correlated\nwith human perception. We propose a new metric, the MAUVE Audio Divergence\n(MAD), computed on representations from a self-supervised audio embedding\nmodel. We find that this metric effectively captures diverse musical desiderata\n(average rank correlation 0.84 for MAD vs. 0.49 for FAD and also correlates\nmore strongly with MusicPrefs (0.62 vs. 0.14).\n", "date": "2025-03-20", "categories": []}
{"id": "2407.05782", "title": "Sequential Contrastive Audio-Visual Learning", "abstract": "  Contrastive learning has emerged as a powerful technique in audio-visual\nrepresentation learning, leveraging the natural co-occurrence of audio and\nvisual modalities in webscale video datasets. However, conventional contrastive\naudio-visual learning (CAV) methodologies often rely on aggregated\nrepresentations derived through temporal aggregation, neglecting the intrinsic\nsequential nature of the data. This oversight raises concerns regarding the\nability of standard approaches to capture and utilize fine-grained information\nwithin sequences. In response to this limitation, we propose sequential\ncontrastive audiovisual learning (SCAV), which contrasts examples based on\ntheir non-aggregated representation space using multidimensional sequential\ndistances. Audio-visual retrieval experiments with the VGGSound and Music\ndatasets demonstrate the effectiveness of SCAV, with up to 3.5x relative\nimprovements in recall against traditional aggregation-based contrastive\nlearning and other previously proposed methods, which utilize more parameters\nand data. We also show that models trained with SCAV exhibit a significant\ndegree of flexibility regarding the metric employed for retrieval, allowing us\nto use a hybrid retrieval approach that is both effective and efficient.\n", "date": "2024-07-08", "categories": []}
{"id": "2411.17698", "title": "Video-Guided Foley Sound Generation with Multimodal Controls", "abstract": "  Generating sound effects for videos often requires creating artistic sound\neffects that diverge significantly from real-life sources and flexible control\nin the sound design. To address this problem, we introduce MultiFoley, a model\ndesigned for video-guided sound generation that supports multimodal\nconditioning through text, audio, and video. Given a silent video and a text\nprompt, MultiFoley allows users to create clean sounds (e.g., skateboard wheels\nspinning without wind noise) or more whimsical sounds (e.g., making a lion's\nroar sound like a cat's meow). MultiFoley also allows users to choose reference\naudio from sound effects (SFX) libraries or partial videos for conditioning. A\nkey novelty of our model lies in its joint training on both internet video\ndatasets with low-quality audio and professional SFX recordings, enabling\nhigh-quality, full-bandwidth (48kHz) audio generation. Through automated\nevaluations and human studies, we demonstrate that MultiFoley successfully\ngenerates synchronized high-quality sounds across varied conditional inputs and\noutperforms existing methods. Please see our project page for video results:\nhttps://ificl.github.io/MultiFoley/\n", "date": "2024-11-26", "categories": []}
{"id": "2501.09291", "title": "LAVCap: LLM-based Audio-Visual Captioning using Optimal Transport", "abstract": "  Automated audio captioning is a task that generates textual descriptions for\naudio content, and recent studies have explored using visual information to\nenhance captioning quality. However, current methods often fail to effectively\nfuse audio and visual data, missing important semantic cues from each modality.\nTo address this, we introduce LAVCap, a large language model (LLM)-based\naudio-visual captioning framework that effectively integrates visual\ninformation with audio to improve audio captioning performance. LAVCap employs\nan optimal transport-based alignment loss to bridge the modality gap between\naudio and visual features, enabling more effective semantic extraction.\nAdditionally, we propose an optimal transport attention module that enhances\naudio-visual fusion using an optimal transport assignment map. Combined with\nthe optimal training strategy, experimental results demonstrate that each\ncomponent of our framework is effective. LAVCap outperforms existing\nstate-of-the-art methods on the AudioCaps dataset, without relying on large\ndatasets or post-processing. Code is available at\nhttps://github.com/NAVER-INTEL-Co-Lab/gaudi-lavcap.\n", "date": "2025-01-16", "categories": []}
{"id": "2503.10719", "title": "Long-Video Audio Synthesis with Multi-Agent Collaboration", "abstract": "  Video-to-audio synthesis, which generates synchronized audio for visual\ncontent, critically enhances viewer immersion and narrative coherence in film\nand interactive media. However, video-to-audio dubbing for long-form content\nremains an unsolved challenge due to dynamic semantic shifts, temporal\nmisalignment, and the absence of dedicated datasets. While existing methods\nexcel in short videos, they falter in long scenarios (e.g., movies) due to\nfragmented synthesis and inadequate cross-scene consistency. We propose\nLVAS-Agent, a novel multi-agent framework that emulates professional dubbing\nworkflows through collaborative role specialization. Our approach decomposes\nlong-video synthesis into four steps including scene segmentation, script\ngeneration, sound design and audio synthesis. Central innovations include a\ndiscussion-correction mechanism for scene/script refinement and a\ngeneration-retrieval loop for temporal-semantic alignment. To enable systematic\nevaluation, we introduce LVAS-Bench, the first benchmark with 207\nprofessionally curated long videos spanning diverse scenarios. Experiments\ndemonstrate superior audio-visual alignment over baseline methods. Project\npage: https://lvas-agent.github.io\n", "date": "2025-03-13", "categories": []}
{"id": "2503.12840", "title": "Dynamic Derivation and Elimination: Audio Visual Segmentation with\n  Enhanced Audio Semantics", "abstract": "  Sound-guided object segmentation has drawn considerable attention for its\npotential to enhance multimodal perception. Previous methods primarily focus on\ndeveloping advanced architectures to facilitate effective audio-visual\ninteractions, without fully addressing the inherent challenges posed by audio\nnatures, \\emph{\\ie}, (1) feature confusion due to the overlapping nature of\naudio signals, and (2) audio-visual matching difficulty from the varied sounds\nproduced by the same object. To address these challenges, we propose Dynamic\nDerivation and Elimination (DDESeg): a novel audio-visual segmentation\nframework. Specifically, to mitigate feature confusion, DDESeg reconstructs the\nsemantic content of the mixed audio signal by enriching the distinct semantic\ninformation of each individual source, deriving representations that preserve\nthe unique characteristics of each sound. To reduce the matching difficulty, we\nintroduce a discriminative feature learning module, which enhances the semantic\ndistinctiveness of generated audio representations. Considering that not all\nderived audio representations directly correspond to visual features (e.g.,\noff-screen sounds), we propose a dynamic elimination module to filter out\nnon-matching elements. This module facilitates targeted interaction between\nsounding regions and relevant audio semantics. By scoring the interacted\nfeatures, we identify and filter out irrelevant audio information, ensuring\naccurate audio-visual alignment. Comprehensive experiments demonstrate that our\nframework achieves superior performance in AVS datasets.\n", "date": "2025-03-17", "categories": []}
{"id": "2305.07499", "title": "Device-Robust Acoustic Scene Classification via Impulse Response\n  Augmentation", "abstract": "  The ability to generalize to a wide range of recording devices is a crucial\nperformance factor for audio classification models. The characteristics of\ndifferent types of microphones introduce distributional shifts in the digitized\naudio signals due to their varying frequency responses. If this domain shift is\nnot taken into account during training, the model's performance could degrade\nseverely when it is applied to signals recorded by unseen devices. In\nparticular, training a model on audio signals recorded with a small number of\ndifferent microphones can make generalization to unseen devices difficult. To\ntackle this problem, we convolve audio signals in the training set with\npre-recorded device impulse responses (DIRs) to artificially increase the\ndiversity of recording devices. We systematically study the effect of DIR\naugmentation on the task of Acoustic Scene Classification using CNNs and Audio\nSpectrogram Transformers. The results show that DIR augmentation in isolation\nperforms similarly to the state-of-the-art method Freq-MixStyle. However, we\nalso show that DIR augmentation and Freq-MixStyle are complementary, achieving\na new state-of-the-art performance on signals recorded by devices unseen during\ntraining.\n", "date": "2023-05-12", "categories": []}
{"id": "2503.10700", "title": "TA-V2A: Textually Assisted Video-to-Audio Generation", "abstract": "  As artificial intelligence-generated content (AIGC) continues to evolve,\nvideo-to-audio (V2A) generation has emerged as a key area with promising\napplications in multimedia editing, augmented reality, and automated content\ncreation. While Transformer and Diffusion models have advanced audio\ngeneration, a significant challenge persists in extracting precise semantic\ninformation from videos, as current models often lose sequential context by\nrelying solely on frame-based features. To address this, we present TA-V2A, a\nmethod that integrates language, audio, and video features to improve semantic\nrepresentation in latent space. By incorporating large language models for\nenhanced video comprehension, our approach leverages text guidance to enrich\nsemantic expression. Our diffusion model-based system utilizes automated text\nmodulation to enhance inference quality and efficiency, providing personalized\ncontrol through text-guided interfaces. This integration enhances semantic\nexpression while ensuring temporal alignment, leading to more accurate and\ncoherent video-to-audio generation.\n", "date": "2025-03-12", "categories": []}
{"id": "2408.10330", "title": "Meta-Learning in Audio and Speech Processing: An End to End\n  Comprehensive Review", "abstract": "  This survey overviews various meta-learning approaches used in audio and\nspeech processing scenarios. Meta-learning is used where model performance\nneeds to be maximized with minimum annotated samples, making it suitable for\nlow-sample audio processing. Although the field has made some significant\ncontributions, audio meta-learning still lacks the presence of comprehensive\nsurvey papers. We present a systematic review of meta-learning methodologies in\naudio processing. This includes audio-specific discussions on data\naugmentation, feature extraction, preprocessing techniques, meta-learners, task\nselection strategies and also presents important datasets in audio, together\nwith crucial real-world use cases. Through this extensive review, we aim to\nprovide valuable insights and identify future research directions in the\nintersection of meta-learning and audio processing.\n", "date": "2024-08-19", "categories": []}
{"id": "2410.02056", "title": "Synthio: Augmenting Small-Scale Audio Classification Datasets with\n  Synthetic Data", "abstract": "  We present Synthio, a novel approach for augmenting small-scale audio\nclassification datasets with synthetic data. Our goal is to improve audio\nclassification accuracy with limited labeled data. Traditional data\naugmentation techniques, which apply artificial transformations (e.g., adding\nrandom noise or masking segments), struggle to create data that captures the\ntrue diversity present in real-world audios. To address this shortcoming, we\npropose to augment the dataset with synthetic audio generated from\ntext-to-audio (T2A) diffusion models. However, synthesizing effective\naugmentations is challenging because not only should the generated data be\nacoustically consistent with the underlying small-scale dataset, but they\nshould also have sufficient compositional diversity. To overcome the first\nchallenge, we align the generations of the T2A model with the small-scale\ndataset using preference optimization. This ensures that the acoustic\ncharacteristics of the generated data remain consistent with the small-scale\ndataset. To address the second challenge, we propose a novel caption generation\ntechnique that leverages the reasoning capabilities of Large Language Models to\n(1) generate diverse and meaningful audio captions and (2) iteratively refine\ntheir quality. The generated captions are then used to prompt the aligned T2A\nmodel. We extensively evaluate Synthio on ten datasets and four simulated\nlimited-data settings. Results indicate our method consistently outperforms all\nbaselines by 0.1%-39% using a T2A model trained only on weakly-captioned\nAudioSet.\n", "date": "2024-10-02", "categories": []}
{"id": "2501.17202", "title": "Audio Large Language Models Can Be Descriptive Speech Quality Evaluators", "abstract": "  An ideal multimodal agent should be aware of the quality of its input\nmodalities. Recent advances have enabled large language models (LLMs) to\nincorporate auditory systems for handling various speech-related tasks.\nHowever, most audio LLMs remain unaware of the quality of the speech they\nprocess. This limitation arises because speech quality evaluation is typically\nexcluded from multi-task training due to the lack of suitable datasets. To\naddress this, we introduce the first natural language-based speech evaluation\ncorpus, generated from authentic human ratings. In addition to the overall Mean\nOpinion Score (MOS), this corpus offers detailed analysis across multiple\ndimensions and identifies causes of quality degradation. It also enables\ndescriptive comparisons between two speech samples (A/B tests) with human-like\njudgment. Leveraging this corpus, we propose an alignment approach with LLM\ndistillation (ALLD) to guide the audio LLM in extracting relevant information\nfrom raw speech and generating meaningful responses. Experimental results\ndemonstrate that ALLD outperforms the previous state-of-the-art regression\nmodel in MOS prediction, with a mean square error of 0.17 and an A/B test\naccuracy of 98.6%. Additionally, the generated responses achieve BLEU scores of\n25.8 and 30.2 on two tasks, surpassing the capabilities of task-specific\nmodels. This work advances the comprehensive perception of speech signals by\naudio LLMs, contributing to the development of real-world auditory and sensory\nintelligent agents.\n", "date": "2025-01-27", "categories": []}
{"id": "2407.07464", "title": "Video-to-Audio Generation with Hidden Alignment", "abstract": "  Generating semantically and temporally aligned audio content in accordance\nwith video input has become a focal point for researchers, particularly\nfollowing the remarkable breakthrough in text-to-video generation. In this\nwork, we aim to offer insights into the video-to-audio generation paradigm,\nfocusing on three crucial aspects: vision encoders, auxiliary embeddings, and\ndata augmentation techniques. Beginning with a foundational model built on a\nsimple yet surprisingly effective intuition, we explore various vision encoders\nand auxiliary embeddings through ablation studies. Employing a comprehensive\nevaluation pipeline that emphasizes generation quality and video-audio\nsynchronization alignment, we demonstrate that our model exhibits\nstate-of-the-art video-to-audio generation capabilities. Furthermore, we\nprovide critical insights into the impact of different data augmentation\nmethods on enhancing the generation framework's overall capacity. We showcase\npossibilities to advance the challenge of generating synchronized audio from\nsemantic and temporal perspectives. We hope these insights will serve as a\nstepping stone toward developing more realistic and accurate audio-visual\ngeneration models.\n", "date": "2024-07-10", "categories": []}
{"id": "2501.04926", "title": "FLowHigh: Towards Efficient and High-Quality Audio Super-Resolution with\n  Single-Step Flow Matching", "abstract": "  Audio super-resolution is challenging owing to its ill-posed nature.\nRecently, the application of diffusion models in audio super-resolution has\nshown promising results in alleviating this challenge. However, diffusion-based\nmodels have limitations, primarily the necessity for numerous sampling steps,\nwhich causes significantly increased latency when synthesizing high-quality\naudio samples. In this paper, we propose FLowHigh, a novel approach that\nintegrates flow matching, a highly efficient generative model, into audio\nsuper-resolution. We also explore probability paths specially tailored for\naudio super-resolution, which effectively capture high-resolution audio\ndistributions, thereby enhancing reconstruction quality. The proposed method\ngenerates high-fidelity, high-resolution audio through a single-step sampling\nprocess across various input sampling rates. The experimental results on the\nVCTK benchmark dataset demonstrate that FLowHigh achieves state-of-the-art\nperformance in audio super-resolution, as evaluated by log-spectral distance\nand ViSQOL while maintaining computational efficiency with only a single-step\nsampling process.\n", "date": "2025-01-09", "categories": []}
{"id": "2503.08540", "title": "Mellow: a small audio language model for reasoning", "abstract": "  Multimodal Audio-Language Models (ALMs) can understand and reason over both\naudio and text. Typically, reasoning performance correlates with model size,\nwith the best results achieved by models exceeding 8 billion parameters.\nHowever, no prior work has explored enabling small audio-language models to\nperform reasoning tasks, despite the potential applications for edge devices.\nTo address this gap, we introduce Mellow, a small Audio-Language Model\nspecifically designed for reasoning. Mellow achieves state-of-the-art\nperformance among existing small audio-language models and surpasses several\nlarger models in reasoning capabilities. For instance, Mellow scores 52.11 on\nMMAU, comparable to SoTA Qwen2 Audio (which scores 52.5) while using 50 times\nfewer parameters and being trained on 60 times less data (audio hrs). To train\nMellow, we introduce ReasonAQA, a dataset designed to enhance audio-grounded\nreasoning in models. It consists of a mixture of existing datasets (30% of the\ndata) and synthetically generated data (70%). The synthetic dataset is derived\nfrom audio captioning datasets, where Large Language Models (LLMs) generate\ndetailed and multiple-choice questions focusing on audio events, objects,\nacoustic scenes, signal properties, semantics, and listener emotions. To\nevaluate Mellow's reasoning ability, we benchmark it on a diverse set of tasks,\nassessing on both in-distribution and out-of-distribution data, including audio\nunderstanding, deductive reasoning, and comparative reasoning. Finally, we\nconduct extensive ablation studies to explore the impact of projection layer\nchoices, synthetic data generation methods, and language model pretraining on\nreasoning performance. Our training dataset, findings, and baseline pave the\nway for developing small ALMs capable of reasoning.\n", "date": "2025-03-11", "categories": []}
{"id": "2502.15602", "title": "KAD: No More FAD! An Effective and Efficient Evaluation Metric for Audio\n  Generation", "abstract": "  Although being widely adopted for evaluating generated audio signals, the\nFr\\'echet Audio Distance (FAD) suffers from significant limitations, including\nreliance on Gaussian assumptions, sensitivity to sample size, and high\ncomputational complexity. As an alternative, we introduce the Kernel Audio\nDistance (KAD), a novel, distribution-free, unbiased, and computationally\nefficient metric based on Maximum Mean Discrepancy (MMD). Through analysis and\nempirical validation, we demonstrate KAD's advantages: (1) faster convergence\nwith smaller sample sizes, enabling reliable evaluation with limited data; (2)\nlower computational cost, with scalable GPU acceleration; and (3) stronger\nalignment with human perceptual judgments. By leveraging advanced embeddings\nand characteristic kernels, KAD captures nuanced differences between real and\ngenerated audio. Open-sourced in the kadtk toolkit, KAD provides an efficient,\nreliable, and perceptually aligned benchmark for evaluating generative audio\nmodels.\n", "date": "2025-02-21", "categories": []}
{"id": "2503.05929", "title": "Audio-to-Image Encoding for Improved Voice Characteristic Detection\n  Using Deep Convolutional Neural Networks", "abstract": "  This paper introduces a novel audio-to-image encoding framework that\nintegrates multiple dimensions of voice characteristics into a single RGB image\nfor speaker recognition. In this method, the green channel encodes raw audio\ndata, the red channel embeds statistical descriptors of the voice signal\n(including key metrics such as median and mean values for fundamental\nfrequency, spectral centroid, bandwidth, rolloff, zero-crossing rate, MFCCs,\nRMS energy, spectral flatness, spectral contrast, chroma, and harmonic-to-noise\nratio), and the blue channel comprises subframes representing these features in\na spatially organized format. A deep convolutional neural network trained on\nthese composite images achieves 98% accuracy in speaker classification across\ntwo speakers, suggesting that this integrated multi-channel representation can\nprovide a more discriminative input for voice recognition tasks.\n", "date": "2025-03-07", "categories": []}
{"id": "2503.06247", "title": "Infant Cry Detection Using Causal Temporal Representation", "abstract": "  This paper addresses a major challenge in acoustic event detection, in\nparticular infant cry detection in the presence of other sounds and background\nnoises: the lack of precise annotated data. We present two contributions for\nsupervised and unsupervised infant cry detection. The first is an annotated\ndataset for cry segmentation, which enables supervised models to achieve\nstate-of-the-art performance. Additionally, we propose a novel unsupervised\nmethod, Causal Representation Spare Transition Clustering (CRSTC), based on\ncausal temporal representation, which helps address the issue of data scarcity\nmore generally. By integrating the detected cry segments, we significantly\nimprove the performance of downstream infant cry classification, highlighting\nthe potential of this approach for infant care applications.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.06984", "title": "Synchronized Video-to-Audio Generation via Mel Quantization-Continuum\n  Decomposition", "abstract": "  Video-to-audio generation is essential for synthesizing realistic audio\ntracks that synchronize effectively with silent videos. Following the\nperspective of extracting essential signals from videos that can precisely\ncontrol the mature text-to-audio generative diffusion models, this paper\npresents how to balance the representation of mel-spectrograms in terms of\ncompleteness and complexity through a new approach called Mel\nQuantization-Continuum Decomposition (Mel-QCD). We decompose the\nmel-spectrogram into three distinct types of signals, employing quantization or\ncontinuity to them, we can effectively predict them from video by a devised\nvideo-to-all (V2X) predictor. Then, the predicted signals are recomposed and\nfed into a ControlNet, along with a textual inversion design, to control the\naudio generation process. Our proposed Mel-QCD method demonstrates\nstate-of-the-art performance across eight metrics, evaluating dimensions such\nas quality, synchronization, and semantic consistency. Our codes and demos will\nbe released at \\href{Website}{https://wjc2830.github.io/MelQCD/}.\n", "date": "2025-03-10", "categories": []}
{"id": "2408.12481", "title": "Self-Learning for Personalized Keyword Spotting on Ultra-Low-Power Audio\n  Sensors", "abstract": "  This paper proposes a self-learning method to incrementally train (fine-tune)\na personalized Keyword Spotting (KWS) model after the deployment on ultra-low\npower smart audio sensors. We address the fundamental problem of the absence of\nlabeled training data by assigning pseudo-labels to the new recorded audio\nframes based on a similarity score with respect to few user recordings. By\nexperimenting with multiple KWS models with a number of parameters up to 0.5M\non two public datasets, we show an accuracy improvement of up to +19.2% and\n+16.0% vs. the initial models pretrained on a large set of generic keywords.\nThe labeling task is demonstrated on a sensor system composed of a low-power\nmicrophone and an energy-efficient Microcontroller (MCU). By efficiently\nexploiting the heterogeneous processing engines of the MCU, the always-on\nlabeling task runs in real-time with an average power cost of up to 8.2 mW. On\nthe same platform, we estimate an energy cost for on-device training 10x lower\nthan the labeling energy if sampling a new utterance every 6.1 s or 18.8 s with\na DS-CNN-S or a DS-CNN-M model. Our empirical result paves the way to\nself-adaptive personalized KWS sensors at the extreme edge.\n", "date": "2024-08-22", "categories": []}
{"id": "2503.03797", "title": "VoiceGRPO: Modern MoE Transformers with Group Relative Policy\n  Optimization GRPO for AI Voice Health Care Applications on Voice Pathology\n  Detection", "abstract": "  This research introduces a novel AI techniques as Mixture-of-Experts\nTransformers with Group Relative Policy Optimization (GRPO) for voice health\ncare applications on voice pathology detection. With the architectural\ninnovations, we adopt advanced training paradigms inspired by reinforcement\nlearning, namely Proximal Policy Optimization (PPO) and Group-wise Regularized\nPolicy Optimization (GRPO), to enhance model stability and performance.\nExperiments conducted on a synthetically generated voice pathology dataset\ndemonstrate that our proposed models significantly improve diagnostic accuracy,\nF1 score, and ROC-AUC compared to conventional approaches. These findings\nunderscore the potential of integrating transformer architectures with novel\ntraining strategies to advance automated voice pathology detection and\nultimately contribute to more effective healthcare delivery. The code we used\nto train and evaluate our models is available at\nhttps://github.com/enkhtogtokh/voicegrpo\n", "date": "2025-03-05", "categories": []}
{"id": "2503.03983", "title": "Audio Flamingo 2: An Audio-Language Model with Long-Audio Understanding\n  and Expert Reasoning Abilities", "abstract": "  Understanding and reasoning over non-speech sounds and music are crucial for\nboth humans and AI agents to interact effectively with their environments. In\nthis paper, we introduce Audio Flamingo 2 (AF2), an Audio-Language Model (ALM)\nwith advanced audio understanding and reasoning capabilities. AF2 leverages (i)\na custom CLAP model, (ii) synthetic Audio QA data for fine-grained audio\nreasoning, and (iii) a multi-stage curriculum learning strategy. AF2 achieves\nstate-of-the-art performance with only a 3B parameter small language model,\nsurpassing large open-source and proprietary models across over 20 benchmarks.\nNext, for the first time, we extend audio understanding to long audio segments\n(30 secs to 5 mins) and propose LongAudio, a large and novel dataset for\ntraining ALMs on long audio captioning and question-answering tasks.\nFine-tuning AF2 on LongAudio leads to exceptional performance on our proposed\nLongAudioBench, an expert annotated benchmark for evaluating ALMs on long audio\nunderstanding capabilities. We conduct extensive ablation studies to confirm\nthe efficacy of our approach. Project Website:\nhttps://research.nvidia.com/labs/adlr/AF2/.\n", "date": "2025-03-06", "categories": []}
{"id": "2406.05923", "title": "Contrastive Learning from Synthetic Audio Doppelg\\\"angers", "abstract": "  Learning robust audio representations currently demands extensive datasets of\nreal-world sound recordings. By applying artificial transformations to these\nrecordings, models can learn to recognize similarities despite subtle\nvariations through techniques like contrastive learning. However, these\ntransformations are only approximations of the true diversity found in\nreal-world sounds, which are generated by complex interactions of physical\nprocesses, from vocal cord vibrations to the resonance of musical instruments.\nWe propose a solution to both the data scale and transformation limitations,\nleveraging synthetic audio. By randomly perturbing the parameters of a sound\nsynthesizer, we generate audio doppelg\\\"angers-synthetic positive pairs with\ncausally manipulated variations in timbre, pitch, and temporal envelopes. These\nvariations, difficult to achieve through augmentations of existing audio,\nprovide a rich source of contrastive information. Despite the shift to randomly\ngenerated synthetic data, our method produces strong representations,\noutperforming real data on several standard audio classification tasks.\nNotably, our approach is lightweight, requires no data storage, and has only a\nsingle hyperparameter, which we extensively analyze. We offer this method as a\ncomplement to existing strategies for contrastive learning in audio, using\nsynthesized sounds to reduce the data burden on practitioners.\n", "date": "2024-06-09", "categories": []}
{"id": "2502.00734", "title": "CycleGuardian: A Framework for Automatic RespiratorySound classification\n  Based on Improved Deep clustering and Contrastive Learning", "abstract": "  Auscultation plays a pivotal role in early respiratory and pulmonary disease\ndiagnosis. Despite the emergence of deep learning-based methods for automatic\nrespiratory sound classification post-Covid-19, limited datasets impede\nperformance enhancement. Distinguishing between normal and abnormal respiratory\nsounds poses challenges due to the coexistence of normal respiratory components\nand noise components in both types. Moreover, different abnormal respiratory\nsounds exhibit similar anomalous features, hindering their differentiation.\nBesides, existing state-of-the-art models suffer from excessive parameter size,\nimpeding deployment on resource-constrained mobile platforms. To address these\nissues, we design a lightweight network CycleGuardian and propose a framework\nbased on an improved deep clustering and contrastive learning. We first\ngenerate a hybrid spectrogram for feature diversity and grouping spectrograms\nto facilitating intermittent abnormal sound capture.Then, CycleGuardian\nintegrates a deep clustering module with a similarity-constrained clustering\ncomponent to improve the ability to capture abnormal features and a contrastive\nlearning module with group mixing for enhanced abnormal feature discernment.\nMulti-objective optimization enhances overall performance during training. In\nexperiments we use the ICBHI2017 dataset, following the official split method\nand without any pre-trained weights, our method achieves Sp: 82.06 $\\%$, Se:\n44.47$\\%$, and Score: 63.26$\\%$ with a network model size of 38M, comparing to\nthe current model, our method leads by nearly 7$\\%$, achieving the current best\nperformances. Additionally, we deploy the network on Android devices,\nshowcasing a comprehensive intelligent respiratory sound auscultation system.\n", "date": "2025-02-02", "categories": []}
{"id": "2503.01485", "title": "FlowDec: A flow-based full-band general audio codec with high perceptual\n  quality", "abstract": "  We propose FlowDec, a neural full-band audio codec for general audio sampled\nat 48 kHz that combines non-adversarial codec training with a stochastic\npostfilter based on a novel conditional flow matching method. Compared to the\nprior work ScoreDec which is based on score matching, we generalize from speech\nto general audio and move from 24 kbit/s to as low as 4 kbit/s, while improving\noutput quality and reducing the required postfilter DNN evaluations from 60 to\n6 without any fine-tuning or distillation techniques. We provide theoretical\ninsights and geometric intuitions for our approach in comparison to ScoreDec as\nwell as another recent work that uses flow matching, and conduct ablation\nstudies on our proposed components. We show that FlowDec is a competitive\nalternative to the recent GAN-dominated stream of neural codecs, achieving FAD\nscores better than those of the established GAN-based codec DAC and listening\ntest scores that are on par, and producing qualitatively more natural\nreconstructions for speech and harmonic structures in music.\n", "date": "2025-03-03", "categories": []}
{"id": "2406.05784", "title": "Optimizing Multi-Stuttered Speech Classification: Leveraging Whisper's\n  Encoder for Efficient Parameter Reduction in Automated Assessment", "abstract": "  The automated classification of stuttered speech has significant implications\nfor timely assessments providing assistance to speech language pathologists.\nDespite notable advancements in the field, the cases in which multiple\ndisfluencies occur in speech require attention. We have taken a progressive\napproach to fill this gap by classifying multi-stuttered speech more\nefficiently. The problem has been addressed by firstly curating a dataset of\nmulti-stuttered disfluencies from open source dataset SEP-28k audio clips.\nSecondly, employing Whisper, a state-of-the-art speech recognition model has\nbeen leveraged by using its encoder and taking the problem as multi label\nclassification. Thirdly, using a 6 encoder layer Whisper and experimenting with\nvarious layer freezing strategies, a computationally efficient configuration of\nthe model was identified. The proposed configuration achieved micro, macro, and\nweighted F1-scores of 0.88, 0.85, and 0.87, correspondingly on an external test\ndataset i.e. Fluency-Bank. In addition, through layer freezing strategies, we\nwere able to achieve the aforementioned results by fine-tuning a single encoder\nlayer, consequently, reducing the model's trainable parameters from 20.27\nmillion to 3.29 million. This research study unveils the contribution of the\nlast encoder layer in the identification of disfluencies in stuttered speech.\nConsequently, it has led to a computationally efficient approach, 83.7% less\nparameters to train, making the proposed approach more adaptable for various\ndialects and languages.\n", "date": "2024-06-09", "categories": []}
{"id": "2502.19387", "title": "Residual Speech Embeddings for Tone Classification: Removing Linguistic\n  Content to Enhance Paralinguistic Analysis", "abstract": "  Self-supervised learning models for speech processing, such as wav2vec2,\nHuBERT, WavLM, and Whisper, generate embeddings that capture both linguistic\nand paralinguistic information, making it challenging to analyze tone\nindependently of spoken content. In this work, we introduce a method for\ndisentangling paralinguistic features from linguistic content by regressing\nspeech embeddings onto their corresponding text embeddings and using the\nresiduals as a representation of vocal tone. We evaluate this approach across\nmultiple self-supervised speech embeddings, demonstrating that residual\nembeddings significantly improve tone classification performance compared to\nraw speech embeddings. Our results show that this method enhances linear\nseparability, enabling improved classification even with simple models such as\nlogistic regression. Visualization of the residual embeddings further confirms\nthe successful removal of linguistic information while preserving tone-related\nfeatures. These findings highlight the potential of residual embeddings for\napplications in sentiment analysis, speaker characterization, and\nparalinguistic speech processing.\n", "date": "2025-02-26", "categories": []}
{"id": "2408.14080", "title": "SONICS: Synthetic Or Not -- Identifying Counterfeit Songs", "abstract": "  The recent surge in AI-generated songs presents exciting possibilities and\nchallenges. These innovations necessitate the ability to distinguish between\nhuman-composed and synthetic songs to safeguard artistic integrity and protect\nhuman musical artistry. Existing research and datasets in fake song detection\nonly focus on singing voice deepfake detection (SVDD), where the vocals are\nAI-generated but the instrumental music is sourced from real songs. However,\nthese approaches are inadequate for detecting contemporary end-to-end\nartificial songs where all components (vocals, music, lyrics, and style) could\nbe AI-generated. Additionally, existing datasets lack music-lyrics diversity,\nlong-duration songs, and open-access fake songs. To address these gaps, we\nintroduce SONICS, a novel dataset for end-to-end Synthetic Song Detection\n(SSD), comprising over 97k songs (4,751 hours) with over 49k synthetic songs\nfrom popular platforms like Suno and Udio. Furthermore, we highlight the\nimportance of modeling long-range temporal dependencies in songs for effective\nauthenticity detection, an aspect entirely overlooked in existing methods. To\nutilize long-range patterns, we introduce SpecTTTra, a novel architecture that\nsignificantly improves time and memory efficiency over conventional CNN and\nTransformer-based models. For long songs, our top-performing variant\noutperforms ViT by 8% in F1 score, is 38% faster, and uses 26% less memory,\nwhile also surpassing ConvNeXt with a 1% F1 score gain, 20% speed boost, and\n67% memory reduction.\n", "date": "2024-08-26", "categories": []}
{"id": "2408.16532", "title": "WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio\n  Language Modeling", "abstract": "  Language models have been effectively applied to modeling natural signals,\nsuch as images, video, speech, and audio. A crucial component of these models\nis the codec tokenizer, which compresses high-dimensional natural signals into\nlower-dimensional discrete tokens. In this paper, we introduce WavTokenizer,\nwhich offers several advantages over previous SOTA acoustic codec models in the\naudio domain: 1)extreme compression. By compressing the layers of quantizers\nand the temporal dimension of the discrete codec, one-second audio of 24kHz\nsampling rate requires only a single quantizer with 40 or 75 tokens. 2)improved\nsubjective quality. Despite the reduced number of tokens, WavTokenizer achieves\nstate-of-the-art reconstruction quality with outstanding UTMOS scores and\ninherently contains richer semantic information. Specifically, we achieve these\nresults by designing a broader VQ space, extended contextual windows, and\nimproved attention networks, as well as introducing a powerful multi-scale\ndiscriminator and an inverse Fourier transform structure. We conducted\nextensive reconstruction experiments in the domains of speech, audio, and\nmusic. WavTokenizer exhibited strong performance across various objective and\nsubjective metrics compared to state-of-the-art models. We also tested semantic\ninformation, VQ utilization, and adaptability to generative models.\nComprehensive ablation studies confirm the necessity of each module in\nWavTokenizer. The related code, demos, and pre-trained models are available at\nhttps://github.com/jishengpeng/WavTokenizer.\n", "date": "2024-08-29", "categories": []}
{"id": "2410.10676", "title": "Both Ears Wide Open: Towards Language-Driven Spatial Audio Generation", "abstract": "  Recently, diffusion models have achieved great success in mono-channel audio\ngeneration. However, when it comes to stereo audio generation, the soundscapes\noften have a complex scene of multiple objects and directions. Controlling\nstereo audio with spatial contexts remains challenging due to high data costs\nand unstable generative models. To the best of our knowledge, this work\nrepresents the first attempt to address these issues. We first construct a\nlarge-scale, simulation-based, and GPT-assisted dataset, BEWO-1M, with abundant\nsoundscapes and descriptions even including moving and multiple sources. Beyond\ntext modality, we have also acquired a set of images and rationally paired\nstereo audios through retrieval to advance multimodal generation. Existing\naudio generation models tend to generate rather random and indistinct spatial\naudio. To provide accurate guidance for Latent Diffusion Models, we introduce\nthe SpatialSonic model utilizing spatial-aware encoders and azimuth state\nmatrices to reveal reasonable spatial guidance. By leveraging spatial guidance,\nour model not only achieves the objective of generating immersive and\ncontrollable spatial audio from text but also extends to other modalities as\nthe pioneer attempt. Finally, under fair settings, we conduct subjective and\nobjective evaluations on simulated and real-world data to compare our approach\nwith prevailing methods. The results demonstrate the effectiveness of our\nmethod, highlighting its capability to generate spatial audio that adheres to\nphysical rules.\n", "date": "2024-10-14", "categories": []}
{"id": "2502.17527", "title": "Perceptual Noise-Masking with Music through Deep Spectral Envelope\n  Shaping", "abstract": "  People often listen to music in noisy environments, seeking to isolate\nthemselves from ambient sounds. Indeed, a music signal can mask some of the\nnoise's frequency components due to the effect of simultaneous masking. In this\narticle, we propose a neural network based on a psychoacoustic masking model,\ndesigned to enhance the music's ability to mask ambient noise by reshaping its\nspectral envelope with predicted filter frequency responses. The model is\ntrained with a perceptual loss function that balances two constraints:\neffectively masking the noise while preserving the original music mix and the\nuser's chosen listening level. We evaluate our approach on simulated data\nreplicating a user's experience of listening to music with headphones in a\nnoisy environment. The results, based on defined objective metrics, demonstrate\nthat our system improves the state of the art.\n", "date": "2025-02-24", "categories": []}
{"id": "2502.18328", "title": "From Vision to Sound: Advancing Audio Anomaly Detection with\n  Vision-Based Algorithms", "abstract": "  Recent advances in Visual Anomaly Detection (VAD) have introduced\nsophisticated algorithms leveraging embeddings generated by pre-trained feature\nextractors. Inspired by these developments, we investigate the adaptation of\nsuch algorithms to the audio domain to address the problem of Audio Anomaly\nDetection (AAD). Unlike most existing AAD methods, which primarily classify\nanomalous samples, our approach introduces fine-grained temporal-frequency\nlocalization of anomalies within the spectrogram, significantly improving\nexplainability. This capability enables a more precise understanding of where\nand when anomalies occur, making the results more actionable for end users. We\nevaluate our approach on industrial and environmental benchmarks, demonstrating\nthe effectiveness of VAD techniques in detecting anomalies in audio signals.\nMoreover, they improve explainability by enabling localized anomaly\nidentification, making audio anomaly detection systems more interpretable and\npractical.\n", "date": "2025-02-25", "categories": []}
{"id": "2411.19806", "title": "Zero-shot Musical Stem Retrieval with Joint-Embedding Predictive\n  Architectures", "abstract": "  In this paper, we tackle the task of musical stem retrieval. Given a musical\nmix, it consists in retrieving a stem that would fit with it, i.e., that would\nsound pleasant if played together. To do so, we introduce a new method based on\nJoint-Embedding Predictive Architectures, where an encoder and a predictor are\njointly trained to produce latent representations of a context and predict\nlatent representations of a target. In particular, we design our predictor to\nbe conditioned on arbitrary instruments, enabling our model to perform\nzero-shot stem retrieval. In addition, we discover that pretraining the encoder\nusing contrastive learning drastically improves the model's performance.\n  We validate the retrieval performances of our model using the MUSDB18 and\nMoisesDB datasets. We show that it significantly outperforms previous baselines\non both datasets, showcasing its ability to support more or less precise (and\npossibly unseen) conditioning. We also evaluate the learned embeddings on a\nbeat tracking task, demonstrating that they retain temporal structure and local\ninformation.\n", "date": "2024-11-29", "categories": []}
{"id": "2502.06664", "title": "Evaluation of Deep Audio Representations for Hearables", "abstract": "  Effectively steering hearable devices requires understanding the acoustic\nenvironment around the user. In the computational analysis of sound scenes,\nfoundation models have emerged as the state of the art to produce\nhigh-performance, robust, multi-purpose audio representations. We introduce and\nrelease Deep Evaluation of Audio Representations (DEAR), the first dataset and\nbenchmark to evaluate the efficacy of foundation models in capturing essential\nacoustic properties for hearables. The dataset includes 1,158 audio tracks,\neach 30 seconds long, created by spatially mixing proprietary monologues with\ncommercial, high-quality recordings of everyday acoustic scenes. Our benchmark\nencompasses eight tasks that assess the general context, speech sources, and\ntechnical acoustic properties of the audio scenes. Through our evaluation of\nfour general-purpose audio representation models, we demonstrate that the BEATs\nmodel significantly surpasses its counterparts. This superiority underscores\nthe advantage of models trained on diverse audio collections, confirming their\napplicability to a wide array of auditory tasks, including encoding the\nenvironment properties necessary for hearable steering. The DEAR dataset and\nassociated code are available at https://dear-dataset.github.io.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.15607", "title": "Benchmarking machine learning for bowel sound pattern classification\n  from tabular features to pretrained models", "abstract": "  The development of electronic stethoscopes and wearable recording sensors\nopened the door to the automated analysis of bowel sound (BS) signals. This\nenables a data-driven analysis of bowel sound patterns, their interrelations,\nand their correlation to different pathologies. This work leverages a BS\ndataset collected from 16 healthy subjects that was annotated according to four\nestablished BS patterns. This dataset is used to evaluate the performance of\nmachine learning models to detect and/or classify BS patterns. The selection of\nconsidered models covers models using tabular features, convolutional neural\nnetworks based on spectrograms and models pre-trained on large audio datasets.\nThe results highlight the clear superiority of pre-trained models, particularly\nin detecting classes with few samples, achieving an AUC of 0.89 in\ndistinguishing BS from non-BS using a HuBERT model and an AUC of 0.89 in\ndifferentiating bowel sound patterns using a Wav2Vec 2.0 model. These results\npave the way for an improved understanding of bowel sounds in general and\nfuture machine-learning-driven diagnostic applications for gastrointestinal\nexaminations\n", "date": "2025-02-21", "categories": []}
{"id": "2502.16914", "title": "ENACT-Heart -- ENsemble-based Assessment Using CNN and Transformer on\n  Heart Sounds", "abstract": "  This study explores the application of Vision Transformer (ViT) principles in\naudio analysis, specifically focusing on heart sounds. This paper introduces\nENACT-Heart - a novel ensemble approach that leverages the complementary\nstrengths of Convolutional Neural Networks (CNN) and ViT through a Mixture of\nExperts (MoE) framework, achieving a remarkable classification accuracy of\n97.52%. This outperforms the individual contributions of ViT (93.88%) and CNN\n(95.45%), demonstrating the potential for enhanced diagnostic accuracy in\ncardiovascular health monitoring. These results demonstrate the potential of\nensemble methods in enhancing classification performance for cardiovascular\nhealth monitoring and diagnosis.\n", "date": "2025-02-24", "categories": []}
{"id": "2502.15056", "title": "Fundamental Survey on Neuromorphic Based Audio Classification", "abstract": "  Audio classification is paramount in a variety of applications including\nsurveillance, healthcare monitoring, and environmental analysis. Traditional\nmethods frequently depend on intricate signal processing algorithms and\nmanually crafted features, which may fall short in fully capturing the\ncomplexities of audio patterns. Neuromorphic computing, inspired by the\narchitecture and functioning of the human brain, presents a promising\nalternative for audio classification tasks. This survey provides an exhaustive\nexamination of the current state-of-the-art in neuromorphic-based audio\nclassification. It delves into the crucial components of neuromorphic systems,\nsuch as Spiking Neural Networks (SNNs), memristors, and neuromorphic hardware\nplatforms, highlighting their advantages in audio classification. Furthermore,\nthe survey explores various methodologies and strategies employed in\nneuromorphic audio classification, including event-based processing,\nspike-based learning, and bio-inspired feature extraction. It examines how\nthese approaches address the limitations of traditional audio classification\nmethods, particularly in terms of energy efficiency, real-time processing, and\nrobustness to environmental noise. Additionally, the paper conducts a\ncomparative analysis of different neuromorphic audio classification models and\nbenchmarks, evaluating their performance metrics, computational efficiency, and\nscalability. By providing a comprehensive guide for researchers, engineers and\npractitioners, this survey aims to stimulate further innovation and\nadvancements in the evolving field of neuromorphic audio classification.\n", "date": "2025-02-20", "categories": []}
{"id": "2502.15503", "title": "BAN: Neuroanatomical Aligning in Auditory Recognition between Artificial\n  Neural Network and Human Cortex", "abstract": "  Drawing inspiration from neurosciences, artificial neural networks (ANNs)\nhave evolved from shallow architectures to highly complex, deep structures,\nyielding exceptional performance in auditory recognition tasks. However,\ntraditional ANNs often struggle to align with brain regions due to their\nexcessive depth and lack of biologically realistic features, like recurrent\nconnection. To address this, a brain-like auditory network (BAN) is introduced,\nwhich incorporates four neuroanatomically mapped areas and recurrent\nconnection, guided by a novel metric called the brain-like auditory score\n(BAS). BAS serves as a benchmark for evaluating the similarity between BAN and\nhuman auditory recognition pathway. We further propose that specific areas in\nthe cerebral cortex, mainly the middle and medial superior temporal (T2/T3)\nareas, correspond to the designed network structure, drawing parallels with the\nbrain's auditory perception pathway. Our findings suggest that the\nneuroanatomical similarity in the cortex and auditory classification abilities\nof the ANN are well-aligned. In addition to delivering excellent performance on\na music genre classification task, the BAN demonstrates a high BAS score. In\nconclusion, this study presents BAN as a recurrent, brain-inspired ANN,\nrepresenting the first model that mirrors the cortical pathway of auditory\nrecognition.\n", "date": "2025-02-21", "categories": []}
{"id": "2410.16505", "title": "Do Audio-Language Models Understand Linguistic Variations?", "abstract": "  Open-vocabulary audio language models (ALMs), like Contrastive Language Audio\nPretraining (CLAP), represent a promising new paradigm for audio-text retrieval\nusing natural language queries. In this paper, for the first time, we perform\ncontrolled experiments on various benchmarks to show that existing ALMs\nstruggle to generalize to linguistic variations in textual queries. To address\nthis issue, we propose RobustCLAP, a novel and compute-efficient technique to\nlearn audio-language representations agnostic to linguistic variations.\nSpecifically, we reformulate the contrastive loss used in CLAP architectures by\nintroducing a multi-view contrastive learning objective, where paraphrases are\ntreated as different views of the same audio scene and use this for training.\nOur proposed approach improves the text-to-audio retrieval performance of CLAP\nby 0.8%-13% across benchmarks and enhances robustness to linguistic variation.\n", "date": "2024-10-21", "categories": []}
{"id": "2502.12511", "title": "Myna: Masking-Based Contrastive Learning of Musical Representations", "abstract": "  We present Myna, a simple yet effective approach for self-supervised musical\nrepresentation learning. Built on a contrastive learning framework, Myna\nintroduces two key innovations: (1) the use of a Vision Transformer (ViT) on\nmel-spectrograms as the backbone and (2) a novel data augmentation strategy,\ntoken masking, that masks 90 percent of spectrogram tokens. These innovations\ndeliver both effectiveness and efficiency: (i) Token masking enables a\nsignificant increase in per-GPU batch size, from 48 or 120 in prior methods\n(CLMR, MULE) to 4096. (ii) By avoiding traditional augmentations, Myna retains\npitch sensitivity, enhancing performance in tasks like key detection. (iii) The\nuse of vertical patches allows the model to better capture critical features\nfor key detection. Our hybrid model, Myna-22M-Hybrid, processes both 16x16 and\n128x2 patches, achieving state-of-the-art results. Trained on a single GPU, it\noutperforms MULE (62M) on average and rivals MERT-95M, which was trained on 16\nand 64 GPUs, respectively. Additionally, it surpasses MERT-95M-public,\nestablishing itself as the best-performing model trained on publicly available\ndata. We release our code and models to promote reproducibility and facilitate\nfuture research.\n", "date": "2025-02-18", "categories": []}
{"id": "2502.13440", "title": "Semi-supervised classification of bird vocalizations", "abstract": "  Changes in bird populations can indicate broader changes in ecosystems,\nmaking birds one of the most important animal groups to monitor. Combining\nmachine learning and passive acoustics enables continuous monitoring over\nextended periods without direct human involvement. However, most existing\ntechniques require extensive expert-labeled datasets for training and cannot\neasily detect time-overlapping calls in busy soundscapes. We propose a\nsemi-supervised acoustic bird detector designed to allow both the detection of\ntime-overlapping calls (when separated in frequency) and the use of few labeled\ntraining samples. The classifier is trained and evaluated on a combination of\ncommunity-recorded open-source data and long-duration soundscape recordings\nfrom Singapore. It achieves a mean F0.5 score of 0.701 across 315 classes from\n110 bird species on a hold-out test set, with an average of 11 labeled training\nsamples per class. It outperforms the state-of-the-art BirdNET classifier on a\ntest set of 103 bird species despite significantly fewer labeled training\nsamples. The detector is further tested on 144 microphone-hours of continuous\nsoundscape data. The rich soundscape in Singapore makes suppression of false\npositives a challenge on raw, continuous data streams. Nevertheless, we\ndemonstrate that achieving high precision in such environments with minimal\nlabeled training data is possible.\n", "date": "2025-02-19", "categories": []}
{"id": "2502.12759", "title": "High-Fidelity Music Vocoder using Neural Audio Codecs", "abstract": "  While neural vocoders have made significant progress in high-fidelity speech\nsynthesis, their application on polyphonic music has remained underexplored. In\nthis work, we propose DisCoder, a neural vocoder that leverages a generative\nadversarial encoder-decoder architecture informed by a neural audio codec to\nreconstruct high-fidelity 44.1 kHz audio from mel spectrograms. Our approach\nfirst transforms the mel spectrogram into a lower-dimensional representation\naligned with the Descript Audio Codec (DAC) latent space before reconstructing\nit to an audio signal using a fine-tuned DAC decoder. DisCoder achieves\nstate-of-the-art performance in music synthesis on several objective metrics\nand in a MUSHRA listening study. Our approach also shows competitive\nperformance in speech synthesis, highlighting its potential as a universal\nvocoder.\n", "date": "2025-02-18", "categories": []}
{"id": "2502.13004", "title": "Language Barriers: Evaluating Cross-Lingual Performance of CNN and\n  Transformer Architectures for Speech Quality Estimation", "abstract": "  Objective speech quality models aim to predict human-perceived speech quality\nusing automated methods. However, cross-lingual generalization remains a major\nchallenge, as Mean Opinion Scores (MOS) vary across languages due to\nlinguistic, perceptual, and dataset-specific differences. A model trained\nprimarily on English data may struggle to generalize to languages with\ndifferent phonetic, tonal, and prosodic characteristics, leading to\ninconsistencies in objective assessments. This study investigates the\ncross-lingual performance of two speech quality models: NISQA, a CNN-based\nmodel, and a Transformer-based Audio Spectrogram Transformer (AST) model. Both\nmodels were trained exclusively on English datasets containing over 49,000\nspeech samples and subsequently evaluated on speech in German, French,\nMandarin, Swedish, and Dutch. We analyze model performance using Pearson\nCorrelation Coefficient (PCC) and Root Mean Square Error (RMSE) across five\nspeech quality dimensions: coloration, discontinuity, loudness, noise, and MOS.\nOur findings show that while AST achieves a more stable cross-lingual\nperformance, both models exhibit noticeable biases. Notably, Mandarin speech\nquality predictions correlate highly with human MOS scores, whereas Swedish and\nDutch present greater prediction challenges. Discontinuities remain difficult\nto model across all languages. These results highlight the need for more\nbalanced multilingual datasets and architecture-specific adaptations to improve\ncross-lingual generalization.\n", "date": "2025-02-18", "categories": []}
{"id": "2409.10995", "title": "SynthSOD: Developing an Heterogeneous Dataset for Orchestra Music Source\n  Separation", "abstract": "  Recent advancements in music source separation have significantly progressed,\nparticularly in isolating vocals, drums, and bass elements from mixed tracks.\nThese developments owe much to the creation and use of large-scale, multitrack\ndatasets dedicated to these specific components. However, the challenge of\nextracting similarly sounding sources from orchestra recordings has not been\nextensively explored, largely due to a scarcity of comprehensive and clean (i.e\nbleed-free) multitrack datasets. In this paper, we introduce a novel multitrack\ndataset called SynthSOD, developed using a set of simulation techniques to\ncreate a realistic (i.e. using high-quality soundfonts), musically motivated,\nand heterogeneous training set comprising different dynamics, natural tempo\nchanges, styles, and conditions. Moreover, we demonstrate the application of a\nwidely used baseline music separation model trained on our synthesized dataset\nw.r.t to the well-known EnsembleSet, and evaluate its performance under both\nsynthetic and real-world conditions.\n", "date": "2024-09-17", "categories": []}
{"id": "2502.11840", "title": "ChordFormer: A Conformer-Based Architecture for Large-Vocabulary Audio\n  Chord Recognition", "abstract": "  Chord recognition serves as a critical task in music information retrieval\ndue to the abstract and descriptive nature of chords in music analysis. While\naudio chord recognition systems have achieved significant accuracy for small\nvocabularies (e.g., major/minor chords), large-vocabulary chord recognition\nremains a challenging problem. This complexity also arises from the inherent\nlong-tail distribution of chords, where rare chord types are underrepresented\nin most datasets, leading to insufficient training samples. Effective chord\nrecognition requires leveraging contextual information from audio sequences,\nyet existing models, such as combinations of convolutional neural networks,\nbidirectional long short-term memory networks, and bidirectional transformers,\nface limitations in capturing long-term dependencies and exhibit suboptimal\nperformance on large-vocabulary chord recognition tasks. This work proposes\nChordFormer, a novel conformer-based architecture designed to tackle structural\nchord recognition (e.g., triads, bass, sevenths) for large vocabularies.\nChordFormer leverages conformer blocks that integrate convolutional neural\nnetworks with transformers, thus enabling the model to capture both local\npatterns and global dependencies effectively. By addressing challenges such as\nclass imbalance through a reweighted loss function and structured chord\nrepresentations, ChordFormer outperforms state-of-the-art models, achieving a\n2% improvement in frame-wise accuracy and a 6% increase in class-wise accuracy\non large-vocabulary chord datasets. Furthermore, ChordFormer excels in handling\nclass imbalance, providing robust and balanced recognition across chord types.\nThis approach bridges the gap between theoretical music knowledge and practical\napplications, advancing the field of large-vocabulary chord recognition.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.12007", "title": "Demographic Attributes Prediction from Speech Using WavLM Embeddings", "abstract": "  This paper introduces a general classifier based on WavLM features, to infer\ndemographic characteristics, such as age, gender, native language, education,\nand country, from speech. Demographic feature prediction plays a crucial role\nin applications like language learning, accessibility, and digital forensics,\nenabling more personalized and inclusive technologies. Leveraging pretrained\nmodels for embedding extraction, the proposed framework identifies key acoustic\nand linguistic fea-tures associated with demographic attributes, achieving a\nMean Absolute Error (MAE) of 4.94 for age prediction and over 99.81% accuracy\nfor gender classification across various datasets. Our system improves upon\nexisting models by up to relative 30% in MAE and up to relative 10% in accuracy\nand F1 scores across tasks, leveraging a diverse range of datasets and large\npretrained models to ensure robustness and generalizability. This study offers\nnew insights into speaker diversity and provides a strong foundation for future\nresearch in speech-based demographic profiling.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.10011", "title": "InterGridNet: An Electric Network Frequency Approach for Audio Source\n  Location Classification Using Convolutional Neural Networks", "abstract": "  A novel framework, called InterGridNet, is introduced, leveraging a shallow\nRawNet model for geolocation classification of Electric Network Frequency (ENF)\nsignatures in the SP Cup 2016 dataset. During data preparation, recordings are\nsorted into audio and power groups based on inherent characteristics, further\ndivided into 50 Hz and 60 Hz groups via spectrogram analysis. Residual blocks\nwithin the classification model extract frame-level embeddings, aiding\ndecision-making through softmax activation. The topology and the\nhyperparameters of the shallow RawNet are optimized using a Neural Architecture\nSearch. The overall accuracy of InterGridNet in the test recordings is 92%,\nindicating its effectiveness against the state-of-the-art methods tested in the\nSP Cup 2016. These findings underscore InterGridNet's effectiveness in\naccurately classifying audio recordings from diverse power grids, advancing\nstate-of-the-art geolocation estimation methods.\n", "date": "2025-02-14", "categories": []}
{"id": "2410.02130", "title": "MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers\n  for Open-Domain Sound Generation", "abstract": "  We introduce MDSGen, a novel framework for vision-guided open-domain sound\ngeneration optimized for model parameter size, memory consumption, and\ninference speed. This framework incorporates two key innovations: (1) a\nredundant video feature removal module that filters out unnecessary visual\ninformation, and (2) a temporal-aware masking strategy that leverages temporal\ncontext for enhanced audio generation accuracy. In contrast to existing\nresource-heavy Unet-based models, \\texttt{MDSGen} employs denoising masked\ndiffusion transformers, facilitating efficient generation without reliance on\npre-trained diffusion models. Evaluated on the benchmark VGGSound dataset, our\nsmallest model (5M parameters) achieves $97.9$% alignment accuracy, using\n$172\\times$ fewer parameters, $371$% less memory, and offering $36\\times$\nfaster inference than the current 860M-parameter state-of-the-art model\n($93.9$% accuracy). The larger model (131M parameters) reaches nearly $99$%\naccuracy while requiring $6.5\\times$ fewer parameters. These results highlight\nthe scalability and effectiveness of our approach. The code is available at\nhttps://bit.ly/mdsgen.\n", "date": "2024-10-03", "categories": []}
{"id": "2502.08939", "title": "TokenSynth: A Token-based Neural Synthesizer for Instrument Cloning and\n  Text-to-Instrument", "abstract": "  Recent advancements in neural audio codecs have enabled the use of tokenized\naudio representations in various audio generation tasks, such as\ntext-to-speech, text-to-audio, and text-to-music generation. Leveraging this\napproach, we propose TokenSynth, a novel neural synthesizer that utilizes a\ndecoder-only transformer to generate desired audio tokens from MIDI tokens and\nCLAP (Contrastive Language-Audio Pretraining) embedding, which has\ntimbre-related information. Our model is capable of performing instrument\ncloning, text-to-instrument synthesis, and text-guided timbre manipulation\nwithout any fine-tuning. This flexibility enables diverse sound design and\nintuitive timbre control. We evaluated the quality of the synthesized audio,\nthe timbral similarity between synthesized and target audio/text, and synthesis\naccuracy (i.e., how accurately it follows the input MIDI) using objective\nmeasures. TokenSynth demonstrates the potential of leveraging advanced neural\naudio codecs and transformers to create powerful and versatile neural\nsynthesizers. The source code, model weights, and audio demos are available at:\nhttps://github.com/KyungsuKim42/tokensynth\n", "date": "2025-02-13", "categories": []}
{"id": "2405.02119", "title": "EnvId: A Metric Learning Approach for Forensic Few-Shot Identification\n  of Unseen Environments", "abstract": "  Audio recordings may provide important evidence in criminal investigations.\nOne such case is the forensic association of a recorded audio to its recording\nlocation. For example, a voice message may be the only investigative cue to\nnarrow down the candidate sites for a crime. Up to now, several works provide\nsupervised classification tools for closed-set recording environment\nidentification under relatively clean recording conditions. However, in\nforensic investigations, the candidate locations are case-specific. Thus,\nsupervised learning techniques are not applicable without retraining a\nclassifier on a sufficient amount of training samples for each case and\nrespective candidate set. In addition, a forensic tool has to deal with audio\nmaterial from uncontrolled sources with variable properties and quality. In\nthis work, we therefore attempt a major step towards practical forensic\napplication scenarios. We propose a representation learning framework called\nEnvId, short for environment identification. EnvId avoids case-specific\nretraining by modeling the task as a few-shot classification problem. We\ndemonstrate that EnvId can handle forensically challenging material. It\nprovides good quality predictions even under unseen signal degradations,\nout-of-distribution reverberation characteristics or recording position\nmismatches.\n", "date": "2024-05-03", "categories": []}
{"id": "2409.02266", "title": "LSTMSE-Net: Long Short Term Speech Enhancement Network for Audio-visual\n  Speech Enhancement", "abstract": "  In this paper, we propose long short term memory speech enhancement network\n(LSTMSE-Net), an audio-visual speech enhancement (AVSE) method. This innovative\nmethod leverages the complementary nature of visual and audio information to\nboost the quality of speech signals. Visual features are extracted with\nVisualFeatNet (VFN), and audio features are processed through an encoder and\ndecoder. The system scales and concatenates visual and audio features, then\nprocesses them through a separator network for optimized speech enhancement.\nThe architecture highlights advancements in leveraging multi-modal data and\ninterpolation techniques for robust AVSE challenge systems. The performance of\nLSTMSE-Net surpasses that of the baseline model from the COG-MHEAR AVSE\nChallenge 2024 by a margin of 0.06 in scale-invariant signal-to-distortion\nratio (SISDR), $0.03$ in short-time objective intelligibility (STOI), and\n$1.32$ in perceptual evaluation of speech quality (PESQ). The source code of\nthe proposed LSTMSE-Net is available at\n\\url{https://github.com/mtanveer1/AVSEC-3-Challenge}.\n", "date": "2024-09-03", "categories": []}
{"id": "2409.11228", "title": "Learning Source Disentanglement in Neural Audio Codec", "abstract": "  Neural audio codecs have significantly advanced audio compression by\nefficiently converting continuous audio signals into discrete tokens. These\ncodecs preserve high-quality sound and enable sophisticated sound generation\nthrough generative models trained on these tokens. However, existing neural\ncodec models are typically trained on large, undifferentiated audio datasets,\nneglecting the essential discrepancies between sound domains like speech,\nmusic, and environmental sound effects. This oversight complicates data\nmodeling and poses additional challenges to the controllability of sound\ngeneration. To tackle these issues, we introduce the Source-Disentangled Neural\nAudio Codec (SD-Codec), a novel approach that combines audio coding and source\nseparation. By jointly learning audio resynthesis and separation, SD-Codec\nexplicitly assigns audio signals from different domains to distinct codebooks,\nsets of discrete representations. Experimental results indicate that SD-Codec\nnot only maintains competitive resynthesis quality but also, supported by the\nseparation results, demonstrates successful disentanglement of different\nsources in the latent space, thereby enhancing interpretability in audio codec\nand providing potential finer control over the audio generation process.\n", "date": "2024-09-17", "categories": []}
{"id": "2412.10857", "title": "Robust Persian Digit Recognition in Noisy Environments Using Hybrid\n  CNN-BiGRU Model", "abstract": "  Artificial intelligence (AI) has significantly advanced speech recognition\napplications. However, many existing neural network-based methods struggle with\nnoise, reducing accuracy in real-world environments. This study addresses\nisolated spoken Persian digit recognition (zero to nine) under noisy\nconditions, particularly for phonetically similar numbers. A hybrid model\ncombining residual convolutional neural networks and bidirectional gated\nrecurrent units (BiGRU) is proposed, utilizing word units instead of phoneme\nunits for speaker-independent recognition. The FARSDIGIT1 dataset, augmented\nwith various approaches, is processed using Mel-Frequency Cepstral Coefficients\n(MFCC) for feature extraction. Experimental results demonstrate the model's\neffectiveness, achieving 98.53%, 96.10%, and 95.92% accuracy on training,\nvalidation, and test sets, respectively. In noisy conditions, the proposed\napproach improves recognition by 26.88% over phoneme unit-based LSTM models and\nsurpasses the Mel-scale Two Dimension Root Cepstrum Coefficients (MTDRCC)\nfeature extraction technique along with MLP model (MTDRCC+MLP) by 7.61%.\n", "date": "2024-12-14", "categories": []}
{"id": "2211.02874", "title": "Improved Techniques for the Conditional Generative Augmentation of\n  Clinical Audio Data", "abstract": "  Data augmentation is a valuable tool for the design of deep learning systems\nto overcome data limitations and stabilize the training process. Especially in\nthe medical domain, where the collection of large-scale data sets is\nchallenging and expensive due to limited access to patient data, relevant\nenvironments, as well as strict regulations, community-curated large-scale\npublic datasets, pretrained models, and advanced data augmentation methods are\nthe main factors for developing reliable systems to improve patient care.\nHowever, for the development of medical acoustic sensing systems, an emerging\nfield of research, the community lacks large-scale publicly available data sets\nand pretrained models. To address the problem of limited data, we propose a\nconditional generative adversarial neural network-based augmentation method\nwhich is able to synthesize mel spectrograms from a learned data distribution\nof a source data set. In contrast to previously proposed fully convolutional\nmodels, the proposed model implements residual Squeeze and Excitation modules\nin the generator architecture. We show that our method outperforms all\nclassical audio augmentation techniques and previously published generative\nmethods in terms of generated sample quality and a performance improvement of\n2.84% of Macro F1-Score for a classifier trained on the augmented data set, an\nenhancement of $1.14\\%$ in relation to previous work. By analyzing the\ncorrelation of intermediate feature spaces, we show that the residual Squeeze\nand Excitation modules help the model to reduce redundancy in the latent\nfeatures. Therefore, the proposed model advances the state-of-the-art in the\naugmentation of clinical audio data and improves the data bottleneck for the\ndesign of clinical acoustic sensing systems.\n", "date": "2022-11-05", "categories": []}
{"id": "2406.02345", "title": "Progressive Confident Masking Attention Network for Audio-Visual\n  Segmentation", "abstract": "  Audio and visual signals typically occur simultaneously, and humans possess\nan innate ability to correlate and synchronize information from these two\nmodalities. Recently, a challenging problem known as Audio-Visual Segmentation\n(AVS) has emerged, intending to produce segmentation maps for sounding objects\nwithin a scene. However, the methods proposed so far have not sufficiently\nintegrated audio and visual information, and the computational costs have been\nextremely high. Additionally, the outputs of different stages have not been\nfully utilized. To facilitate this research, we introduce a novel Progressive\nConfident Masking Attention Network (PMCANet). It leverages attention\nmechanisms to uncover the intrinsic correlations between audio signals and\nvisual frames. Furthermore, we design an efficient and effective\ncross-attention module to enhance semantic perception by selecting query\ntokens. This selection is determined through confidence-driven units based on\nthe network's multi-stage predictive outputs. Experiments demonstrate that our\nnetwork outperforms other AVS methods while requiring less computational\nresources. The code is available at: https://github.com/PrettyPlate/PCMANet.\n", "date": "2024-06-04", "categories": []}
{"id": "2410.14945", "title": "ImmerseDiffusion: A Generative Spatial Audio Latent Diffusion Model", "abstract": "  We introduce ImmerseDiffusion, an end-to-end generative audio model that\nproduces 3D immersive soundscapes conditioned on the spatial, temporal, and\nenvironmental conditions of sound objects. ImmerseDiffusion is trained to\ngenerate first-order ambisonics (FOA) audio, which is a conventional spatial\naudio format comprising four channels that can be rendered to multichannel\nspatial output. The proposed generative system is composed of a spatial audio\ncodec that maps FOA audio to latent components, a latent diffusion model\ntrained based on various user input types, namely, text prompts, spatial,\ntemporal and environmental acoustic parameters, and optionally a spatial audio\nand text encoder trained in a Contrastive Language and Audio Pretraining (CLAP)\nstyle. We propose metrics to evaluate the quality and spatial adherence of the\ngenerated spatial audio. Finally, we assess the model performance in terms of\ngeneration quality and spatial conformance, comparing the two proposed modes:\n``descriptive\", which uses spatial text prompts) and ``parametric\", which uses\nnon-spatial text prompts and spatial parameters. Our evaluations demonstrate\npromising results that are consistent with the user conditions and reflect\nreliable spatial fidelity.\n", "date": "2024-10-19", "categories": []}
{"id": "2502.06285", "title": "End-to-End Multi-Microphone Speaker Extraction Using Relative Transfer\n  Functions", "abstract": "  This paper introduces a multi-microphone method for extracting a desired\nspeaker from a mixture involving multiple speakers and directional noise in a\nreverberant environment. In this work, we propose leveraging the instantaneous\nrelative transfer function (RTF), estimated from a reference utterance recorded\nin the same position as the desired source. The effectiveness of the RTF-based\nspatial cue is compared with direction of arrival (DOA)-based spatial cue and\nthe conventional spectral embedding. Experimental results in challenging\nacoustic scenarios demonstrate that using spatial cues yields better\nperformance than the spectral-based cue and that the instantaneous RTF\noutperforms the DOA-based spatial cue.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.06364", "title": "Automatic Identification of Samples in Hip-Hop Music via Multi-Loss\n  Training and an Artificial Dataset", "abstract": "  Sampling, the practice of reusing recorded music or sounds from another\nsource in a new work, is common in popular music genres like hip-hop and rap.\nNumerous services have emerged that allow users to identify connections between\nsamples and the songs that incorporate them, with the goal of enhancing music\ndiscovery. Designing a system that can perform the same task automatically is\nchallenging, as samples are commonly altered with audio effects like pitch- and\ntime-stretching and may only be seconds long. Progress on this task has been\nminimal and is further blocked by the limited availability of training data.\nHere, we show that a convolutional neural network trained on an artificial\ndataset can identify real-world samples in commercial hip-hop music. We extract\nvocal, harmonic, and percussive elements from several databases of\nnon-commercial music recordings using audio source separation, and train the\nmodel to fingerprint a subset of these elements in transformed versions of the\noriginal audio. We optimize the model using a joint classification and metric\nlearning loss and show that it achieves 13% greater precision on real-world\ninstances of sampling than a fingerprinting system using acoustic landmarks,\nand that it can recognize samples that have been both pitch shifted and time\nstretched. We also show that, for half of the commercial music recordings we\ntested, our model is capable of locating the position of a sample to within\nfive seconds.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.04476", "title": "ADIFF: Explaining audio difference using natural language", "abstract": "  Understanding and explaining differences between audio recordings is crucial\nfor fields like audio forensics, quality assessment, and audio generation. This\ninvolves identifying and describing audio events, acoustic scenes, signal\ncharacteristics, and their emotional impact on listeners. This paper stands out\nas the first work to comprehensively study the task of explaining audio\ndifferences and then propose benchmark, baselines for the task. First, we\npresent two new datasets for audio difference explanation derived from the\nAudioCaps and Clotho audio captioning datasets. Using Large Language Models\n(LLMs), we generate three levels of difference explanations: (1) concise\ndescriptions of audio events and objects, (2) brief sentences about audio\nevents, acoustic scenes, and signal properties, and (3) comprehensive\nexplanations that include semantics and listener emotions. For the baseline, we\nuse prefix tuning where audio embeddings from two audio files are used to\nprompt a frozen language model. Our empirical analysis and ablation studies\nreveal that the naive baseline struggles to distinguish perceptually similar\nsounds and generate detailed tier 3 explanations. To address these limitations,\nwe propose ADIFF, which introduces a cross-projection module, position\ncaptioning, and a three-step training process to enhance the model's ability to\nproduce detailed explanations. We evaluate our model using objective metrics\nand human evaluation and show our model enhancements lead to significant\nimprovements in performance over naive baseline and SoTA Audio-Language Model\n(ALM) Qwen Audio. Lastly, we conduct multiple ablation studies to study the\neffects of cross-projection, language model parameters, position captioning,\nthird stage fine-tuning, and present our findings. Our benchmarks, findings,\nand strong baseline pave the way for nuanced and human-like explanations of\naudio differences.\n", "date": "2025-02-06", "categories": []}
{"id": "2502.05139", "title": "Meta Audiobox Aesthetics: Unified Automatic Quality Assessment for\n  Speech, Music, and Sound", "abstract": "  The quantification of audio aesthetics remains a complex challenge in audio\nprocessing, primarily due to its subjective nature, which is influenced by\nhuman perception and cultural context. Traditional methods often depend on\nhuman listeners for evaluation, leading to inconsistencies and high resource\ndemands. This paper addresses the growing need for automated systems capable of\npredicting audio aesthetics without human intervention. Such systems are\ncrucial for applications like data filtering, pseudo-labeling large datasets,\nand evaluating generative audio models, especially as these models become more\nsophisticated. In this work, we introduce a novel approach to audio aesthetic\nevaluation by proposing new annotation guidelines that decompose human\nlistening perspectives into four distinct axes. We develop and train\nno-reference, per-item prediction models that offer a more nuanced assessment\nof audio quality. Our models are evaluated against human mean opinion scores\n(MOS) and existing methods, demonstrating comparable or superior performance.\nThis research not only advances the field of audio aesthetics but also provides\nopen-source models and datasets to facilitate future work and benchmarking. We\nrelease our code and pre-trained model at:\nhttps://github.com/facebookresearch/audiobox-aesthetics\n", "date": "2025-02-07", "categories": []}
{"id": "2405.11831", "title": "SSAMBA: Self-Supervised Audio Representation Learning with Mamba State\n  Space Model", "abstract": "  Transformers have revolutionized deep learning across various tasks,\nincluding audio representation learning, due to their powerful modeling\ncapabilities. However, they often suffer from quadratic complexity in both GPU\nmemory usage and computational inference time, affecting their efficiency.\nRecently, state space models (SSMs) like Mamba have emerged as a promising\nalternative, offering a more efficient approach by avoiding these complexities.\nGiven these advantages, we explore the potential of SSM-based models in audio\ntasks. In this paper, we introduce Self-Supervised Audio Mamba (SSAMBA), the\nfirst self-supervised, attention-free, and SSM-based model for audio\nrepresentation learning. SSAMBA leverages the bidirectional Mamba to capture\ncomplex audio patterns effectively. We incorporate a self-supervised\npretraining framework that optimizes both discriminative and generative\nobjectives, enabling the model to learn robust audio representations from\nlarge-scale, unlabeled datasets. We evaluated SSAMBA on various tasks such as\naudio classification, keyword spotting, and speaker identification. Our results\ndemonstrate that SSAMBA outperforms the Self-Supervised Audio Spectrogram\nTransformer (SSAST) in most tasks. Notably, SSAMBA is approximately 92.7%\nfaster in batch inference speed and 95.4% more memory-efficient than SSAST for\nthe tiny model size with an input token size of 22k. These efficiency gains,\ncombined with superior performance, underscore the effectiveness of SSAMBA's\narchitectural innovation, making it a compelling choice for a wide range of\naudio processing applications.\n", "date": "2024-05-20", "categories": []}
{"id": "2409.19689", "title": "InfantCryNet: A Data-driven Framework for Intelligent Analysis of Infant\n  Cries", "abstract": "  Understanding the meaning of infant cries is a significant challenge for\nyoung parents in caring for their newborns. The presence of background noise\nand the lack of labeled data present practical challenges in developing systems\nthat can detect crying and analyze its underlying reasons. In this paper, we\npresent a novel data-driven framework, \"InfantCryNet,\" for accomplishing these\ntasks. To address the issue of data scarcity, we employ pre-trained audio\nmodels to incorporate prior knowledge into our model. We propose the use of\nstatistical pooling and multi-head attention pooling techniques to extract\nfeatures more effectively. Additionally, knowledge distillation and model\nquantization are applied to enhance model efficiency and reduce the model size,\nbetter supporting industrial deployment in mobile devices. Experiments on\nreal-life datasets demonstrate the superior performance of the proposed\nframework, outperforming state-of-the-art baselines by 4.4% in classification\naccuracy. The model compression effectively reduces the model size by 7%\nwithout compromising performance and by up to 28% with only an 8% decrease in\naccuracy, offering practical insights for model selection and system design.\n", "date": "2024-09-29", "categories": []}
{"id": "2502.01709", "title": "Adapter-Based Multi-Agent AVSR Extension for Pre-Trained ASR Models", "abstract": "  We present an approach to Audio-Visual Speech Recognition that builds on a\npre-trained Whisper model. To infuse visual information into this audio-only\nmodel, we extend it with an AV fusion module and LoRa adapters, one of the most\nup-to-date adapter approaches. One advantage of adapter-based approaches, is\nthat only a relatively small number of parameters are trained, while the basic\nmodel remains unchanged. Common AVSR approaches train single models to handle\nseveral noise categories and noise levels simultaneously. Taking advantage of\nthe lightweight nature of adapter approaches, we train noise-scenario-specific\nadapter-sets, each covering individual noise-categories or a specific\nnoise-level range. The most suitable adapter-set is selected by previously\nclassifying the noise-scenario. This enables our models to achieve an optimum\ncoverage across different noise-categories and noise-levels, while training\nonly a minimum number of parameters.\n  Compared to a full fine-tuning approach with SOTA performance our models\nachieve almost comparable results over the majority of the tested\nnoise-categories and noise-levels, with up to 88.5% less trainable parameters.\nOur approach can be extended by further noise-specific adapter-sets to cover\nadditional noise scenarios. It is also possible to utilize the underlying\npowerful ASR model when no visual information is available, as it remains\nunchanged.\n", "date": "2025-02-03", "categories": []}
{"id": "2502.02118", "title": "BRIDLE: Generalized Self-supervised Learning with Quantization", "abstract": "  Self-supervised learning has been a powerful approach for learning meaningful\nrepresentations from unlabeled data across various domains, reducing the\nreliance on large labeled datasets. Inspired by BERT's success in capturing\ndeep bidirectional contexts in natural language processing, similar frameworks\nhave been adapted to other modalities such as audio, with models like BEATs\nextending the bidirectional training paradigm to audio signals using vector\nquantization (VQ). However, these frameworks face challenges, notably their\ndependence on a single codebook for quantization, which may not capture the\ncomplex, multifaceted nature of signals. In addition, inefficiencies in\ncodebook utilization lead to underutilized code vectors. To address these\nlimitations, we introduce BRIDLE (Bidirectional Residual Quantization\nInterleaved Discrete Learning Encoder), a self-supervised encoder pretraining\nframework that incorporates residual quantization (RQ) into the bidirectional\ntraining process, and is generalized for pretraining with audio, image, and\nvideo. Using multiple hierarchical codebooks, RQ enables fine-grained\ndiscretization in the latent space, enhancing representation quality. BRIDLE\ninvolves an interleaved training procedure between the encoder and tokenizer.\nWe evaluate BRIDLE on audio understanding tasks using classification\nbenchmarks, achieving state-of-the-art results, and demonstrate competitive\nperformance on image classification and video classification tasks, showing\nconsistent improvements over traditional VQ methods in downstream performance.\n", "date": "2025-02-04", "categories": []}
{"id": "2409.06513", "title": "Sines, Transient, Noise Neural Modeling of Piano Notes", "abstract": "  This paper introduces a novel method for emulating piano sounds. We propose\nto exploit the sines, transient, and noise decomposition to design a\ndifferentiable spectral modeling synthesizer replicating piano notes. Three\nsub-modules learn these components from piano recordings and generate the\ncorresponding harmonic, transient, and noise signals. Splitting the emulation\ninto three independently trainable models reduces the modeling tasks'\ncomplexity. The quasi-harmonic content is produced using a differentiable\nsinusoidal model guided by physics-derived formulas, whose parameters are\nautomatically estimated from audio recordings. The noise sub-module uses a\nlearnable time-varying filter, and the transients are generated using a deep\nconvolutional network. From singular notes, we emulate the coupling between\ndifferent keys in trichords with a convolutional-based network. Results show\nthe model matches the partial distribution of the target while predicting the\nenergy in the higher part of the spectrum presents more challenges. The energy\ndistribution in the spectra of the transient and noise components is accurate\noverall. While the model is more computationally and memory efficient,\nperceptual tests reveal limitations in accurately modeling the attack phase of\nnotes. Despite this, it generally achieves perceptual accuracy in emulating\nsingle notes and trichords.\n", "date": "2024-09-10", "categories": []}
{"id": "2501.19010", "title": "DyPCL: Dynamic Phoneme-level Contrastive Learning for Dysarthric Speech\n  Recognition", "abstract": "  Dysarthric speech recognition often suffers from performance degradation due\nto the intrinsic diversity of dysarthric severity and extrinsic disparity from\nnormal speech. To bridge these gaps, we propose a Dynamic Phoneme-level\nContrastive Learning (DyPCL) method, which leads to obtaining invariant\nrepresentations across diverse speakers. We decompose the speech utterance into\nphoneme segments for phoneme-level contrastive learning, leveraging dynamic\nconnectionist temporal classification alignment. Unlike prior studies focusing\non utterance-level embeddings, our granular learning allows discrimination of\nsubtle parts of speech. In addition, we introduce dynamic curriculum learning,\nwhich progressively transitions from easy negative samples to\ndifficult-to-distinguishable negative samples based on phonetic similarity of\nphoneme. Our approach to training by difficulty levels alleviates the inherent\nvariability of speakers, better identifying challenging speeches. Evaluated on\nthe UASpeech dataset, DyPCL outperforms baseline models, achieving an average\n22.10\\% relative reduction in word error rate (WER) across the overall\ndysarthria group.\n", "date": "2025-01-31", "categories": []}
{"id": "2501.18919", "title": "Deepfake Detection of Singing Voices With Whisper Encodings", "abstract": "  The deepfake generation of singing vocals is a concerning issue for artists\nin the music industry. In this work, we propose a singing voice deepfake\ndetection (SVDD) system, which uses noise-variant encodings of open-AI's\nWhisper model. As counter-intuitive as it may sound, even though the Whisper\nmodel is known to be noise-robust, the encodings are rich in non-speech\ninformation, and are noise-variant. This leads us to evaluate Whisper encodings\nas feature representations for the SVDD task. Therefore, in this work, the SVDD\ntask is performed on vocals and mixtures, and the performance is evaluated in\n\\%EER over varying Whisper model sizes and two classifiers- CNN and ResNet34,\nunder different testing conditions.\n", "date": "2025-01-31", "categories": []}
{"id": "2501.03464", "title": "LHGNN: Local-Higher Order Graph Neural Networks For Audio Classification\n  and Tagging", "abstract": "  Transformers have set new benchmarks in audio processing tasks, leveraging\nself-attention mechanisms to capture complex patterns and dependencies within\naudio data. However, their focus on pairwise interactions limits their ability\nto process the higher-order relations essential for identifying distinct audio\nobjects. To address this limitation, this work introduces the Local- Higher\nOrder Graph Neural Network (LHGNN), a graph based model that enhances feature\nunderstanding by integrating local neighbourhood information with higher-order\ndata from Fuzzy C-Means clusters, thereby capturing a broader spectrum of audio\nrelationships. Evaluation of the model on three publicly available audio\ndatasets shows that it outperforms Transformer-based models across all\nbenchmarks while operating with substantially fewer parameters. Moreover, LHGNN\ndemonstrates a distinct advantage in scenarios lacking ImageNet pretraining,\nestablishing its effectiveness and efficiency in environments where extensive\npretraining data is unavailable.\n", "date": "2025-01-07", "categories": []}
{"id": "2501.17578", "title": "Music2Latent2: Audio Compression with Summary Embeddings and\n  Autoregressive Decoding", "abstract": "  Efficiently compressing high-dimensional audio signals into a compact and\ninformative latent space is crucial for various tasks, including generative\nmodeling and music information retrieval (MIR). Existing audio autoencoders,\nhowever, often struggle to achieve high compression ratios while preserving\naudio fidelity and facilitating efficient downstream applications. We introduce\nMusic2Latent2, a novel audio autoencoder that addresses these limitations by\nleveraging consistency models and a novel approach to representation learning\nbased on unordered latent embeddings, which we call summary embeddings. Unlike\nconventional methods that encode local audio features into ordered sequences,\nMusic2Latent2 compresses audio signals into sets of summary embeddings, where\neach embedding can capture distinct global features of the input sample. This\nenables to achieve higher reconstruction quality at the same compression ratio.\nTo handle arbitrary audio lengths, Music2Latent2 employs an autoregressive\nconsistency model trained on two consecutive audio chunks with causal masking,\nensuring coherent reconstruction across segment boundaries. Additionally, we\npropose a novel two-step decoding procedure that leverages the denoising\ncapabilities of consistency models to further refine the generated audio at no\nadditional cost. Our experiments demonstrate that Music2Latent2 outperforms\nexisting continuous audio autoencoders regarding audio quality and performance\non downstream tasks. Music2Latent2 paves the way for new possibilities in audio\ncompression.\n", "date": "2025-01-29", "categories": []}
{"id": "2501.17612", "title": "VoicePrompter: Robust Zero-Shot Voice Conversion with Voice Prompt and\n  Conditional Flow Matching", "abstract": "  Despite remarkable advancements in recent voice conversion (VC) systems,\nenhancing speaker similarity in zero-shot scenarios remains challenging. This\nchallenge arises from the difficulty of generalizing and adapting speaker\ncharacteristics in speech within zero-shot environments, which is further\ncomplicated by mismatch between the training and inference processes. To\naddress these challenges, we propose VoicePrompter, a robust zero-shot VC model\nthat leverages in-context learning with voice prompts. VoicePrompter is\ncomposed of (1) a factorization method that disentangles speech components and\n(2) a DiT-based conditional flow matching (CFM) decoder that conditions on\nthese factorized features and voice prompts. Additionally, (3) latent mixup is\nused to enhance in-context learning by combining various speaker features. This\napproach improves speaker similarity and naturalness in zero-shot VC by\napplying mixup to latent representations. Experimental results demonstrate that\nVoicePrompter outperforms existing zero-shot VC systems in terms of speaker\nsimilarity, speech intelligibility, and audio quality. Our demo is available at\n\\url{https://hayeong0.github.io/VoicePrompter-demo/}.\n", "date": "2025-01-29", "categories": []}
{"id": "2501.14994", "title": "Robust Cross-Etiology and Speaker-Independent Dysarthric Speech\n  Recognition", "abstract": "  In this paper, we present a speaker-independent dysarthric speech recognition\nsystem, with a focus on evaluating the recently released Speech Accessibility\nProject (SAP-1005) dataset, which includes speech data from individuals with\nParkinson's disease (PD). Despite the growing body of research in dysarthric\nspeech recognition, many existing systems are speaker-dependent and adaptive,\nlimiting their generalizability across different speakers and etiologies. Our\nprimary objective is to develop a robust speaker-independent model capable of\naccurately recognizing dysarthric speech, irrespective of the speaker.\nAdditionally, as a secondary objective, we aim to test the cross-etiology\nperformance of our model by evaluating it on the TORGO dataset, which contains\nspeech samples from individuals with cerebral palsy (CP) and amyotrophic\nlateral sclerosis (ALS). By leveraging the Whisper model, our\nspeaker-independent system achieved a CER of 6.99% and a WER of 10.71% on the\nSAP-1005 dataset. Further, in cross-etiology settings, we achieved a CER of\n25.08% and a WER of 39.56% on the TORGO dataset. These results highlight the\npotential of our approach to generalize across unseen speakers and different\netiologies of dysarthria.\n", "date": "2025-01-25", "categories": []}
{"id": "2501.15613", "title": "Stepback: Enhanced Disentanglement for Voice Conversion via Multi-Task\n  Learning", "abstract": "  Voice conversion (VC) modifies voice characteristics while preserving\nlinguistic content. This paper presents the Stepback network, a novel model for\nconverting speaker identity using non-parallel data. Unlike traditional VC\nmethods that rely on parallel data, our approach leverages deep learning\ntechniques to enhance disentanglement completion and linguistic content\npreservation. The Stepback network incorporates a dual flow of different domain\ndata inputs and uses constraints with self-destructive amendments to optimize\nthe content encoder. Extensive experiments show that our model significantly\nimproves VC performance, reducing training costs while achieving high-quality\nvoice conversion. The Stepback network's design offers a promising solution for\nadvanced voice conversion tasks.\n", "date": "2025-01-26", "categories": []}
{"id": "2501.15900", "title": "Investigating the Sensitivity of Pre-trained Audio Embeddings to Common\n  Effects", "abstract": "  In recent years, foundation models have significantly advanced data-driven\nsystems across various domains. Yet, their underlying properties, especially\nwhen functioning as feature extractors, remain under-explored. In this paper,\nwe investigate the sensitivity to audio effects of audio embeddings extracted\nfrom widely-used foundation models, including OpenL3, PANNs, and CLAP. We focus\non audio effects as the source of sensitivity due to their prevalent presence\nin large audio datasets. By applying parameterized audio effects (gain,\nlow-pass filtering, reverberation, and bitcrushing), we analyze the correlation\nbetween the deformation trajectories and the effect strength in the embedding\nspace. We propose to quantify the dimensionality and linearizability of the\ndeformation trajectories induced by audio effects using canonical correlation\nanalysis. We find that there exists a direction along which the embeddings move\nmonotonically as the audio effect strength increases, but that the subspace\ncontaining the displacements is generally high-dimensional. This shows that\npre-trained audio embeddings do not globally linearize the effects. Our\nempirical results on instrument classification downstream tasks confirm that\nprojecting out the estimated deformation directions cannot generally improve\nthe robustness of pre-trained embeddings to audio effects.\n", "date": "2025-01-27", "categories": []}
{"id": "2501.16171", "title": "Separate This, and All of these Things Around It: Music Source\n  Separation via Hyperellipsoidal Queries", "abstract": "  Music source separation is an audio-to-audio retrieval task of extracting one\nor more constituent components, or composites thereof, from a musical audio\nmixture. Each of these constituent components is often referred to as a \"stem\"\nin literature. Historically, music source separation has been dominated by a\nstem-based paradigm, leading to most state-of-the-art systems being either a\ncollection of single-stem extraction models, or a tightly coupled system with a\nfixed, difficult-to-modify, set of supported stems. Combined with the limited\ndata availability, advances in music source separation have thus been mostly\nlimited to the \"VDBO\" set of stems: \\textit{vocals}, \\textit{drum},\n\\textit{bass}, and the catch-all \\textit{others}. Recent work in music source\nseparation has begun to challenge the fixed-stem paradigm, moving towards\nmodels able to extract any musical sound as long as this target type of sound\ncould be specified to the model as an additional query input. We generalize\nthis idea to a \\textit{query-by-region} source separation system, specifying\nthe target based on the query regardless of how many sound sources or which\nsound classes are contained within it. To do so, we propose the use of\nhyperellipsoidal regions as queries to allow for an intuitive yet easily\nparametrizable approach to specifying both the target (location) as well as its\nspread. Evaluation of the proposed system on the MoisesDB dataset demonstrated\nstate-of-the-art performance of the proposed system both in terms of\nsignal-to-noise ratios and retrieval metrics.\n", "date": "2025-01-27", "categories": []}
{"id": "2501.14610", "title": "Leveraging Spatial Cues from Cochlear Implant Microphones to Efficiently\n  Enhance Speech Separation in Real-World Listening Scenes", "abstract": "  Speech separation approaches for single-channel, dry speech mixtures have\nsignificantly improved. However, real-world spatial and reverberant acoustic\nenvironments remain challenging, limiting the effectiveness of these approaches\nfor assistive hearing devices like cochlear implants (CIs). To address this, we\nquantify the impact of real-world acoustic scenes on speech separation and\nexplore how spatial cues can enhance separation quality efficiently. We analyze\nperformance based on implicit spatial cues (inherent in the acoustic input and\nlearned by the model) and explicit spatial cues (manually calculated spatial\nfeatures added as auxiliary inputs). Our findings show that spatial cues (both\nimplicit and explicit) improve separation for mixtures with spatially separated\nand nearby talkers. Furthermore, spatial cues enhance separation when spectral\ncues are ambiguous, such as when voices are similar. Explicit spatial cues are\nparticularly beneficial when implicit spatial cues are weak. For instance,\nsingle CI microphone recordings provide weaker implicit spatial cues than\nbilateral CIs, but even single CIs benefit from explicit cues. These results\nemphasize the importance of training models on real-world data to improve\ngeneralizability in everyday listening scenarios. Additionally, our statistical\nanalyses offer insights into how data properties influence model performance,\nsupporting the development of efficient speech separation approaches for CIs\nand other assistive devices in real-world settings.\n", "date": "2025-01-24", "categories": []}
{"id": "2501.13884", "title": "Exploring Finetuned Audio-LLM on Heart Murmur Features", "abstract": "  Large language models (LLMs) for audio have excelled in recognizing and\nanalyzing human speech, music, and environmental sounds. However, their\npotential for understanding other types of sounds, particularly biomedical\nsounds, remains largely underexplored despite significant scientific interest.\nIn this study, we focus on diagnosing cardiovascular diseases using\nphonocardiograms, i.e., heart sounds. Most existing deep neural network (DNN)\nparadigms are restricted to heart murmur classification (healthy vs unhealthy)\nand do not predict other acoustic features of the murmur such as timing,\ngrading, harshness, pitch, and quality, which are important in helping\nphysicians diagnose the underlying heart conditions. We propose to finetune an\naudio LLM, Qwen2-Audio, on the PhysioNet CirCor DigiScope phonocardiogram (PCG)\ndataset and evaluate its performance in classifying 11 expert-labeled murmur\nfeatures. Additionally, we aim to achieve more noise-robust and generalizable\nsystem by exploring a preprocessing segmentation algorithm using an audio\nrepresentation model, SSAMBA. Our results indicate that the LLM-based model\noutperforms state-of-the-art methods in 8 of the 11 features and performs\ncomparably in the remaining 3. Moreover, the LLM successfully classifies\nlong-tail murmur features with limited training data, a task that all previous\nmethods have failed to classify. These findings underscore the potential of\naudio LLMs as assistants to human cardiologists in enhancing heart disease\ndiagnosis.\n", "date": "2025-01-23", "categories": []}
{"id": "2501.10666", "title": "Speech Emotion Detection Based on MFCC and CNN-LSTM Architecture", "abstract": "  Emotion detection techniques have been applied to multiple cases mainly from\nfacial image features and vocal audio features, of which the latter aspect is\ndisputed yet not only due to the complexity of speech audio processing but also\nthe difficulties of extracting appropriate features. Part of the SAVEE and\nRAVDESS datasets are selected and combined as the dataset, containing seven\nsorts of common emotions (i.e. happy, neutral, sad, anger, disgust, fear, and\nsurprise) and thousands of samples. Based on the Librosa package, this paper\nprocesses the initial audio input into waveplot and spectrum for analysis and\nconcentrates on multiple features including MFCC as targets for feature\nextraction. The hybrid CNN-LSTM architecture is adopted by virtue of its strong\ncapability to deal with sequential data and time series, which mainly consists\nof four convolutional layers and three long short-term memory layers. As a\nresult, the architecture achieved an accuracy of 61.07% comprehensively for the\ntest set, among which the detection of anger and neutral reaches a performance\nof 75.31% and 71.70% respectively. It can also be concluded that the\nclassification accuracy is dependent on the properties of emotion to some\nextent, with frequently-used and distinct-featured emotions having less\nprobability to be misclassified into other categories. Emotions like surprise\nwhose meaning depends on the specific context are more likely to confuse with\npositive or negative emotions, and negative emotions also have a possibility to\nget mixed with each other.\n", "date": "2025-01-18", "categories": []}
{"id": "2501.10755", "title": "An Experimental Study on Joint Modeling for Sound Event Localization and\n  Detection with Source Distance Estimation", "abstract": "  In traditional sound event localization and detection (SELD) tasks, the focus\nis typically on sound event detection (SED) and direction-of-arrival (DOA)\nestimation, but they fall short of providing full spatial information about the\nsound source. The 3D SELD task addresses this limitation by integrating source\ndistance estimation (SDE), allowing for complete spatial localization. We\npropose three approaches to tackle this challenge: a novel method with\nindependent training and joint prediction, which firstly treats DOA and\ndistance estimation as separate tasks and then combines them to solve 3D SELD;\na dual-branch representation with source Cartesian coordinate used for\nsimultaneous DOA and distance estimation; and a three-branch structure that\njointly models SED, DOA, and SDE within a unified framework. Our proposed\nmethod ranked first in the DCASE 2024 Challenge Task 3, demonstrating the\neffectiveness of joint modeling for addressing the 3D SELD task. The relevant\ncode for this paper will be open-sourced in the future.\n", "date": "2025-01-18", "categories": []}
{"id": "2501.12385", "title": "Audio Texture Manipulation by Exemplar-Based Analogy", "abstract": "  Audio texture manipulation involves modifying the perceptual characteristics\nof a sound to achieve specific transformations, such as adding, removing, or\nreplacing auditory elements. In this paper, we propose an exemplar-based\nanalogy model for audio texture manipulation. Instead of conditioning on\ntext-based instructions, our method uses paired speech examples, where one clip\nrepresents the original sound and another illustrates the desired\ntransformation. The model learns to apply the same transformation to new input,\nallowing for the manipulation of sound textures. We construct a quadruplet\ndataset representing various editing tasks, and train a latent diffusion model\nin a self-supervised manner. We show through quantitative evaluations and\nperceptual studies that our model outperforms text-conditioned baselines and\ngeneralizes to real-world, out-of-distribution, and non-speech scenarios.\nProject page: https://berkeley-speech-group.github.io/audio-texture-analogy/\n", "date": "2025-01-21", "categories": []}
{"id": "2501.09877", "title": "CLAP-S: Support Set Based Adaptation for Downstream Fiber-optic Acoustic\n  Recognition", "abstract": "  Contrastive Language-Audio Pretraining (CLAP) models have demonstrated\nunprecedented performance in various acoustic signal recognition tasks.\nFiber-optic-based acoustic recognition is one of the most important downstream\ntasks and plays a significant role in environmental sensing. Adapting CLAP for\nfiber-optic acoustic recognition has become an active research area. As a\nnon-conventional acoustic sensor, fiber-optic acoustic recognition presents a\nchallenging, domain-specific, low-shot deployment environment with significant\ndomain shifts due to unique frequency response and noise characteristics. To\naddress these challenges, we propose a support-based adaptation method, CLAP-S,\nwhich linearly interpolates a CLAP Adapter with the Support Set, leveraging\nboth implicit knowledge through fine-tuning and explicit knowledge retrieved\nfrom memory for cross-domain generalization. Experimental results show that our\nmethod delivers competitive performance on both laboratory-recorded fiber-optic\nESC-50 datasets and a real-world fiber-optic gunshot-firework dataset. Our\nresearch also provides valuable insights for other downstream acoustic\nrecognition tasks. The code and gunshot-firework dataset are available at\nhttps://github.com/Jingchensun/clap-s.\n", "date": "2025-01-16", "categories": []}
{"id": "2408.12549", "title": "Modeling Time-Variant Responses of Optical Compressors with Selective\n  State Space Models", "abstract": "  This paper presents a method for modeling optical dynamic range compressors\nusing deep neural networks with Selective State Space models. The proposed\napproach surpasses previous methods based on recurrent layers by employing a\nSelective State Space block to encode the input audio. It features a refined\ntechnique integrating Feature-wise Linear Modulation and Gated Linear Units to\nadjust the network dynamically, conditioning the compression's attack and\nrelease phases according to external parameters. The proposed architecture is\nwell-suited for low-latency and real-time applications, crucial in live audio\nprocessing. The method has been validated on the analog optical compressors\nTubeTech CL 1B and Teletronix LA-2A, which possess distinct characteristics.\nEvaluation is performed using quantitative metrics and subjective listening\ntests, comparing the proposed method with other state-of-the-art models.\nResults show that our black-box modeling methods outperform all others,\nachieving accurate emulation of the compression process for both seen and\nunseen settings during training. We further show a correlation between this\naccuracy and the sampling density of the control parameters in the dataset and\nidentify settings with fast attack and slow release as the most challenging to\nemulate.\n", "date": "2024-08-22", "categories": []}
{"id": "2501.09608", "title": "Metric Learning with Progressive Self-Distillation for Audio-Visual\n  Embedding Learning", "abstract": "  Metric learning projects samples into an embedded space, where similarities\nand dissimilarities are quantified based on their learned representations.\nHowever, existing methods often rely on label-guided representation learning,\nwhere representations of different modalities, such as audio and visual data,\nare aligned based on annotated labels. This approach tends to underutilize\nlatent complex features and potential relationships inherent in the\ndistributions of audio and visual data that are not directly tied to the\nlabels, resulting in suboptimal performance in audio-visual embedding learning.\nTo address this issue, we propose a novel architecture that integrates\ncross-modal triplet loss with progressive self-distillation. Our method\nenhances representation learning by leveraging inherent distributions and\ndynamically refining soft audio-visual alignments -- probabilistic alignments\nbetween audio and visual data that capture the inherent relationships beyond\nexplicit labels. Specifically, the model distills audio-visual\ndistribution-based knowledge from annotated labels in a subset of each batch.\nThis self-distilled knowledge is used t\n", "date": "2025-01-16", "categories": []}
{"id": "2410.05301", "title": "Diffusion-based Unsupervised Audio-visual Speech Enhancement", "abstract": "  This paper proposes a new unsupervised audio-visual speech enhancement (AVSE)\napproach that combines a diffusion-based audio-visual speech generative model\nwith a non-negative matrix factorization (NMF) noise model. First, the\ndiffusion model is pre-trained on clean speech conditioned on corresponding\nvideo data to simulate the speech generative distribution. This pre-trained\nmodel is then paired with the NMF-based noise model to estimate clean speech\niteratively. Specifically, a diffusion-based posterior sampling approach is\nimplemented within the reverse diffusion process, where after each iteration, a\nspeech estimate is obtained and used to update the noise parameters.\nExperimental results confirm that the proposed AVSE approach not only\noutperforms its audio-only counterpart but also generalizes better than a\nrecent supervised-generative AVSE method. Additionally, the new inference\nalgorithm offers a better balance between inference speed and performance\ncompared to the previous diffusion-based method. Code and demo available at:\nhttps://jeaneudesayilo.github.io/fast_UdiffSE\n", "date": "2024-10-04", "categories": []}
{"id": "2501.08587", "title": "Sound Scene Synthesis at the DCASE 2024 Challenge", "abstract": "  This paper presents Task 7 at the DCASE 2024 Challenge: sound scene\nsynthesis. Recent advances in sound synthesis and generative models have\nenabled the creation of realistic and diverse audio content. We introduce a\nstandardized evaluation framework for comparing different sound scene synthesis\nsystems, incorporating both objective and subjective metrics. The challenge\nattracted four submissions, which are evaluated using the Fr\\'echet Audio\nDistance (FAD) and human perceptual ratings. Our analysis reveals significant\ninsights into the current capabilities and limitations of sound scene synthesis\nsystems, while also highlighting areas for future improvement in this rapidly\nevolving field.\n", "date": "2025-01-15", "categories": []}
{"id": "2410.03335", "title": "Audio-Agent: Leveraging LLMs For Audio Generation, Editing and\n  Composition", "abstract": "  We introduce Audio-Agent, a multimodal framework for audio generation,\nediting and composition based on text or video inputs. Conventional approaches\nfor text-to-audio (TTA) tasks often make single-pass inferences from text\ndescriptions. While straightforward, this design struggles to produce\nhigh-quality audio when given complex text conditions. In our method, we\nutilize a pre-trained TTA diffusion network as the audio generation agent to\nwork in tandem with GPT-4, which decomposes the text condition into atomic,\nspecific instructions and calls the agent for audio generation. In doing so,\nAudio-Agent can generate high-quality audio that is closely aligned with the\nprovided text or video exhibiting complex and multiple events, while supporting\nvariable-length and variable-volume generation. For video-to-audio (VTA) tasks,\nmost existing methods require training a timestamp detector to synchronize\nvideo events with the generated audio, a process that can be tedious and\ntime-consuming. Instead, we propose a simpler approach by fine-tuning a\npre-trained Large Language Model (LLM), e.g., Gemma2-2B-it, to obtain both\nsemantic and temporal conditions that bridge the video and audio modality.\nConsequently, our framework contributes a comprehensive solution for both TTA\nand VTA tasks without substantial computational overhead in training.\n", "date": "2024-10-04", "categories": []}
{"id": "2501.08047", "title": "Gen-A: Generalizing Ambisonics Neural Encoding to Unseen Microphone\n  Arrays", "abstract": "  Using deep neural networks (DNNs) for encoding of microphone array (MA)\nsignals to the Ambisonics spatial audio format can surpass certain limitations\nof established conventional methods, but existing DNN-based methods need to be\ntrained separately for each MA. This paper proposes a DNN-based method for\nAmbisonics encoding that can generalize to arbitrary MA geometries unseen\nduring training. The method takes as inputs the MA geometry and MA signals and\nuses a multi-level encoder consisting of separate paths for geometry and signal\ndata, where geometry features inform the signal encoder at each level. The\nmethod is validated in simulated anechoic and reverberant conditions with one\nand two sources. The results indicate improvement over conventional encoding\nacross the whole frequency range for dry scenes, while for reverberant scenes\nthe improvement is frequency-dependent.\n", "date": "2025-01-14", "categories": []}
{"id": "2409.14526", "title": "What Are They Doing? Joint Audio-Speech Co-Reasoning", "abstract": "  In audio and speech processing, tasks usually focus on either the audio or\nspeech modality, even when both sounds and human speech are present in the same\naudio clip. Recent Auditory Large Language Models (ALLMs) have made it possible\nto process audio and speech simultaneously within a single model, leading to\nfurther considerations of joint audio-speech tasks.\n  In this paper, we establish a novel benchmark to investigate how well ALLMs\ncan perform joint audio-speech processing. Specifically, we introduce Joint\nAudio-Speech Co-Reasoning (JASCO), a novel task that unifies audio and speech\nprocessing, strictly requiring co-reasoning across both modalities. We also\nrelease a scene-reasoning dataset called \"What Are They Doing\". Additionally,\nwe provide deeper insights into the models' behaviors by analyzing their\ndependence on each modality.\n", "date": "2024-09-22", "categories": []}
{"id": "2411.03172", "title": "Blind Estimation of Sub-band Acoustic Parameters from Ambisonics\n  Recordings using Spectro-Spatial Covariance Features", "abstract": "  Estimating frequency-varying acoustic parameters is essential for enhancing\nimmersive perception in realistic spatial audio creation. In this paper, we\npropose a unified framework that blindly estimates reverberation time (T60),\ndirect-to-reverberant ratio (DRR), and clarity (C50) across 10 frequency bands\nusing first-order Ambisonics (FOA) speech recordings as inputs. The proposed\nframework utilizes a novel feature named Spectro-Spatial Covariance Vector\n(SSCV), efficiently representing temporal, spectral as well as spatial\ninformation of the FOA signal. Our models significantly outperform existing\nsingle-channel methods with only spectral information, reducing estimation\nerrors by more than half for all three acoustic parameters. Additionally, we\nintroduce FOA-Conv3D, a novel back-end network for effectively utilising the\nSSCV feature with a 3D convolutional encoder. FOA-Conv3D outperforms the\nconvolutional neural network (CNN) and recurrent convolutional neural network\n(CRNN) backends, achieving lower estimation errors and accounting for a higher\nproportion of variance (PoV) for all 3 acoustic parameters.\n", "date": "2024-11-05", "categories": []}
{"id": "2501.05068", "title": "D3RM: A Discrete Denoising Diffusion Refinement Model for Piano\n  Transcription", "abstract": "  Diffusion models have been widely used in the generative domain due to their\nconvincing performance in modeling complex data distributions. Moreover, they\nhave shown competitive results on discriminative tasks, such as image\nsegmentation. While diffusion models have also been explored for automatic\nmusic transcription, their performance has yet to reach a competitive level. In\nthis paper, we focus on discrete diffusion model's refinement capabilities and\npresent a novel architecture for piano transcription. Our model utilizes\nNeighborhood Attention layers as the denoising module, gradually predicting the\ntarget high-resolution piano roll, conditioned on the finetuned features of a\npretrained acoustic model. To further enhance refinement, we devise a novel\nstrategy which applies distinct transition states during training and inference\nstage of discrete diffusion models. Experiments on the MAESTRO dataset show\nthat our approach outperforms previous diffusion-based piano transcription\nmodels and the baseline model in terms of F1 score. Our code is available in\nhttps://github.com/hanshounsu/d3rm.\n", "date": "2025-01-09", "categories": []}
{"id": "2501.06478", "title": "Speech Recognition for Automatically Assessing Afrikaans and isiXhosa\n  Preschool Oral Narratives", "abstract": "  We develop automatic speech recognition (ASR) systems for stories told by\nAfrikaans and isiXhosa preschool children. Oral narratives provide a way to\nassess children's language development before they learn to read. We consider a\nrange of prior child-speech ASR strategies to determine which is best suited to\nthis unique setting. Using Whisper and only 5 minutes of transcribed in-domain\nchild speech, we find that additional in-domain adult data (adult speech\nmatching the story domain) provides the biggest improvement, especially when\ncoupled with voice conversion. Semi-supervised learning also helps for both\nlanguages, while parameter-efficient fine-tuning helps on Afrikaans but not on\nisiXhosa (which is under-represented in the Whisper model). Few child-speech\nstudies look at non-English data, and even fewer at the preschool ages of 4 and\n5. Our work therefore represents a unique validation of a wide range of\nprevious child-speech ASR strategies in an under-explored setting.\n", "date": "2025-01-11", "categories": []}
{"id": "2501.06514", "title": "Neural Codec Source Tracing: Toward Comprehensive Attribution in\n  Open-Set Condition", "abstract": "  Current research in audio deepfake detection is gradually transitioning from\nbinary classification to multi-class tasks, referred as audio deepfake source\ntracing task. However, existing studies on source tracing consider only\nclosed-set scenarios and have not considered the challenges posed by open-set\nconditions. In this paper, we define the Neural Codec Source Tracing (NCST)\ntask, which is capable of performing open-set neural codec classification and\ninterpretable ALM detection. Specifically, we constructed the ST-Codecfake\ndataset for the NCST task, which includes bilingual audio samples generated by\n11 state-of-the-art neural codec methods and ALM-based out-ofdistribution (OOD)\ntest samples. Furthermore, we establish a comprehensive source tracing\nbenchmark to assess NCST models in open-set conditions. The experimental\nresults reveal that although the NCST models perform well in in-distribution\n(ID) classification and OOD detection, they lack robustness in classifying\nunseen real audio. The ST-codecfake dataset and code are available.\n", "date": "2025-01-11", "categories": []}
{"id": "2501.07474", "title": "Estimating Musical Surprisal in Audio", "abstract": "  In modeling musical surprisal expectancy with computational methods, it has\nbeen proposed to use the information content (IC) of one-step predictions from\nan autoregressive model as a proxy for surprisal in symbolic music. With an\nappropriately chosen model, the IC of musical events has been shown to\ncorrelate with human perception of surprise and complexity aspects, including\ntonal and rhythmic complexity. This work investigates whether an analogous\nmethodology can be applied to music audio. We train an autoregressive\nTransformer model to predict compressed latent audio representations of a\npretrained autoencoder network. We verify learning effects by estimating the\ndecrease in IC with repetitions. We investigate the mean IC of musical segment\ntypes (e.g., A or B) and find that segment types appearing later in a piece\nhave a higher IC than earlier ones on average. We investigate the IC's relation\nto audio and musical features and find it correlated with timbral variations\nand loudness and, to a lesser extent, dissonance, rhythmic complexity, and\nonset density related to audio and musical features. Finally, we investigate if\nthe IC can predict EEG responses to songs and thus model humans' surprisal in\nmusic. We provide code for our method on github.com/sonycslparis/audioic.\n", "date": "2025-01-13", "categories": []}
{"id": "2410.01350", "title": "Takin-VC: Expressive Zero-Shot Voice Conversion via Adaptive Hybrid\n  Content Encoding and Enhanced Timbre Modeling", "abstract": "  Expressive zero-shot voice conversion (VC) is a critical and challenging task\nthat aims to transform the source timbre into an arbitrary unseen speaker while\npreserving the original content and expressive qualities. Despite recent\nprogress in zero-shot VC, there remains considerable potential for improvements\nin speaker similarity and speech naturalness. Moreover, existing zero-shot VC\nsystems struggle to fully reproduce paralinguistic information in highly\nexpressive speech, such as breathing, crying, and emotional nuances, limiting\ntheir practical applicability. To address these issues, we propose Takin-VC, a\nnovel expressive zero-shot VC framework via adaptive hybrid content encoding\nand memory-augmented context-aware timbre modeling. Specifically, we introduce\nan innovative hybrid content encoder that incorporates an adaptive fusion\nmodule, capable of effectively integrating quantized features of the\npre-trained WavLM and HybridFormer in an implicit manner, so as to extract\nprecise linguistic features while enriching paralinguistic elements. For timbre\nmodeling, we propose advanced memory-augmented and context-aware modules to\ngenerate high-quality target timbre features and fused representations that\nseamlessly align source content with target timbre. To enhance real-time\nperformance, we advocate a conditional flow matching model to reconstruct the\nMel-spectrogram of the source speech. Experimental results show that our\nTakin-VC consistently surpasses state-of-the-art VC systems, achieving notable\nimprovements in terms of speech naturalness, speech expressiveness, and speaker\nsimilarity, while offering enhanced inference speed.\n", "date": "2024-10-02", "categories": []}
{"id": "2401.01145", "title": "HAAQI-Net: A Non-intrusive Neural Music Audio Quality Assessment Model\n  for Hearing Aids", "abstract": "  This paper introduces HAAQI-Net, a non-intrusive deep learning-based music\naudio quality assessment model for hearing aid users. Unlike traditional\nmethods like the Hearing Aid Audio Quality Index (HAAQI) that require intrusive\nreference signal comparisons, HAAQI-Net offers a more accessible and\ncomputationally efficient alternative. By utilizing a Bidirectional Long\nShort-Term Memory (BLSTM) architecture with attention mechanisms and features\nextracted from the pre-trained BEATs model, it can predict HAAQI scores\ndirectly from music audio clips and hearing loss patterns. Experimental results\ndemonstrate HAAQI-Net's effectiveness, achieving a Linear Correlation\nCoefficient (LCC) of 0.9368 , a Spearman's Rank Correlation Coefficient (SRCC)\nof 0.9486 , and a Mean Squared Error (MSE) of 0.0064 and inference time\nsignificantly reduces from 62.52 to 2.54 seconds. To address computational\noverhead, a knowledge distillation strategy was applied, reducing parameters by\n75.85% and inference time by 96.46%, while maintaining strong performance (LCC:\n0.9071 , SRCC: 0.9307 , MSE: 0.0091 ). To expand its capabilities, HAAQI-Net\nwas adapted to predict subjective human scores like the Mean Opinion Score\n(MOS) through fine-tuning. This adaptation significantly improved prediction\naccuracy, validated through statistical analysis. Furthermore, the robustness\nof HAAQI-Net was evaluated under varying Sound Pressure Level (SPL) conditions,\nrevealing optimal performance at a reference SPL of 65 dB, with accuracy\ngradually decreasing as SPL deviated from this point. The advancements in\nsubjective score prediction, SPL robustness, and computational efficiency\nposition HAAQI-Net as a scalable solution for music audio quality assessment in\nhearing aid applications, contributing to efficient and accurate models in\naudio signal processing and hearing aid technology.\n", "date": "2024-01-02", "categories": []}
{"id": "2404.16969", "title": "COCOLA: Coherence-Oriented Contrastive Learning of Musical Audio\n  Representations", "abstract": "  We present COCOLA (Coherence-Oriented Contrastive Learning for Audio), a\ncontrastive learning method for musical audio representations that captures the\nharmonic and rhythmic coherence between samples. Our method operates at the\nlevel of the stems composing music tracks and can input features obtained via\nHarmonic-Percussive Separation (HPS). COCOLA allows the objective evaluation of\ngenerative models for music accompaniment generation, which are difficult to\nbenchmark with established metrics. In this regard, we evaluate recent music\naccompaniment generation models, demonstrating the effectiveness of the\nproposed method. We release the model checkpoints trained on public datasets\ncontaining separate stems (MUSDB18-HQ, MoisesDB, Slakh2100, and CocoChorales).\n", "date": "2024-04-25", "categories": []}
{"id": "2501.04819", "title": "Planing It by Ear: Convolutional Neural Networks for Acoustic Anomaly\n  Detection in Industrial Wood Planers", "abstract": "  In recent years, the wood product industry has been facing a skilled labor\nshortage. The result is more frequent sudden failures, resulting in additional\ncosts for these companies already operating in a very competitive market.\nMoreover, sawmills are challenging environments for machinery and sensors.\nGiven that experienced machine operators may be able to diagnose defects or\nmalfunctions, one possible way of assisting novice operators is through\nacoustic monitoring. As a step towards the automation of wood-processing\nequipment and decision support systems for machine operators, in this paper, we\nexplore using a deep convolutional autoencoder for acoustic anomaly detection\nof wood planers on a new real-life dataset. Specifically, our convolutional\nautoencoder with skip connections (Skip-CAE) and our Skip-CAE transformer\noutperform the DCASE autoencoder baseline, one-class SVM, isolation forest and\na published convolutional autoencoder architecture, respectively obtaining an\narea under the ROC curve of 0.846 and 0.875 on a dataset of real-factory planer\nsounds. Moreover, we show that adding skip connections and attention mechanism\nunder the form of a transformer encoder-decoder helps to further improve the\nanomaly detection capabilities.\n", "date": "2025-01-08", "categories": []}
{"id": "2408.06954", "title": "Neural Speech and Audio Coding: Modern AI Technology Meets Traditional\n  Codecs", "abstract": "  This paper explores the integration of model-based and data-driven approaches\nwithin the realm of neural speech and audio coding systems. It highlights the\nchallenges posed by the subjective evaluation processes of speech and audio\ncodecs and discusses the limitations of purely data-driven approaches, which\noften require inefficiently large architectures to match the performance of\nmodel-based methods. The study presents hybrid systems as a viable solution,\noffering significant improvements to the performance of conventional codecs\nthrough meticulously chosen design enhancements. Specifically, it introduces a\nneural network-based signal enhancer designed to post-process existing codecs'\noutput, along with the autoencoder-based end-to-end models and LPCNet--hybrid\nsystems that combine linear predictive coding (LPC) with neural networks.\nFurthermore, the paper delves into predictive models operating within custom\nfeature spaces (TF-Codec) or predefined transform domains (MDCTNet) and\nexamines the use of psychoacoustically calibrated loss functions to train\nend-to-end neural audio codecs. Through these investigations, the paper\ndemonstrates the potential of hybrid systems to advance the field of speech and\naudio coding by bridging the gap between traditional model-based approaches and\nmodern data-driven techniques.\n", "date": "2024-08-13", "categories": []}
{"id": "2409.08514", "title": "Apollo: Band-sequence Modeling for High-Quality Audio Restoration", "abstract": "  Audio restoration has become increasingly significant in modern society, not\nonly due to the demand for high-quality auditory experiences enabled by\nadvanced playback devices, but also because the growing capabilities of\ngenerative audio models necessitate high-fidelity audio. Typically, audio\nrestoration is defined as a task of predicting undistorted audio from damaged\ninput, often trained using a GAN framework to balance perception and\ndistortion. Since audio degradation is primarily concentrated in mid- and\nhigh-frequency ranges, especially due to codecs, a key challenge lies in\ndesigning a generator capable of preserving low-frequency information while\naccurately reconstructing high-quality mid- and high-frequency content.\nInspired by recent advancements in high-sample-rate music separation, speech\nenhancement, and audio codec models, we propose Apollo, a generative model\ndesigned for high-sample-rate audio restoration. Apollo employs an explicit\nfrequency band split module to model the relationships between different\nfrequency bands, allowing for more coherent and higher-quality restored audio.\nEvaluated on the MUSDB18-HQ and MoisesDB datasets, Apollo consistently\noutperforms existing SR-GAN models across various bit rates and music genres,\nparticularly excelling in complex scenarios involving mixtures of multiple\ninstruments and vocals. Apollo significantly improves music restoration quality\nwhile maintaining computational efficiency. The source code for Apollo is\npublicly available at https://github.com/JusperLee/Apollo.\n", "date": "2024-09-13", "categories": []}
{"id": "2501.03038", "title": "Piano Transcription by Hierarchical Language Modeling with Pretrained\n  Roll-based Encoders", "abstract": "  Automatic Music Transcription (AMT), aiming to get musical notes from raw\naudio, typically uses frame-level systems with piano-roll outputs or language\nmodel (LM)-based systems with note-level predictions. However, frame-level\nsystems require manual thresholding, while the LM-based systems struggle with\nlong sequences. In this paper, we propose a hybrid method combining pre-trained\nroll-based encoders with an LM decoder to leverage the strengths of both\nmethods. Besides, our approach employs a hierarchical prediction strategy,\nfirst predicting onset and pitch, then velocity, and finally offset. The\nhierarchical prediction strategy reduces computational costs by breaking down\nlong sequences into different hierarchies. Evaluated on two benchmark\nroll-based encoders, our method outperforms traditional piano-roll outputs 0.01\nand 0.022 in onset-offset-velocity F1 score, demonstrating its potential as a\nperformance-enhancing plug-in for arbitrary roll-based music transcription\nencoder.\n", "date": "2025-01-06", "categories": []}
{"id": "2501.03689", "title": "MAJL: A Model-Agnostic Joint Learning Framework for Music Source\n  Separation and Pitch Estimation", "abstract": "  Music source separation and pitch estimation are two vital tasks in music\ninformation retrieval. Typically, the input of pitch estimation is obtained\nfrom the output of music source separation. Therefore, existing methods have\ntried to perform these two tasks simultaneously, so as to leverage the mutually\nbeneficial relationship between both tasks. However, these methods still face\ntwo critical challenges that limit the improvement of both tasks: the lack of\nlabeled data and joint learning optimization. To address these challenges, we\npropose a Model-Agnostic Joint Learning (MAJL) framework for both tasks. MAJL\nis a generic framework and can use variant models for each task. It includes a\ntwo-stage training method and a dynamic weighting method named Dynamic Weights\non Hard Samples (DWHS), which addresses the lack of labeled data and joint\nlearning optimization, respectively. Experimental results on public music\ndatasets show that MAJL outperforms state-of-the-art methods on both tasks,\nwith significant improvements of 0.92 in Signal-to-Distortion Ratio (SDR) for\nmusic source separation and 2.71% in Raw Pitch Accuracy (RPA) for pitch\nestimation. Furthermore, comprehensive studies not only validate the\neffectiveness of each component of MAJL, but also indicate the great generality\nof MAJL in adapting to different model architectures.\n", "date": "2025-01-07", "categories": []}
{"id": "2501.03805", "title": "Detecting the Undetectable: Assessing the Efficacy of Current Spoof\n  Detection Methods Against Seamless Speech Edits", "abstract": "  Neural speech editing advancements have raised concerns about their misuse in\nspoofing attacks. Traditional partially edited speech corpora primarily focus\non cut-and-paste edits, which, while maintaining speaker consistency, often\nintroduce detectable discontinuities. Recent methods, like\nA\\textsuperscript{3}T and Voicebox, improve transitions by leveraging\ncontextual information. To foster spoofing detection research, we introduce the\nSpeech INfilling Edit (SINE) dataset, created with Voicebox. We detailed the\nprocess of re-implementing Voicebox training and dataset creation. Subjective\nevaluations confirm that speech edited using this novel technique is more\nchallenging to detect than conventional cut-and-paste methods. Despite human\ndifficulty, experimental results demonstrate that self-supervised-based\ndetectors can achieve remarkable performance in detection, localization, and\ngeneralization across different edit methods. The dataset and related models\nwill be made publicly available.\n", "date": "2025-01-07", "categories": []}
{"id": "2410.09472", "title": "DRCap: Decoding CLAP Latents with Retrieval-Augmented Generation for\n  Zero-shot Audio Captioning", "abstract": "  While automated audio captioning (AAC) has made notable progress, traditional\nfully supervised AAC models still face two critical challenges: the need for\nexpensive audio-text pair data for training and performance degradation when\ntransferring across domains. To overcome these limitations, we present DRCap, a\ndata-efficient and flexible zero-shot audio captioning system that requires\ntext-only data for training and can quickly adapt to new domains without\nadditional fine-tuning. DRCap integrates a contrastive language-audio\npre-training (CLAP) model and a large-language model (LLM) as its backbone.\nDuring training, the model predicts the ground-truth caption with a fixed text\nencoder from CLAP, whereas, during inference, the text encoder is replaced with\nthe audio encoder to generate captions for audio clips in a zero-shot manner.\nTo mitigate the modality gap of the CLAP model, we use both the projection\nstrategy from the encoder side and the retrieval-augmented generation strategy\nfrom the decoder side. Specifically, audio embeddings are first projected onto\na text embedding support to absorb extensive semantic information within the\njoint multi-modal space of CLAP. At the same time, similar captions retrieved\nfrom a datastore are fed as prompts to instruct the LLM, incorporating external\nknowledge to take full advantage of its strong generative capability.\nConditioned on both the projected CLAP embedding and the retrieved similar\ncaptions, the model is able to produce a more accurate and semantically rich\ntextual description. By tailoring the text embedding support and the caption\ndatastore to the target domain, DRCap acquires a robust ability to adapt to new\ndomains in a training-free manner. Experimental results demonstrate that DRCap\noutperforms all other zero-shot models in in-domain scenarios and achieves\nstate-of-the-art performance in cross-domain scenarios.\n", "date": "2024-10-12", "categories": []}
{"id": "2501.02030", "title": "Detecting Music Performance Errors with Transformers", "abstract": "  Beginner musicians often struggle to identify specific errors in their\nperformances, such as playing incorrect notes or rhythms. There are two\nlimitations in existing tools for music error detection: (1) Existing\napproaches rely on automatic alignment; therefore, they are prone to errors\ncaused by small deviations between alignment targets.; (2) There is a lack of\nsufficient data to train music error detection models, resulting in\nover-reliance on heuristics. To address (1), we propose a novel transformer\nmodel, Polytune, that takes audio inputs and outputs annotated music scores.\nThis model can be trained end-to-end to implicitly align and compare\nperformance audio with music scores through latent space representations. To\naddress (2), we present a novel data generation technique capable of creating\nlarge-scale synthetic music error datasets. Our approach achieves a 64.1%\naverage Error Detection F1 score, improving upon prior work by 40 percentage\npoints across 14 instruments. Additionally, compared with existing\ntranscription methods repurposed for music error detection, our model can\nhandle multiple instruments. Our source code and datasets are available at\nhttps://github.com/ben2002chou/Polytune.\n", "date": "2025-01-03", "categories": []}
{"id": "2501.03045", "title": "Single-Channel Distance-Based Source Separation for Mobile GPU in\n  Outdoor and Indoor Environments", "abstract": "  This study emphasizes the significance of exploring distance-based source\nseparation (DSS) in outdoor environments. Unlike existing studies that\nprimarily focus on indoor settings, the proposed model is designed to capture\nthe unique characteristics of outdoor audio sources. It incorporates advanced\ntechniques, including a two-stage conformer block, a linear relation-aware\nself-attention (RSA), and a TensorFlow Lite GPU delegate. While the linear RSA\nmay not capture physical cues as explicitly as the quadratic RSA, the linear\nRSA enhances the model's context awareness, leading to improved performance on\nthe DSS that requires an understanding of physical cues in outdoor and indoor\nenvironments. The experimental results demonstrated that the proposed model\novercomes the limitations of existing approaches and considerably enhances\nenergy efficiency and real-time inference speed on mobile devices.\n", "date": "2025-01-06", "categories": []}
{"id": "2501.01108", "title": "MuQ: Self-Supervised Music Representation Learning with Mel Residual\n  Vector Quantization", "abstract": "  Recent years have witnessed the success of foundation models pre-trained with\nself-supervised learning (SSL) in various music informatics understanding\ntasks, including music tagging, instrument classification, key detection, and\nmore. In this paper, we propose a self-supervised music representation learning\nmodel for music understanding. Distinguished from previous studies adopting\nrandom projection or existing neural codec, the proposed model, named MuQ, is\ntrained to predict tokens generated by Mel Residual Vector Quantization\n(Mel-RVQ). Our Mel-RVQ utilizes residual linear projection structure for Mel\nspectrum quantization to enhance the stability and efficiency of target\nextraction and lead to better performance. Experiments in a large variety of\ndownstream tasks demonstrate that MuQ outperforms previous self-supervised\nmusic representation models with only 0.9K hours of open-source pre-training\ndata. Scaling up the data to over 160K hours and adopting iterative training\nconsistently improve the model performance. To further validate the strength of\nour model, we present MuQ-MuLan, a joint music-text embedding model based on\ncontrastive learning, which achieves state-of-the-art performance in the\nzero-shot music tagging task on the MagnaTagATune dataset. Code and checkpoints\nare open source in https://github.com/tencent-ailab/MuQ.\n", "date": "2025-01-02", "categories": []}
{"id": "2410.16130", "title": "Can Large Audio-Language Models Truly Hear? Tackling Hallucinations with\n  Multi-Task Assessment and Stepwise Audio Reasoning", "abstract": "  Recent advancements in large audio-language models (LALMs) have shown\nimpressive capabilities in understanding and reasoning about audio and speech\ninformation. However, these models still face challenges, including\nhallucinating non-existent sound events, misidentifying the order of sound\nevents, and incorrectly attributing sound sources, which undermine their\nreliability and real-world application. To systematically evaluate these\nissues, we propose three distinct tasks: object existence, temporal order, and\nobject attribute within audio. These tasks assess the models' comprehension of\ncritical audio information aspects. Our experimental results reveal limitations\nin these fundamental tasks, underscoring the need for better models in\nrecognizing specific sound events, determining event sequences, and identifying\nsound sources. To improve performance in these areas, we introduce a multi-turn\nchain-of-thought approach, which demonstrates significantly improved model\nperformance across the proposed tasks.\n", "date": "2024-10-21", "categories": []}
{"id": "2501.00064", "title": "Lungmix: A Mixup-Based Strategy for Generalization in Respiratory Sound\n  Classification", "abstract": "  Respiratory sound classification plays a pivotal role in diagnosing\nrespiratory diseases. While deep learning models have shown success with\nvarious respiratory sound datasets, our experiments indicate that models\ntrained on one dataset often fail to generalize effectively to others, mainly\ndue to data collection and annotation \\emph{inconsistencies}. To address this\nlimitation, we introduce \\emph{Lungmix}, a novel data augmentation technique\ninspired by Mixup. Lungmix generates augmented data by blending waveforms using\nloudness and random masks while interpolating labels based on their semantic\nmeaning, helping the model learn more generalized representations.\nComprehensive evaluations across three datasets, namely ICBHI, SPR, and HF,\ndemonstrate that Lungmix significantly enhances model generalization to unseen\ndata. In particular, Lungmix boosts the 4-class classification score by up to\n3.55\\%, achieving performance comparable to models trained directly on the\ntarget dataset.\n", "date": "2024-12-29", "categories": []}
{"id": "2501.01117", "title": "Robust COVID-19 Detection from Cough Sounds using Deep Neural Decision\n  Tree and Forest: A Comprehensive Cross-Datasets Evaluation", "abstract": "  This research presents a robust approach to classifying COVID-19 cough sounds\nusing cutting-edge machine-learning techniques. Leveraging deep neural decision\ntrees and deep neural decision forests, our methodology demonstrates consistent\nperformance across diverse cough sound datasets. We begin with a comprehensive\nextraction of features to capture a wide range of audio features from\nindividuals, whether COVID-19 positive or negative. To determine the most\nimportant features, we use recursive feature elimination along with\ncross-validation. Bayesian optimization fine-tunes hyper-parameters of deep\nneural decision tree and deep neural decision forest models. Additionally, we\nintegrate the SMOTE during training to ensure a balanced representation of\npositive and negative data. Model performance refinement is achieved through\nthreshold optimization, maximizing the ROC-AUC score. Our approach undergoes a\ncomprehensive evaluation in five datasets: Cambridge, Coswara, COUGHVID,\nVirufy, and the combined Virufy with the NoCoCoDa dataset. Consistently\noutperforming state-of-the-art methods, our proposed approach yields notable\nAUC scores of 0.97, 0.98, 0.92, 0.93, 0.99, and 0.99 across the respective\ndatasets. Merging all datasets into a combined dataset, our method, using a\ndeep neural decision forest classifier, achieves an AUC of 0.97. Also, our\nstudy includes a comprehensive cross-datasets analysis, revealing demographic\nand geographic differences in the cough sounds associated with COVID-19. These\ndifferences highlight the challenges in transferring learned features across\ndiverse datasets and underscore the potential benefits of dataset integration,\nimproving generalizability and enhancing COVID-19 detection from audio signals.\n", "date": "2025-01-02", "categories": []}
{"id": "2412.16861", "title": "SoundLoc3D: Invisible 3D Sound Source Localization and Classification\n  Using a Multimodal RGB-D Acoustic Camera", "abstract": "  Accurately localizing 3D sound sources and estimating their semantic labels\n-- where the sources may not be visible, but are assumed to lie on the physical\nsurface of objects in the scene -- have many real applications, including\ndetecting gas leak and machinery malfunction. The audio-visual weak-correlation\nin such setting poses new challenges in deriving innovative methods to answer\nif or how we can use cross-modal information to solve the task. Towards this\nend, we propose to use an acoustic-camera rig consisting of a pinhole RGB-D\ncamera and a coplanar four-channel microphone array~(Mic-Array). By using this\nrig to record audio-visual signals from multiviews, we can use the cross-modal\ncues to estimate the sound sources 3D locations. Specifically, our framework\nSoundLoc3D treats the task as a set prediction problem, each element in the set\ncorresponds to a potential sound source. Given the audio-visual\nweak-correlation, the set representation is initially learned from a single\nview microphone array signal, and then refined by actively incorporating\nphysical surface cues revealed from multiview RGB-D images. We demonstrate the\nefficiency and superiority of SoundLoc3D on large-scale simulated dataset, and\nfurther show its robustness to RGB-D measurement inaccuracy and ambient noise\ninterference.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.19279", "title": "Improving Generalization for AI-Synthesized Voice Detection", "abstract": "  AI-synthesized voice technology has the potential to create realistic human\nvoices for beneficial applications, but it can also be misused for malicious\npurposes. While existing AI-synthesized voice detection models excel in\nintra-domain evaluation, they face challenges in generalizing across different\ndomains, potentially becoming obsolete as new voice generators emerge. Current\nsolutions use diverse data and advanced machine learning techniques (e.g.,\ndomain-invariant representation, self-supervised learning), but are limited by\npredefined vocoders and sensitivity to factors like background noise and\nspeaker identity. In this work, we introduce an innovative disentanglement\nframework aimed at extracting domain-agnostic artifact features related to\nvocoders. Utilizing these features, we enhance model learning in a flat loss\nlandscape, enabling escape from suboptimal solutions and improving\ngeneralization. Extensive experiments on benchmarks show our approach\noutperforms state-of-the-art methods, achieving up to 5.12% improvement in the\nequal error rate metric in intra-domain and 7.59% in cross-domain evaluations.\n", "date": "2024-12-26", "categories": []}
{"id": "2412.20378", "title": "Tri-Ergon: Fine-grained Video-to-Audio Generation with Multi-modal\n  Conditions and LUFS Control", "abstract": "  Video-to-audio (V2A) generation utilizes visual-only video features to\nproduce realistic sounds that correspond to the scene. However, current V2A\nmodels often lack fine-grained control over the generated audio, especially in\nterms of loudness variation and the incorporation of multi-modal conditions. To\novercome these limitations, we introduce Tri-Ergon, a diffusion-based V2A model\nthat incorporates textual, auditory, and pixel-level visual prompts to enable\ndetailed and semantically rich audio synthesis. Additionally, we introduce\nLoudness Units relative to Full Scale (LUFS) embedding, which allows for\nprecise manual control of the loudness changes over time for individual audio\nchannels, enabling our model to effectively address the intricate correlation\nof video and audio in real-world Foley workflows. Tri-Ergon is capable of\ncreating 44.1 kHz high-fidelity stereo audio clips of varying lengths up to 60\nseconds, which significantly outperforms existing state-of-the-art V2A methods\nthat typically generate mono audio for a fixed duration.\n", "date": "2024-12-29", "categories": []}
{"id": "2412.20619", "title": "Audiopedia: Audio QA with Knowledge", "abstract": "  In this paper, we introduce Audiopedia, a novel task called Audio Question\nAnswering with Knowledge, which requires both audio comprehension and external\nknowledge reasoning. Unlike traditional Audio Question Answering (AQA)\nbenchmarks that focus on simple queries answerable from audio alone, Audiopedia\ntargets knowledge-intensive questions. We define three sub-tasks: (i) Single\nAudio Question Answering (s-AQA), where questions are answered based on a\nsingle audio sample, (ii) Multi-Audio Question Answering (m-AQA), which\nrequires reasoning over multiple audio samples, and (iii) Retrieval-Augmented\nAudio Question Answering (r-AQA), which involves retrieving relevant audio to\nanswer the question. We benchmark large audio language models (LALMs) on these\nsub-tasks and observe suboptimal performance. To address this, we propose a\ngeneric framework that can be adapted to any LALM, equipping them with\nknowledge reasoning capabilities. Our framework has two components: (i) Audio\nEntity Linking (AEL) and (ii) Knowledge-Augmented Audio Large Multimodal Model\n(KA2LM), which together improve performance on knowledge-intensive AQA tasks.\nTo our knowledge, this is the first work to address advanced audio\nunderstanding via knowledge-intensive tasks like Audiopedia.\n", "date": "2024-12-29", "categories": []}
{"id": "2305.14032", "title": "Patch-Mix Contrastive Learning with Audio Spectrogram Transformer on\n  Respiratory Sound Classification", "abstract": "  Respiratory sound contains crucial information for the early diagnosis of\nfatal lung diseases. Since the COVID-19 pandemic, there has been a growing\ninterest in contact-free medical care based on electronic stethoscopes. To this\nend, cutting-edge deep learning models have been developed to diagnose lung\ndiseases; however, it is still challenging due to the scarcity of medical data.\nIn this study, we demonstrate that the pretrained model on large-scale visual\nand audio datasets can be generalized to the respiratory sound classification\ntask. In addition, we introduce a straightforward Patch-Mix augmentation, which\nrandomly mixes patches between different samples, with Audio Spectrogram\nTransformer (AST). We further propose a novel and effective Patch-Mix\nContrastive Learning to distinguish the mixed representations in the latent\nspace. Our method achieves state-of-the-art performance on the ICBHI dataset,\noutperforming the prior leading score by an improvement of 4.08%.\n", "date": "2023-05-23", "categories": []}
{"id": "2306.00107", "title": "MERT: Acoustic Music Understanding Model with Large-Scale\n  Self-supervised Training", "abstract": "  Self-supervised learning (SSL) has recently emerged as a promising paradigm\nfor training generalisable models on large-scale data in the fields of vision,\ntext, and speech. Although SSL has been proven effective in speech and audio,\nits application to music audio has yet to be thoroughly explored. This is\npartially due to the distinctive challenges associated with modelling musical\nknowledge, particularly tonal and pitched characteristics of music. To address\nthis research gap, we propose an acoustic Music undERstanding model with\nlarge-scale self-supervised Training (MERT), which incorporates teacher models\nto provide pseudo labels in the masked language modelling (MLM) style acoustic\npre-training. In our exploration, we identified an effective combination of\nteacher models, which outperforms conventional speech and audio approaches in\nterms of performance. This combination includes an acoustic teacher based on\nResidual Vector Quantisation - Variational AutoEncoder (RVQ-VAE) and a musical\nteacher based on the Constant-Q Transform (CQT). Furthermore, we explore a wide\nrange of settings to overcome the instability in acoustic language model\npre-training, which allows our designed paradigm to scale from 95M to 330M\nparameters. Experimental results indicate that our model can generalise and\nperform well on 14 music understanding tasks and attain state-of-the-art (SOTA)\noverall scores.\n", "date": "2023-05-31", "categories": []}
{"id": "2405.04880", "title": "The Codecfake Dataset and Countermeasures for the Universally Detection\n  of Deepfake Audio", "abstract": "  With the proliferation of Audio Language Model (ALM) based deepfake audio,\nthere is an urgent need for generalized detection methods. ALM-based deepfake\naudio currently exhibits widespread, high deception, and type versatility,\nposing a significant challenge to current audio deepfake detection (ADD) models\ntrained solely on vocoded data. To effectively detect ALM-based deepfake audio,\nwe focus on the mechanism of the ALM-based audio generation method, the\nconversion from neural codec to waveform. We initially constructed the\nCodecfake dataset, an open-source, large-scale collection comprising over 1\nmillion audio samples in both English and Chinese, focus on ALM-based audio\ndetection. As countermeasure, to achieve universal detection of deepfake audio\nand tackle domain ascent bias issue of original sharpness aware minimization\n(SAM), we propose the CSAM strategy to learn a domain balanced and generalized\nminima. In our experiments, we first demonstrate that ADD model training with\nthe Codecfake dataset can effectively detects ALM-based audio. Furthermore, our\nproposed generalization countermeasure yields the lowest average equal error\nrate (EER) of 0.616% across all test conditions compared to baseline models.\nThe dataset and associated code are available online.\n", "date": "2024-05-08", "categories": []}
{"id": "2407.05551", "title": "Read, Watch and Scream! Sound Generation from Text and Video", "abstract": "  Despite the impressive progress of multimodal generative models,\nvideo-to-audio generation still suffers from limited performance and limits the\nflexibility to prioritize sound synthesis for specific objects within the\nscene. Conversely, text-to-audio generation methods generate high-quality audio\nbut pose challenges in ensuring comprehensive scene depiction and time-varying\ncontrol. To tackle these challenges, we propose a novel video-and-text-to-audio\ngeneration method, called \\ours, where video serves as a conditional control\nfor a text-to-audio generation model. Especially, our method estimates the\nstructural information of sound (namely, energy) from the video while receiving\nkey content cues from a user prompt. We employ a well-performing text-to-audio\nmodel to consolidate the video control, which is much more efficient for\ntraining multimodal diffusion models with massive triplet-paired\n(audio-video-text) data. In addition, by separating the generative components\nof audio, it becomes a more flexible system that allows users to freely adjust\nthe energy, surrounding environment, and primary sound source according to\ntheir preferences. Experimental results demonstrate that our method shows\nsuperiority in terms of quality, controllability, and training efficiency. Code\nand demo are available at https://naver-ai.github.io/rewas.\n", "date": "2024-07-08", "categories": []}
{"id": "2410.00070", "title": "Mamba for Streaming ASR Combined with Unimodal Aggregation", "abstract": "  This paper works on streaming automatic speech recognition (ASR). Mamba, a\nrecently proposed state space model, has demonstrated the ability to match or\nsurpass Transformers in various tasks while benefiting from a linear complexity\nadvantage. We explore the efficiency of Mamba encoder for streaming ASR and\npropose an associated lookahead mechanism for leveraging controllable future\ninformation. Additionally, a streaming-style unimodal aggregation (UMA) method\nis implemented, which automatically detects token activity and streamingly\ntriggers token output, and meanwhile aggregates feature frames for better\nlearning token representation. Based on UMA, an early termination (ET) method\nis proposed to further reduce recognition latency. Experiments conducted on two\nMandarin Chinese datasets demonstrate that the proposed model achieves\ncompetitive ASR performance in terms of both recognition accuracy and latency.\n", "date": "2024-09-30", "categories": []}
{"id": "2412.18217", "title": "U-Mamba-Net: A highly efficient Mamba-based U-net style network for\n  noisy and reverberant speech separation", "abstract": "  The topic of speech separation involves separating mixed speech with multiple\noverlapping speakers into several streams, with each stream containing speech\nfrom only one speaker. Many highly effective models have emerged and\nproliferated rapidly over time. However, the size and computational load of\nthese models have also increased accordingly. This is a disaster for the\ncommunity, as researchers need more time and computational resources to\nreproduce and compare existing models. In this paper, we propose U-mamba-net: a\nlightweight Mamba-based U-style model for speech separation in complex\nenvironments. Mamba is a state space sequence model that incorporates feature\nselection capabilities. U-style network is a fully convolutional neural network\nwhose symmetric contracting and expansive paths are able to learn\nmulti-resolution features. In our work, Mamba serves as a feature filter,\nalternating with U-Net. We test the proposed model on Libri2mix. The results\nshow that U-Mamba-Net achieves improved performance with quite low\ncomputational cost.\n", "date": "2024-12-24", "categories": []}
{"id": "2206.08297", "title": "A Language Model With Million Context Length For Raw Audio", "abstract": "  Modeling long-term dependencies for audio signals is a particularly\nchallenging problem, as even small-time scales yield on the order of a hundred\nthousand samples. With the recent advent of Transformers, neural architectures\nbecame good at modeling dependencies over longer time scales, but they suffered\nfrom quadratic constraints to scale them. We propose a generative\nauto-regressive architecture that can model audio waveforms over quite a large\ncontext, greater than 500,000 samples. Our work is adapted to learn time\ndependencies by learning a latent representation by a CNN front-end, and then\nlearning dependencies over these representations using Transformer encoders,\nfully trained end-to-end: thereby allowing to learn representations as it deems\nfit for the next sample. Unlike previous works that compared different time\nscales to show improvement, we use a standard dataset, with the same number of\nparameters/context to show improvements. We achieve a state-of-the-art\nperformance as compared to other approaches such as Wavenet, SaSHMI, and\nSample-RNN on a standard dataset for modeling long-term structure. This work\ngives very exciting direction for the field, given improvements in context\nmodeling that can be scaled with more data, as well as potentially better\nresults by using billions/trillions of parameters.\n", "date": "2022-06-16", "categories": []}
{"id": "2303.10446", "title": "Content Adaptive Front End For Audio Classification", "abstract": "  We propose a learnable content adaptive front end for audio signal\nprocessing. Before the modern advent of deep learning, we used fixed\nrepresentation non-learnable front-ends like spectrogram or mel-spectrogram\nwith/without neural architectures. With convolutional architectures supporting\nvarious applications such as ASR and acoustic scene understanding, a shift to a\nlearnable front ends occurred in which both the type of basis functions and the\nweight were learned from scratch and optimized for the particular task of\ninterest. With the shift to transformer-based architectures with no\nconvolutional blocks present, a linear layer projects small waveform patches\nonto a small latent dimension before feeding them to a transformer\narchitecture. In this work, we propose a way of computing a content-adaptive\nlearnable time-frequency representation. We pass each audio signal through a\nbank of convolutional filters, each giving a fixed-dimensional vector. It is\nakin to learning a bank of finite impulse-response filterbanks and passing the\ninput signal through the optimum filter bank depending on the content of the\ninput signal. A content-adaptive learnable time-frequency representation may be\nmore broadly applicable, beyond the experiments in this paper.\n", "date": "2023-03-18", "categories": []}
{"id": "2406.15668", "title": "PI-Whisper: Designing an Adaptive and Incremental Automatic Speech\n  Recognition System for Edge Devices", "abstract": "  Edge-based automatic speech recognition (ASR) technologies are increasingly\nprevalent in the development of intelligent and personalized assistants.\nHowever, resource-constrained ASR models face significant challenges in\nadaptivity, incrementality, and inclusivity when faced with a diverse\npopulation. To tackle those challenges, we propose PI-Whisper, a novel ASR\nsystem that adaptively enhances recognition capabilities by identifying\nspeakers' characteristics in real-time. In this work, we show how the design of\nPI-Whisper allows for incremental adaptation of new characteristics without the\nneed for repetitive retraining, enhances recognition capabilities, and improves\nequity and fairness across diverse speaker groups. PI-Whisper demonstrates\nthese advantages by achieving state-of-the-art accuracy, reducing the word\nerror rate (WER) by up to 13.7% relative to baselines while scaling linearly to\ncomputing resources.\n", "date": "2024-06-21", "categories": []}
{"id": "2411.14100", "title": "BEST-STD: Bidirectional Mamba-Enhanced Speech Tokenization for Spoken\n  Term Detection", "abstract": "  Spoken term detection (STD) is often hindered by reliance on frame-level\nfeatures and the computationally intensive DTW-based template matching,\nlimiting its practicality. To address these challenges, we propose a novel\napproach that encodes speech into discrete, speaker-agnostic semantic tokens.\nThis facilitates fast retrieval using text-based search algorithms and\neffectively handles out-of-vocabulary terms. Our approach focuses on generating\nconsistent token sequences across varying utterances of the same term. We also\npropose a bidirectional state space modeling within the Mamba encoder, trained\nin a self-supervised learning framework, to learn contextual frame-level\nfeatures that are further encoded into discrete tokens. Our analysis shows that\nour speech tokens exhibit greater speaker invariance than those from existing\ntokenizers, making them more suitable for STD tasks. Empirical evaluation on\nLibriSpeech and TIMIT databases indicates that our method outperforms existing\nSTD baselines while being more efficient.\n", "date": "2024-11-21", "categories": []}
{"id": "2412.16182", "title": "Decoding Poultry Vocalizations -- Natural Language Processing and\n  Transformer Models for Semantic and Emotional Analysis", "abstract": "  Deciphering the acoustic language of chickens offers new opportunities in\nanimal welfare and ecological informatics. Their subtle vocal signals encode\nhealth conditions, emotional states, and dynamic interactions within\necosystems. Understanding the semantics of these calls provides a valuable tool\nfor interpreting their functional vocabulary and clarifying how each sound\nserves a specific purpose in social and environmental contexts. We apply\nadvanced Natural Language Processing and transformer based models to translate\nbioacoustic data into meaningful insights. Our method integrates Wave2Vec 2.0\nfor raw audio feature extraction with a fine tuned Bidirectional Encoder\nRepresentations from Transformers model, pretrained on a broad corpus of animal\nsounds and adapted to poultry tasks. This pipeline decodes poultry\nvocalizations into interpretable categories including distress calls, feeding\nsignals, and mating vocalizations, revealing emotional nuances often overlooked\nby conventional analyses. Achieving 92 percent accuracy in classifying key\nvocalization types, our approach demonstrates the feasibility of real time\nautomated monitoring of flock health and stress. By tracking this functional\nvocabulary, farmers can respond proactively to environmental or behavioral\nchanges, improving poultry welfare, reducing stress related productivity\nlosses, and supporting more sustainable farm management. Beyond agriculture,\nthis research enhances our understanding of computational ecology. Accessing\nthe semantic foundation of animal calls may indicate biodiversity,\nenvironmental stressors, and species interactions, informing integrative\necosystem level decision making.\n", "date": "2024-12-11", "categories": []}
{"id": "2311.02369", "title": "TACNET: Temporal Audio Source Counting Network", "abstract": "  In this paper, we introduce the Temporal Audio Source Counting Network\n(TaCNet), an innovative architecture that addresses limitations in audio source\ncounting tasks. TaCNet operates directly on raw audio inputs, eliminating\ncomplex preprocessing steps and simplifying the workflow. Notably, it excels in\nreal-time speaker counting, even with truncated input windows. Our extensive\nevaluation, conducted using the LibriCount dataset, underscores TaCNet's\nexceptional performance, positioning it as a state-of-the-art solution for\naudio source counting tasks. With an average accuracy of 74.18 percentage over\n11 classes, TaCNet demonstrates its effectiveness across diverse scenarios,\nincluding applications involving Chinese and Persian languages. This\ncross-lingual adaptability highlights its versatility and potential impact.\n", "date": "2023-11-04", "categories": []}
{"id": "2412.15602", "title": "Music Genre Classification: Ensemble Learning with Subcomponents-level\n  Attention", "abstract": "  Music Genre Classification is one of the most popular topics in the fields of\nMusic Information Retrieval (MIR) and digital signal processing. Deep Learning\nhas emerged as the top performer for classifying music genres among various\nmethods. The letter introduces a novel approach by combining ensemble learning\nwith attention to sub-components, aiming to enhance the accuracy of identifying\nmusic genres. The core innovation of our work is the proposal to classify the\nsubcomponents of the music pieces separately, allowing our model to capture\ndistinct characteristics from those sub components. By applying ensemble\nlearning techniques to these individual classifications, we make the final\nclassification decision on the genre of the music. The proposed method has\nsuperior advantages in terms of accuracy compared to the other state-of-the-art\ntechniques trained and tested on the GTZAN dataset.\n", "date": "2024-12-20", "categories": []}
{"id": "2408.09027", "title": "Efficient Autoregressive Audio Modeling via Next-Scale Prediction", "abstract": "  Audio generation has achieved remarkable progress with the advance of\nsophisticated generative models, such as diffusion models (DMs) and\nautoregressive (AR) models. However, due to the naturally significant sequence\nlength of audio, the efficiency of audio generation remains an essential issue\nto be addressed, especially for AR models that are incorporated in large\nlanguage models (LLMs). In this paper, we analyze the token length of audio\ntokenization and propose a novel \\textbf{S}cale-level \\textbf{A}udio\n\\textbf{T}okenizer (SAT), with improved residual quantization. Based on SAT, a\nscale-level \\textbf{A}coustic \\textbf{A}uto\\textbf{R}egressive (AAR) modeling\nframework is further proposed, which shifts the next-token AR prediction to\nnext-scale AR prediction, significantly reducing the training cost and\ninference time. To validate the effectiveness of the proposed approach, we\ncomprehensively analyze design choices and demonstrate the proposed AAR\nframework achieves a remarkable \\textbf{35}$\\times$ faster inference speed and\n+\\textbf{1.33} Fr\\'echet Audio Distance (FAD) against baselines on the AudioSet\nbenchmark. Code: \\url{https://github.com/qiuk2/AAR}.\n", "date": "2024-08-16", "categories": []}
{"id": "2412.12111", "title": "Voice Biomarker Analysis and Automated Severity Classification of\n  Dysarthric Speech in a Multilingual Context", "abstract": "  Dysarthria, a motor speech disorder, severely impacts voice quality,\npronunciation, and prosody, leading to diminished speech intelligibility and\nreduced quality of life. Accurate assessment is crucial for effective\ntreatment, but traditional perceptual assessments are limited by their\nsubjectivity and resource intensity. To mitigate the limitations, automatic\ndysarthric speech assessment methods have been proposed to support clinicians\non their decision-making. While these methods have shown promising results,\nmost research has focused on monolingual environments. However, multilingual\napproaches are necessary to address the global burden of dysarthria and ensure\nequitable access to accurate diagnosis. This thesis proposes a novel\nmultilingual dysarthria severity classification method, by analyzing three\nlanguages: English, Korean, and Tamil.\n", "date": "2024-12-01", "categories": []}
{"id": "2412.12395", "title": "Sound Classification of Four Insect Classes", "abstract": "  The goal of this project is to classify four different insect sounds: cicada,\nbeetle, termite, and cricket. One application of this project is for pest\ncontrol to monitor and protect our ecosystem. Our project leverages data\naugmentation, including pitch shifting and speed changing, to improve model\ngeneralization. This project will test the performance of Decision Tree, Random\nForest, SVM RBF, XGBoost, and k-NN models, combined with MFCC feature. A\npotential novelty of this project is that various data augmentation techniques\nare used and created 6 data along with the original sound. The dataset consists\nof the sound recordings of these four insects. This project aims to achieve a\nhigh classification accuracy and to reduce the over-fitting problem.\n", "date": "2024-12-16", "categories": []}
{"id": "2412.12619", "title": "Phoneme-Level Feature Discrepancies: A Key to Detecting Sophisticated\n  Speech Deepfakes", "abstract": "  Recent advancements in text-to-speech and speech conversion technologies have\nenabled the creation of highly convincing synthetic speech. While these\ninnovations offer numerous practical benefits, they also cause significant\nsecurity challenges when maliciously misused. Therefore, there is an urgent\nneed to detect these synthetic speech signals. Phoneme features provide a\npowerful speech representation for deepfake detection. However, previous\nphoneme-based detection approaches typically focused on specific phonemes,\noverlooking temporal inconsistencies across the entire phoneme sequence. In\nthis paper, we develop a new mechanism for detecting speech deepfakes by\nidentifying the inconsistencies of phoneme-level speech features. We design an\nadaptive phoneme pooling technique that extracts sample-specific phoneme-level\nfeatures from frame-level speech data. By applying this technique to features\nextracted by pre-trained audio models on previously unseen deepfake datasets,\nwe demonstrate that deepfake samples often exhibit phoneme-level\ninconsistencies when compared to genuine speech. To further enhance detection\naccuracy, we propose a deepfake detector that uses a graph attention network to\nmodel the temporal dependencies of phoneme-level features. Additionally, we\nintroduce a random phoneme substitution augmentation technique to increase\nfeature diversity during training. Extensive experiments on four benchmark\ndatasets demonstrate the superior performance of our method over existing\nstate-of-the-art detection methods.\n", "date": "2024-12-17", "categories": []}
{"id": "2407.20592", "title": "EgoSonics: Generating Synchronized Audio for Silent Egocentric Videos", "abstract": "  We introduce EgoSonics, a method to generate semantically meaningful and\nsynchronized audio tracks conditioned on silent egocentric videos. Generating\naudio for silent egocentric videos could open new applications in virtual\nreality, assistive technologies, or for augmenting existing datasets. Existing\nwork has been limited to domains like speech, music, or impact sounds and\ncannot capture the broad range of audio frequencies found in egocentric videos.\nEgoSonics addresses these limitations by building on the strengths of latent\ndiffusion models for conditioned audio synthesis. We first encode and process\npaired audio-video data to make them suitable for generation. The encoded data\nis then used to train a model that can generate an audio track that captures\nthe semantics of the input video. Our proposed SyncroNet builds on top of\nControlNet to provide control signals that enables generation of temporally\nsynchronized audio. Extensive evaluations and a comprehensive user study show\nthat our model outperforms existing work in audio quality, and in our proposed\nsynchronization evaluation method. Furthermore, we demonstrate downstream\napplications of our model in improving video summarization.\n", "date": "2024-07-30", "categories": []}
{"id": "2408.15771", "title": "wav2pos: Sound Source Localization using Masked Autoencoders", "abstract": "  We present a novel approach to the 3D sound source localization task for\ndistributed ad-hoc microphone arrays by formulating it as a set-to-set\nregression problem. By training a multi-modal masked autoencoder model that\noperates on audio recordings and microphone coordinates, we show that such a\nformulation allows for accurate localization of the sound source, by\nreconstructing coordinates masked in the input. Our approach is flexible in the\nsense that a single model can be used with an arbitrary number of microphones,\neven when a subset of audio recordings and microphone coordinates are missing.\nWe test our method on simulated and real-world recordings of music and speech\nin indoor environments, and demonstrate competitive performance compared to\nboth classical and other learning based localization methods.\n", "date": "2024-08-28", "categories": []}
{"id": "2410.02144", "title": "SoundMorpher: Perceptually-Uniform Sound Morphing with Diffusion Model", "abstract": "  We present SoundMorpher, an open-world sound morphing method designed to\ngenerate perceptually uniform morphing trajectories. Traditional sound morphing\ntechniques typically assume a linear relationship between the morphing factor\nand sound perception, achieving smooth transitions by linearly interpolating\nthe semantic features of source and target sounds while gradually adjusting the\nmorphing factor. However, these methods oversimplify the complexities of sound\nperception, resulting in limitations in morphing quality. In contrast,\nSoundMorpher explores an explicit relationship between the morphing factor and\nthe perception of morphed sounds, leveraging log Mel-spectrogram features. This\napproach further refines the morphing sequence by ensuring a constant target\nperceptual difference for each transition and determining the corresponding\nmorphing factors using binary search. To address the lack of a formal\nquantitative evaluation framework for sound morphing, we propose a set of\nmetrics based on three established objective criteria. These metrics enable\ncomprehensive assessment of morphed results and facilitate direct comparisons\nbetween methods, fostering advancements in sound morphing research. Extensive\nexperiments demonstrate the effectiveness and versatility of SoundMorpher in\nreal-world scenarios, showcasing its potential in applications such as creative\nmusic composition, film post-production, and interactive audio technologies.\nOur demonstration and codes are available\nat~\\url{https://xinleiniu.github.io/SoundMorpher-demo/}.\n", "date": "2024-10-03", "categories": []}
{"id": "2410.09289", "title": "Multimodal Audio-based Disease Prediction with Transformer-based\n  Hierarchical Fusion Network", "abstract": "  Audio-based disease prediction is emerging as a promising supplement to\ntraditional medical diagnosis methods, facilitating early, convenient, and\nnon-invasive disease detection and prevention. Multimodal fusion, which\nintegrates features from various domains within or across bio-acoustic\nmodalities, has proven effective in enhancing diagnostic performance. However,\nmost existing methods in the field employ unilateral fusion strategies that\nfocus solely on either intra-modal or inter-modal fusion. This approach limits\nthe full exploitation of the complementary nature of diverse acoustic feature\ndomains and bio-acoustic modalities. Additionally, the inadequate and isolated\nexploration of latent dependencies within modality-specific and modality-shared\nspaces curtails their capacity to manage the inherent heterogeneity in\nmultimodal data. To fill these gaps, we propose a transformer-based\nhierarchical fusion network designed for general multimodal audio-based disease\nprediction. Specifically, we seamlessly integrate intra-modal and inter-modal\nfusion in a hierarchical manner and proficiently encode the necessary\nintra-modal and inter-modal complementary correlations, respectively.\nComprehensive experiments demonstrate that our model achieves state-of-the-art\nperformance in predicting three diseases: COVID-19, Parkinson's disease, and\npathological dysarthria, showcasing its promising potential in a broad context\nof audio-based disease prediction tasks. Additionally, extensive ablation\nstudies and qualitative analyses highlight the significant benefits of each\nmain component within our model.\n", "date": "2024-10-11", "categories": []}
{"id": "2412.10649", "title": "Hidden Echoes Survive Training in Audio To Audio Generative Instrument\n  Models", "abstract": "  As generative techniques pervade the audio domain, there has been increasing\ninterest in tracing back through these complicated models to understand how\nthey draw on their training data to synthesize new examples, both to ensure\nthat they use properly licensed data and also to elucidate their black box\nbehavior. In this paper, we show that if imperceptible echoes are hidden in the\ntraining data, a wide variety of audio to audio architectures (differentiable\ndigital signal processing (DDSP), Realtime Audio Variational autoEncoder\n(RAVE), and ``Dance Diffusion'') will reproduce these echoes in their outputs.\nHiding a single echo is particularly robust across all architectures, but we\nalso show promising results hiding longer time spread echo patterns for an\nincreased information capacity. We conclude by showing that echoes make their\nway into fine tuned models, that they survive mixing/demixing, and that they\nsurvive pitch shift augmentation during training. Hence, this simple, classical\nidea in watermarking shows significant promise for tagging generative audio\nmodels.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.11551", "title": "Region-Based Optimization in Continual Learning for Audio Deepfake\n  Detection", "abstract": "  Rapid advancements in speech synthesis and voice conversion bring convenience\nbut also new security risks, creating an urgent need for effective audio\ndeepfake detection. Although current models perform well, their effectiveness\ndiminishes when confronted with the diverse and evolving nature of real-world\ndeepfakes. To address this issue, we propose a continual learning method named\nRegion-Based Optimization (RegO) for audio deepfake detection. Specifically, we\nuse the Fisher information matrix to measure important neuron regions for real\nand fake audio detection, dividing them into four regions. First, we directly\nfine-tune the less important regions to quickly adapt to new tasks. Next, we\napply gradient optimization in parallel for regions important only to real\naudio detection, and in orthogonal directions for regions important only to\nfake audio detection. For regions that are important to both, we use sample\nproportion-based adaptive gradient optimization. This region-adaptive\noptimization ensures an appropriate trade-off between memory stability and\nlearning plasticity. Additionally, to address the increase of redundant neurons\nfrom old tasks, we further introduce the Ebbinghaus forgetting mechanism to\nrelease them, thereby promoting the capability of the model to learn more\ngeneralized discriminative features. Experimental results show our method\nachieves a 21.3% improvement in EER over the state-of-the-art continual\nlearning approach RWM for audio deepfake detection. Moreover, the effectiveness\nof RegO extends beyond the audio deepfake detection domain, showing potential\nsignificance in other tasks, such as image recognition. The code is available\nat https://github.com/cyjie429/RegO\n", "date": "2024-12-16", "categories": []}
{"id": "2412.11896", "title": "Classification of Spontaneous and Scripted Speech for Multilingual Audio", "abstract": "  Distinguishing scripted from spontaneous speech is an essential tool for\nbetter understanding how speech styles influence speech processing research. It\ncan also improve recommendation systems and discovery experiences for media\nusers through better segmentation of large recorded speech catalogues. This\npaper addresses the challenge of building a classifier that generalises well\nacross different formats and languages. We systematically evaluate models\nranging from traditional, handcrafted acoustic and prosodic features to\nadvanced audio transformers, utilising a large, multilingual proprietary\npodcast dataset for training and validation. We break down the performance of\neach model across 11 language groups to evaluate cross-lingual biases. Our\nexperimental analysis extends to publicly available datasets to assess the\nmodels' generalisability to non-podcast domains. Our results indicate that\ntransformer-based models consistently outperform traditional feature-based\ntechniques, achieving state-of-the-art performance in distinguishing between\nscripted and spontaneous speech across various languages.\n", "date": "2024-12-16", "categories": []}
{"id": "2409.06223", "title": "Enhancing Temporal Understanding in Audio Question Answering for Large\n  Audio Language Models", "abstract": "  The Audio Question Answering (AQA) task includes audio event classification,\naudio captioning, and open-ended reasoning. Recently, AQA has garnered\nattention due to the advent of Large Audio Language Models (LALMs). Current\nliterature focuses on constructing LALMs by integrating audio encoders with\ntext-only Large Language Models (LLMs) through a projection module. While LALMs\nexcel in general audio understanding, they are limited in temporal reasoning,\nwhich may hinder their commercial applications and on-device deployment. This\npaper addresses these challenges and limitations in audio temporal reasoning.\nFirst, we introduce a data augmentation technique for generating reliable audio\ntemporal questions and answers using an LLM. Second, we perform a further\nfine-tuning of an existing baseline using curriculum learning strategy to\nspecialize in temporal reasoning without compromising performance on fine-tuned\ntasks. We demonstrate the performance of our model using state-of-the-art LALMs\non public audio benchmark datasets. Third, we implement our AQA model on-device\nlocally and investigate its CPU inference for edge applications.\n", "date": "2024-09-10", "categories": []}
{"id": "2412.01408", "title": "Towards Cross-Lingual Audio Abuse Detection in Low-Resource Settings\n  with Few-Shot Learning", "abstract": "  Online abusive content detection, particularly in low-resource settings and\nwithin the audio modality, remains underexplored. We investigate the potential\nof pre-trained audio representations for detecting abusive language in\nlow-resource languages, in this case, in Indian languages using Few Shot\nLearning (FSL). Leveraging powerful representations from models such as Wav2Vec\nand Whisper, we explore cross-lingual abuse detection using the ADIMA dataset\nwith FSL. Our approach integrates these representations within the\nModel-Agnostic Meta-Learning (MAML) framework to classify abusive language in\n10 languages. We experiment with various shot sizes (50-200) evaluating the\nimpact of limited data on performance. Additionally, a feature visualization\nstudy was conducted to better understand model behaviour. This study highlights\nthe generalization ability of pre-trained models in low-resource scenarios and\noffers valuable insights into detecting abusive language in multilingual\ncontexts.\n", "date": "2024-12-02", "categories": []}
{"id": "2411.12008", "title": "Compression of Higher Order Ambisonics with Multichannel RVQGAN", "abstract": "  A multichannel extension to the RVQGAN neural coding method is proposed, and\nrealized for data-driven compression of third-order Ambisonics audio. The\ninput- and output layers of the generator and discriminator models are modified\nto accept multiple (16) channels without increasing the model bitrate. We also\npropose a loss function for accounting for spatial perception in immersive\nreproduction, and transfer learning from single-channel models. Listening test\nresults with 7.1.4 immersive playback show that the proposed extension is\nsuitable for coding scene-based, 16-channel Ambisonics content with good\nquality at 16 kbps when trained and tested on the EigenScape database. The\nmodel has potential applications for learning other types of content and\nmultichannel formats.\n", "date": "2024-11-18", "categories": []}
{"id": "2412.09168", "title": "YingSound: Video-Guided Sound Effects Generation with Multi-modal\n  Chain-of-Thought Controls", "abstract": "  Generating sound effects for product-level videos, where only a small amount\nof labeled data is available for diverse scenes, requires the production of\nhigh-quality sounds in few-shot settings. To tackle the challenge of limited\nlabeled data in real-world scenes, we introduce YingSound, a foundation model\ndesigned for video-guided sound generation that supports high-quality audio\ngeneration in few-shot settings. Specifically, YingSound consists of two major\nmodules. The first module uses a conditional flow matching transformer to\nachieve effective semantic alignment in sound generation across audio and\nvisual modalities. This module aims to build a learnable audio-visual\naggregator (AVA) that integrates high-resolution visual features with\ncorresponding audio features at multiple stages. The second module is developed\nwith a proposed multi-modal visual-audio chain-of-thought (CoT) approach to\ngenerate finer sound effects in few-shot settings. Finally, an\nindustry-standard video-to-audio (V2A) dataset that encompasses various\nreal-world scenarios is presented. We show that YingSound effectively generates\nhigh-quality synchronized sounds across diverse conditional inputs through\nautomated evaluations and human studies. Project Page:\n\\url{https://giantailab.github.io/yingsound/}\n", "date": "2024-12-12", "categories": []}
{"id": "2412.09467", "title": "Audios Don't Lie: Multi-Frequency Channel Attention Mechanism for Audio\n  Deepfake Detection", "abstract": "  With the rapid development of artificial intelligence technology, the\napplication of deepfake technology in the audio field has gradually increased,\nresulting in a wide range of security risks. Especially in the financial and\nsocial security fields, the misuse of deepfake audios has raised serious\nconcerns. To address this challenge, this study proposes an audio deepfake\ndetection method based on multi-frequency channel attention mechanism (MFCA)\nand 2D discrete cosine transform (DCT). By processing the audio signal into a\nmelspectrogram, using MobileNet V2 to extract deep features, and combining it\nwith the MFCA module to weight different frequency channels in the audio\nsignal, this method can effectively capture the fine-grained frequency domain\nfeatures in the audio signal and enhance the Classification capability of fake\naudios. Experimental results show that compared with traditional methods, the\nmodel proposed in this study shows significant advantages in accuracy,\nprecision,recall, F1 score and other indicators. Especially in complex audio\nscenarios, this method shows stronger robustness and generalization\ncapabilities and provides a new idea for audio deepfake detection and has\nimportant practical application value. In the future, more advanced audio\ndetection technologies and optimization strategies will be explored to further\nimprove the accuracy and generalization capabilities of audio deepfake\ndetection.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.07406", "title": "Learning Self-Supervised Audio-Visual Representations for Sound\n  Recommendations", "abstract": "  We propose a novel self-supervised approach for learning audio and visual\nrepresentations from unlabeled videos, based on their correspondence. The\napproach uses an attention mechanism to learn the relative importance of\nconvolutional features extracted at different resolutions from the audio and\nvisual streams and uses the attention features to encode the audio and visual\ninput based on their correspondence. We evaluated the representations learned\nby the model to classify audio-visual correlation as well as to recommend sound\neffects for visual scenes. Our results show that the representations generated\nby the attention model improves the correlation accuracy compared to the\nbaseline, by 18% and the recommendation accuracy by 10% for VGG-Sound, which is\na public video dataset. Additionally, audio-visual representations learned by\ntraining the attention model with cross-modal contrastive learning further\nimproves the recommendation performance, based on our evaluation using\nVGG-Sound and a more challenging dataset consisting of gameplay video\nrecordings.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.05951", "title": "When Vision Models Meet Parameter Efficient Look-Aside Adapters Without\n  Large-Scale Audio Pretraining", "abstract": "  Recent studies show that pretrained vision models can boost performance in\naudio downstream tasks. To enhance the performance further, an additional\npretraining stage with large scale audio data is typically required to infuse\naudio specific knowledge into the vision model. However, such approaches\nrequire extensive audio data and a carefully designed objective function. In\nthis work, we propose bypassing the pretraining stage by directly fine-tuning\nthe vision model with our Look Aside Adapter (LoAA) designed for efficient\naudio understanding. Audio spectrum data is represented across two\nheterogeneous dimensions time and frequency and we refine adapters to\nfacilitate interactions between tokens across these dimensions. Our experiments\ndemonstrate that our adapters allow vision models to reach or surpass the\nperformance of pretrained audio models in various audio and speech tasks,\noffering a resource efficient and effective solution for leveraging vision\nmodels in audio applications.\n", "date": "2024-12-08", "categories": []}
{"id": "2412.06703", "title": "Source Separation & Automatic Transcription for Music", "abstract": "  Source separation is the process of isolating individual sounds in an\nauditory mixture of multiple sounds [1], and has a variety of applications\nranging from speech enhancement and lyric transcription [2] to digital audio\nproduction for music. Furthermore, Automatic Music Transcription (AMT) is the\nprocess of converting raw music audio into sheet music that musicians can read\n[3]. Historically, these tasks have faced challenges such as significant audio\nnoise, long training times, and lack of free-use data due to copyright\nrestrictions. However, recent developments in deep learning have brought new\npromising approaches to building low-distortion stems and generating sheet\nmusic from audio signals [4]. Using spectrogram masking, deep neural networks,\nand the MuseScore API, we attempt to create an end-to-end pipeline that allows\nfor an initial music audio mixture (e.g...wav file) to be separated into\ninstrument stems, converted into MIDI files, and transcribed into sheet music\nfor each component instrument.\n", "date": "2024-12-09", "categories": []}
{"id": "2401.05531", "title": "VI-PANN: Harnessing Transfer Learning and Uncertainty-Aware Variational\n  Inference for Improved Generalization in Audio Pattern Recognition", "abstract": "  Transfer learning (TL) is an increasingly popular approach to training deep\nlearning (DL) models that leverages the knowledge gained by training a\nfoundation model on diverse, large-scale datasets for use on downstream tasks\nwhere less domain- or task-specific data is available. The literature is rich\nwith TL techniques and applications; however, the bulk of the research makes\nuse of deterministic DL models which are often uncalibrated and lack the\nability to communicate a measure of epistemic (model) uncertainty in\nprediction. Unlike their deterministic counterparts, Bayesian DL (BDL) models\nare often well-calibrated, provide access to epistemic uncertainty for a\nprediction, and are capable of achieving competitive predictive performance. In\nthis study, we propose variational inference pre-trained audio neural networks\n(VI-PANNs). VI-PANNs are a variational inference variant of the popular\nResNet-54 architecture which are pre-trained on AudioSet, a large-scale audio\nevent detection dataset. We evaluate the quality of the resulting uncertainty\nwhen transferring knowledge from VI-PANNs to other downstream acoustic\nclassification tasks using the ESC-50, UrbanSound8K, and DCASE2013 datasets. We\ndemonstrate, for the first time, that it is possible to transfer calibrated\nuncertainty information along with knowledge from upstream tasks to enhance a\nmodel's capability to perform downstream tasks.\n", "date": "2024-01-10", "categories": []}
{"id": "2412.03267", "title": "Detecting abnormal heart sound using mobile phones and on-device IConNet", "abstract": "  Given the global prevalence of cardiovascular diseases, there is a pressing\nneed for easily accessible early screening methods. Typically, this requires\nmedical practitioners to investigate heart auscultations for irregular sounds,\nfollowed by echocardiography and electrocardiography tests. To democratize\nearly diagnosis, we present a user-friendly solution for abnormal heart sound\ndetection, utilizing mobile phones and a lightweight neural network optimized\nfor on-device inference. Unlike previous approaches reliant on specialized\nstethoscopes, our method directly analyzes audio recordings, facilitated by a\nnovel architecture known as IConNet. IConNet, an Interpretable Convolutional\nNeural Network, harnesses insights from audio signal processing, enhancing\nefficiency and providing transparency in neural pattern extraction from raw\nwaveform signals. This is a significant step towards trustworthy AI in\nhealthcare, aiding in remote health monitoring efforts.\n", "date": "2024-12-04", "categories": []}
{"id": "2308.05037", "title": "Separate Anything You Describe", "abstract": "  Language-queried audio source separation (LASS) is a new paradigm for\ncomputational auditory scene analysis (CASA). LASS aims to separate a target\nsound from an audio mixture given a natural language query, which provides a\nnatural and scalable interface for digital audio applications. Recent works on\nLASS, despite attaining promising separation performance on specific sources\n(e.g., musical instruments, limited classes of audio events), are unable to\nseparate audio concepts in the open domain. In this work, we introduce\nAudioSep, a foundation model for open-domain audio source separation with\nnatural language queries. We train AudioSep on large-scale multimodal datasets\nand extensively evaluate its capabilities on numerous tasks including audio\nevent separation, musical instrument separation, and speech enhancement.\nAudioSep demonstrates strong separation performance and impressive zero-shot\ngeneralization ability using audio captions or text labels as queries,\nsubstantially outperforming previous audio-queried and language-queried sound\nseparation models. For reproducibility of this work, we will release the source\ncode, evaluation benchmark and pre-trained model at:\nhttps://github.com/Audio-AGI/AudioSep.\n", "date": "2023-08-09", "categories": []}
{"id": "2412.00312", "title": "Raw Audio Classification with Cosine Convolutional Neural Network\n  (CosCovNN)", "abstract": "  This study explores the field of audio classification from raw waveform using\nConvolutional Neural Networks (CNNs), a method that eliminates the need for\nextracting specialised features in the pre-processing step. Unlike recent\ntrends in literature, which often focuses on designing frontends or filters for\nonly the initial layers of CNNs, our research introduces the Cosine\nConvolutional Neural Network (CosCovNN) replacing the traditional CNN filters\nwith Cosine filters. The CosCovNN surpasses the accuracy of the equivalent CNN\narchitectures with approximately $77\\%$ less parameters. Our research further\nprogresses with the development of an augmented CosCovNN named Vector Quantised\nCosine Convolutional Neural Network with Memory (VQCCM), incorporating a memory\nand vector quantisation layer VQCCM achieves state-of-the-art (SOTA)\nperformance across five different datasets in comparison with existing\nliterature. Our findings show that cosine filters can greatly improve the\nefficiency and accuracy of CNNs in raw audio classification.\n", "date": "2024-11-30", "categories": []}
{"id": "2412.01425", "title": "Reject Threshold Adaptation for Open-Set Model Attribution of Deepfake\n  Audio", "abstract": "  Open environment oriented open set model attribution of deepfake audio is an\nemerging research topic, aiming to identify the generation models of deepfake\naudio. Most previous work requires manually setting a rejection threshold for\nunknown classes to compare with predicted probabilities. However, models often\noverfit training instances and generate overly confident predictions. Moreover,\nthresholds that effectively distinguish unknown categories in the current\ndataset may not be suitable for identifying known and unknown categories in\nanother data distribution. To address the issues, we propose a novel framework\nfor open set model attribution of deepfake audio with rejection threshold\nadaptation (ReTA). Specifically, the reconstruction error learning module\ntrains by combining the representation of system fingerprints with labels\ncorresponding to either the target class or a randomly chosen other class\nlabel. This process generates matching and non-matching reconstructed samples,\nestablishing the reconstruction error distributions for each class and laying\nthe foundation for the reject threshold calculation module. The reject\nthreshold calculation module utilizes gaussian probability estimation to fit\nthe distributions of matching and non-matching reconstruction errors. It then\ncomputes adaptive reject thresholds for all classes through probability\nminimization criteria. The experimental results demonstrate the effectiveness\nof ReTA in improving the open set model attributes of deepfake audio.\n", "date": "2024-12-02", "categories": []}
{"id": "2405.00233", "title": "SemantiCodec: An Ultra Low Bitrate Semantic Audio Codec for General\n  Sound", "abstract": "  Large language models (LLMs) have significantly advanced audio processing\nthrough audio codecs that convert audio into discrete tokens, enabling the\napplication of language modelling techniques to audio data. However,\ntraditional codecs often operate at high bitrates or within narrow domains such\nas speech and lack the semantic clues required for efficient language\nmodelling. Addressing these challenges, we introduce SemantiCodec, a novel\ncodec designed to compress audio into fewer than a hundred tokens per second\nacross diverse audio types, including speech, general sound, and music, without\ncompromising quality. SemantiCodec features a dual-encoder architecture: a\nsemantic encoder using a self-supervised pre-trained Audio Masked Autoencoder\n(AudioMAE), discretized using k-means clustering on extensive audio data, and\nan acoustic encoder to capture the remaining details. The semantic and acoustic\nencoder outputs are used to reconstruct audio via a diffusion-model-based\ndecoder. SemantiCodec is presented in three variants with token rates of 25,\n50, and 100 per second, supporting a range of ultra-low bit rates between 0.31\nkbps and 1.40 kbps. Experimental results demonstrate that SemantiCodec\nsignificantly outperforms the state-of-the-art Descript codec on reconstruction\nquality. Our results also suggest that SemantiCodec contains significantly\nricher semantic information than all evaluated state-of-the-art audio codecs,\neven at significantly lower bitrates. Our code and demos are available at\nhttps://haoheliu.github.io/SemantiCodec/.\n", "date": "2024-04-30", "categories": []}
{"id": "2411.19842", "title": "Scaling Transformers for Low-Bitrate High-Quality Speech Coding", "abstract": "  The tokenization of speech with neural audio codec models is a vital part of\nmodern AI pipelines for the generation or understanding of speech, alone or in\na multimodal context. Traditionally such tokenization models have concentrated\non low parameter-count architectures using only components with strong\ninductive biases. In this work we show that by scaling a transformer\narchitecture with large parameter count to this problem, and applying a\nflexible Finite Scalar Quantization (FSQ) based bottleneck, it is possible to\nreach state-of-the-art speech quality at extremely low bit-rates of $400$ or\n$700$ bits-per-second. The trained models strongly out-perform existing\nbaselines in both objective and subjective tests.\n", "date": "2024-11-29", "categories": []}
{"id": "2406.11248", "title": "Performance Improvement of Language-Queried Audio Source Separation\n  Based on Caption Augmentation From Large Language Models for DCASE Challenge\n  2024 Task 9", "abstract": "  We present a prompt-engineering-based text-augmentation approach applied to a\nlanguage-queried audio source separation (LASS) task. To enhance the\nperformance of LASS, the proposed approach utilizes large language models\n(LLMs) to generate multiple captions corresponding to each sentence of the\ntraining dataset. To this end, we first perform experiments to identify the\nmost effective prompts for caption augmentation with a smaller number of\ncaptions. A LASS model trained with these augmented captions demonstrates\nimproved performance on the DCASE 2024 Task 9 validation set compared to that\ntrained without augmentation. This study highlights the effectiveness of\nLLM-based caption augmentation in advancing language-queried audio source\nseparation.\n", "date": "2024-06-17", "categories": []}
{"id": "2409.11369", "title": "Learning Spatially-Aware Language and Audio Embeddings", "abstract": "  Humans can picture a sound scene given an imprecise natural language\ndescription. For example, it is easy to imagine an acoustic environment given a\nphrase like \"the lion roar came from right behind me!\". For a machine to have\nthe same degree of comprehension, the machine must know what a lion is\n(semantic attribute), what the concept of \"behind\" is (spatial attribute) and\nhow these pieces of linguistic information align with the semantic and spatial\nattributes of the sound (what a roar sounds like when its coming from behind).\nState-of-the-art audio foundation models which learn to map between audio\nscenes and natural textual descriptions, are trained on non-spatial audio and\ntext pairs, and hence lack spatial awareness. In contrast, sound event\nlocalization and detection models are limited to recognizing sounds from a\nfixed number of classes, and they localize the source to absolute position\n(e.g., 0.2m) rather than a position described using natural language (e.g.,\n\"next to me\"). To address these gaps, we present ELSA a spatially aware-audio\nand text embedding model trained using multimodal contrastive learning. ELSA\nsupports non-spatial audio, spatial audio, and open vocabulary text captions\ndescribing both the spatial and semantic components of sound. To train ELSA:\n(a) we spatially augment the audio and captions of three open-source audio\ndatasets totaling 4,738 hours of audio, and (b) we design an encoder to capture\nthe semantics of non-spatial audio, and the semantics and spatial attributes of\nspatial audio using contrastive learning. ELSA is competitive with\nstate-of-the-art for both semantic retrieval and 3D source localization. In\nparticular, ELSA achieves +2.8% mean audio-to-text and text-to-audio R@1 above\nthe baseline, and outperforms by -11.6{\\deg} mean-absolute-error in 3D source\nlocalization over the baseline.\n", "date": "2024-09-17", "categories": []}
{"id": "2411.15685", "title": "State-Space Large Audio Language Models", "abstract": "  Large Audio Language Models (LALM) combine the audio perception models and\nthe Large Language Models (LLM) and show a remarkable ability to reason about\nthe input audio, infer the meaning, and understand the intent. However, these\nsystems rely on Transformers which scale quadratically with the input sequence\nlengths which poses computational challenges in deploying these systems in\nmemory and time-constrained scenarios. Recently, the state-space models (SSMs)\nhave emerged as an alternative to transformer networks.\n  While there have been successful attempts to replace transformer-based audio\nperception models with state-space ones, state-space-based LALMs remain\nunexplored. First, we begin by replacing the transformer-based audio perception\nmodule and then replace the transformer-based LLM and propose the first\nstate-space-based LALM. Experimental results demonstrate that space-based LALM\ndespite having a significantly lower number of parameters performs\ncompetitively with transformer-based LALMs on close-ended tasks on a variety of\ndatasets.\n", "date": "2024-11-24", "categories": []}
{"id": "2411.14474", "title": "Attention-guided Spectrogram Sequence Modeling with CNNs for Music Genre\n  Classification", "abstract": "  Music genre classification is a critical component of music recommendation\nsystems, generation algorithms, and cultural analytics. In this work, we\npresent an innovative model for classifying music genres using attention-based\ntemporal signature modeling. By processing spectrogram sequences through\nConvolutional Neural Networks (CNNs) and multi-head attention layers, our\napproach captures the most temporally significant moments within each piece,\ncrafting a unique \"signature\" for genre identification. This temporal focus not\nonly enhances classification accuracy but also reveals insights into\ngenre-specific characteristics that can be intuitively mapped to listener\nperceptions. Our findings offer potential applications in personalized music\nrecommendation systems by highlighting cross-genre similarities and\ndistinctiveness, aligning closely with human musical intuition. This work\nbridges the gap between technical classification tasks and the nuanced, human\nexperience of genre.\n", "date": "2024-11-18", "categories": []}
{"id": "2411.14627", "title": "Generative AI for Music and Audio", "abstract": "  Generative AI has been transforming the way we interact with technology and\nconsume content. In the next decade, AI technology will reshape how we create\naudio content in various media, including music, theater, films, games,\npodcasts, and short videos. In this dissertation, I introduce the three main\ndirections of my research centered around generative AI for music and audio: 1)\nmultitrack music generation, 2) assistive music creation tools, and 3)\nmultimodal learning for audio and music. Through my research, I aim to answer\nthe following two fundamental questions: 1) How can AI help professionals or\namateurs create music and audio content? 2) Can AI learn to create music in a\nway similar to how humans learn music? My long-term goal is to lower the\nbarrier of entry for music composition and democratize audio content creation\n", "date": "2024-11-21", "categories": []}
{"id": "2411.14972", "title": "Open-Amp: Synthetic Data Framework for Audio Effect Foundation Models", "abstract": "  This paper introduces Open-Amp, a synthetic data framework for generating\nlarge-scale and diverse audio effects data. Audio effects are relevant to many\nmusical audio processing and Music Information Retrieval (MIR) tasks, such as\nmodelling of analog audio effects, automatic mixing, tone matching and\ntranscription. Existing audio effects datasets are limited in scope, usually\nincluding relatively few audio effects processors and a limited amount of input\naudio signals. Our proposed framework overcomes these issues, by crowdsourcing\nneural network emulations of guitar amplifiers and effects, created by users of\nopen-source audio effects emulation software. This allows users of Open-Amp\ncomplete control over the input signals to be processed by the effects models,\nas well as providing high-quality emulations of hundreds of devices. Open-Amp\ncan render audio online during training, allowing great flexibility in data\naugmentation. Our experiments show that using Open-Amp to train a guitar\neffects encoder achieves new state-of-the-art results on multiple guitar\neffects classification tasks. Furthermore, we train a one-to-many guitar\neffects model using Open-Amp, and use it to emulate unseen analog effects via\nmanipulation of its learned latent space, indicating transferability to analog\nguitar effects data.\n", "date": "2024-11-22", "categories": []}
{"id": "2411.15082", "title": "Towards Speaker Identification with Minimal Dataset and Constrained\n  Resources using 1D-Convolution Neural Network", "abstract": "  Voice recognition and speaker identification are vital for applications in\nsecurity and personal assistants. This paper presents a lightweight\n1D-Convolutional Neural Network (1D-CNN) designed to perform speaker\nidentification on minimal datasets. Our approach achieves a validation accuracy\nof 97.87%, leveraging data augmentation techniques to handle background noise\nand limited training samples. Future improvements include testing on larger\ndatasets and integrating transfer learning methods to enhance generalizability.\nWe provide all code, the custom dataset, and the trained models to facilitate\nreproducibility. These resources are available on our GitHub repository:\nhttps://github.com/IrfanNafiz/RecMe.\n", "date": "2024-11-22", "categories": []}
{"id": "2404.02702", "title": "PSCodec: A Series of High-Fidelity Low-bitrate Neural Speech Codecs\n  Leveraging Prompt Encoders", "abstract": "  Neural speech codecs have recently emerged as a focal point in the fields of\nspeech compression and generation. Despite this progress, achieving\nhigh-quality speech reconstruction under low-bitrate scenarios remains a\nsignificant challenge. In this paper, we propose PSCodec, a series of neural\nspeech codecs based on prompt encoders, comprising PSCodec-Base,\nPSCodec-DRL-ICT, and PSCodec-CasAN, which are capable of delivering\nhigh-performance speech reconstruction with low bandwidths. Specifically, we\nfirst introduce PSCodec-Base, which leverages a pretrained speaker verification\nmodel-based prompt encoder (VPP-Enc) and a learnable Mel-spectrogram-based\nprompt encoder (MelP-Enc) to effectively disentangle and integrate voiceprint\nand Mel-related features in utterances. To further enhance feature utilization\nefficiency, we propose PSCodec-DRL-ICT, incorporating a structural similarity\n(SSIM) based disentangled representation loss (DRL) and an incremental\ncontinuous training (ICT) strategy. While PSCodec-DRL-ICT demonstrates\nimpressive performance, its reliance on extensive hyperparameter tuning and\nmulti-stage training makes it somewhat labor-intensive. To circumvent these\nlimitations, we propose PSCodec-CasAN, utilizing an advanced cascaded attention\nnetwork (CasAN) to enhance representational capacity of the entire system.\nExtensive experiments show that our proposed PSCodec-Base, PSCodec-DRL-ICT, and\nPSCodec-CasAN all significantly outperform several state-of-the-art neural\ncodecs, exhibiting substantial improvements in both speech reconstruction\nquality and speaker similarity under low-bitrate conditions.\n", "date": "2024-04-03", "categories": []}
{"id": "2411.13592", "title": "A Novel Speech Analysis and Correction Tool for Arabic-Speaking Children", "abstract": "  This paper introduces a new application named ArPA for Arabic kids who have\ntrouble with pronunciation. Our application comprises two key components: the\ndiagnostic module and the therapeutic module. The diagnostic process involves\ncapturing the child's speech signal, preprocessing, and analyzing it using\ndifferent machine learning classifiers like K-Nearest Neighbors (KNN), Support\nVector Machine (SVM), and Decision Trees as well as deep neural network\nclassifiers like ResNet18. The therapeutic module offers eye-catching gamified\ninterfaces in which each correctly spoken letter earns a higher avatar level,\nproviding positive reinforcement for the child's pronunciation improvement. Two\ndatasets were used for experimental evaluation: one from a childcare centre and\nthe other including Arabic alphabet pronunciation recordings. Our work uses a\nnovel technique for speech recognition using Melspectrogram and MFCC images.\nThe results show that the ResNet18 classifier on speech-to-image converted data\neffectively identifies mispronunciations in Arabic speech with an accuracy of\n99.015\\% with Mel-Spectrogram images outperforming ResNet18 with MFCC images.\n", "date": "2024-11-18", "categories": []}
{"id": "2411.13179", "title": "SONNET: Enhancing Time Delay Estimation by Leveraging Simulated Audio", "abstract": "  Time delay estimation or Time-Difference-Of-Arrival estimates is a critical\ncomponent for multiple localization applications such as multilateration,\ndirection of arrival, and self-calibration. The task is to estimate the time\ndifference between a signal arriving at two different sensors. For the audio\nsensor modality, most current systems are based on classical methods such as\nthe Generalized Cross-Correlation Phase Transform (GCC-PHAT) method. In this\npaper we demonstrate that learning based methods can, even based on synthetic\ndata, significantly outperform GCC-PHAT on novel real world data. To overcome\nthe lack of data with ground truth for the task, we train our model on a\nsimulated dataset which is sufficiently large and varied, and that captures the\nrelevant characteristics of the real world problem. We provide our trained\nmodel, SONNET (Simulation Optimized Neural Network Estimator of Timeshifts),\nwhich is runnable in real-time and works on novel data out of the box for many\nreal data applications, i.e. without re-training. We further demonstrate\ngreatly improved performance on the downstream task of self-calibration when\nusing our model compared to classical methods.\n", "date": "2024-11-20", "categories": []}
{"id": "2208.10489", "title": "Audio Deepfake Attribution: An Initial Dataset and Investigation", "abstract": "  The rapid progress of deep speech synthesis models has posed significant\nthreats to society such as malicious manipulation of content. This has led to\nan increase in studies aimed at detecting so-called deepfake audio. However,\nexisting works focus on the binary detection of real audio and fake audio. In\nreal-world scenarios such as model copyright protection and digital evidence\nforensics, binary classification alone is insufficient. It is essential to\nidentify the source of deepfake audio. Therefore, audio deepfake attribution\nhas emerged as a new challenge. To this end, we designed the first deepfake\naudio dataset for the attribution of audio generation tools, called Audio\nDeepfake Attribution (ADA), and conducted a comprehensive investigation on\nsystem fingerprints. To address the challenges of attribution of continuously\nemerging unknown audio generation tools in the real world, we propose the\nClass-Representation Multi-Center Learning (CRML) method for open-set audio\ndeepfake attribution (OSADA). CRML enhances the global directional variation of\nrepresentations, ensuring the learning of discriminative representations with\nstrong intra-class similarity and inter-class discrepancy among known classes.\nFinally, the strong class discrimination capability learned from known classes\nis extended to both known and unknown classes. Experimental results demonstrate\nthat the CRML method effectively addresses open-set risks in real-world\nscenarios. The dataset is publicly available at:\nhttps://zenodo.org/records/13318702, and https://zenodo.org/records/13340666.\n", "date": "2022-08-21", "categories": []}
{"id": "2404.10420", "title": "AudioProtoPNet: An interpretable deep learning model for bird sound\n  classification", "abstract": "  Deep learning models have significantly advanced acoustic bird monitoring by\nbeing able to recognize numerous bird species based on their vocalizations.\nHowever, traditional deep learning models are black boxes that provide no\ninsight into their underlying computations, limiting their usefulness to\nornithologists and machine learning engineers. Explainable models could\nfacilitate debugging, knowledge discovery, trust, and interdisciplinary\ncollaboration. This study introduces AudioProtoPNet, an adaptation of the\nPrototypical Part Network (ProtoPNet) for multi-label bird sound\nclassification. It is an inherently interpretable model that uses a ConvNeXt\nbackbone to extract embeddings, with the classification layer replaced by a\nprototype learning classifier trained on these embeddings. The classifier\nlearns prototypical patterns of each bird species' vocalizations from\nspectrograms of training instances. During inference, audio recordings are\nclassified by comparing them to the learned prototypes in the embedding space,\nproviding explanations for the model's decisions and insights into the most\ninformative embeddings of each bird species. The model was trained on the\nBirdSet training dataset, which consists of 9,734 bird species and over 6,800\nhours of recordings. Its performance was evaluated on the seven test datasets\nof BirdSet, covering different geographical regions. AudioProtoPNet\noutperformed the state-of-the-art model Perch, achieving an average AUROC of\n0.90 and a cmAP of 0.42, with relative improvements of 7.1% and 16.7% over\nPerch, respectively. These results demonstrate that even for the challenging\ntask of multi-label bird sound classification, it is possible to develop\npowerful yet inherently interpretable deep learning models that provide\nvaluable insights for ornithologists and machine learning engineers.\n", "date": "2024-04-16", "categories": []}
{"id": "2402.09821", "title": "Diffusion Models for Audio Restoration", "abstract": "  With the development of audio playback devices and fast data transmission,\nthe demand for high sound quality is rising for both entertainment and\ncommunications. In this quest for better sound quality, challenges emerge from\ndistortions and interferences originating at the recording side or caused by an\nimperfect transmission pipeline. To address this problem, audio restoration\nmethods aim to recover clean sound signals from the corrupted input data. We\npresent here audio restoration algorithms based on diffusion models, with a\nfocus on speech enhancement and music restoration tasks. Traditional\napproaches, often grounded in handcrafted rules and statistical heuristics,\nhave shaped our understanding of audio signals. In the past decades, there has\nbeen a notable shift towards data-driven methods that exploit the modeling\ncapabilities of DNNs. Deep generative models, and among them diffusion models,\nhave emerged as powerful techniques for learning complex data distributions.\nHowever, relying solely on DNN-based learning approaches carries the risk of\nreducing interpretability, particularly when employing end-to-end models.\nNonetheless, data-driven approaches allow more flexibility in comparison to\nstatistical model-based frameworks, whose performance depends on distributional\nand statistical assumptions that can be difficult to guarantee. Here, we aim to\nshow that diffusion models can combine the best of both worlds and offer the\nopportunity to design audio restoration algorithms with a good degree of\ninterpretability and a remarkable performance in terms of sound quality. We\nexplain the diffusion formalism and its application to the conditional\ngeneration of clean audio signals. We believe that diffusion models open an\nexciting field of research with the potential to spawn new audio restoration\nalgorithms that are natural-sounding and remain robust in difficult acoustic\nsituations.\n", "date": "2024-02-15", "categories": []}
{"id": "2404.09466", "title": "Scoring Time Intervals using Non-Hierarchical Transformer For Automatic\n  Piano Transcription", "abstract": "  The neural semi-Markov Conditional Random Field (semi-CRF) framework has\ndemonstrated promise for event-based piano transcription. In this framework,\nall events (notes or pedals) are represented as closed time intervals tied to\nspecific event types. The neural semi-CRF approach requires an interval scoring\nmatrix that assigns a score for every candidate interval. However, designing an\nefficient and expressive architecture for scoring intervals is not trivial.\nThis paper introduces a simple method for scoring intervals using scaled inner\nproduct operations that resemble how attention scoring is done in transformers.\nWe show theoretically that, due to the special structure from encoding the\nnon-overlapping intervals, under a mild condition, the inner product operations\nare expressive enough to represent an ideal scoring matrix that can yield the\ncorrect transcription result. We then demonstrate that an encoder-only\nstructured non-hierarchical transformer backbone, operating only on a\nlow-time-resolution feature map, is capable of transcribing piano notes and\npedals with high accuracy and time precision. The experiment shows that our\napproach achieves the new state-of-the-art performance across all subtasks in\nterms of the F1 measure on the Maestro dataset.\n", "date": "2024-04-15", "categories": []}
{"id": "2411.05969", "title": "Toward Transdisciplinary Approaches to Audio Deepfake Discernment", "abstract": "  This perspective calls for scholars across disciplines to address the\nchallenge of audio deepfake detection and discernment through an\ninterdisciplinary lens across Artificial Intelligence methods and linguistics.\nWith an avalanche of tools for the generation of realistic-sounding fake speech\non one side, the detection of deepfakes is lagging on the other. Particularly\nhindering audio deepfake detection is the fact that current AI models lack a\nfull understanding of the inherent variability of language and the complexities\nand uniqueness of human speech. We see the promising potential in recent\ntransdisciplinary work that incorporates linguistic knowledge into AI\napproaches to provide pathways for expert-in-the-loop and to move beyond expert\nagnostic AI-based methods for more robust and comprehensive deepfake detection.\n", "date": "2024-11-08", "categories": []}
{"id": "2406.16148", "title": "Towards Open Respiratory Acoustic Foundation Models: Pretraining and\n  Benchmarking", "abstract": "  Respiratory audio, such as coughing and breathing sounds, has predictive\npower for a wide range of healthcare applications, yet is currently\nunder-explored. The main problem for those applications arises from the\ndifficulty in collecting large labeled task-specific data for model\ndevelopment. Generalizable respiratory acoustic foundation models pretrained\nwith unlabeled data would offer appealing advantages and possibly unlock this\nimpasse. However, given the safety-critical nature of healthcare applications,\nit is pivotal to also ensure openness and replicability for any proposed\nfoundation model solution. To this end, we introduce OPERA, an OPEn Respiratory\nAcoustic foundation model pretraining and benchmarking system, as the first\napproach answering this need. We curate large-scale respiratory audio datasets\n(~136K samples, over 400 hours), pretrain three pioneering foundation models,\nand build a benchmark consisting of 19 downstream respiratory health tasks for\nevaluation. Our pretrained models demonstrate superior performance (against\nexisting acoustic models pretrained with general audio on 16 out of 19 tasks)\nand generalizability (to unseen datasets and new respiratory audio modalities).\nThis highlights the great promise of respiratory acoustic foundation models and\nencourages more studies using OPERA as an open resource to accelerate research\non respiratory audio for health. The system is accessible from\nhttps://github.com/evelyn0414/OPERA.\n", "date": "2024-06-23", "categories": []}
{"id": "2411.04142", "title": "Unified Pathological Speech Analysis with Prompt Tuning", "abstract": "  Pathological speech analysis has been of interest in the detection of certain\ndiseases like depression and Alzheimer's disease and attracts much interest\nfrom researchers. However, previous pathological speech analysis models are\ncommonly designed for a specific disease while overlooking the connection\nbetween diseases, which may constrain performance and lower training\nefficiency. Instead of fine-tuning deep models for different tasks, prompt\ntuning is a much more efficient training paradigm. We thus propose a unified\npathological speech analysis system for as many as three diseases with the\nprompt tuning technique. This system uses prompt tuning to adjust only a small\npart of the parameters to detect different diseases from speeches of possible\npatients. Our system leverages a pre-trained spoken language model and\ndemonstrates strong performance across multiple disorders while only\nfine-tuning a fraction of the parameters. This efficient training approach\nleads to faster convergence and improved F1 scores by allowing knowledge to be\nshared across tasks. Our experiments on Alzheimer's disease, Depression, and\nParkinson's disease show competitive results, highlighting the effectiveness of\nour method in pathological speech analysis.\n", "date": "2024-11-05", "categories": []}
{"id": "2407.11745", "title": "Universal Sound Separation with Self-Supervised Audio Masked Autoencoder", "abstract": "  Universal sound separation (USS) is a task of separating mixtures of\narbitrary sound sources. Typically, universal separation models are trained\nfrom scratch in a supervised manner, using labeled data. Self-supervised\nlearning (SSL) is an emerging deep learning approach that leverages unlabeled\ndata to obtain task-agnostic representations, which can benefit many downstream\ntasks. In this paper, we propose integrating a self-supervised pre-trained\nmodel, namely the audio masked autoencoder (A-MAE), into a universal sound\nseparation system to enhance its separation performance. We employ two\nstrategies to utilize SSL embeddings: freezing or updating the parameters of\nA-MAE during fine-tuning. The SSL embeddings are concatenated with the\nshort-time Fourier transform (STFT) to serve as input features for the\nseparation model. We evaluate our methods on the AudioSet dataset, and the\nexperimental results indicate that the proposed methods successfully enhance\nthe separation performance of a state-of-the-art ResUNet-based USS model.\n", "date": "2024-07-16", "categories": []}
{"id": "2409.18680", "title": "Beyond Single-Audio: Advancing Multi-Audio Processing in Audio Large\n  Language Models", "abstract": "  Various audio-LLMs (ALLMs) have been explored recently for tackling different\naudio tasks simultaneously using a single, unified model. While existing\nevaluations of ALLMs primarily focus on single-audio tasks, real-world\napplications often involve processing multiple audio streams simultaneously. To\nbridge this gap, we propose the first multi-audio evaluation (MAE) benchmark\nthat consists of 20 datasets from 11 multi-audio tasks encompassing both speech\nand sound scenarios. Comprehensive experiments on MAE demonstrate that the\nexisting ALLMs, while being powerful in comprehending primary audio elements in\nindividual audio inputs, struggling to handle multi-audio scenarios. To this\nend, we propose a novel multi-audio-LLM (MALLM) to capture audio context among\nmultiple similar audios using discriminative learning on our proposed synthetic\ndata. The results demonstrate that the proposed MALLM outperforms all baselines\nand achieves high data efficiency using synthetic data without requiring human\nannotations. The proposed MALLM opens the door for ALLMs towards multi-audio\nprocessing era and brings us closer to replicating human auditory capabilities\nin machines.\n", "date": "2024-09-27", "categories": []}
{"id": "2411.02964", "title": "Speaker Emotion Recognition: Leveraging Self-Supervised Models for\n  Feature Extraction Using Wav2Vec2 and HuBERT", "abstract": "  Speech is the most natural way of expressing ourselves as humans. Identifying\nemotion from speech is a nontrivial task due to the ambiguous definition of\nemotion itself. Speaker Emotion Recognition (SER) is essential for\nunderstanding human emotional behavior. The SER task is challenging due to the\nvariety of speakers, background noise, complexity of emotions, and speaking\nstyles. It has many applications in education, healthcare, customer service,\nand Human-Computer Interaction (HCI). Previously, conventional machine learning\nmethods such as SVM, HMM, and KNN have been used for the SER task. In recent\nyears, deep learning methods have become popular, with convolutional neural\nnetworks and recurrent neural networks being used for SER tasks. The input of\nthese methods is mostly spectrograms and hand-crafted features. In this work,\nwe study the use of self-supervised transformer-based models, Wav2Vec2 and\nHuBERT, to determine the emotion of speakers from their voice. The models\nautomatically extract features from raw audio signals, which are then used for\nthe classification task. The proposed solution is evaluated on reputable\ndatasets, including RAVDESS, SHEMO, SAVEE, AESDD, and Emo-DB. The results show\nthe effectiveness of the proposed method on different datasets. Moreover, the\nmodel has been used for real-world applications like call center conversations,\nand the results demonstrate that the model accurately predicts emotions.\n", "date": "2024-11-05", "categories": []}
{"id": "2411.02402", "title": "Optimal Transport Maps are Good Voice Converters", "abstract": "  Recently, neural network-based methods for computing optimal transport maps\nhave been effectively applied to style transfer problems. However, the\napplication of these methods to voice conversion is underexplored. In our\npaper, we fill this gap by investigating optimal transport as a framework for\nvoice conversion. We present a variety of optimal transport algorithms designed\nfor different data representations, such as mel-spectrograms and latent\nrepresentation of self-supervised speech models. For the mel-spectogram data\nrepresentation, we achieve strong results in terms of Frechet Audio Distance\n(FAD). This performance is consistent with our theoretical analysis, which\nsuggests that our method provides an upper bound on the FAD between the target\nand generated distributions. Within the latent space of the WavLM encoder, we\nachived state-of-the-art results and outperformed existing methods even with\nlimited reference speaker data.\n", "date": "2024-10-17", "categories": []}
{"id": "2411.02758", "title": "DEMONet: Underwater Acoustic Target Recognition based on Multi-Expert\n  Network and Cross-Temporal Variational Autoencoder", "abstract": "  Building a robust underwater acoustic recognition system in real-world\nscenarios is challenging due to the complex underwater environment and the\ndynamic motion states of targets. A promising optimization approach is to\nleverage the intrinsic physical characteristics of targets, which remain\ninvariable regardless of environmental conditions, to provide robust insights.\nHowever, our study reveals that while physical characteristics exhibit robust\nproperties, they may lack class-specific discriminative patterns. Consequently,\ndirectly incorporating physical characteristics into model training can\npotentially introduce unintended inductive biases, leading to performance\ndegradation. To utilize the benefits of physical characteristics while\nmitigating possible detrimental effects, we propose DEMONet in this study,\nwhich utilizes the detection of envelope modulation on noise (DEMON) to provide\nrobust insights into the shaft frequency or blade counts of targets. DEMONet is\na multi-expert network that allocates various underwater signals to their\nbest-matched expert layer based on DEMON spectra for fine-grained signal\nprocessing. Thereinto, DEMON spectra are solely responsible for providing\nimplicit physical characteristics without establishing a mapping relationship\nwith the target category. Furthermore, to mitigate noise and spurious\nmodulation spectra in DEMON features, we introduce a cross-temporal alignment\nstrategy and employ a variational autoencoder (VAE) to reconstruct\nnoise-resistant DEMON spectra to replace the raw DEMON features. The\neffectiveness of the proposed DEMONet with cross-temporal VAE was primarily\nevaluated on the DeepShip dataset and our proprietary datasets. Experimental\nresults demonstrated that our approach could achieve state-of-the-art\nperformance on both datasets.\n", "date": "2024-11-05", "categories": []}
{"id": "2411.02860", "title": "Continual Audio-Visual Sound Separation", "abstract": "  In this paper, we introduce a novel continual audio-visual sound separation\ntask, aiming to continuously separate sound sources for new classes while\npreserving performance on previously learned classes, with the aid of visual\nguidance. This problem is crucial for practical visually guided auditory\nperception as it can significantly enhance the adaptability and robustness of\naudio-visual sound separation models, making them more applicable for\nreal-world scenarios where encountering new sound sources is commonplace. The\ntask is inherently challenging as our models must not only effectively utilize\ninformation from both modalities in current tasks but also preserve their\ncross-modal association in old tasks to mitigate catastrophic forgetting during\naudio-visual continual learning. To address these challenges, we propose a\nnovel approach named ContAV-Sep (\\textbf{Cont}inual\n\\textbf{A}udio-\\textbf{V}isual Sound \\textbf{Sep}aration). ContAV-Sep presents\na novel Cross-modal Similarity Distillation Constraint (CrossSDC) to uphold the\ncross-modal semantic similarity through incremental tasks and retain previously\nacquired knowledge of semantic similarity in old models, mitigating the risk of\ncatastrophic forgetting. The CrossSDC can seamlessly integrate into the\ntraining process of different audio-visual sound separation frameworks.\nExperiments demonstrate that ContAV-Sep can effectively mitigate catastrophic\nforgetting and achieve significantly better performance compared to other\ncontinual learning baselines for audio-visual sound separation. Code is\navailable at: \\url{https://github.com/weiguoPian/ContAV-Sep_NeurIPS2024}.\n", "date": "2024-11-05", "categories": []}
{"id": "2309.08030", "title": "AV2Wav: Diffusion-Based Re-synthesis from Continuous Self-supervised\n  Features for Audio-Visual Speech Enhancement", "abstract": "  Speech enhancement systems are typically trained using pairs of clean and\nnoisy speech. In audio-visual speech enhancement (AVSE), there is not as much\nground-truth clean data available; most audio-visual datasets are collected in\nreal-world environments with background noise and reverberation, hampering the\ndevelopment of AVSE. In this work, we introduce AV2Wav, a resynthesis-based\naudio-visual speech enhancement approach that can generate clean speech despite\nthe challenges of real-world training data. We obtain a subset of nearly clean\nspeech from an audio-visual corpus using a neural quality estimator, and then\ntrain a diffusion model on this subset to generate waveforms conditioned on\ncontinuous speech representations from AV-HuBERT with noise-robust training. We\nuse continuous rather than discrete representations to retain prosody and\nspeaker information. With this vocoding task alone, the model can perform\nspeech enhancement better than a masking-based baseline. We further fine-tune\nthe diffusion model on clean/noisy utterance pairs to improve the performance.\nOur approach outperforms a masking-based baseline in terms of both automatic\nmetrics and a human listening test and is close in quality to the target speech\nin the listening test. Audio samples can be found at\nhttps://home.ttic.edu/~jcchou/demo/avse/avse_demo.html.\n", "date": "2023-09-14", "categories": []}
{"id": "2411.00121", "title": "I Can Hear You: Selective Robust Training for Deepfake Audio Detection", "abstract": "  Recent advances in AI-generated voices have intensified the challenge of\ndetecting deepfake audio, posing risks for scams and the spread of\ndisinformation. To tackle this issue, we establish the largest public voice\ndataset to date, named DeepFakeVox-HQ, comprising 1.3 million samples,\nincluding 270,000 high-quality deepfake samples from 14 diverse sources.\nDespite previously reported high accuracy, existing deepfake voice detectors\nstruggle with our diversely collected dataset, and their detection success\nrates drop even further under realistic corruptions and adversarial attacks. We\nconduct a holistic investigation into factors that enhance model robustness and\nshow that incorporating a diversified set of voice augmentations is beneficial.\nMoreover, we find that the best detection models often rely on high-frequency\nfeatures, which are imperceptible to humans and can be easily manipulated by an\nattacker. To address this, we propose the F-SAT: Frequency-Selective\nAdversarial Training method focusing on high-frequency components. Empirical\nresults demonstrate that using our training dataset boosts baseline model\nperformance (without robust training) by 33%, and our robust training further\nimproves accuracy by 7.7% on clean samples and by 29.3% on corrupted and\nattacked samples, over the state-of-the-art RawNet3 model.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.00195", "title": "Machine Learning Framework for Audio-Based Content Evaluation using\n  MFCC, Chroma, Spectral Contrast, and Temporal Feature Engineering", "abstract": "  This study presents a machine learning framework for assessing similarity\nbetween audio content and predicting sentiment score. We construct a dataset\ncontaining audio samples from music covers on YouTube along with the audio of\nthe original song, and sentiment scores derived from user comments, serving as\nproxy labels for content quality. Our approach involves extensive\npre-processing, segmenting audio signals into 30-second windows, and extracting\nhigh-dimensional feature representations through Mel-Frequency Cepstral\nCoefficients (MFCC), Chroma, Spectral Contrast, and Temporal characteristics.\nLeveraging these features, we train regression models to predict sentiment\nscores on a 0-100 scale, achieving root mean square error (RMSE) values of\n3.420, 5.482, 2.783, and 4.212, respectively. Improvements over a baseline\nmodel based on absolute difference metrics are observed. These results\ndemonstrate the potential of machine learning to capture sentiment and\nsimilarity in audio, offering an adaptable framework for AI applications in\nmedia analysis.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.00275", "title": "Improving Musical Instrument Classification with Advanced Machine\n  Learning Techniques", "abstract": "  Musical instrument classification, a key area in Music Information Retrieval,\nhas gained considerable interest due to its applications in education, digital\nmusic production, and consumer media. Recent advances in machine learning,\nspecifically deep learning, have enhanced the capability to identify and\nclassify musical instruments from audio signals. This study applies various\nmachine learning methods, including Naive Bayes, Support Vector Machines,\nRandom Forests, Boosting techniques like AdaBoost and XGBoost, as well as deep\nlearning models such as Convolutional Neural Networks and Artificial Neural\nNetworks. The effectiveness of these methods is evaluated on the NSynth\ndataset, a large repository of annotated musical sounds. By comparing these\napproaches, the analysis aims to showcase the advantages and limitations of\neach method, providing guidance for developing more accurate and efficient\nclassification systems. Additionally, hybrid model testing and discussion are\nincluded. This research aims to support further studies in instrument\nclassification by proposing new approaches and future research directions.\n", "date": "2024-11-01", "categories": []}
{"id": "2410.23230", "title": "Aligning Audio-Visual Joint Representations with an Agentic Workflow", "abstract": "  Visual content and accompanied audio signals naturally formulate a joint\nrepresentation to improve audio-visual (AV) related applications. While studies\ndevelop various AV representation learning frameworks, the importance of AV\ndata alignment is usually undermined for achieving high-quality representation.\nWe observe that an audio signal may contain background noise interference.\nAlso, non-synchronization may appear between audio and video streams. These\nnon-strict data alignment limits representation quality and downgrade\napplication performance. In this paper, we propose to improve AV joint\nrepresentations from a data-centric perspective by aligning audio signals to\nvisual data. Our alignment is conducted in an agentic workflow controlled by an\nLLM-based assistant named AVAgent. For each input AV data pair, our AVAgent\nuses a multi-modal LLM to convert audio and visual data into language\ndescriptions separately (i.e., tool use). Then, AVAgent reasons whether this\npaired data is aligned well and plans to edit the audio signal if needed (i.e.,\nplanning). The audio editing is executed by predefined actions that filter\nnoise or augment data. Moreover, we use a VLM to evaluate how modified audio\nsignals match the visual content and provide feedback to AVAgent (i.e.,\nreflection). The tool use, planning, and reflection steps operate cyclically to\nbecome an agentic workflow where audio signals are gradually aligned to visual\ncontent. To this end, existing methods can directly leverage the aligned AV\ndata via our agentic workflow to improve AV joint representations. The\nexperimental results comprehensively demonstrate the state-of-the-art\nperformance of the proposed approach against previous baselines in diverse\ndownstream tasks.\n", "date": "2024-10-30", "categories": []}
{"id": "2410.23796", "title": "Improving snore detection under limited dataset through\n  harmonic/percussive source separation and convolutional neural networks", "abstract": "  Snoring, an acoustic biomarker commonly observed in individuals with\nObstructive Sleep Apnoea Syndrome (OSAS), holds significant potential for\ndiagnosing and monitoring this recognized clinical disorder. Irrespective of\nsnoring types, most snoring instances exhibit identifiable harmonic patterns\nmanifested through distinctive energy distributions over time. In this work, we\npropose a novel method to differentiate monaural snoring from non-snoring\nsounds by analyzing the harmonic content of the input sound using\nharmonic/percussive sound source separation (HPSS). The resulting feature,\nbased on the harmonic spectrogram from HPSS, is employed as input data for\nconventional neural network architectures, aiming to enhance snoring detection\nperformance even under a limited data learning framework. To evaluate the\nperformance of our proposal, we studied two different scenarios: 1) using a\nlarge dataset of snoring and interfering sounds, and 2) using a reduced\ntraining set composed of around 1% of the data material. In the former\nscenario, the proposed HPSS-based feature provides competitive results compared\nto other input features from the literature. However, the key advantage of the\nproposed method lies in the superior performance of the harmonic spectrogram\nderived from HPSS in a limited data learning context. In this particular\nscenario, using the proposed harmonic feature significantly enhances the\nperformance of all the studied architectures in comparison to the classical\ninput features documented in the existing literature. This finding clearly\ndemonstrates that incorporating harmonic content enables more reliable learning\nof the essential time-frequency characteristics that are prevalent in most\nsnoring sounds, even in scenarios where the amount of training data is limited.\n", "date": "2024-10-31", "categories": []}
{"id": "2410.23815", "title": "The NPU-HWC System for the ISCSLP 2024 Inspirational and Convincing\n  Audio Generation Challenge", "abstract": "  This paper presents the NPU-HWC system submitted to the ISCSLP 2024\nInspirational and Convincing Audio Generation Challenge 2024 (ICAGC). Our\nsystem consists of two modules: a speech generator for Track 1 and a background\naudio generator for Track 2. In Track 1, we employ Single-Codec to tokenize the\nspeech into discrete tokens and use a language-model-based approach to achieve\nzero-shot speaking style cloning. The Single-Codec effectively decouples timbre\nand speaking style at the token level, reducing the acoustic modeling burden on\nthe autoregressive language model. Additionally, we use DSPGAN to upsample 16\nkHz mel-spectrograms to high-fidelity 48 kHz waveforms. In Track 2, we propose\na background audio generator based on large language models (LLMs). This system\nproduces scene-appropriate accompaniment descriptions, synthesizes background\naudio with Tango 2, and integrates it with the speech generated by our Track 1\nsystem. Our submission achieves the second place and the first place in Track 1\nand Track 2 respectively.\n", "date": "2024-10-31", "categories": []}
{"id": "2410.24177", "title": "DC-Spin: A Speaker-invariant Speech Tokenizer for Spoken Language Models", "abstract": "  Spoken language models (SLMs) have gained increasing attention with\nadvancements in text-based, decoder-only language models. SLMs process text and\nspeech, enabling simultaneous speech understanding and generation. This paper\npresents Double-Codebook Speaker-invariant Clustering (DC-Spin), which aims to\nimprove speech tokenization by bridging audio signals and SLM tokens. DC-Spin\nextracts speaker-invariant tokens rich in phonetic information and resilient to\ninput variations, enhancing zero-shot SLM tasks and speech resynthesis. We\npropose a chunk-wise approach to enable streamable DC-Spin without retraining\nand degradation. Comparisons of tokenization methods (self-supervised and\nneural audio codecs), model scalability, and downstream task proxies show that\ntokens easily modeled by an n-gram LM or aligned with phonemes offer strong\nperformance, providing insights for designing speech tokenizers for SLMs.\n", "date": "2024-10-31", "categories": []}
{"id": "2406.08384", "title": "Diff-A-Riff: Musical Accompaniment Co-creation via Latent Diffusion\n  Models", "abstract": "  Recent advancements in deep generative models present new opportunities for\nmusic production but also pose challenges, such as high computational demands\nand limited audio quality. Moreover, current systems frequently rely solely on\ntext input and typically focus on producing complete musical pieces, which is\nincompatible with existing workflows in music production. To address these\nissues, we introduce \"Diff-A-Riff,\" a Latent Diffusion Model designed to\ngenerate high-quality instrumental accompaniments adaptable to any musical\ncontext. This model offers control through either audio references, text\nprompts, or both, and produces 48kHz pseudo-stereo audio while significantly\nreducing inference time and memory usage. We demonstrate the model's\ncapabilities through objective metrics and subjective listening tests, with\nextensive examples available on the accompanying website:\nsonycslparis.github.io/diffariff-companion/\n", "date": "2024-06-12", "categories": []}
{"id": "2410.21557", "title": "A Novel Score-CAM based Denoiser for Spectrographic Signature Extraction\n  without Ground Truth", "abstract": "  Sonar based audio classification techniques are a growing area of research in\nthe field of underwater acoustics. Usually, underwater noise picked up by\npassive sonar transducers contains all types of signals that travel through the\nocean and is transformed into spectrographic images. As a result, the\ncorresponding spectrograms intended to display the temporal-frequency data of a\ncertain object often include the tonal regions of abundant extraneous noise\nthat can effectively interfere with a 'contact'. So, a majority of\nspectrographic samples extracted from underwater audio signals are rendered\nunusable due to their clutter and lack the required indistinguishability\nbetween different objects. With limited clean true data for supervised\ntraining, creating classification models for these audio signals is severely\nbottlenecked.\n  This paper derives several new techniques to combat this problem by\ndeveloping a novel Score-CAM based denoiser to extract an object's signature\nfrom noisy spectrographic data without being given any ground truth data. In\nparticular, this paper proposes a novel generative adversarial network\narchitecture for learning and producing spectrographic training data in similar\ndistributions to low-feature spectrogram inputs. In addition, this paper also a\ngeneralizable class activation mapping based denoiser for different\ndistributions of acoustic data, even real-world data distributions. Utilizing\nthese novel architectures and proposed denoising techniques, these experiments\ndemonstrate state-of-the-art noise reduction accuracy and improved\nclassification accuracy than current audio classification standards. As such,\nthis approach has applications not only to audio data but for countless data\ndistributions used all around the world for machine learning.\n", "date": "2024-10-28", "categories": []}
{"id": "2410.22448", "title": "A Closer Look at Neural Codec Resynthesis: Bridging the Gap between\n  Codec and Waveform Generation", "abstract": "  Neural Audio Codecs, initially designed as a compression technique, have\ngained more attention recently for speech generation. Codec models represent\neach audio frame as a sequence of tokens, i.e., discrete embeddings. The\ndiscrete and low-frequency nature of neural codecs introduced a new way to\ngenerate speech with token-based models. As these tokens encode information at\nvarious levels of granularity, from coarse to fine, most existing works focus\non how to better generate the coarse tokens. In this paper, we focus on an\nequally important but often overlooked question: How can we better resynthesize\nthe waveform from coarse tokens? We point out that both the choice of learning\ntarget and resynthesis approach have a dramatic impact on the generated audio\nquality. Specifically, we study two different strategies based on token\nprediction and regression, and introduce a new method based on Schr\\\"odinger\nBridge. We examine how different design choices affect machine and human\nperception.\n", "date": "2024-10-29", "categories": []}
{"id": "2410.21561", "title": "Audio Classification of Low Feature Spectrograms Utilizing Convolutional\n  Neural Networks", "abstract": "  Modern day audio signal classification techniques lack the ability to\nclassify low feature audio signals in the form of spectrographic temporal\nfrequency data representations. Additionally, currently utilized techniques\nrely on full diverse data sets that are often not representative of real-world\ndistributions. This paper derives several first-of-its-kind machine learning\nmethodologies to analyze these low feature audio spectrograms given data\ndistributions that may have normalized, skewed, or even limited training sets.\nIn particular, this paper proposes several novel customized convolutional\narchitectures to extract identifying features using binary, one-class, and\nsiamese approaches to identify the spectrographic signature of a given audio\nsignal. Utilizing these novel convolutional architectures as well as the\nproposed classification methods, these experiments demonstrate state-of-the-art\nclassification accuracy and improved efficiency than traditional audio\nclassification methods.\n", "date": "2024-10-28", "categories": []}
{"id": "2410.22208", "title": "Drone Acoustic Analysis for Predicting Psychoacoustic Annoyance via\n  Artificial Neural Networks", "abstract": "  Unmanned Aerial Vehicles (UAVs) have become widely used in various fields and\nindustrial applications thanks to their low operational cost, compact size and\nwide accessibility. However, the noise generated by drone propellers has\nemerged as a significant concern. This may affect the public willingness to\nimplement these vehicles in services that require operation in proximity to\nresidential areas. The standard approaches to address this challenge include\nsound pressure measurements and noise characteristic analyses. The integration\nof Artificial Intelligence models in recent years has further streamlined the\nprocess by enhancing complex feature detection in drone acoustics data. This\nstudy builds upon prior research by examining the efficacy of various Deep\nLearning models in predicting Psychoacoustic Annoyance, an effective index for\nmeasuring perceived annoyance by human ears, based on multiple drone\ncharacteristics as input. This is accomplished by constructing a training\ndataset using precise measurements of various drone models with multiple\nmicrophones and analyzing flight data, maneuvers, drone physical\ncharacteristics, and perceived annoyance under realistic conditions. The aim of\nthis research is to improve our understanding of drone noise, aid in the\ndevelopment of noise reduction techniques, and encourage the acceptance of\ndrone usage on public spaces.\n", "date": "2024-10-29", "categories": []}
{"id": "2410.22271", "title": "Leveraging Reverberation and Visual Depth Cues for Sound Event\n  Localization and Detection with Distance Estimation", "abstract": "  This report describes our systems submitted for the DCASE2024 Task 3\nchallenge: Audio and Audiovisual Sound Event Localization and Detection with\nSource Distance Estimation (Track B). Our main model is based on the\naudio-visual (AV) Conformer, which processes video and audio embeddings\nextracted with ResNet50 and with an audio encoder pre-trained on SELD,\nrespectively. This model outperformed the audio-visual baseline of the\ndevelopment set of the STARSS23 dataset by a wide margin, halving its DOAE and\nimproving the F1 by more than 3x. Our second system performs a temporal\nensemble from the outputs of the AV-Conformer. We then extended the model with\nfeatures for distance estimation, such as direct and reverberant signal\ncomponents extracted from the omnidirectional audio channel, and depth maps\nextracted from the video frames. While the new system improved the RDE of our\nprevious model by about 3 percentage points, it achieved a lower F1 score. This\nmay be caused by sound classes that rarely appear in the training set and that\nthe more complex system does not detect, as analysis can determine. To overcome\nthis problem, our fourth and final system consists of an ensemble strategy\ncombining the predictions of the other three. Many opportunities to refine the\nsystem and training strategy can be tested in future ablation experiments, and\nlikely achieve incremental performance gains for this audio-visual task.\n", "date": "2024-10-29", "categories": []}
{"id": "2408.06804", "title": "Deep Learning for Speaker Identification: Architectural Insights from\n  AB-1 Corpus Analysis and Performance Evaluation", "abstract": "  In the fields of security systems, forensic investigations, and personalized\nservices, the importance of speech as a fundamental human input outweighs\ntext-based interactions. This research delves deeply into the complex field of\nSpeaker Identification (SID), examining its essential components and\nemphasising Mel Spectrogram and Mel Frequency Cepstral Coefficients (MFCC) for\nfeature extraction. Moreover, this study evaluates six slightly distinct model\narchitectures using extensive analysis to evaluate their performance, with\nhyperparameter tuning applied to the best-performing model. This work performs\na linguistic analysis to verify accent and gender accuracy, in addition to bias\nevaluation within the AB-1 Corpus dataset.\n", "date": "2024-08-13", "categories": []}
{"id": "2410.20352", "title": "An approach to hummed-tune and song sequences matching", "abstract": "  Melody stuck in your head, also known as \"earworm\", is tough to get rid of,\nunless you listen to it again or sing it out loud. But what if you can not find\nthe name of that song? It must be an intolerable feeling. Recognizing a song\nname base on humming sound is not an easy task for a human being and should be\ndone by machines. However, there is no research paper published about hum tune\nrecognition. Adapting from Hum2Song Zalo AI Challenge 2021 - a competition\nabout querying the name of a song by user's giving humming tune, which is\nsimilar to Google's Hum to Search. This paper covers details about the\npre-processed data from the original type (mp3) to usable form for training and\ninference. In training an embedding model for the feature extraction phase, we\nran experiments with some states of the art, such as ResNet, VGG, AlexNet,\nMobileNetV2. And for the inference phase, we use the Faiss module to\neffectively search for a song that matched the sequence of humming sound. The\nresult comes at nearly 94\\% in MRR@10 metric on the public test set, along with\nthe top 1 result on the public leaderboard.\n", "date": "2024-10-27", "categories": []}
{"id": "2410.20773", "title": "An Ensemble Approach to Music Source Separation: A Comparative Analysis\n  of Conventional and Hierarchical Stem Separation", "abstract": "  Music source separation (MSS) is a task that involves isolating individual\nsound sources, or stems, from mixed audio signals. This paper presents an\nensemble approach to MSS, combining several state-of-the-art architectures to\nachieve superior separation performance across traditional Vocal, Drum, and\nBass (VDB) stems, as well as expanding into second-level hierarchical\nseparation for sub-stems like kick, snare, lead vocals, and background vocals.\nOur method addresses the limitations of relying on a single model by utilising\nthe complementary strengths of various models, leading to more balanced results\nacross stems. For stem selection, we used the harmonic mean of Signal-to-Noise\nRatio (SNR) and Signal-to-Distortion Ratio (SDR), ensuring that extreme values\ndo not skew the results and that both metrics are weighted effectively. In\naddition to consistently high performance across the VDB stems, we also\nexplored second-level hierarchical separation, revealing important insights\ninto the complexities of MSS and how factors like genre and instrumentation can\ninfluence model performance. While the second-level separation results show\nroom for improvement, the ability to isolate sub-stems marks a significant\nadvancement. Our findings pave the way for further research in MSS,\nparticularly in expanding model capabilities beyond VDB and improving niche\nstem separations such as guitar and piano.\n", "date": "2024-10-28", "categories": []}
{"id": "2410.20997", "title": "SepMamba: State-space models for speaker separation using Mamba", "abstract": "  Deep learning-based single-channel speaker separation has improved\nsignificantly in recent years largely due to the introduction of the\ntransformer-based attention mechanism. However, these improvements come at the\nexpense of intense computational demands, precluding their use in many\npractical applications. As a computationally efficient alternative with similar\nmodeling capabilities, Mamba was recently introduced. We propose SepMamba, a\nU-Net-based architecture composed primarily of bidirectional Mamba layers. We\nfind that our approach outperforms similarly-sized prominent models - including\ntransformer-based models - on the WSJ0 2-speaker dataset while enjoying a\nsignificant reduction in computational cost, memory usage, and forward pass\ntime. We additionally report strong results for causal variants of SepMamba.\nOur approach provides a computationally favorable alternative to\ntransformer-based architectures for deep speech separation.\n", "date": "2024-10-28", "categories": []}
{"id": "2410.21269", "title": "OmniSep: Unified Omni-Modality Sound Separation with Query-Mixup", "abstract": "  The scaling up has brought tremendous success in the fields of vision and\nlanguage in recent years. When it comes to audio, however, researchers\nencounter a major challenge in scaling up the training data, as most natural\naudio contains diverse interfering signals. To address this limitation, we\nintroduce Omni-modal Sound Separation (OmniSep), a novel framework capable of\nisolating clean soundtracks based on omni-modal queries, encompassing both\nsingle-modal and multi-modal composed queries. Specifically, we introduce the\nQuery-Mixup strategy, which blends query features from different modalities\nduring training. This enables OmniSep to optimize multiple modalities\nconcurrently, effectively bringing all modalities under a unified framework for\nsound separation. We further enhance this flexibility by allowing queries to\ninfluence sound separation positively or negatively, facilitating the retention\nor removal of specific sounds as desired. Finally, OmniSep employs a\nretrieval-augmented approach known as Query-Aug, which enables open-vocabulary\nsound separation. Experimental evaluations on MUSIC, VGGSOUND-CLEAN+, and\nMUSIC-CLEAN+ datasets demonstrate effectiveness of OmniSep, achieving\nstate-of-the-art performance in text-, image-, and audio-queried sound\nseparation tasks. For samples and further information, please visit the demo\npage at \\url{https://omnisep.github.io/}.\n", "date": "2024-10-28", "categories": []}
{"id": "2307.16584", "title": "Audio-visual video-to-speech synthesis with synthesized input audio", "abstract": "  Video-to-speech synthesis involves reconstructing the speech signal of a\nspeaker from a silent video. The implicit assumption of this task is that the\nsound signal is either missing or contains a high amount of noise/corruption\nsuch that it is not useful for processing. Previous works in the literature\neither use video inputs only or employ both video and audio inputs during\ntraining, and discard the input audio pathway during inference. In this work we\ninvestigate the effect of using video and audio inputs for video-to-speech\nsynthesis during both training and inference. In particular, we use pre-trained\nvideo-to-speech models to synthesize the missing speech signals and then train\nan audio-visual-to-speech synthesis model, using both the silent video and the\nsynthesized speech as inputs, to predict the final reconstructed speech. Our\nexperiments demonstrate that this approach is successful with both raw\nwaveforms and mel spectrograms as target outputs.\n", "date": "2023-07-31", "categories": []}
{"id": "2410.19168", "title": "MMAU: A Massive Multi-Task Audio Understanding and Reasoning Benchmark", "abstract": "  The ability to comprehend audio--which includes speech, non-speech sounds,\nand music--is crucial for AI agents to interact effectively with the world. We\npresent MMAU, a novel benchmark designed to evaluate multimodal audio\nunderstanding models on tasks requiring expert-level knowledge and complex\nreasoning. MMAU comprises 10k carefully curated audio clips paired with\nhuman-annotated natural language questions and answers spanning speech,\nenvironmental sounds, and music. It includes information extraction and\nreasoning questions, requiring models to demonstrate 27 distinct skills across\nunique and challenging tasks. Unlike existing benchmarks, MMAU emphasizes\nadvanced perception and reasoning with domain-specific knowledge, challenging\nmodels to tackle tasks akin to those faced by experts. We assess 18 open-source\nand proprietary (Large) Audio-Language Models, demonstrating the significant\nchallenges posed by MMAU. Notably, even the most advanced Gemini Pro v1.5\nachieves only 52.97% accuracy, and the state-of-the-art open-source Qwen2-Audio\nachieves only 52.50%, highlighting considerable room for improvement. We\nbelieve MMAU will drive the audio and multimodal research community to develop\nmore advanced audio understanding models capable of solving complex audio\ntasks.\n", "date": "2024-10-24", "categories": []}
{"id": "2410.19719", "title": "Arabic Music Classification and Generation using Deep Learning", "abstract": "  This paper proposes a machine learning approach for classifying classical and\nnew Egyptian music by composer and generating new similar music. The proposed\nsystem utilizes a convolutional neural network (CNN) for classification and a\nCNN autoencoder for generation. The dataset used in this project consists of\nnew and classical Egyptian music pieces composed by different composers.\n  To classify the music by composer, each sample is normalized and transformed\ninto a mel spectrogram. The CNN model is trained on the dataset using the mel\nspectrograms as input features and the composer labels as output classes. The\nmodel achieves 81.4\\% accuracy in classifying the music by composer,\ndemonstrating the effectiveness of the proposed approach.\n  To generate new music similar to the original pieces, a CNN autoencoder is\ntrained on a similar dataset. The model is trained to encode the mel\nspectrograms of the original pieces into a lower-dimensional latent space and\nthen decode them back into the original mel spectrogram. The generated music is\nproduced by sampling from the latent space and decoding the samples back into\nmel spectrograms, which are then transformed into audio.\n  In conclusion, the proposed system provides a promising approach to\nclassifying and generating classical Egyptian music, which can be applied in\nvarious musical applications, such as music recommendation systems, music\nproduction, and music education.\n", "date": "2024-10-25", "categories": []}
{"id": "2410.19722", "title": "Temporal Convolution-based Hybrid Model Approach with Representation\n  Learning for Real-Time Acoustic Anomaly Detection", "abstract": "  The early detection of potential failures in industrial machinery components\nis paramount for ensuring the reliability and safety of operations, thereby\npreserving Machine Condition Monitoring (MCM). This research addresses this\nimperative by introducing an innovative approach to Real-Time Acoustic Anomaly\nDetection. Our method combines semi-supervised temporal convolution with\nrepresentation learning and a hybrid model strategy with Temporal Convolutional\nNetworks (TCN) to handle various intricate anomaly patterns found in acoustic\ndata effectively. The proposed model demonstrates superior performance compared\nto established research in the field, underscoring the effectiveness of this\napproach. Not only do we present quantitative evidence of its superiority, but\nwe also employ visual representations, such as t-SNE plots, to further\nsubstantiate the model's efficacy.\n", "date": "2024-10-25", "categories": []}
{"id": "2410.18628", "title": "Wavetable Synthesis Using CVAE for Timbre Control Based on Semantic\n  Label", "abstract": "  Synthesizers are essential in modern music production. However, their complex\ntimbre parameters, often filled with technical terms, require expertise. This\nresearch introduces a method of timbre control in wavetable synthesis that is\nintuitive and sensible and utilizes semantic labels. Using a conditional\nvariational autoencoder (CVAE), users can select a wavetable and define the\ntimbre with labels such as bright, warm, and rich. The CVAE model, featuring\nconvolutional and upsampling layers, effectively captures the wavetable\nnuances, ensuring real-time performance owing to their processing in the time\ndomain. Experiments demonstrate that this approach allows for real-time,\neffective control of the timbre of the wavetable using semantic inputs and aims\nfor intuitive timbre control through data-based semantic control.\n", "date": "2024-10-24", "categories": []}
{"id": "2410.15608", "title": "Moonshine: Speech Recognition for Live Transcription and Voice Commands", "abstract": "  This paper introduces Moonshine, a family of speech recognition models\noptimized for live transcription and voice command processing. Moonshine is\nbased on an encoder-decoder transformer architecture and employs Rotary\nPosition Embedding (RoPE) instead of traditional absolute position embeddings.\nThe model is trained on speech segments of various lengths, but without using\nzero-padding, leading to greater efficiency for the encoder during inference\ntime. When benchmarked against OpenAI's Whisper tiny-en, Moonshine Tiny\ndemonstrates a 5x reduction in compute requirements for transcribing a\n10-second speech segment while incurring no increase in word error rates across\nstandard evaluation datasets. These results highlight Moonshine's potential for\nreal-time and resource-constrained applications.\n", "date": "2024-10-21", "categories": []}
{"id": "2410.17209", "title": "Audio-to-Score Conversion Model Based on Whisper methodology", "abstract": "  This thesis develops a Transformer model based on Whisper, which extracts\nmelodies and chords from music audio and records them into ABC notation. A\ncomprehensive data processing workflow is customized for ABC notation,\nincluding data cleansing, formatting, and conversion, and a mutation mechanism\nis implemented to increase the diversity and quality of training data. This\nthesis innovatively introduces the \"Orpheus' Score\", a custom notation system\nthat converts music information into tokens, designs a custom vocabulary\nlibrary, and trains a corresponding custom tokenizer. Experiments show that\ncompared to traditional algorithms, the model has significantly improved\naccuracy and performance. While providing a convenient audio-to-score tool for\nmusic enthusiasts, this work also provides new ideas and tools for research in\nmusic information processing.\n", "date": "2024-10-22", "categories": []}
{"id": "2404.07970", "title": "Differentiable All-pole Filters for Time-varying Audio Systems", "abstract": "  Infinite impulse response filters are an essential building block of many\ntime-varying audio systems, such as audio effects and synthesisers. However,\ntheir recursive structure impedes end-to-end training of these systems using\nautomatic differentiation. Although non-recursive filter approximations like\nfrequency sampling and frame-based processing have been proposed and widely\nused in previous works, they cannot accurately reflect the gradient of the\noriginal system. We alleviate this difficulty by re-expressing a time-varying\nall-pole filter to backpropagate the gradients through itself, so the filter\nimplementation is not bound to the technical limitations of automatic\ndifferentiation frameworks. This implementation can be employed within audio\nsystems containing filters with poles for efficient gradient evaluation. We\ndemonstrate its training efficiency and expressive capabilities for modelling\nreal-world dynamic audio systems on a phaser, time-varying subtractive\nsynthesiser, and compressor. We make our code and audio samples available and\nprovide the trained audio effect and synth models in a VST plugin at\nhttps://diffapf.github.io/web/.\n", "date": "2024-04-11", "categories": []}
{"id": "2410.14990", "title": "Audio Processing using Pattern Recognition for Music Genre\n  Classification", "abstract": "  This project explores the application of machine learning techniques for\nmusic genre classification using the GTZAN dataset, which contains 100 audio\nfiles per genre. Motivated by the growing demand for personalized music\nrecommendations, we focused on classifying five genres-Blues, Classical, Jazz,\nHip Hop, and Country-using a variety of algorithms including Logistic\nRegression, K-Nearest Neighbors (KNN), Random Forest, and Artificial Neural\nNetworks (ANN) implemented via Keras. The ANN model demonstrated the best\nperformance, achieving a validation accuracy of 92.44%. We also analyzed key\naudio features such as spectral roll-off, spectral centroid, and MFCCs, which\nhelped enhance the model's accuracy. Future work will expand the model to cover\nall ten genres, investigate advanced methods like Long Short-Term Memory (LSTM)\nnetworks and ensemble approaches, and develop a web application for real-time\ngenre classification and playlist generation. This research aims to contribute\nto improving music recommendation systems and content curation.\n", "date": "2024-10-19", "categories": []}
{"id": "2410.14411", "title": "SNAC: Multi-Scale Neural Audio Codec", "abstract": "  Neural audio codecs have recently gained popularity because they can\nrepresent audio signals with high fidelity at very low bitrates, making it\nfeasible to use language modeling approaches for audio generation and\nunderstanding. Residual Vector Quantization (RVQ) has become the standard\ntechnique for neural audio compression using a cascade of VQ codebooks. This\npaper proposes the Multi-Scale Neural Audio Codec, a simple extension of RVQ\nwhere the quantizers can operate at different temporal resolutions. By applying\na hierarchy of quantizers at variable frame rates, the codec adapts to the\naudio structure across multiple timescales. This leads to more efficient\ncompression, as demonstrated by extensive objective and subjective evaluations.\nThe code and model weights are open-sourced at\nhttps://github.com/hubertsiuzdak/snac.\n", "date": "2024-10-18", "categories": []}
{"id": "2410.11522", "title": "Leveraging LLM Embeddings for Cross Dataset Label Alignment and Zero\n  Shot Music Emotion Prediction", "abstract": "  In this work, we present a novel method for music emotion recognition that\nleverages Large Language Model (LLM) embeddings for label alignment across\nmultiple datasets and zero-shot prediction on novel categories. First, we\ncompute LLM embeddings for emotion labels and apply non-parametric clustering\nto group similar labels, across multiple datasets containing disjoint labels.\nWe use these cluster centers to map music features (MERT) to the LLM embedding\nspace. To further enhance the model, we introduce an alignment regularization\nthat enables dissociation of MERT embeddings from different clusters. This\nfurther enhances the model's ability to better adaptation to unseen datasets.\nWe demonstrate the effectiveness of our approach by performing zero-shot\ninference on a new dataset, showcasing its ability to generalize to unseen\nlabels without additional training.\n", "date": "2024-10-15", "categories": []}
{"id": "2312.14069", "title": "EmphAssess : a Prosodic Benchmark on Assessing Emphasis Transfer in\n  Speech-to-Speech Models", "abstract": "  We introduce EmphAssess, a prosodic benchmark designed to evaluate the\ncapability of speech-to-speech models to encode and reproduce prosodic\nemphasis. We apply this to two tasks: speech resynthesis and speech-to-speech\ntranslation. In both cases, the benchmark evaluates the ability of the model to\nencode emphasis in the speech input and accurately reproduce it in the output,\npotentially across a change of speaker and language. As part of the evaluation\npipeline, we introduce EmphaClass, a new model that classifies emphasis at the\nframe or word level.\n", "date": "2023-12-21", "categories": []}
{"id": "2410.09360", "title": "Towards the Synthesis of Non-speech Vocalizations", "abstract": "  In this report, we focus on the unconditional generation of infant cry sounds\nusing the DiffWave framework, which has shown great promise in generating\nhigh-quality audio from noise. We use two distinct datasets of infant cries:\nthe Baby Chillanto and the deBarbaro cry dataset. These datasets are used to\ntrain the DiffWave model to generate new cry sounds that maintain high fidelity\nand diversity. The focus here is on DiffWave's capability to handle the\nunconditional generation task.\n", "date": "2024-10-12", "categories": []}
{"id": "2410.09869", "title": "Prompt Tuning for Audio Deepfake Detection: Computationally Efficient\n  Test-time Domain Adaptation with Limited Target Dataset", "abstract": "  We study test-time domain adaptation for audio deepfake detection (ADD),\naddressing three challenges: (i) source-target domain gaps, (ii) limited target\ndataset size, and (iii) high computational costs. We propose an ADD method\nusing prompt tuning in a plug-in style. It bridges domain gaps by integrating\nit seamlessly with state-of-the-art transformer models and/or with other\nfine-tuning methods, boosting their performance on target data (challenge (i)).\nIn addition, our method can fit small target datasets because it does not\nrequire a large number of extra parameters (challenge (ii)). This feature also\ncontributes to computational efficiency, countering the high computational\ncosts typically associated with large-scale pre-trained models in ADD\n(challenge (iii)). We conclude that prompt tuning for ADD under domain gaps\npresents a promising avenue for enhancing accuracy with minimal target data and\nnegligible extra computational burden.\n", "date": "2024-10-13", "categories": []}
{"id": "2308.00122", "title": "High-Quality Visually-Guided Sound Separation from Diverse Categories", "abstract": "  We propose DAVIS, a Diffusion-based Audio-VIsual Separation framework that\nsolves the audio-visual sound source separation task through generative\nlearning. Existing methods typically frame sound separation as a mask-based\nregression problem, achieving significant progress. However, they face\nlimitations in capturing the complex data distribution required for\nhigh-quality separation of sounds from diverse categories. In contrast, DAVIS\nleverages a generative diffusion model and a Separation U-Net to synthesize\nseparated sounds directly from Gaussian noise, conditioned on both the audio\nmixture and the visual information. With its generative objective, DAVIS is\nbetter suited to achieving the goal of high-quality sound separation across\ndiverse sound categories. We compare DAVIS to existing state-of-the-art\ndiscriminative audio-visual separation methods on the AVE and MUSIC datasets,\nand results show that DAVIS outperforms other methods in separation quality,\ndemonstrating the advantages of our framework for tackling the audio-visual\nsource separation task.\n", "date": "2023-07-31", "categories": []}
{"id": "2410.08235", "title": "A Recurrent Neural Network Approach to the Answering Machine Detection\n  Problem", "abstract": "  In the field of telecommunications and cloud communications, accurately and\nin real-time detecting whether a human or an answering machine has answered an\noutbound call is of paramount importance. This problem is of particular\nsignificance during campaigns as it enhances service quality, efficiency and\ncost reduction through precise caller identification. Despite the significance\nof the field, it remains inadequately explored in the existing literature. This\npaper presents an innovative approach to answering machine detection that\nleverages transfer learning through the YAMNet model for feature extraction.\nThe YAMNet architecture facilitates the training of a recurrent-based\nclassifier, enabling real-time processing of audio streams, as opposed to\nfixed-length recordings. The results demonstrate an accuracy of over 96% on the\ntest set. Furthermore, we conduct an in-depth analysis of misclassified samples\nand reveal that an accuracy exceeding 98% can be achieved with the integration\nof a silence detection algorithm, such as the one provided by FFmpeg.\n", "date": "2024-10-07", "categories": []}
{"id": "2410.08250", "title": "Exploring ASR-Based Wav2Vec2 for Automated Speech Disorder Assessment:\n  Insights and Analysis", "abstract": "  With the rise of SSL and ASR technologies, the Wav2Vec2 ASR-based model has\nbeen fine-tuned for automated speech disorder quality assessment tasks,\nyielding impressive results and setting a new baseline for Head and Neck Cancer\nspeech contexts. This demonstrates that the ASR dimension from Wav2Vec2 closely\naligns with assessment dimensions. Despite its effectiveness, this system\nremains a black box with no clear interpretation of the connection between the\nmodel ASR dimension and clinical assessments. This paper presents the first\nanalysis of this baseline model for speech quality assessment, focusing on\nintelligibility and severity tasks. We conduct a layer-wise analysis to\nidentify key layers and compare different SSL and ASR Wav2Vec2 models based on\npre-trained data. Additionally, post-hoc XAI methods, including Canonical\nCorrelation Analysis (CCA) and visualization techniques, are used to track\nmodel evolution and visualize embeddings for enhanced interpretability.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.08321", "title": "Music Genre Classification using Large Language Models", "abstract": "  This paper exploits the zero-shot capabilities of pre-trained large language\nmodels (LLMs) for music genre classification. The proposed approach splits\naudio signals into 20 ms chunks and processes them through convolutional\nfeature encoders, a transformer encoder, and additional layers for coding audio\nunits and generating feature vectors. The extracted feature vectors are used to\ntrain a classification head. During inference, predictions on individual chunks\nare aggregated for a final genre classification. We conducted a comprehensive\ncomparison of LLMs, including WavLM, HuBERT, and wav2vec 2.0, with traditional\ndeep learning architectures like 1D and 2D convolutional neural networks (CNNs)\nand the audio spectrogram transformer (AST). Our findings demonstrate the\nsuperior performance of the AST model, achieving an overall accuracy of 85.5%,\nsurpassing all other models evaluated. These results highlight the potential of\nLLMs and transformer-based architectures for advancing music information\nretrieval tasks, even in zero-shot scenarios.\n", "date": "2024-10-10", "categories": []}
{"id": "2406.18731", "title": "WavRx: a Disease-Agnostic, Generalizable, and Privacy-Preserving Speech\n  Health Diagnostic Model", "abstract": "  Speech is known to carry health-related attributes, which has emerged as a\nnovel venue for remote and long-term health monitoring. However, existing\nmodels are usually tailored for a specific type of disease, and have been shown\nto lack generalizability across datasets. Furthermore, concerns have been\nraised recently towards the leakage of speaker identity from health embeddings.\nTo mitigate these limitations, we propose WavRx, a speech health diagnostics\nmodel that captures the respiration and articulation related dynamics from a\nuniversal speech representation. Our in-domain and cross-domain experiments on\nsix pathological speech datasets demonstrate WavRx as a new state-of-the-art\nhealth diagnostic model. Furthermore, we show that the amount of speaker\nidentity entailed in the WavRx health embeddings is significantly reduced\nwithout extra guidance during training. An in-depth analysis of the model was\nperformed, thus providing physiological interpretation of its improved\ngeneralizability and privacy-preserving ability.\n", "date": "2024-06-26", "categories": []}
{"id": "2410.07379", "title": "Learn from Real: Reality Defender's Submission to ASVspoof5 Challenge", "abstract": "  Audio deepfake detection is crucial to combat the malicious use of\nAI-synthesized speech. Among many efforts undertaken by the community, the\nASVspoof challenge has become one of the benchmarks to evaluate the\ngeneralizability and robustness of detection models. In this paper, we present\nReality Defender's submission to the ASVspoof5 challenge, highlighting a novel\npretraining strategy which significantly improves generalizability while\nmaintaining low computational cost during training. Our system SLIM learns the\nstyle-linguistics dependency embeddings from various types of bonafide speech\nusing self-supervised contrastive learning. The learned embeddings help to\ndiscriminate spoof from bonafide speech by focusing on the relationship between\nthe style and linguistics aspects. We evaluated our system on ASVspoof5,\nASV2019, and In-the-wild. Our submission achieved minDCF of 0.1499 and EER of\n5.5% on ASVspoof5 Track 1, and EER of 7.4% and 10.8% on ASV2019 and In-the-wild\nrespectively.\n", "date": "2024-10-09", "categories": []}
{"id": "2410.07436", "title": "Toward Robust Real-World Audio Deepfake Detection: Closing the\n  Explainability Gap", "abstract": "  The rapid proliferation of AI-manipulated or generated audio deepfakes poses\nserious challenges to media integrity and election security. Current AI-driven\ndetection solutions lack explainability and underperform in real-world\nsettings. In this paper, we introduce novel explainability methods for\nstate-of-the-art transformer-based audio deepfake detectors and open-source a\nnovel benchmark for real-world generalizability. By narrowing the\nexplainability gap between transformer-based audio deepfake detectors and\ntraditional methods, our results not only build trust with human experts, but\nalso pave the way for unlocking the potential of citizen intelligence to\novercome the scalability issue in audio deepfake detection.\n", "date": "2024-10-09", "categories": []}
{"id": "2410.07530", "title": "Audio Explanation Synthesis with Generative Foundation Models", "abstract": "  The increasing success of audio foundation models across various tasks has\nled to a growing need for improved interpretability to understand their\nintricate decision-making processes better. Existing methods primarily focus on\nexplaining these models by attributing importance to elements within the input\nspace based on their influence on the final decision. In this paper, we\nintroduce a novel audio explanation method that capitalises on the generative\ncapacity of audio foundation models. Our method leverages the intrinsic\nrepresentational power of the embedding space within these models by\nintegrating established feature attribution techniques to identify significant\nfeatures in this space. The method then generates listenable audio explanations\nby prioritising the most important features. Through rigorous benchmarking\nagainst standard datasets, including keyword spotting and speech emotion\nrecognition, our model demonstrates its efficacy in producing audio\nexplanations.\n", "date": "2024-10-10", "categories": []}
{"id": "2304.11648", "title": "Towards Controllable Audio Texture Morphing", "abstract": "  In this paper, we propose a data-driven approach to train a Generative\nAdversarial Network (GAN) conditioned on \"soft-labels\" distilled from the\npenultimate layer of an audio classifier trained on a target set of audio\ntexture classes. We demonstrate that interpolation between such conditions or\ncontrol vectors provides smooth morphing between the generated audio textures,\nand shows similar or better audio texture morphing capability compared to the\nstate-of-the-art methods. The proposed approach results in a well-organized\nlatent space that generates novel audio outputs while remaining consistent with\nthe semantics of the conditioning parameters. This is a step towards a general\ndata-driven approach to designing generative audio models with customized\ncontrols capable of traversing out-of-distribution regions for novel sound\nsynthesis.\n", "date": "2023-04-23", "categories": []}
{"id": "2308.11859", "title": "Example-Based Framework for Perceptually Guided Audio Texture Generation", "abstract": "  Controllable generation using StyleGANs is usually achieved by training the\nmodel using labeled data. For audio textures, however, there is currently a\nlack of large semantically labeled datasets. Therefore, to control generation,\nwe develop a method for semantic control over an unconditionally trained\nStyleGAN in the absence of such labeled datasets. In this paper, we propose an\nexample-based framework to determine guidance vectors for audio texture\ngeneration based on user-defined semantic attributes. Our approach leverages\nthe semantically disentangled latent space of an unconditionally trained\nStyleGAN. By using a few synthetic examples to indicate the presence or absence\nof a semantic attribute, we infer the guidance vectors in the latent space of\nthe StyleGAN to control that attribute during generation. Our results show that\nour framework can find user-defined and perceptually relevant guidance vectors\nfor controllable generation for audio textures. Furthermore, we demonstrate an\napplication of our framework to other tasks, such as selective semantic\nattribute transfer.\n", "date": "2023-08-23", "categories": []}
{"id": "2310.07284", "title": "Typing to Listen at the Cocktail Party: Text-Guided Target Speaker\n  Extraction", "abstract": "  Humans can easily isolate a single speaker from a complex acoustic\nenvironment, a capability referred to as the \"Cocktail Party Effect.\" However,\nreplicating this ability has been a significant challenge in the field of\ntarget speaker extraction (TSE). Traditional TSE approaches predominantly rely\non voiceprints, which raise privacy concerns and face issues related to the\nquality and availability of enrollment samples, as well as intra-speaker\nvariability. To address these issues, this work introduces a novel text-guided\nTSE paradigm named LLM-TSE. In this paradigm, a state-of-the-art large language\nmodel, LLaMA 2, processes typed text input from users to extract semantic cues.\nWe demonstrate that textual descriptions alone can effectively serve as cues\nfor extraction, thus addressing privacy concerns and reducing dependency on\nvoiceprints. Furthermore, our approach offers flexibility by allowing the user\nto specify the extraction or suppression of a speaker and enhances robustness\nagainst intra-speaker variability by incorporating context-dependent textual\ninformation. Experimental results show competitive performance with text-based\ncues alone and demonstrate the effectiveness of using text as a task selector.\nAdditionally, they achieve a new state-of-the-art when combining text-based\ncues with pre-registered cues. This work represents the first integration of\nLLMs with TSE, potentially establishing a new benchmark in solving the cocktail\nparty problem and expanding the scope of TSE applications by providing a\nversatile, privacy-conscious solution.\n", "date": "2023-10-11", "categories": []}
{"id": "2409.11439", "title": "Machine listening in a neonatal intensive care unit", "abstract": "  Oxygenators, alarm devices, and footsteps are some of the most common sound\nsources in a hospital. Detecting them has scientific value for environmental\npsychology but comes with challenges of its own: namely, privacy preservation\nand limited labeled data. In this paper, we address these two challenges via a\ncombination of edge computing and cloud computing. For privacy preservation, we\nhave designed an acoustic sensor which computes third-octave spectrograms on\nthe fly instead of recording audio waveforms. For sample-efficient machine\nlearning, we have repurposed a pretrained audio neural network (PANN) via\nspectral transcoding and label space adaptation. A small-scale study in a\nneonatological intensive care unit (NICU) confirms that the time series of\ndetected events align with another modality of measurement: i.e., electronic\nbadges for parents and healthcare professionals. Hence, this paper demonstrates\nthe feasibility of polyphonic machine listening in a hospital ward while\nguaranteeing privacy by design.\n", "date": "2024-09-16", "categories": []}
{"id": "2410.03904", "title": "Did You Hear That? Introducing AADG: A Framework for Generating\n  Benchmark Data in Audio Anomaly Detection", "abstract": "  We introduce a novel, general-purpose audio generation framework specifically\ndesigned for anomaly detection and localization. Unlike existing datasets that\npredominantly focus on industrial and machine-related sounds, our framework\nfocuses a broader range of environments, particularly useful in real-world\nscenarios where only audio data are available, such as in video-derived or\ntelephonic audio. To generate such data, we propose a new method inspired by\nthe LLM-Modulo framework, which leverages large language models(LLMs) as world\nmodels to simulate such real-world scenarios. This tool is modular allowing a\nplug-and-play approach. It operates by first using LLMs to predict plausible\nreal-world scenarios. An LLM further extracts the constituent sounds, the order\nand the way in which these should be merged to create coherent wholes. Much\nlike the LLM-Modulo framework, we include rigorous verification of each output\nstage, ensuring the reliability of the generated data. The data produced using\nthe framework serves as a benchmark for anomaly detection applications,\npotentially enhancing the performance of models trained on audio data,\nparticularly in handling out-of-distribution cases. Our contributions thus fill\na critical void in audio anomaly detection resources and provide a scalable\ntool for generating diverse, realistic audio data.\n", "date": "2024-10-04", "categories": []}
{"id": "2410.04990", "title": "Stage-Wise and Prior-Aware Neural Speech Phase Prediction", "abstract": "  This paper proposes a novel Stage-wise and Prior-aware Neural Speech Phase\nPrediction (SP-NSPP) model, which predicts the phase spectrum from input\namplitude spectrum by two-stage neural networks. In the initial\nprior-construction stage, we preliminarily predict a rough prior phase spectrum\nfrom the amplitude spectrum. The subsequent refinement stage transforms the\namplitude spectrum into a refined high-quality phase spectrum conditioned on\nthe prior phase. Networks in both stages use ConvNeXt v2 blocks as the backbone\nand adopt adversarial training by innovatively introducing a phase spectrum\ndiscriminator (PSD). To further improve the continuity of the refined phase, we\nalso incorporate a time-frequency integrated difference (TFID) loss in the\nrefinement stage. Experimental results confirm that, compared to neural\nnetwork-based no-prior phase prediction methods, the proposed SP-NSPP achieves\nhigher phase prediction accuracy, thanks to introducing the coarse phase priors\nand diverse training criteria. Compared to iterative phase estimation\nalgorithms, our proposed SP-NSPP does not require multiple rounds of staged\niterations, resulting in higher generation efficiency.\n", "date": "2024-10-07", "categories": []}
{"id": "2409.11262", "title": "The Sounds of Home: A Speech-Removed Residential Audio Dataset for Sound\n  Event Detection", "abstract": "  This paper presents a residential audio dataset to support sound event\ndetection research for smart home applications aimed at promoting wellbeing for\nolder adults. The dataset is constructed by deploying audio recording systems\nin the homes of 8 participants aged 55-80 years for a 7-day period. Acoustic\ncharacteristics are documented through detailed floor plans and construction\nmaterial information to enable replication of the recording environments for AI\nmodel deployment. A novel automated speech removal pipeline is developed, using\npre-trained audio neural networks to detect and remove segments containing\nspoken voice, while preserving segments containing other sound events. The\nresulting dataset consists of privacy-compliant audio recordings that\naccurately capture the soundscapes and activities of daily living within\nresidential spaces. The paper details the dataset creation methodology, the\nspeech removal pipeline utilizing cascaded model architectures, and an analysis\nof the vocal label distribution to validate the speech removal process. This\ndataset enables the development and benchmarking of sound event detection\nmodels tailored specifically for in-home applications.\n", "date": "2024-09-17", "categories": []}
{"id": "2308.09302", "title": "Robust Audio Anti-Spoofing with Fusion-Reconstruction Learning on\n  Multi-Order Spectrograms", "abstract": "  Robust audio anti-spoofing has been increasingly challenging due to the\nrecent advancements on deepfake techniques. While spectrograms have\ndemonstrated their capability for anti-spoofing, complementary information\npresented in multi-order spectral patterns have not been well explored, which\nlimits their effectiveness for varying spoofing attacks. Therefore, we propose\na novel deep learning method with a spectral fusion-reconstruction strategy,\nnamely S2pecNet, to utilise multi-order spectral patterns for robust audio\nanti-spoofing representations. Specifically, spectral patterns up to\nsecond-order are fused in a coarse-to-fine manner and two branches are designed\nfor the fine-level fusion from the spectral and temporal contexts. A\nreconstruction from the fused representation to the input spectrograms further\nreduces the potential fused information loss. Our method achieved the\nstate-of-the-art performance with an EER of 0.77% on a widely used dataset:\nASVspoof2019 LA Challenge.\n", "date": "2023-08-18", "categories": []}
{"id": "2410.02271", "title": "CoLLAP: Contrastive Long-form Language-Audio Pretraining with Musical\n  Temporal Structure Augmentation", "abstract": "  Modeling temporal characteristics plays a significant role in the\nrepresentation learning of audio waveform. We propose Contrastive Long-form\nLanguage-Audio Pretraining (\\textbf{CoLLAP}) to significantly extend the\nperception window for both the input audio (up to 5 minutes) and the language\ndescriptions (exceeding 250 words), while enabling contrastive learning across\nmodalities and temporal dynamics. Leveraging recent Music-LLMs to generate\nlong-form music captions for full-length songs, augmented with musical temporal\nstructures, we collect 51.3K audio-text pairs derived from the large-scale\nAudioSet training dataset, where the average audio length reaches 288 seconds.\nWe propose a novel contrastive learning architecture that fuses language\nrepresentations with structured audio representations by segmenting each song\ninto clips and extracting their embeddings. With an attention mechanism, we\ncapture multimodal temporal correlations, allowing the model to automatically\nweigh and enhance the final fusion score for improved contrastive alignment.\nFinally, we develop two variants of the CoLLAP model with different types of\nbackbone language models. Through comprehensive experiments on multiple\nlong-form music-text retrieval datasets, we demonstrate consistent performance\nimprovement in retrieval accuracy compared with baselines. We also show the\npretrained CoLLAP models can be transferred to various music information\nretrieval tasks, with heterogeneous long-form multimodal contexts.\n", "date": "2024-10-03", "categories": []}
{"id": "2410.00980", "title": "Heterogeneous sound classification with the Broad Sound Taxonomy and\n  Dataset", "abstract": "  Automatic sound classification has a wide range of applications in machine\nlistening, enabling context-aware sound processing and understanding. This\npaper explores methodologies for automatically classifying heterogeneous sounds\ncharacterized by high intra-class variability. Our study evaluates the\nclassification task using the Broad Sound Taxonomy, a two-level taxonomy\ncomprising 28 classes designed to cover a heterogeneous range of sounds with\nsemantic distinctions tailored for practical user applications. We construct a\ndataset through manual annotation to ensure accuracy, diverse representation\nwithin each class and relevance in real-world scenarios. We compare a variety\nof both traditional and modern machine learning approaches to establish a\nbaseline for the task of heterogeneous sound classification. We investigate the\nrole of input features, specifically examining how acoustically derived sound\nrepresentations compare to embeddings extracted with pre-trained deep neural\nnetworks that capture both acoustic and semantic information about sounds.\nExperimental results illustrate that audio embeddings encoding acoustic and\nsemantic information achieve higher accuracy in the classification task. After\ncareful analysis of classification errors, we identify some underlying reasons\nfor failure and propose actions to mitigate them. The paper highlights the need\nfor deeper exploration of all stages of classification, understanding the data\nand adopting methodologies capable of effectively handling data complexity and\ngeneralizing in real-world sound environments.\n", "date": "2024-10-01", "categories": []}
{"id": "2410.01562", "title": "HRTF Estimation using a Score-based Prior", "abstract": "  We present a head-related transfer function (HRTF) estimation method which\nrelies on a data-driven prior given by a score-based diffusion model. The HRTF\nis estimated in reverberant environments using natural excitation signals, e.g.\nhuman speech. The impulse response of the room is estimated along with the HRTF\nby optimizing a parametric model of reverberation based on the statistical\nbehaviour of room acoustics. The posterior distribution of HRTF given the\nreverberant measurement and excitation signal is modelled using the score-based\nHRTF prior and a log-likelihood approximation. We show that the resulting\nmethod outperforms several baselines, including an oracle recommender system\nthat assigns the optimal HRTF in our training set based on the smallest\ndistance to the true HRTF at the given direction of arrival. In particular, we\nshow that the diffusion prior can account for the large variability of\nhigh-frequency content in HRTFs.\n", "date": "2024-10-02", "categories": []}
{"id": "2410.00511", "title": "Pre-training with Synthetic Patterns for Audio", "abstract": "  In this paper, we propose to pre-train audio encoders using synthetic\npatterns instead of real audio data. Our proposed framework consists of two key\nelements. The first one is Masked Autoencoder (MAE), a self-supervised learning\nframework that learns from reconstructing data from randomly masked\ncounterparts. MAEs tend to focus on low-level information such as visual\npatterns and regularities within data. Therefore, it is unimportant what is\nportrayed in the input, whether it be images, audio mel-spectrograms, or even\nsynthetic patterns. This leads to the second key element, which is synthetic\ndata. Synthetic data, unlike real audio, is free from privacy and licensing\ninfringement issues. By combining MAEs and synthetic patterns, our framework\nenables the model to learn generalized feature representations without real\ndata, while addressing the issues related to real audio. To evaluate the\nefficacy of our framework, we conduct extensive experiments across a total of\n13 audio tasks and 17 synthetic datasets. The experiments provide insights into\nwhich types of synthetic patterns are effective for audio. Our results\ndemonstrate that our framework achieves performance comparable to models\npre-trained on AudioSet-2M and partially outperforms image-based pre-training\nmethods.\n", "date": "2024-10-01", "categories": []}
{"id": "2409.19132", "title": "From Vision to Audio and Beyond: A Unified Model for Audio-Visual\n  Representation and Generation", "abstract": "  Video encompasses both visual and auditory data, creating a perceptually rich\nexperience where these two modalities complement each other. As such, videos\nare a valuable type of media for the investigation of the interplay between\naudio and visual elements. Previous studies of audio-visual modalities\nprimarily focused on either audio-visual representation learning or generative\nmodeling of a modality conditioned on the other, creating a disconnect between\nthese two branches. A unified framework that learns representation and\ngenerates modalities has not been developed yet. In this work, we introduce a\nnovel framework called Vision to Audio and Beyond (VAB) to bridge the gap\nbetween audio-visual representation learning and vision-to-audio generation.\nThe key approach of VAB is that rather than working with raw video frames and\naudio data, VAB performs representation learning and generative modeling within\nlatent spaces. In particular, VAB uses a pre-trained audio tokenizer and an\nimage encoder to obtain audio tokens and visual features, respectively. It then\nperforms the pre-training task of visual-conditioned masked audio token\nprediction. This training strategy enables the model to engage in contextual\nlearning and simultaneous video-to-audio generation. After the pre-training\nphase, VAB employs the iterative-decoding approach to rapidly generate audio\ntokens conditioned on visual features. Since VAB is a unified model, its\nbackbone can be fine-tuned for various audio-visual downstream tasks. Our\nexperiments showcase the efficiency of VAB in producing high-quality audio from\nvideo, and its capability to acquire semantic audio-visual features, leading to\ncompetitive results in audio-visual retrieval and classification.\n", "date": "2024-09-27", "categories": []}
{"id": "2409.19270", "title": "OpenSep: Leveraging Large Language Models with Textual Inversion for\n  Open World Audio Separation", "abstract": "  Audio separation in real-world scenarios, where mixtures contain a variable\nnumber of sources, presents significant challenges due to limitations of\nexisting models, such as over-separation, under-separation, and dependence on\npredefined training sources. We propose OpenSep, a novel framework that\nleverages large language models (LLMs) for automated audio separation,\neliminating the need for manual intervention and overcoming source limitations.\nOpenSep uses textual inversion to generate captions from audio mixtures with\noff-the-shelf audio captioning models, effectively parsing the sound sources\npresent. It then employs few-shot LLM prompting to extract detailed audio\nproperties of each parsed source, facilitating separation in unseen mixtures.\nAdditionally, we introduce a multi-level extension of the mix-and-separate\ntraining framework to enhance modality alignment by separating single source\nsounds and mixtures simultaneously. Extensive experiments demonstrate OpenSep's\nsuperiority in precisely separating new, unseen, and variable sources in\nchallenging mixtures, outperforming SOTA baseline methods. Code is released at\nhttps://github.com/tanvir-utexas/OpenSep.git\n", "date": "2024-09-28", "categories": []}
{"id": "2409.19595", "title": "Solution for Temporal Sound Localisation Task of ECCV Second Perception\n  Test Challenge 2024", "abstract": "  This report proposes an improved method for the Temporal Sound Localisation\n(TSL) task, which localizes and classifies the sound events occurring in the\nvideo according to a predefined set of sound classes. The champion solution\nfrom last year's first competition has explored the TSL by fusing audio and\nvideo modalities with the same weight. Considering the TSL task aims to\nlocalize sound events, we conduct relevant experiments that demonstrated the\nsuperiority of sound features (Section 3). Based on our findings, to enhance\naudio modality features, we employ various models to extract audio features,\nsuch as InterVideo, CaVMAE, and VideoMAE models. Our approach ranks first in\nthe final test with a score of 0.4925.\n", "date": "2024-09-29", "categories": []}
{"id": "2409.19806", "title": "PALM: Few-Shot Prompt Learning for Audio Language Models", "abstract": "  Audio-Language Models (ALMs) have recently achieved remarkable success in\nzero-shot audio recognition tasks, which match features of audio waveforms with\nclass-specific text prompt features, inspired by advancements in\nVision-Language Models (VLMs). Given the sensitivity of zero-shot performance\nto the choice of hand-crafted text prompts, many prompt learning techniques\nhave been developed for VLMs. We explore the efficacy of these approaches in\nALMs and propose a novel method, Prompt Learning in Audio Language Models\n(PALM), which optimizes the feature space of the text encoder branch. Unlike\nexisting methods that work in the input space, our approach results in greater\ntraining efficiency. We demonstrate the effectiveness of our approach on 11\naudio recognition datasets, encompassing a variety of speech-processing tasks,\nand compare the results with three baselines in a few-shot learning setup. Our\nmethod is either on par with or outperforms other approaches while being\ncomputationally less demanding. Code is available at\nhttps://asif-hanif.github.io/palm/\n", "date": "2024-09-29", "categories": []}
{"id": "2409.18542", "title": "MIMII-Gen: Generative Modeling Approach for Simulated Evaluation of\n  Anomalous Sound Detection System", "abstract": "  Insufficient recordings and the scarcity of anomalies present significant\nchallenges in developing and validating robust anomaly detection systems for\nmachine sounds. To address these limitations, we propose a novel approach for\ngenerating diverse anomalies in machine sound using a latent diffusion-based\nmodel that integrates an encoder-decoder framework. Our method utilizes the\nFlan-T5 model to encode captions derived from audio file metadata, enabling\nconditional generation through a carefully designed U-Net architecture. This\napproach aids our model in generating audio signals within the EnCodec latent\nspace, ensuring high contextual relevance and quality. We objectively evaluated\nthe quality of our generated sounds using the Fr\\'echet Audio Distance (FAD)\nscore and other metrics, demonstrating that our approach surpasses existing\nmodels in generating reliable machine audio that closely resembles actual\nabnormal conditions. The evaluation of the anomaly detection system using our\ngenerated data revealed a strong correlation, with the area under the curve\n(AUC) score differing by 4.8\\% from the original, validating the effectiveness\nof our generated data. These results demonstrate the potential of our approach\nto enhance the evaluation and robustness of anomaly detection systems across\nvaried and previously unseen conditions. Audio samples can be found at\n\\url{https://hpworkhub.github.io/MIMII-Gen.github.io/}.\n", "date": "2024-09-27", "categories": []}
{"id": "2406.07923", "title": "CTC-aligned Audio-Text Embedding for Streaming Open-vocabulary Keyword\n  Spotting", "abstract": "  This paper introduces a novel approach for streaming openvocabulary keyword\nspotting (KWS) with text-based keyword enrollment. For every input frame, the\nproposed method finds the optimal alignment ending at the frame using\nconnectionist temporal classification (CTC) and aggregates the frame-level\nacoustic embedding (AE) to obtain higher-level (i.e., character, word, or\nphrase) AE that aligns with the text embedding (TE) of the target keyword text.\nAfter that, we calculate the similarity of the aggregated AE and the TE. To the\nbest of our knowledge, this is the first attempt to dynamically align the audio\nand the keyword text on-the-fly to attain the joint audio-text embedding for\nKWS. Despite operating in a streaming fashion, our approach achieves\ncompetitive performance on the LibriPhrase dataset compared to the\nnon-streaming methods with a mere 155K model parameters and a decoding\nalgorithm with time complexity O(U), where U is the length of the target\nkeyword at inference time.\n", "date": "2024-06-12", "categories": []}
{"id": "2409.17656", "title": "Prototype based Masked Audio Model for Self-Supervised Learning of Sound\n  Event Detection", "abstract": "  A significant challenge in sound event detection (SED) is the effective\nutilization of unlabeled data, given the limited availability of labeled data\ndue to high annotation costs. Semi-supervised algorithms rely on labeled data\nto learn from unlabeled data, and the performance is constrained by the quality\nand size of the former. In this paper, we introduce the Prototype based Masked\nAudio Model~(PMAM) algorithm for self-supervised representation learning in\nSED, to better exploit unlabeled data. Specifically, semantically rich\nframe-level pseudo labels are constructed from a Gaussian mixture model (GMM)\nbased prototypical distribution modeling. These pseudo labels supervise the\nlearning of a Transformer-based masked audio model, in which binary\ncross-entropy loss is employed instead of the widely used InfoNCE loss, to\nprovide independent loss contributions from different prototypes, which is\nimportant in real scenarios in which multiple labels may apply to unsupervised\ndata frames. A final stage of fine-tuning with just a small amount of labeled\ndata yields a very high performing SED model. On like-for-like tests using the\nDESED task, our method achieves a PSDS1 score of 62.5\\%, surpassing current\nstate-of-the-art models and demonstrating the superiority of the proposed\ntechnique.\n", "date": "2024-09-26", "categories": []}
{"id": "2409.17069", "title": "The Effect of Perceptual Metrics on Music Representation Learning for\n  Genre Classification", "abstract": "  The subjective quality of natural signals can be approximated with objective\nperceptual metrics. Designed to approximate the perceptual behaviour of human\nobservers, perceptual metrics often reflect structures found in natural signals\nand neurological pathways. Models trained with perceptual metrics as loss\nfunctions can capture perceptually meaningful features from the structures held\nwithin these metrics. We demonstrate that using features extracted from\nautoencoders trained with perceptual losses can improve performance on music\nunderstanding tasks, i.e. genre classification, over using these metrics\ndirectly as distances when learning a classifier. This result suggests improved\ngeneralisation to novel signals when using perceptual metrics as loss functions\nfor representation learning.\n", "date": "2024-09-25", "categories": []}
{"id": "2309.07115", "title": "Getting More for Less: Using Weak Labels and AV-Mixup for Robust\n  Audio-Visual Speaker Verification", "abstract": "  Distance Metric Learning (DML) has typically dominated the audio-visual\nspeaker verification problem space, owing to strong performance in new and\nunseen classes. In our work, we explored multitask learning techniques to\nfurther enhance DML, and show that an auxiliary task with even weak labels can\nincrease the quality of the learned speaker representation without increasing\nmodel complexity during inference. We also extend the Generalized End-to-End\nLoss (GE2E) to multimodal inputs and demonstrate that it can achieve\ncompetitive performance in an audio-visual space. Finally, we introduce\nAV-Mixup, a multimodal augmentation technique during training time that has\nshown to reduce speaker overfit. Our network achieves state of the art\nperformance for speaker verification, reporting 0.244%, 0.252%, 0.441% Equal\nError Rate (EER) on the VoxCeleb1-O/E/H test sets, which is to our knowledge,\nthe best published results on VoxCeleb1-E and VoxCeleb1-H.\n", "date": "2023-09-13", "categories": []}
{"id": "2311.14517", "title": "tinyCLAP: Distilling Constrastive Language-Audio Pretrained Models", "abstract": "  Contrastive Language-Audio Pretraining (CLAP) became of crucial importance in\nthe field of audio and speech processing. Its employment ranges from sound\nevent detection to text-to-audio generation. However, one of the main\nlimitations is the considerable amount of data required in the training process\nand the overall computational complexity during inference. This paper\ninvestigates how we can reduce the complexity of contrastive language-audio\npre-trained models, yielding an efficient model that we call tinyCLAP. We\nderive an unimodal distillation loss from first principles and explore how the\ndimensionality of the shared, multimodal latent space can be reduced via\npruning. TinyCLAP uses only 6% of the original Microsoft CLAP parameters with a\nminimal reduction (less than 5%) in zero-shot classification performance across\nthe three sound event detection datasets on which it was tested\n", "date": "2023-11-24", "categories": []}
{"id": "2406.03882", "title": "Spontaneous Speech-Based Suicide Risk Detection Using Whisper and Large\n  Language Models", "abstract": "  The early detection of suicide risk is important since it enables the\nintervention to prevent potential suicide attempts. This paper studies the\nautomatic detection of suicide risk based on spontaneous speech from\nadolescents, and collects a Mandarin dataset with 15 hours of suicide speech\nfrom more than a thousand adolescents aged from ten to eighteen for our\nexperiments. To leverage the diverse acoustic and linguistic features embedded\nin spontaneous speech, both the Whisper speech model and textual large language\nmodels (LLMs) are used for suicide risk detection. Both all-parameter\nfinetuning and parameter-efficient finetuning approaches are used to adapt the\npre-trained models for suicide risk detection, and multiple audio-text fusion\napproaches are evaluated to combine the representations of Whisper and the LLM.\nThe proposed system achieves a detection accuracy of 0.807 and an F1-score of\n0.846 on the test set with 119 subjects, indicating promising potential for\nreal suicide risk detection applications.\n", "date": "2024-06-06", "categories": []}
{"id": "2409.15283", "title": "Equivariance-based self-supervised learning for audio signal recovery\n  from clipped measurements", "abstract": "  In numerous inverse problems, state-of-the-art solving strategies involve\ntraining neural networks from ground truth and associated measurement datasets\nthat, however, may be expensive or impossible to collect. Recently,\nself-supervised learning techniques have emerged, with the major advantage of\nno longer requiring ground truth data. Most theoretical and experimental\nresults on self-supervised learning focus on linear inverse problems. The\npresent work aims to study self-supervised learning for the non-linear inverse\nproblem of recovering audio signals from clipped measurements. An\nequivariance-based selfsupervised loss is proposed and studied. Performance is\nassessed on simulated clipped measurements with controlled and varied levels of\nclipping, and further reported on standard real music signals. We show that the\nperformance of the proposed equivariance-based self-supervised declipping\nstrategy compares favorably to fully supervised learning while only requiring\nclipped measurements alone for training.\n", "date": "2024-09-03", "categories": []}
{"id": "2409.15383", "title": "Generalization in birdsong classification: impact of transfer learning\n  methods and dataset characteristics", "abstract": "  Animal sounds can be recognised automatically by machine learning, and this\nhas an important role to play in biodiversity monitoring. Yet despite\nincreasingly impressive capabilities, bioacoustic species classifiers still\nexhibit imbalanced performance across species and habitats, especially in\ncomplex soundscapes. In this study, we explore the effectiveness of transfer\nlearning in large-scale bird sound classification across various conditions,\nincluding single- and multi-label scenarios, and across different model\narchitectures such as CNNs and Transformers. Our experiments demonstrate that\nboth fine-tuning and knowledge distillation yield strong performance, with\ncross-distillation proving particularly effective in improving in-domain\nperformance on Xeno-canto data. However, when generalizing to soundscapes,\nshallow fine-tuning exhibits superior performance compared to knowledge\ndistillation, highlighting its robustness and constrained nature. Our study\nfurther investigates how to use multi-species labels, in cases where these are\npresent but incomplete. We advocate for more comprehensive labeling practices\nwithin the animal sound community, including annotating background species and\nproviding temporal details, to enhance the training of robust bird sound\nclassifiers. These findings provide insights into the optimal reuse of\npretrained models for advancing automatic bioacoustic recognition.\n", "date": "2024-09-21", "categories": []}
{"id": "2409.15732", "title": "Hypothesis Clustering and Merging: Novel MultiTalker Speech Recognition\n  with Speaker Tokens", "abstract": "  In many real-world scenarios, such as meetings, multiple speakers are present\nwith an unknown number of participants, and their utterances often overlap. We\naddress these multi-speaker challenges by a novel attention-based\nencoder-decoder method augmented with special speaker class tokens obtained by\nspeaker clustering. During inference, we select multiple recognition hypotheses\nconditioned on predicted speaker cluster tokens, and these hypotheses are\nmerged by agglomerative hierarchical clustering (AHC) based on the normalized\nedit distance. The clustered hypotheses result in the multi-speaker\ntranscriptions with the appropriate number of speakers determined by AHC. Our\nexperiments on the LibriMix dataset demonstrate that our proposed method was\nparticularly effective in complex 3-mix environments, achieving a 55% relative\nerror reduction on clean data and a 36% relative error reduction on noisy data\ncompared with conventional serialized output training.\n", "date": "2024-09-24", "categories": []}
{"id": "2409.15957", "title": "ASD-Diffusion: Anomalous Sound Detection with Diffusion Models", "abstract": "  Unsupervised Anomalous Sound Detection (ASD) aims to design a generalizable\nmethod that can be used to detect anomalies when only normal sounds are given.\nIn this paper, Anomalous Sound Detection based on Diffusion Models\n(ASD-Diffusion) is proposed for ASD in real-world factories. In our pipeline,\nthe anomalies in acoustic features are reconstructed from their noisy corrupted\nfeatures into their approximate normal pattern. Secondly, a post-processing\nanomalies filter algorithm is proposed to detect anomalies that exhibit\nsignificant deviation from the original input after reconstruction.\nFurthermore, denoising diffusion implicit model is introduced to accelerate the\ninference speed by a longer sampling interval of the denoising process. The\nproposed method is innovative in the application of diffusion models as a new\nscheme. Experimental results on the development set of DCASE 2023 challenge\ntask 2 outperform the baseline by 7.75%, demonstrating the effectiveness of the\nproposed method.\n", "date": "2024-09-24", "categories": []}
{"id": "2409.16077", "title": "Leveraging Mixture of Experts for Improved Speech Deepfake Detection", "abstract": "  Speech deepfakes pose a significant threat to personal security and content\nauthenticity. Several detectors have been proposed in the literature, and one\nof the primary challenges these systems have to face is the generalization over\nunseen data to identify fake signals across a wide range of datasets. In this\npaper, we introduce a novel approach for enhancing speech deepfake detection\nperformance using a Mixture of Experts architecture. The Mixture of Experts\nframework is well-suited for the speech deepfake detection task due to its\nability to specialize in different input types and handle data variability\nefficiently. This approach offers superior generalization and adaptability to\nunseen data compared to traditional single models or ensemble methods.\nAdditionally, its modular structure supports scalable updates, making it more\nflexible in managing the evolving complexity of deepfake techniques while\nmaintaining high detection accuracy. We propose an efficient, lightweight\ngating mechanism to dynamically assign expert weights for each input,\noptimizing detection performance. Experimental results across multiple datasets\ndemonstrate the effectiveness and potential of our proposed approach.\n", "date": "2024-09-24", "categories": []}
{"id": "2409.14043", "title": "ECHO: Environmental Sound Classification with Hierarchical\n  Ontology-guided Semi-Supervised Learning", "abstract": "  Environment Sound Classification has been a well-studied research problem in\nthe field of signal processing and up till now more focus has been laid on\nfully supervised approaches. Over the last few years, focus has moved towards\nsemi-supervised methods which concentrate on the utilization of unlabeled data,\nand self-supervised methods which learn the intermediate representation through\npretext task or contrastive learning. However, both approaches require a vast\namount of unlabelled data to improve performance. In this work, we propose a\nnovel framework called Environmental Sound Classification with Hierarchical\nOntology-guided semi-supervised Learning (ECHO) that utilizes label\nontology-based hierarchy to learn semantic representation by defining a novel\npretext task. In the pretext task, the model tries to predict coarse labels\ndefined by the Large Language Model (LLM) based on ground truth label ontology.\nThe trained model is further fine-tuned in a supervised way to predict the\nactual task. Our proposed novel semi-supervised framework achieves an accuracy\nimprovement in the range of 1\\% to 8\\% over baseline systems across three\ndatasets namely UrbanSound8K, ESC-10, and ESC-50.\n", "date": "2024-09-21", "categories": []}
{"id": "2409.14086", "title": "AMT-APC: Automatic Piano Cover by Fine-Tuning an Automatic Music\n  Transcription Model", "abstract": "  There have been several studies on automatically generating piano covers, and\nrecent advancements in deep learning have enabled the creation of more\nsophisticated covers. However, existing automatic piano cover models still have\nroom for improvement in terms of expressiveness and fidelity to the original.\nTo address these issues, we propose a learning algorithm called AMT-APC, which\nleverages the capabilities of automatic music transcription models. By\nutilizing the strengths of well-established automatic music transcription\nmodels, we aim to improve the accuracy of piano cover generation. Our\nexperiments demonstrate that the AMT-APC model reproduces original tracks more\naccurately than any existing models.\n", "date": "2024-09-21", "categories": []}
{"id": "2409.14340", "title": "Self-Supervised Audio-Visual Soundscape Stylization", "abstract": "  Speech sounds convey a great deal of information about the scenes, resulting\nin a variety of effects ranging from reverberation to additional ambient\nsounds. In this paper, we manipulate input speech to sound as though it was\nrecorded within a different scene, given an audio-visual conditional example\nrecorded from that scene. Our model learns through self-supervision, taking\nadvantage of the fact that natural video contains recurring sound events and\ntextures. We extract an audio clip from a video and apply speech enhancement.\nWe then train a latent diffusion model to recover the original speech, using\nanother audio-visual clip taken from elsewhere in the video as a conditional\nhint. Through this process, the model learns to transfer the conditional\nexample's sound properties to the input speech. We show that our model can be\nsuccessfully trained using unlabeled, in-the-wild videos, and that an\nadditional visual signal can improve its sound prediction abilities. Please see\nour project webpage for video results:\nhttps://tinglok.netlify.app/files/avsoundscape/\n", "date": "2024-09-22", "categories": []}
{"id": "2409.14348", "title": "A Feature Engineering Approach for Literary and Colloquial Tamil Speech\n  Classification using 1D-CNN", "abstract": "  In ideal human computer interaction (HCI), the colloquial form of a language\nwould be preferred by most users, since it is the form used in their day-to-day\nconversations. However, there is also an undeniable necessity to preserve the\nformal literary form. By embracing the new and preserving the old, both service\nto the common man (practicality) and service to the language itself\n(conservation) can be rendered. Hence, it is ideal for computers to have the\nability to accept, process, and converse in both forms of the language, as\nrequired. To address this, it is first necessary to identify the form of the\ninput speech, which in the current work is between literary and colloquial\nTamil speech. Such a front-end system must consist of a simple, effective, and\nlightweight classifier that is trained on a few effective features that are\ncapable of capturing the underlying patterns of the speech signal. To\naccomplish this, a one-dimensional convolutional neural network (1D-CNN) that\nlearns the envelope of features across time, is proposed. The network is\ntrained on a select number of handcrafted features initially, and then on Mel\nfrequency cepstral coefficients (MFCC) for comparison. The handcrafted features\nwere selected to address various aspects of speech such as the spectral and\ntemporal characteristics, prosody, and voice quality. The features are\ninitially analyzed by considering ten parallel utterances and observing the\ntrend of each feature with respect to time. The proposed 1D-CNN, trained using\nthe handcrafted features, offers an F1 score of 0.9803, while that trained on\nthe MFCC offers an F1 score of 0.9895. In light of this, feature ablation and\nfeature combination are explored. When the best ranked handcrafted features,\nfrom the feature ablation study, are combined with the MFCC, they offer the\nbest results with an F1 score of 0.9946.\n", "date": "2024-09-22", "categories": []}
{"id": "2409.14971", "title": "Blind Spatial Impulse Response Generation from Separate Room- and\n  Scene-Specific Information", "abstract": "  For audio in augmented reality (AR), knowledge of the users' real acoustic\nenvironment is crucial for rendering virtual sounds that seamlessly blend into\nthe environment. As acoustic measurements are usually not feasible in practical\nAR applications, information about the room needs to be inferred from available\nsound sources. Then, additional sound sources can be rendered with the same\nroom acoustic qualities. Crucially, these are placed at different positions\nthan the sources available for estimation. Here, we propose to use an encoder\nnetwork trained using a contrastive loss that maps input sounds to a\nlow-dimensional feature space representing only room-specific information.\nThen, a diffusion-based spatial room impulse response generator is trained to\ntake the latent space and generate a new response, given a new source-receiver\nposition. We show how both room- and position-specific parameters are\nconsidered in the final output.\n", "date": "2024-09-23", "categories": []}
{"id": "2404.04904", "title": "Cross-Domain Audio Deepfake Detection: Dataset and Analysis", "abstract": "  Audio deepfake detection (ADD) is essential for preventing the misuse of\nsynthetic voices that may infringe on personal rights and privacy. Recent\nzero-shot text-to-speech (TTS) models pose higher risks as they can clone\nvoices with a single utterance. However, the existing ADD datasets are\noutdated, leading to suboptimal generalization of detection models. In this\npaper, we construct a new cross-domain ADD dataset comprising over 300 hours of\nspeech data that is generated by five advanced zero-shot TTS models. To\nsimulate real-world scenarios, we employ diverse attack methods and audio\nprompts from different datasets. Experiments show that, through novel\nattack-augmented training, the Wav2Vec2-large and Whisper-medium models achieve\nequal error rates of 4.1\\% and 6.5\\% respectively. Additionally, we demonstrate\nour models' outstanding few-shot ADD ability by fine-tuning with just one\nminute of target-domain data. Nonetheless, neural codec compressors greatly\naffect the detection accuracy, necessitating further research.\n", "date": "2024-04-07", "categories": []}
{"id": "2409.13676", "title": "A sound description: Exploring prompt templates and class descriptions\n  to enhance zero-shot audio classification", "abstract": "  Audio-text models trained via contrastive learning offer a practical approach\nto perform audio classification through natural language prompts, such as \"this\nis a sound of\" followed by category names. In this work, we explore alternative\nprompt templates for zero-shot audio classification, demonstrating the\nexistence of higher-performing options. First, we find that the formatting of\nthe prompts significantly affects performance so that simply prompting the\nmodels with properly formatted class labels performs competitively with\noptimized prompt templates and even prompt ensembling. Moreover, we look into\ncomplementing class labels by audio-centric descriptions. By leveraging large\nlanguage models, we generate textual descriptions that prioritize acoustic\nfeatures of sound events to disambiguate between classes, without extensive\nprompt engineering. We show that prompting with class descriptions leads to\nstate-of-the-art results in zero-shot audio classification across major ambient\nsound datasets. Remarkably, this method requires no additional training and\nremains fully zero-shot.\n", "date": "2024-09-19", "categories": []}
{"id": "2409.13689", "title": "Temporally Aligned Audio for Video with Autoregression", "abstract": "  We introduce V-AURA, the first autoregressive model to achieve high temporal\nalignment and relevance in video-to-audio generation. V-AURA uses a\nhigh-framerate visual feature extractor and a cross-modal audio-visual feature\nfusion strategy to capture fine-grained visual motion events and ensure precise\ntemporal alignment. Additionally, we propose VisualSound, a benchmark dataset\nwith high audio-visual relevance. VisualSound is based on VGGSound, a video\ndataset consisting of in-the-wild samples extracted from YouTube. During the\ncuration, we remove samples where auditory events are not aligned with the\nvisual ones. V-AURA outperforms current state-of-the-art models in temporal\nalignment and semantic relevance while maintaining comparable audio quality.\nCode, samples, VisualSound and models are available at\nhttps://v-aura.notion.site\n", "date": "2024-09-20", "categories": []}
{"id": "2409.12634", "title": "Exploring bat song syllable representations in self-supervised audio\n  encoders", "abstract": "  How well can deep learning models trained on human-generated sounds\ndistinguish between another species' vocalization types? We analyze the\nencoding of bat song syllables in several self-supervised audio encoders, and\nfind that models pre-trained on human speech generate the most distinctive\nrepresentations of different syllable types. These findings form first steps\ntowards the application of cross-species transfer learning in bat bioacoustics,\nas well as an improved understanding of out-of-distribution signal processing\nin audio encoder models.\n", "date": "2024-09-19", "categories": []}
{"id": "2409.11498", "title": "Augment, Drop & Swap: Improving Diversity in LLM Captions for Efficient\n  Music-Text Representation Learning", "abstract": "  Audio-text contrastive models have become a powerful approach in music\nrepresentation learning. Despite their empirical success, however, little is\nknown about the influence of key design choices on the quality of music-text\nrepresentations learnt through this framework. In this work, we expose these\ndesign choices within the constraints of limited data and computation budgets,\nand establish a more solid understanding of their impact grounded in empirical\nobservations along three axes: the choice of base encoders, the level of\ncuration in training data, and the use of text augmentation. We find that data\ncuration is the single most important factor for music-text contrastive\ntraining in resource-constrained scenarios. Motivated by this insight, we\nintroduce two novel techniques, Augmented View Dropout and TextSwap, which\nincrease the diversity and descriptiveness of text inputs seen in training.\nThrough our experiments we demonstrate that these are effective at boosting\nperformance across different pre-training regimes, model architectures, and\ndownstream data distributions, without incurring higher computational costs or\nrequiring additional training data.\n", "date": "2024-09-17", "categories": []}
{"id": "2409.11729", "title": "DETECLAP: Enhancing Audio-Visual Representation Learning with Object\n  Information", "abstract": "  Current audio-visual representation learning can capture rough object\ncategories (e.g., ``animals'' and ``instruments''), but it lacks the ability to\nrecognize fine-grained details, such as specific categories like ``dogs'' and\n``flutes'' within animals and instruments. To address this issue, we introduce\nDETECLAP, a method to enhance audio-visual representation learning with object\ninformation. Our key idea is to introduce an audio-visual label prediction loss\nto the existing Contrastive Audio-Visual Masked AutoEncoder to enhance its\nobject awareness. To avoid costly manual annotations, we prepare object labels\nfrom both audio and visual inputs using state-of-the-art language-audio models\nand object detectors. We evaluate the method of audio-visual retrieval and\nclassification using the VGGSound and AudioSet20K datasets. Our method achieves\nimprovements in recall@10 of +1.5% and +1.2% for audio-to-visual and\nvisual-to-audio retrieval, respectively, and an improvement in accuracy of\n+0.6% for audio-visual classification.\n", "date": "2024-09-18", "categories": []}
{"id": "2409.11964", "title": "Data Efficient Acoustic Scene Classification using Teacher-Informed\n  Confusing Class Instruction", "abstract": "  In this technical report, we describe the SNTL-NTU team's submission for Task\n1 Data-Efficient Low-Complexity Acoustic Scene Classification of the detection\nand classification of acoustic scenes and events (DCASE) 2024 challenge. Three\nsystems are introduced to tackle training splits of different sizes. For small\ntraining splits, we explored reducing the complexity of the provided baseline\nmodel by reducing the number of base channels. We introduce data augmentation\nin the form of mixup to increase the diversity of training samples. For the\nlarger training splits, we use FocusNet to provide confusing class information\nto an ensemble of multiple Patchout faSt Spectrogram Transformer (PaSST) models\nand baseline models trained on the original sampling rate of 44.1 kHz. We use\nKnowledge Distillation to distill the ensemble model to the baseline student\nmodel. Training the systems on the TAU Urban Acoustic Scene 2022 Mobile\ndevelopment dataset yielded the highest average testing accuracy of (62.21,\n59.82, 56.81, 53.03, 47.97)% on split (100, 50, 25, 10, 5)% respectively over\nthe three systems.\n", "date": "2024-09-18", "categories": []}
{"id": "2409.11308", "title": "SpMis: An Investigation of Synthetic Spoken Misinformation Detection", "abstract": "  In recent years, speech generation technology has advanced rapidly, fueled by\ngenerative models and large-scale training techniques. While these developments\nhave enabled the production of high-quality synthetic speech, they have also\nraised concerns about the misuse of this technology, particularly for\ngenerating synthetic misinformation. Current research primarily focuses on\ndistinguishing machine-generated speech from human-produced speech, but the\nmore urgent challenge is detecting misinformation within spoken content. This\ntask requires a thorough analysis of factors such as speaker identity, topic,\nand synthesis. To address this need, we conduct an initial investigation into\nsynthetic spoken misinformation detection by introducing an open-source\ndataset, SpMis. SpMis includes speech synthesized from over 1,000 speakers\nacross five common topics, utilizing state-of-the-art text-to-speech systems.\nAlthough our results show promising detection capabilities, they also reveal\nsubstantial challenges for practical implementation, underscoring the\nimportance of ongoing research in this critical area.\n", "date": "2024-09-17", "categories": []}
{"id": "2405.00934", "title": "Benchmarking Representations for Speech, Music, and Acoustic Events", "abstract": "  Limited diversity in standardized benchmarks for evaluating audio\nrepresentation learning (ARL) methods may hinder systematic comparison of\ncurrent methods' capabilities. We present ARCH, a comprehensive benchmark for\nevaluating ARL methods on diverse audio classification domains, covering\nacoustic events, music, and speech. ARCH comprises 12 datasets, that allow us\nto thoroughly assess pre-trained SSL models of different sizes. ARCH\nstreamlines benchmarking of ARL techniques through its unified access to a wide\nrange of domains and its ability to readily incorporate new datasets and\nmodels. To address the current lack of open-source, pre-trained models for\nnon-speech audio, we also release new pre-trained models that demonstrate\nstrong performance on non-speech datasets. We argue that the presented\nwide-ranging evaluation provides valuable insights into state-of-the-art ARL\nmethods, and is useful to pinpoint promising research directions.\n", "date": "2024-05-02", "categories": []}
{"id": "2406.14693", "title": "Voice Disorder Analysis: a Transformer-based Approach", "abstract": "  Voice disorders are pathologies significantly affecting patient quality of\nlife. However, non-invasive automated diagnosis of these pathologies is still\nunder-explored, due to both a shortage of pathological voice data, and\ndiversity of the recording types used for the diagnosis. This paper proposes a\nnovel solution that adopts transformers directly working on raw voice signals\nand addresses data shortage through synthetic data generation and data\naugmentation. Further, we consider many recording types at the same time, such\nas sentence reading and sustained vowel emission, by employing a Mixture of\nExpert ensemble to align the predictions on different data types. The\nexperimental results, obtained on both public and private datasets, show the\neffectiveness of our solution in the disorder detection and classification\ntasks and largely improve over existing approaches.\n", "date": "2024-06-20", "categories": []}
{"id": "2409.02489", "title": "NeuroSpex: Neuro-Guided Speaker Extraction with Cross-Modal Attention", "abstract": "  In the study of auditory attention, it has been revealed that there exists a\nrobust correlation between attended speech and elicited neural responses,\nmeasurable through electroencephalography (EEG). Therefore, it is possible to\nuse the attention information available within EEG signals to guide the\nextraction of the target speaker in a cocktail party computationally. In this\npaper, we present a neuro-guided speaker extraction model, i.e. NeuroSpex,\nusing the EEG response of the listener as the sole auxiliary reference cue to\nextract attended speech from monaural speech mixtures. We propose a novel EEG\nsignal encoder that captures the attention information. Additionally, we\npropose a cross-attention (CA) mechanism to enhance the speech feature\nrepresentations, generating a speaker extraction mask. Experimental results on\na publicly available dataset demonstrate that our proposed model outperforms\ntwo baseline models across various evaluation metrics.\n", "date": "2024-09-04", "categories": []}
{"id": "2409.09213", "title": "ReCLAP: Improving Zero Shot Audio Classification by Describing Sounds", "abstract": "  Open-vocabulary audio-language models, like CLAP, offer a promising approach\nfor zero-shot audio classification (ZSAC) by enabling classification with any\narbitrary set of categories specified with natural language prompts. In this\npaper, we propose a simple but effective method to improve ZSAC with CLAP.\nSpecifically, we shift from the conventional method of using prompts with\nabstract category labels (e.g., Sound of an organ) to prompts that describe\nsounds using their inherent descriptive features in a diverse context (e.g.,The\norgan's deep and resonant tones filled the cathedral.). To achieve this, we\nfirst propose ReCLAP, a CLAP model trained with rewritten audio captions for\nimproved understanding of sounds in the wild. These rewritten captions describe\neach sound event in the original caption using their unique discriminative\ncharacteristics. ReCLAP outperforms all baselines on both multi-modal\naudio-text retrieval and ZSAC. Next, to improve zero-shot audio classification\nwith ReCLAP, we propose prompt augmentation. In contrast to the traditional\nmethod of employing hand-written template prompts, we generate custom prompts\nfor each unique label in the dataset. These custom prompts first describe the\nsound event in the label and then employ them in diverse scenes. Our proposed\nmethod improves ReCLAP's performance on ZSAC by 1%-18% and outperforms all\nbaselines by 1% - 55%.\n", "date": "2024-09-13", "categories": []}
{"id": "2409.09272", "title": "SafeEar: Content Privacy-Preserving Audio Deepfake Detection", "abstract": "  Text-to-Speech (TTS) and Voice Conversion (VC) models have exhibited\nremarkable performance in generating realistic and natural audio. However,\ntheir dark side, audio deepfake poses a significant threat to both society and\nindividuals. Existing countermeasures largely focus on determining the\ngenuineness of speech based on complete original audio recordings, which\nhowever often contain private content. This oversight may refrain deepfake\ndetection from many applications, particularly in scenarios involving sensitive\ninformation like business secrets. In this paper, we propose SafeEar, a novel\nframework that aims to detect deepfake audios without relying on accessing the\nspeech content within. Our key idea is to devise a neural audio codec into a\nnovel decoupling model that well separates the semantic and acoustic\ninformation from audio samples, and only use the acoustic information (e.g.,\nprosody and timbre) for deepfake detection. In this way, no semantic content\nwill be exposed to the detector. To overcome the challenge of identifying\ndiverse deepfake audio without semantic clues, we enhance our deepfake detector\nwith real-world codec augmentation. Extensive experiments conducted on four\nbenchmark datasets demonstrate SafeEar's effectiveness in detecting various\ndeepfake techniques with an equal error rate (EER) down to 2.02%.\nSimultaneously, it shields five-language speech content from being deciphered\nby both machine and human auditory analysis, demonstrated by word error rates\n(WERs) all above 93.93% and our user study. Furthermore, our benchmark\nconstructed for anti-deepfake and anti-content recovery evaluation helps\nprovide a basis for future research in the realms of audio privacy preservation\nand deepfake detection.\n", "date": "2024-09-14", "categories": []}
{"id": "2409.09305", "title": "The T05 System for The VoiceMOS Challenge 2024: Transfer Learning from\n  Deep Image Classifier to Naturalness MOS Prediction of High-Quality Synthetic\n  Speech", "abstract": "  We present our system (denoted as T05) for the VoiceMOS Challenge (VMC) 2024.\nOur system was designed for the VMC 2024 Track 1, which focused on the accurate\nprediction of naturalness mean opinion score (MOS) for high-quality synthetic\nspeech. In addition to a pretrained self-supervised learning (SSL)-based speech\nfeature extractor, our system incorporates a pretrained image feature extractor\nto capture the difference of synthetic speech observed in speech spectrograms.\nWe first separately train two MOS predictors that use either of an SSL-based or\nspectrogram-based feature. Then, we fine-tune the two predictors for better MOS\nprediction using the fusion of two extracted features. In the VMC 2024 Track 1,\nour T05 system achieved first place in 7 out of 16 evaluation metrics and\nsecond place in the remaining 9 metrics, with a significant difference compared\nto those ranked third and below. We also report the results of our ablation\nstudy to investigate essential factors of our system.\n", "date": "2024-09-14", "categories": []}
{"id": "2409.09381", "title": "Text Prompt is Not Enough: Sound Event Enhanced Prompt Adapter for\n  Target Style Audio Generation", "abstract": "  Current mainstream audio generation methods primarily rely on simple text\nprompts, often failing to capture the nuanced details necessary for multi-style\naudio generation. To address this limitation, the Sound Event Enhanced Prompt\nAdapter is proposed. Unlike traditional static global style transfer, this\nmethod extracts style embedding through cross-attention between text and\nreference audio for adaptive style control. Adaptive layer normalization is\nthen utilized to enhance the model's capacity to express multiple styles.\nAdditionally, the Sound Event Reference Style Transfer Dataset (SERST) is\nintroduced for the proposed target style audio generation task, enabling\ndual-prompt audio generation using both text and audio references. Experimental\nresults demonstrate the robustness of the model, achieving state-of-the-art\nFr\\'echet Distance of 26.94 and KL Divergence of 1.82, surpassing Tango,\nAudioLDM, and AudioGen. Furthermore, the generated audio shows high similarity\nto its corresponding audio reference. The demo, code, and dataset are publicly\navailable.\n", "date": "2024-09-14", "categories": []}
{"id": "2409.08655", "title": "LMAC-TD: Producing Time Domain Explanations for Audio Classifiers", "abstract": "  Neural networks are typically black-boxes that remain opaque with regards to\ntheir decision mechanisms. Several works in the literature have proposed\npost-hoc explanation methods to alleviate this issue. This paper proposes\nLMAC-TD, a post-hoc explanation method that trains a decoder to produce\nexplanations directly in the time domain. This methodology builds upon the\nfoundation of L-MAC, Listenable Maps for Audio Classifiers, a method that\nproduces faithful and listenable explanations. We incorporate SepFormer, a\npopular transformer-based time-domain source separation architecture. We show\nthrough a user study that LMAC-TD significantly improves the audio quality of\nthe produced explanations while not sacrificing from faithfulness.\n", "date": "2024-09-13", "categories": []}
{"id": "2409.08673", "title": "Acoustic identification of individual animals with hierarchical\n  contrastive learning", "abstract": "  Acoustic identification of individual animals (AIID) is closely related to\naudio-based species classification but requires a finer level of detail to\ndistinguish between individual animals within the same species. In this work,\nwe frame AIID as a hierarchical multi-label classification task and propose the\nuse of hierarchy-aware loss functions to learn robust representations of\nindividual identities that maintain the hierarchical relationships among\nspecies and taxa. Our results demonstrate that hierarchical embeddings not only\nenhance identification accuracy at the individual level but also at higher\ntaxonomic levels, effectively preserving the hierarchical structure in the\nlearned representations. By comparing our approach with non-hierarchical\nmodels, we highlight the advantage of enforcing this structure in the embedding\nspace. Additionally, we extend the evaluation to the classification of novel\nindividual classes, demonstrating the potential of our method in open-set\nclassification scenarios.\n", "date": "2024-09-13", "categories": []}
{"id": "2409.08997", "title": "Biomimetic Frontend for Differentiable Audio Processing", "abstract": "  While models in audio and speech processing are becoming deeper and more\nend-to-end, they as a consequence need expensive training on large data, and\nare often brittle. We build on a classical model of human hearing and make it\ndifferentiable, so that we can combine traditional explainable biomimetic\nsignal processing approaches with deep-learning frameworks. This allows us to\narrive at an expressive and explainable model that is easily trained on modest\namounts of data. We apply this model to audio processing tasks, including\nclassification and enhancement. Results show that our differentiable model\nsurpasses black-box approaches in terms of computational efficiency and\nrobustness, even with little training data. We also discuss other potential\napplications.\n", "date": "2024-09-13", "categories": []}
{"id": "2303.05896", "title": "Distribution Preserving Source Separation With Time Frequency Predictive\n  Models", "abstract": "  We provide an example of a distribution preserving source separation method,\nwhich aims at addressing perceptual shortcomings of state-of-the-art methods.\nOur approach uses unconditioned generative models of signal sources.\nReconstruction is achieved by means of mix-consistent sampling from a\ndistribution conditioned on a realization of a mix. The separated signals\nfollow their respective source distributions, which provides an advantage when\nseparation results are evaluated in a listening test.\n", "date": "2023-03-10", "categories": []}
{"id": "2409.07858", "title": "Audio Decoding by Inverse Problem Solving", "abstract": "  We consider audio decoding as an inverse problem and solve it through\ndiffusion posterior sampling. Explicit conditioning functions are developed for\ninput signal measurements provided by an example of a transform domain\nperceptual audio codec. Viability is demonstrated by evaluating arbitrary\npairings of a set of bitrates and task-agnostic prior models. For instance, we\nobserve significant improvements on piano while maintaining speech performance\nwhen a speech model is replaced by a joint model trained on both speech and\npiano. With a more general music model, improved decoding compared to legacy\nmethods is obtained for a broad range of content types and bitrates. The noisy\nmean model, underlying the proposed derivation of conditioning, enables a\nsignificant reduction of gradient evaluations for diffusion posterior sampling,\ncompared to methods based on Tweedie's mean. Combining Tweedie's mean with our\nconditioning functions improves the objective performance. An audio demo is\navailable at https://dpscodec-demo.github.io/.\n", "date": "2024-09-12", "categories": []}
{"id": "2409.06196", "title": "MTDA-HSED: Mutual-Assistance Tuning and Dual-Branch Aggregating for\n  Heterogeneous Sound Event Detection", "abstract": "  Sound Event Detection (SED) plays a vital role in comprehending and\nperceiving acoustic scenes. Previous methods have demonstrated impressive\ncapabilities. However, they are deficient in learning features of complex\nscenes from heterogeneous dataset. In this paper, we introduce a novel\ndual-branch architecture named Mutual-Assistance Tuning and Dual-Branch\nAggregating for Heterogeneous Sound Event Detection (MTDA-HSED). The MTDA-HSED\narchitecture employs the Mutual-Assistance Audio Adapter (M3A) to effectively\ntackle the multi-scenario problem and uses the Dual-Branch Mid-Fusion (DBMF)\nmodule to tackle the multi-granularity problem. Specifically, M3A is integrated\ninto the BEATs block as an adapter to improve the BEATs' performance by\nfine-tuning it on the multi-scenario dataset. The DBMF module connects BEATs\nand CNN branches, which facilitates the deep fusion of information from the\nBEATs and the CNN branches. Experimental results show that the proposed methods\nexceed the baseline of mpAUC by \\textbf{$5\\%$} on the DESED and MAESTRO Real\ndatasets. Code is available at https://github.com/Visitor-W/MTDA.\n", "date": "2024-09-10", "categories": []}
{"id": "2409.06033", "title": "Investigating Causal Cues: Strengthening Spoofed Audio Detection with\n  Human-Discernible Linguistic Features", "abstract": "  Several types of spoofed audio, such as mimicry, replay attacks, and\ndeepfakes, have created societal challenges to information integrity. Recently,\nresearchers have worked with sociolinguistics experts to label spoofed audio\nsamples with Expert Defined Linguistic Features (EDLFs) that can be discerned\nby the human ear: pitch, pause, word-initial and word-final release bursts of\nconsonant stops, audible intake or outtake of breath, and overall audio\nquality. It is established that there is an improvement in several deepfake\ndetection algorithms when they augmented the traditional and common features of\naudio data with these EDLFs. In this paper, using a hybrid dataset comprised of\nmultiple types of spoofed audio augmented with sociolinguistic annotations, we\ninvestigate causal discovery and inferences between the discernible linguistic\nfeatures and the label in the audio clips, comparing the findings of the causal\nmodels with the expert ground truth validation labeling process. Our findings\nsuggest that the causal models indicate the utility of incorporating linguistic\nfeatures to help discern spoofed audio, as well as the overall need and\nopportunity to incorporate human knowledge into models and techniques for\nstrengthening AI models. The causal discovery and inference can be used as a\nfoundation of training humans to discern spoofed audio as well as automating\nEDLFs labeling for the purpose of performance improvement of the common\nAI-based spoofed audio detectors.\n", "date": "2024-09-09", "categories": []}
{"id": "2409.06348", "title": "VoiceWukong: Benchmarking Deepfake Voice Detection", "abstract": "  With the rapid advancement of technologies like text-to-speech (TTS) and\nvoice conversion (VC), detecting deepfake voices has become increasingly\ncrucial. However, both academia and industry lack a comprehensive and intuitive\nbenchmark for evaluating detectors. Existing datasets are limited in language\ndiversity and lack many manipulations encountered in real-world production\nenvironments.\n  To fill this gap, we propose VoiceWukong, a benchmark designed to evaluate\nthe performance of deepfake voice detectors. To build the dataset, we first\ncollected deepfake voices generated by 19 advanced and widely recognized\ncommercial tools and 15 open-source tools. We then created 38 data variants\ncovering six types of manipulations, constructing the evaluation dataset for\ndeepfake voice detection. VoiceWukong thus includes 265,200 English and 148,200\nChinese deepfake voice samples. Using VoiceWukong, we evaluated 12\nstate-of-the-art detectors. AASIST2 achieved the best equal error rate (EER) of\n13.50%, while all others exceeded 20%. Our findings reveal that these detectors\nface significant challenges in real-world applications, with dramatically\ndeclining performance. In addition, we conducted a user study with more than\n300 participants. The results are compared with the performance of the 12\ndetectors and a multimodel large language model (MLLM), i.e., Qwen2-Audio,\nwhere different detectors and humans exhibit varying identification\ncapabilities for deepfake voices at different deception levels, while the LALM\ndemonstrates no detection ability at all. Furthermore, we provide a leaderboard\nfor deepfake voice detection, publicly available at\n{https://voicewukong.github.io}.\n", "date": "2024-09-10", "categories": []}
{"id": "2409.06372", "title": "SpeechTaxi: On Multilingual Semantic Speech Classification", "abstract": "  Recent advancements in multilingual speech encoding as well as transcription\nraise the question of the most effective approach to semantic speech\nclassification. Concretely, can (1) end-to-end (E2E) classifiers obtained by\nfine-tuning state-of-the-art multilingual speech encoders (MSEs) match or\nsurpass the performance of (2) cascading (CA), where speech is first\ntranscribed into text and classification is delegated to a text-based\nclassifier. To answer this, we first construct SpeechTaxi, an 80-hour\nmultilingual dataset for semantic speech classification of Bible verses,\ncovering 28 diverse languages. We then leverage SpeechTaxi to conduct a wide\nrange of experiments comparing E2E and CA in monolingual semantic speech\nclassification as well as in cross-lingual transfer. We find that E2E based on\nMSEs outperforms CA in monolingual setups, i.e., when trained on in-language\ndata. However, MSEs seem to have poor cross-lingual transfer abilities, with\nE2E substantially lagging CA both in (1) zero-shot transfer to languages unseen\nin training and (2) multilingual training, i.e., joint training on multiple\nlanguages. Finally, we devise a novel CA approach based on transcription to\nRomanized text as a language-agnostic intermediate representation and show that\nit represents a robust solution for languages without native ASR support. Our\nSpeechTaxi dataset is publicly available at: https://huggingface.co/\ndatasets/LennartKeller/SpeechTaxi/.\n", "date": "2024-09-10", "categories": []}
{"id": "2309.11500", "title": "Auto-ACD: A Large-scale Dataset for Audio-Language Representation\n  Learning", "abstract": "  Recently, the AI community has made significant strides in developing\npowerful foundation models, driven by large-scale multimodal datasets. However,\nfor audio representation learning, existing datasets suffer from limitations in\nthe following aspects: insufficient volume, simplistic content, and arduous\ncollection procedures. To establish an audio dataset with high-quality\ncaptions, we propose an innovative, automatic approach leveraging multimodal\ninputs, such as video frames, audio streams. Specifically, we construct a\nlarge-scale, high-quality, audio-language dataset, named as Auto-ACD,\ncomprising over 1.5M audio-text pairs. We exploit a series of pre-trained\nmodels or APIs, to determine audio-visual synchronisation, generate image\ncaptions, object detection, or audio tags for specific videos. Subsequently, we\nemploy LLM to paraphrase a congruent caption for each audio, guided by the\nextracted multi-modality clues. To demonstrate the effectiveness of the\nproposed dataset, we train widely used models on our dataset and show\nperformance improvement on various downstream tasks, for example,\naudio-language retrieval, audio captioning, zero-shot classification. In\naddition, we establish a novel benchmark with environmental information and\nprovide a benchmark for audio-text tasks.\n", "date": "2023-09-20", "categories": []}
{"id": "2409.04949", "title": "Attention-Based Efficient Breath Sound Removal in Studio Audio\n  Recordings", "abstract": "  In this research, we present an innovative, parameter-efficient model that\nutilizes the attention U-Net architecture for the automatic detection and\neradication of non-speech vocal sounds, specifically breath sounds, in vocal\nrecordings. This task is of paramount importance in the field of sound\nengineering, despite being relatively under-explored. The conventional manual\nprocess for detecting and eliminating these sounds requires significant\nexpertise and is extremely time-intensive. Existing automated detection and\nremoval methods often fall short in terms of efficiency and precision. Our\nproposed model addresses these limitations by offering a streamlined process\nand superior accuracy, achieved through the application of advanced deep\nlearning techniques. A unique dataset, derived from Device and Produced Speech\n(DAPS), was employed for this purpose. The training phase of the model\nemphasizes a log spectrogram and integrates an early stopping mechanism to\nprevent overfitting. Our model not only conserves precious time for sound\nengineers but also enhances the quality and consistency of audio production.\nThis constitutes a significant breakthrough, as evidenced by its comparative\nefficiency, necessitating only 1.9M parameters and a training duration of 3.2\nhours - markedly less than the top-performing models in this domain. The model\nis capable of generating identical outputs as previous models with drastically\nimproved precision, making it an optimal choice.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.04953", "title": "Evaluating Neural Networks Architectures for Spring Reverb Modelling", "abstract": "  Reverberation is a key element in spatial audio perception, historically\nachieved with the use of analogue devices, such as plate and spring reverb, and\nin the last decades with digital signal processing techniques that have allowed\ndifferent approaches for Virtual Analogue Modelling (VAM). The\nelectromechanical functioning of the spring reverb makes it a nonlinear system\nthat is difficult to fully emulate in the digital domain with white-box\nmodelling techniques. In this study, we compare five different neural network\narchitectures, including convolutional and recurrent models, to assess their\neffectiveness in replicating the characteristics of this audio effect. The\nevaluation is conducted on two datasets at sampling rates of 16 kHz and 48 kHz.\nThis paper specifically focuses on neural audio architectures that offer\nparametric control, aiming to advance the boundaries of current black-box\nmodelling techniques in the domain of spring reverberation.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.05035", "title": "Deep Generic Representations for Domain-Generalized Anomalous Sound\n  Detection", "abstract": "  Developing a reliable anomalous sound detection (ASD) system requires\nrobustness to noise, adaptation to domain shifts, and effective performance\nwith limited training data. Current leading methods rely on extensive labeled\ndata for each target machine type to train feature extractors using\nOutlier-Exposure (OE) techniques, yet their performance on the target domain\nremains sub-optimal. In this paper, we present \\textit{GenRep}, which utilizes\ngeneric feature representations from a robust, large-scale pre-trained feature\nextractor combined with kNN for domain-generalized ASD, without the need for\nfine-tuning. \\textit{GenRep} incorporates MemMixup, a simple approach for\naugmenting the target memory bank using nearest source samples, paired with a\ndomain normalization technique to address the imbalance between source and\ntarget domains. \\textit{GenRep} outperforms the best OE-based approach without\na need for labeled data with an Official Score of 73.79\\% on the DCASE2023T2\nEval set and demonstrates robustness under limited data scenarios. The code is\navailable open-source.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.05095", "title": "The first Cadenza challenges: using machine learning competitions to\n  improve music for listeners with a hearing loss", "abstract": "  It is well established that listening to music is an issue for those with\nhearing loss, and hearing aids are not a universal solution. How can machine\nlearning be used to address this? This paper details the first application of\nthe open challenge methodology to use machine learning to improve audio quality\nof music for those with hearing loss. The first challenge was a stand-alone\ncompetition (CAD1) and had 9 entrants. The second was an 2024 ICASSP grand\nchallenge (ICASSP24) and attracted 17 entrants. The challenge tasks concerned\ndemixing and remixing pop/rock music to allow a personalised rebalancing of the\ninstruments in the mix, along with amplification to correct for raised hearing\nthresholds. The software baselines provided for entrants to build upon used two\nstate-of-the-art demix algorithms: Hybrid Demucs and Open-Unmix. Evaluation of\nsystems was done using the objective metric HAAQI, the Hearing-Aid Audio\nQuality Index. No entrants improved on the best baseline in CAD1 because there\nwas insufficient room for improvement. Consequently, for ICASSP24 the scenario\nwas made more difficult by using loudspeaker reproduction and specified gains\nto be applied before remixing. This also made the scenario more useful for\nlistening through hearing aids. 9 entrants scored better than the the best\nICASSP24 baseline. Most entrants used a refined version of Hybrid Demucs and\nNAL-R amplification. The highest scoring system combined the outputs of several\ndemixing algorithms in an ensemble approach. These challenges are now open\nbenchmarks for future research with the software and data being freely\navailable.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.05148", "title": "Better Spanish Emotion Recognition In-the-wild: Bringing Attention to\n  Deep Spectrum Voice Analysis", "abstract": "  Within the context of creating new Socially Assistive Robots, emotion\nrecognition has become a key development factor, as it allows the robot to\nadapt to the user's emotional state in the wild. In this work, we focused on\nthe analysis of two voice recording Spanish datasets: ELRA-S0329 and\nEmoMatchSpanishDB. Specifically, we centered our work in the paralanguage,\ne.~g. the vocal characteristics that go along with the message and clarifies\nthe meaning. We proposed the use of the DeepSpectrum method, which consists of\nextracting a visual representation of the audio tracks and feeding them to a\npretrained CNN model. For the classification task, DeepSpectrum is often paired\nwith a Support Vector Classifier --DS-SVC--, or a Fully-Connected deep-learning\nclassifier --DS-FC--. We compared the results of the DS-SVC and DS-FC\narchitectures with the state-of-the-art (SOTA) for ELRA-S0329 and\nEmoMatchSpanishDB. Moreover, we proposed our own classifier based upon\nAttention Mechanisms, namely DS-AM. We trained all models against both\ndatasets, and we found that our DS-AM model outperforms the SOTA models for the\ndatasets and the SOTA DeepSpectrum architectures. Finally, we trained our DS-AM\nmodel in one dataset and tested it in the other, to simulate real-world\nconditions on how biased is the model to the dataset.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.05319", "title": "Machine Anomalous Sound Detection Using Spectral-temporal Modulation\n  Representations Derived from Machine-specific Filterbanks", "abstract": "  Early detection of factory machinery malfunctions is crucial in industrial\napplications. In machine anomalous sound detection (ASD), different machines\nexhibit unique vibration-frequency ranges based on their physical properties.\nMeanwhile, the human auditory system is adept at tracking both temporal and\nspectral dynamics of machine sounds. Consequently, integrating the\ncomputational auditory models of the human auditory system with\nmachine-specific properties can be an effective approach to machine ASD. We\nfirst quantified the frequency importances of four types of machines using the\nFisher ratio (F-ratio). The quantified frequency importances were then used to\ndesign machine-specific non-uniform filterbanks (NUFBs), which extract the log\nnon-uniform spectrum (LNS) feature. The designed NUFBs have a narrower\nbandwidth and higher filter distribution density in frequency regions with\nrelatively high F-ratios. Finally, spectral and temporal modulation\nrepresentations derived from the LNS feature were proposed. These proposed LNS\nfeature and modulation representations are input into an autoencoder\nneural-network-based detector for ASD. The quantification results from the\ntraining set of the Malfunctioning Industrial Machine Investigation and\nInspection dataset with a signal-to-noise (SNR) of 6 dB reveal that the\ndistinguishing information between normal and anomalous sounds of different\nmachines is encoded non-uniformly in the frequency domain. By highlighting\nthese important frequency regions using NUFBs, the LNS feature can\nsignificantly enhance performance using the metric of AUC (area under the\nreceiver operating characteristic curve) under various SNR conditions.\nFurthermore, modulation representations can further improve performance.\nSpecifically, temporal modulation is effective for fans, pumps, and sliders,\nwhile spectral modulation is particularly effective for valves.\n", "date": "2024-09-09", "categories": []}
{"id": "2408.15667", "title": "Towards reliable respiratory disease diagnosis based on cough sounds and\n  vision transformers", "abstract": "  Recent advancements in deep learning techniques have sparked performance\nboosts in various real-world applications including disease diagnosis based on\nmulti-modal medical data. Cough sound data-based respiratory disease (e.g.,\nCOVID-19 and Chronic Obstructive Pulmonary Disease) diagnosis has also\nattracted much attention. However, existing works usually utilise traditional\nmachine learning or deep models of moderate scales. On the other hand, the\ndeveloped approaches are trained and evaluated on small-scale data due to the\ndifficulty of curating and annotating clinical data on scale. To address these\nissues in prior works, we create a unified framework to evaluate various deep\nmodels from lightweight Convolutional Neural Networks (e.g., ResNet18) to\nmodern vision transformers and compare their performance in respiratory disease\nclassification. Based on the observations from such an extensive empirical\nstudy, we propose a novel approach to cough-based disease classification based\non both self-supervised and supervised learning on a large-scale cough data\nset. Experimental results demonstrate our proposed approach outperforms prior\narts consistently on two benchmark datasets for COVID-19 diagnosis and a\nproprietary dataset for COPD/non-COPD classification with an AUROC of 92.5%.\n", "date": "2024-08-28", "categories": []}
{"id": "2409.00408", "title": "Multi-label Zero-Shot Audio Classification with Temporal Attention", "abstract": "  Zero-shot learning models are capable of classifying new classes by\ntransferring knowledge from the seen classes using auxiliary information. While\nmost of the existing zero-shot learning methods focused on single-label\nclassification tasks, the present study introduces a method to perform\nmulti-label zero-shot audio classification. To address the challenge of\nclassifying multi-label sounds while generalizing to unseen classes, we adapt\ntemporal attention. The temporal attention mechanism assigns importance weights\nto different audio segments based on their acoustic and semantic compatibility,\nthus enabling the model to capture the varying dominance of different sound\nclasses within an audio sample by focusing on the segments most relevant for\neach class. This leads to more accurate multi-label zero-shot classification\nthan methods employing temporally aggregated acoustic features without\nweighting, which treat all audio segments equally. We evaluate our approach on\na subset of AudioSet against a zero-shot model using uniformly aggregated\nacoustic features, a zero-rule baseline, and the proposed method in the\nsupervised scenario. Our results show that temporal attention enhances the\nzero-shot audio classification performance in multi-label scenario.\n", "date": "2024-08-31", "categories": []}
{"id": "2409.00724", "title": "BUET Multi-disease Heart Sound Dataset: A Comprehensive Auscultation\n  Dataset for Developing Computer-Aided Diagnostic Systems", "abstract": "  Cardiac auscultation, an integral tool in diagnosing cardiovascular diseases\n(CVDs), often relies on the subjective interpretation of clinicians, presenting\na limitation in consistency and accuracy. Addressing this, we introduce the\nBUET Multi-disease Heart Sound (BMD-HS) dataset - a comprehensive and\nmeticulously curated collection of heart sound recordings. This dataset,\nencompassing 864 recordings across five distinct classes of common heart\nsounds, represents a broad spectrum of valvular heart diseases, with a focus on\ndiagnostically challenging cases. The standout feature of the BMD-HS dataset is\nits innovative multi-label annotation system, which captures a diverse range of\ndiseases and unique disease states. This system significantly enhances the\ndataset's utility for developing advanced machine learning models in automated\nheart sound classification and diagnosis. By bridging the gap between\ntraditional auscultation practices and contemporary data-driven diagnostic\nmethods, the BMD-HS dataset is poised to revolutionize CVD diagnosis and\nmanagement, providing an invaluable resource for the advancement of cardiac\nhealth research. The dataset is publicly available at this link:\nhttps://github.com/mHealthBuet/BMD-HS-Dataset.\n", "date": "2024-09-01", "categories": []}
{"id": "2409.01160", "title": "Expanding on EnCLAP with Auxiliary Retrieval Model for Automated Audio\n  Captioning", "abstract": "  In this technical report, we describe our submission to DCASE2024 Challenge\nTask6 (Automated Audio Captioning) and Task8 (Language-based Audio Retrieval).\nWe develop our approach building upon the EnCLAP audio captioning framework and\noptimizing it for Task6 of the challenge. Notably, we outline the changes in\nthe underlying components and the incorporation of the reranking process.\nAdditionally, we submit a supplementary retriever model, a byproduct of our\nmodified framework, to Task8. Our proposed systems achieve FENSE score of 0.542\non Task6 and mAP@10 score of 0.386 on Task8, significantly outperforming the\nbaseline models.\n", "date": "2024-09-02", "categories": []}
{"id": "2409.01201", "title": "EnCLAP++: Analyzing the EnCLAP Framework for Optimizing Automated Audio\n  Captioning Performance", "abstract": "  In this work, we aim to analyze and optimize the EnCLAP framework, a\nstate-of-the-art model in automated audio captioning. We investigate the impact\nof modifying the acoustic encoder components, explore pretraining with\ndifferent dataset scales, and study the effectiveness of a reranking scheme.\nThrough extensive experimentation and quantitative analysis of generated\ncaptions, we develop EnCLAP++, an enhanced version that significantly surpasses\nthe original.\n", "date": "2024-09-02", "categories": []}
{"id": "2409.01695", "title": "USTC-KXDIGIT System Description for ASVspoof5 Challenge", "abstract": "  This paper describes the USTC-KXDIGIT system submitted to the ASVspoof5\nChallenge for Track 1 (speech deepfake detection) and Track 2 (spoofing-robust\nautomatic speaker verification, SASV). Track 1 showcases a diverse range of\ntechnical qualities from potential processing algorithms and includes both open\nand closed conditions. For these conditions, our system consists of a cascade\nof a frontend feature extractor and a back-end classifier. We focus on\nextensive embedding engineering and enhancing the generalization of the\nback-end classifier model. Specifically, the embedding engineering is based on\nhand-crafted features and speech representations from a self-supervised model,\nused for closed and open conditions, respectively. To detect spoof attacks\nunder various adversarial conditions, we trained multiple systems on an\naugmented training set. Additionally, we used voice conversion technology to\nsynthesize fake audio from genuine audio in the training set to enrich the\nsynthesis algorithms. To leverage the complementary information learned by\ndifferent model architectures, we employed activation ensemble and fused scores\nfrom different systems to obtain the final decision score for spoof detection.\nDuring the evaluation phase, the proposed methods achieved 0.3948 minDCF and\n14.33% EER in the close condition, and 0.0750 minDCF and 2.59% EER in the open\ncondition, demonstrating the robustness of our submitted systems under\nadversarial conditions. In Track 2, we continued using the CM system from Track\n1 and fused it with a CNN-based ASV system. This approach achieved 0.2814\nmin-aDCF in the closed condition and 0.0756 min-aDCF in the open condition,\nshowcasing superior performance in the SASV system.\n", "date": "2024-09-03", "categories": []}
{"id": "2409.01885", "title": "Activity-Guided Industrial Anomalous Sound Detection against\n  Interferences", "abstract": "  We address a practical scenario of anomaly detection for industrial sound\ndata, where the sound of a target machine is corrupted by background noise and\ninterference from neighboring machines. Overcoming this challenge is difficult\nsince the interference is often virtually indistinguishable from the target\nmachine without additional information. To address the issue, we propose SSAD,\na framework of source separation (SS) followed by anomaly detection (AD), which\nleverages machine activity information, often readily available in practical\nsettings. SSAD consists of two components: (i) activity-informed SS, enabling\neffective source separation even given interference with similar timbre, and\n(ii) two-step masking, robustifying anomaly detection by emphasizing anomalies\naligned with the machine activity. Our experiments demonstrate that SSAD\nachieves comparable accuracy to a baseline with full access to clean signals,\nwhile SSAD is provided only a corrupted signal and activity information. In\naddition, thanks to the activity-informed SS and AD with the two-step masking,\nSSAD outperforms standard approaches, particularly in cases with interference.\nIt highlights the practical efficacy of SSAD in addressing the complexities of\nanomaly detection in industrial sound data.\n", "date": "2024-09-03", "categories": []}
{"id": "2211.09944", "title": "MelHuBERT: A simplified HuBERT on Mel spectrograms", "abstract": "  Self-supervised models have had great success in learning speech\nrepresentations that can generalize to various downstream tasks. However, most\nself-supervised models require a large amount of compute and multiple GPUs to\ntrain, significantly hampering the development of self-supervised learning. In\nan attempt to reduce the computation of training, we revisit the training of\nHuBERT, a highly successful self-supervised model. We improve and simplify\nseveral key components, including the loss function, input representation, and\ntraining in multiple stages. Our model, MelHuBERT, is able to achieve favorable\nperformance on phone recognition, speaker identification, and automatic speech\nrecognition against HuBERT, while saving 31.2% of the pre-training time, or\nequivalently 33.5% MACs per one second speech. The code and pre-trained models\nare available in https://github.com/nervjack2/MelHuBERT.\n", "date": "2022-11-17", "categories": []}
{"id": "2408.17166", "title": "Learning Multi-Target TDOA Features for Sound Event Localization and\n  Detection", "abstract": "  Sound event localization and detection (SELD) systems using audio recordings\nfrom a microphone array rely on spatial cues for determining the location of\nsound events. As a consequence, the localization performance of such systems is\nto a large extent determined by the quality of the audio features that are used\nas inputs to the system. We propose a new feature, based on neural generalized\ncross-correlations with phase-transform (NGCC-PHAT), that learns audio\nrepresentations suitable for localization. Using permutation invariant training\nfor the time-difference of arrival (TDOA) estimation problem enables NGCC-PHAT\nto learn TDOA features for multiple overlapping sound events. These features\ncan be used as a drop-in replacement for GCC-PHAT inputs to a SELD-network. We\ntest our method on the STARSS23 dataset and demonstrate improved localization\nperformance compared to using standard GCC-PHAT or SALSA-Lite input features.\n", "date": "2024-08-30", "categories": []}
{"id": "2406.08800", "title": "Can Synthetic Audio From Generative Foundation Models Assist Audio\n  Recognition and Speech Modeling?", "abstract": "  Recent advances in foundation models have enabled audio-generative models\nthat produce high-fidelity sounds associated with music, events, and human\nactions. Despite the success achieved in modern audio-generative models, the\nconventional approach to assessing the quality of the audio generation relies\nheavily on distance metrics like Frechet Audio Distance. In contrast, we aim to\nevaluate the quality of audio generation by examining the effectiveness of\nusing them as training data. Specifically, we conduct studies to explore the\nuse of synthetic audio for audio recognition. Moreover, we investigate whether\nsynthetic audio can serve as a resource for data augmentation in speech-related\nmodeling. Our comprehensive experiments demonstrate the potential of using\nsynthetic audio for audio recognition and speech-related modeling. Our code is\navailable at https://github.com/usc-sail/SynthAudio.\n", "date": "2024-06-13", "categories": []}
{"id": "2408.16448", "title": "Enhancing Sound Source Localization via False Negative Elimination", "abstract": "  Sound source localization aims to localize objects emitting the sound in\nvisual scenes. Recent works obtaining impressive results typically rely on\ncontrastive learning. However, the common practice of randomly sampling\nnegatives in prior arts can lead to the false negative issue, where the sounds\nsemantically similar to visual instance are sampled as negatives and\nincorrectly pushed away from the visual anchor/query. As a result, this\nmisalignment of audio and visual features could yield inferior performance. To\naddress this issue, we propose a novel audio-visual learning framework which is\ninstantiated with two individual learning schemes: self-supervised predictive\nlearning (SSPL) and semantic-aware contrastive learning (SACL). SSPL explores\nimage-audio positive pairs alone to discover semantically coherent similarities\nbetween audio and visual features, while a predictive coding module for feature\nalignment is introduced to facilitate the positive-only learning. In this\nregard SSPL acts as a negative-free method to eliminate false negatives. By\ncontrast, SACL is designed to compact visual features and remove false\nnegatives, providing reliable visual anchor and audio negatives for contrast.\nDifferent from SSPL, SACL releases the potential of audio-visual contrastive\nlearning, offering an effective alternative to achieve the same goal.\nComprehensive experiments demonstrate the superiority of our approach over the\nstate-of-the-arts. Furthermore, we highlight the versatility of the learned\nrepresentation by extending the approach to audio-visual event classification\nand object detection tasks. Code and models are available at:\nhttps://github.com/zjsong/SACL.\n", "date": "2024-08-29", "categories": []}
{"id": "2408.15296", "title": "Feature Representations for Automatic Meerkat Vocalization\n  Classification", "abstract": "  Understanding evolution of vocal communication in social animals is an\nimportant research problem. In that context, beyond humans, there is an\ninterest in analyzing vocalizations of other social animals such as, meerkats,\nmarmosets, apes. While existing approaches address vocalizations of certain\nspecies, a reliable method tailored for meerkat calls is lacking. To that\nextent, this paper investigates feature representations for automatic meerkat\nvocalization analysis. Both traditional signal processing-based representations\nand data-driven representations facilitated by advances in deep learning are\nexplored. Call type classification studies conducted on two data sets reveal\nthat feature extraction methods developed for human speech processing can be\neffectively employed for automatic meerkat call analysis.\n", "date": "2024-08-27", "categories": []}
{"id": "2408.14753", "title": "CoopASD: Cooperative Machine Anomalous Sound Detection with Privacy\n  Concerns", "abstract": "  Machine anomalous sound detection (ASD) has emerged as one of the most\npromising applications in the Industrial Internet of Things (IIoT) due to its\nunprecedented efficacy in mitigating risks of malfunctions and promoting\nproduction efficiency. Previous works mainly investigated the machine ASD task\nunder centralized settings. However, developing the ASD system under\ndecentralized settings is crucial in practice, since the machine data are\ndispersed in various factories and the data should not be explicitly shared due\nto privacy concerns. To enable these factories to cooperatively develop a\nscalable ASD model while preserving their privacy, we propose a novel framework\nnamed CoopASD, where each factory trains an ASD model on its local dataset, and\na central server aggregates these local models periodically. We employ a\npre-trained model as the backbone of the ASD model to improve its robustness\nand develop specialized techniques to stabilize the model under a completely\nnon-iid and domain shift setting. Compared with previous state-of-the-art\n(SOTA) models trained in centralized settings, CoopASD showcases competitive\nresults with negligible degradation of 0.08%. We also conduct extensive\nablation studies to demonstrate the effectiveness of CoopASD.\n", "date": "2024-08-27", "categories": []}
{"id": "2408.14777", "title": "Quartered Chirp Spectral Envelope for Whispered vs Normal Speech\n  Classification", "abstract": "  Whispered speech as an acceptable form of human-computer interaction is\ngaining traction. Systems that address multiple modes of speech require a\nrobust front-end speech classifier. Performance of whispered vs normal speech\nclassification drops in the presence of additive white Gaussian noise, since\nnormal speech takes on some of the characteristics of whispered speech. In this\nwork, we propose a new feature named the quartered chirp spectral envelope, a\ncombination of the chirp spectrum and the quartered spectral envelope, to\nclassify whispered and normal speech. The chirp spectrum can be fine-tuned to\nobtain customized features for a given task, and the quartered spectral\nenvelope has been proven to work especially well for the current task. The\nfeature is trained on a one dimensional convolutional neural network, that\ncaptures the trends in the spectral envelope. The proposed system performs\nbetter than the state of the art, in the presence of white noise.\n", "date": "2024-08-27", "categories": []}
{"id": "2408.14797", "title": "MaskCycleGAN-based Whisper to Normal Speech Conversion", "abstract": "  Whisper to normal speech conversion is an active area of research. Various\narchitectures based on generative adversarial networks have been proposed in\nthe recent past. Especially, recent study shows that MaskCycleGAN, which is a\nmask guided, and cyclic consistency keeping, generative adversarial network,\nperforms really well for voice conversion from spectrogram representations. In\nthe current work we present a MaskCycleGAN approach for the conversion of\nwhispered speech to normal speech. We find that tuning the mask parameters, and\npre-processing the signal with a voice activity detector provides superior\nperformance when compared to the existing approach. The wTIMIT dataset is used\nfor evaluation. Objective metrics such as PESQ and G-Loss are used to evaluate\nthe converted speech, along with subjective evaluation using mean opinion\nscore. The results show that the proposed approach offers considerable\nbenefits.\n", "date": "2024-08-27", "categories": []}
{"id": "2408.14887", "title": "Literary and Colloquial Dialect Identification for Tamil using Acoustic\n  Features", "abstract": "  The evolution and diversity of a language is evident from it's various\ndialects. If the various dialects are not addressed in technological\nadvancements like automatic speech recognition and speech synthesis, there is a\nchance that these dialects may disappear. Speech technology plays a role in\npreserving various dialects of a language from going extinct. In order to build\na full fledged automatic speech recognition system that addresses various\ndialects, an Automatic Dialect Identification (ADI) system acting as the front\nend is required. This is similar to how language identification systems act as\nfront ends to automatic speech recognition systems that handle multiple\nlanguages. The current work proposes a way to identify two popular and broadly\nclassified Tamil dialects, namely literary and colloquial Tamil. Acoustical\ncharacteristics rather than phonetics and phonotactics are used, alleviating\nthe requirement of language-dependant linguistic tools. Hence one major\nadvantage of the proposed method is that it does not require an annotated\ncorpus, hence it can be easily adapted to other languages. Gaussian Mixture\nModels (GMM) using Mel Frequency Cepstral Coefficient (MFCC) features are used\nto perform the classification task. The experiments yielded an error rate of\n12%. Vowel nasalization, as being the reason for this good performance, is\ndiscussed. The number of mixture models for the GMM is varied and the\nperformance is analysed.\n", "date": "2024-08-27", "categories": []}
{"id": "2309.02539", "title": "A Generalized Bandsplit Neural Network for Cinematic Audio Source\n  Separation", "abstract": "  Cinematic audio source separation is a relatively new subtask of audio source\nseparation, with the aim of extracting the dialogue, music, and effects stems\nfrom their mixture. In this work, we developed a model generalizing the\nBandsplit RNN for any complete or overcomplete partitions of the frequency\naxis. Psychoacoustically motivated frequency scales were used to inform the\nband definitions which are now defined with redundancy for more reliable\nfeature extraction. A loss function motivated by the signal-to-noise ratio and\nthe sparsity-promoting property of the 1-norm was proposed. We additionally\nexploit the information-sharing property of a common-encoder setup to reduce\ncomputational complexity during both training and inference, improve separation\nperformance for hard-to-generalize classes of sounds, and allow flexibility\nduring inference time with detachable decoders. Our best model sets the state\nof the art on the Divide and Remaster dataset with performance above the ideal\nratio mask for the dialogue stem.\n", "date": "2023-09-05", "categories": []}
{"id": "2403.08525", "title": "From Weak to Strong Sound Event Labels using Adaptive Change-Point\n  Detection and Active Learning", "abstract": "  We propose an adaptive change point detection method (A-CPD) for machine\nguided weak label annotation of audio recording segments. The goal is to\nmaximize the amount of information gained about the temporal activations of the\ntarget sounds. For each unlabeled audio recording, we use a prediction model to\nderive a probability curve used to guide annotation. The prediction model is\ninitially pre-trained on available annotated sound event data with classes that\nare disjoint from the classes in the unlabeled dataset. The prediction model\nthen gradually adapts to the annotations provided by the annotator in an active\nlearning loop. We derive query segments to guide the weak label annotator\ntowards strong labels, using change point detection on these probabilities. We\nshow that it is possible to derive strong labels of high quality with a limited\nannotation budget, and show favorable results for A-CPD when compared to two\nbaseline query segment strategies.\n", "date": "2024-03-13", "categories": []}
{"id": "2406.17246", "title": "Beyond Silence: Bias Analysis through Loss and Asymmetric Approach in\n  Audio Anti-Spoofing", "abstract": "  Current trends in audio anti-spoofing detection research strive to improve\nmodels' ability to generalize across unseen attacks by learning to identify a\nvariety of spoofing artifacts. This emphasis has primarily focused on the spoof\nclass. Recently, several studies have noted that the distribution of silence\ndiffers between the two classes, which can serve as a shortcut. In this paper,\nwe extend class-wise interpretations beyond silence. We employ loss analysis\nand asymmetric methodologies to move away from traditional attack-focused and\nresult-oriented evaluations towards a deeper examination of model behaviors.\nOur investigations highlight the significant differences in training dynamics\nbetween the two classes, emphasizing the need for future research to focus on\nrobust modeling of the bonafide class.\n", "date": "2024-06-25", "categories": []}
{"id": "2406.18747", "title": "A Stem-Agnostic Single-Decoder System for Music Source Separation Beyond\n  Four Stems", "abstract": "  Despite significant recent progress across multiple subtasks of audio source\nseparation, few music source separation systems support separation beyond the\nfour-stem vocals, drums, bass, and other (VDBO) setup. Of the very few current\nsystems that support source separation beyond this setup, most continue to rely\non an inflexible decoder setup that can only support a fixed pre-defined set of\nstems. Increasing stem support in these inflexible systems correspondingly\nrequires increasing computational complexity, rendering extensions of these\nsystems computationally infeasible for long-tail instruments. In this work, we\npropose Banquet, a system that allows source separation of multiple stems using\njust one decoder. A bandsplit source separation model is extended to work in a\nquery-based setup in tandem with a music instrument recognition PaSST model. On\nthe MoisesDB dataset, Banquet, at only 24.9 M trainable parameters, approached\nthe performance level of the significantly more complex 6-stem Hybrid\nTransformer Demucs on VDBO stems and outperformed it on guitar and piano. The\nquery-based setup allows for the separation of narrow instrument classes such\nas clean acoustic guitars, and can be successfully applied to the extraction of\nless common stems such as reeds and organs. Implementation is available at\nhttps://github.com/kwatcharasupat/query-bandit.\n", "date": "2024-06-26", "categories": []}
{"id": "2407.07275", "title": "Remastering Divide and Remaster: A Cinematic Audio Source Separation\n  Dataset with Multilingual Support", "abstract": "  Cinematic audio source separation (CASS), as a problem of extracting the\ndialogue, music, and effects stems from their mixture, is a relatively new\nsubtask of audio source separation. To date, only one publicly available\ndataset exists for CASS, that is, the Divide and Remaster (DnR) dataset, which\nis currently at version 2. While DnR v2 has been an incredibly useful resource\nfor CASS, several areas of improvement have been identified, particularly\nthrough its use in the 2023 Sound Demixing Challenge. In this work, we develop\nversion 3 of the DnR dataset, addressing issues relating to vocal content in\nnon-dialogue stems, loudness distributions, mastering process, and linguistic\ndiversity. In particular, the dialogue stem of DnR v3 includes speech content\nfrom more than 30 languages from multiple families including but not limited to\nthe Germanic, Romance, Indo-Aryan, Dravidian, Malayo-Polynesian, and Bantu\nfamilies. Benchmark results using the Bandit model indicated that training on\nmultilingual data yields significant generalizability to the model even in\nlanguages with low data availability. Even in languages with high data\navailability, the multilingual model often performs on par or better than\ndedicated models trained on monolingual CASS datasets. Dataset and model\nimplementation will be made available at\nhttps://github.com/kwatcharasupat/source-separation-landing.\n", "date": "2024-07-09", "categories": []}
{"id": "2408.03588", "title": "Facing the Music: Tackling Singing Voice Separation in Cinematic Audio\n  Source Separation", "abstract": "  Cinematic audio source separation (CASS), as a standalone problem of\nextracting individual stems from their mixture, is a fairly new subtask of\naudio source separation. A typical setup of CASS is a three-stem problem, with\nthe aim of separating the mixture into the dialogue (DX), music (MX), and\neffects (FX) stems. Given the creative nature of cinematic sound production,\nhowever, several edge cases exist; some sound sources do not fit neatly in any\nof these three stems, necessitating the use of additional auxiliary stems in\nproduction. One very common edge case is the singing voice in film audio, which\nmay belong in either the DX or MX or neither, depending heavily on the\ncinematic context. In this work, we demonstrate a very straightforward\nextension of the dedicated-decoder Bandit and query-based single-decoder\nBanquet models to a four-stem problem, treating non-musical dialogue,\ninstrumental music, singing voice, and effects as separate stems.\nInterestingly, the query-based Banquet model outperformed the dedicated-decoder\nBandit model. We hypothesized that this is due to a better feature alignment at\nthe bottleneck as enforced by the band-agnostic FiLM layer. Dataset and model\nimplementation will be made available at\nhttps://github.com/kwatcharasupat/source-separation-landing.\n", "date": "2024-08-07", "categories": []}
{"id": "2408.12658", "title": "Hierarchical Generative Modeling of Melodic Vocal Contours in Hindustani\n  Classical Music", "abstract": "  Hindustani music is a performance-driven oral tradition that exhibits the\nrendition of rich melodic patterns. In this paper, we focus on generative\nmodeling of singers' vocal melodies extracted from audio recordings, as the\nvoice is musically prominent within the tradition. Prior generative work in\nHindustani music models melodies as coarse discrete symbols which fails to\ncapture the rich expressive melodic intricacies of singing. Thus, we propose to\nuse a finely quantized pitch contour, as an intermediate representation for\nhierarchical audio modeling. We propose GaMaDHaNi, a modular two-level\nhierarchy, consisting of a generative model on pitch contours, and a pitch\ncontour to audio synthesis model. We compare our approach to non-hierarchical\naudio models and hierarchical models that use a self-supervised intermediate\nrepresentation, through a listening test and qualitative analysis. We also\nevaluate audio model's ability to faithfully represent the pitch contour input\nusing Pearson correlation coefficient. By using pitch contours as an\nintermediate representation, we show that our model may be better equipped to\nlisten and respond to musicians in a human-AI collaborative setting by\nhighlighting two potential interaction use cases (1) primed generation, and (2)\ncoarse pitch conditioning.\n", "date": "2024-08-22", "categories": []}
{"id": "2408.13341", "title": "Toward Improving Synthetic Audio Spoofing Detection Robustness via\n  Meta-Learning and Disentangled Training With Adversarial Examples", "abstract": "  Advances in automatic speaker verification (ASV) promote research into the\nformulation of spoofing detection systems for real-world applications. The\nperformance of ASV systems can be degraded severely by multiple types of\nspoofing attacks, namely, synthetic speech (SS), voice conversion (VC), replay,\ntwins and impersonation, especially in the case of unseen synthetic spoofing\nattacks. A reliable and robust spoofing detection system can act as a security\ngate to filter out spoofing attacks instead of having them reach the ASV\nsystem. A weighted additive angular margin loss is proposed to address the data\nimbalance issue, and different margins has been assigned to improve\ngeneralization to unseen spoofing attacks in this study. Meanwhile, we\nincorporate a meta-learning loss function to optimize differences between the\nembeddings of support versus query set in order to learn a\nspoofing-category-independent embedding space for utterances. Furthermore, we\ncraft adversarial examples by adding imperceptible perturbations to spoofing\nspeech as a data augmentation strategy, then we use an auxiliary batch\nnormalization (BN) to guarantee that corresponding normalization statistics are\nperformed exclusively on the adversarial examples. Additionally, A simple\nattention module is integrated into the residual block to refine the feature\nextraction process. Evaluation results on the Logical Access (LA) track of the\nASVspoof 2019 corpus provides confirmation of our proposed approaches'\neffectiveness in terms of a pooled EER of 0.87%, and a min t-DCF of 0.0277.\nThese advancements offer effective options to reduce the impact of spoofing\nattacks on voice recognition/authentication systems.\n", "date": "2024-08-23", "categories": []}
{"id": "2408.13644", "title": "Studying the Effect of Audio Filters in Pre-Trained Models for\n  Environmental Sound Classification", "abstract": "  Environmental Sound Classification is an important problem of sound\nrecognition and is more complicated than speech recognition problems as\nenvironmental sounds are not well structured with respect to time and\nfrequency. Researchers have used various CNN models to learn audio features\nfrom different audio features like log mel spectrograms, gammatone spectral\ncoefficients, mel-frequency spectral coefficients, generated from the audio\nfiles, over the past years. In this paper, we propose a new methodology :\nTwo-Level Classification; the Level 1 Classifier will be responsible to\nclassify the audio signal into a broader class and the Level 2 Classifiers will\nbe responsible to find the actual class to which the audio belongs, based on\nthe output of the Level 1 Classifier. We have also shown the effects of\ndifferent audio filters, among which a new method of Audio Crop is introduced\nin this paper, which gave the highest accuracies in most of the cases. We have\nused the ESC-50 dataset for our experiment and obtained a maximum accuracy of\n78.75% in case of Level 1 Classification and 98.04% in case of Level 2\nClassifications.\n", "date": "2024-08-24", "categories": []}
{"id": "2408.13746", "title": "Quartered Spectral Envelope and 1D-CNN-based Classification of Normally\n  Phonated and Whispered Speech", "abstract": "  Whisper, as a form of speech, is not sufficiently addressed by mainstream\nspeech applications. This is due to the fact that systems built for normal\nspeech do not work as expected for whispered speech. A first step to building a\nspeech application that is inclusive of whispered speech, is the successful\nclassification of whispered speech and normal speech. Such a front-end\nclassification system is expected to have high accuracy and low computational\noverhead, which is the scope of this paper. One of the characteristics of\nwhispered speech is the absence of the fundamental frequency (or pitch), and\nhence the pitch harmonics as well. The presence of the pitch and pitch\nharmonics in normal speech, and its absence in whispered speech, is evident in\nthe spectral envelope of the Fourier transform. We observe that this\ncharacteristic is predominant in the first quarter of the spectrum, and exploit\nthe same as a feature. We propose the use of one dimensional convolutional\nneural networks (1D-CNN) to capture these features from the quartered spectral\nenvelope (QSE). The system yields an accuracy of 99.31% when trained and tested\non the wTIMIT dataset, and 100% on the CHAINS dataset. The proposed feature is\ncompared with Mel frequency cepstral coefficients (MFCC), a staple in the\nspeech domain. The proposed classification system is also compared with the\nstate-of-the-art system based on log-filterbank energy (LFBE) features trained\non long short-term memory (LSTM) network. The proposed system based on 1D-CNN\nperforms better than, or as good as, the state-of-the-art across multiple\nexperiments. It also converges sooner, with lesser computational overhead.\nFinally, the proposed system is evaluated under the presence of white noise at\nvarious signal-to-noise ratios and found to be robust.\n", "date": "2024-08-25", "categories": []}
{"id": "2408.13201", "title": "EAViT: External Attention Vision Transformer for Audio Classification", "abstract": "  This paper presents the External Attention Vision Transformer (EAViT) model,\na novel approach designed to enhance audio classification accuracy. As digital\naudio resources proliferate, the demand for precise and efficient audio\nclassification systems has intensified, driven by the need for improved\nrecommendation systems and user personalization in various applications,\nincluding music streaming platforms and environmental sound recognition.\nAccurate audio classification is crucial for organizing vast audio libraries\ninto coherent categories, enabling users to find and interact with their\npreferred audio content more effectively. In this study, we utilize the GTZAN\ndataset, which comprises 1,000 music excerpts spanning ten diverse genres. Each\n30-second audio clip is segmented into 3-second excerpts to enhance dataset\nrobustness and mitigate overfitting risks, allowing for more granular feature\nanalysis. The EAViT model integrates multi-head external attention (MEA)\nmechanisms into the Vision Transformer (ViT) framework, effectively capturing\nlong-range dependencies and potential correlations between samples. This\nexternal attention (EA) mechanism employs learnable memory units that enhance\nthe network's capacity to process complex audio features efficiently. The study\ndemonstrates that EAViT achieves a remarkable overall accuracy of 93.99%,\nsurpassing state-of-the-art models.\n", "date": "2024-08-23", "categories": []}
{"id": "2408.11641", "title": "Estimated Audio-Caption Correspondences Improve Language-Based Audio\n  Retrieval", "abstract": "  Dual-encoder-based audio retrieval systems are commonly optimized with\ncontrastive learning on a set of matching and mismatching audio-caption pairs.\nThis leads to a shared embedding space in which corresponding items from the\ntwo modalities end up close together. Since audio-caption datasets typically\nonly contain matching pairs of recordings and descriptions, it has become\ncommon practice to create mismatching pairs by pairing the audio with a caption\nrandomly drawn from the dataset. This is not ideal because the randomly sampled\ncaption could, just by chance, partly or entirely describe the audio recording.\nHowever, correspondence information for all possible pairs is costly to\nannotate and thus typically unavailable; we, therefore, suggest substituting it\nwith estimated correspondences. To this end, we propose a two-staged training\nprocedure in which multiple retrieval models are first trained as usual, i.e.,\nwithout estimated correspondences. In the second stage, the audio-caption\ncorrespondences predicted by these models then serve as prediction targets. We\nevaluate our method on the ClothoV2 and the AudioCaps benchmark and show that\nit improves retrieval performance, even in a restricting self-distillation\nsetting where a single model generates and then learns from the estimated\ncorrespondences. We further show that our method outperforms the current state\nof the art by 1.6 pp. mAP@10 on the ClothoV2 benchmark.\n", "date": "2024-08-21", "categories": []}
{"id": "2408.10383", "title": "BrewCLIP: A Bifurcated Representation Learning Framework for\n  Audio-Visual Retrieval", "abstract": "  Previous methods for audio-image matching generally fall into one of two\ncategories: pipeline models or End-to-End models. Pipeline models first\ntranscribe speech and then encode the resulting text; End-to-End models encode\nspeech directly. Generally, pipeline models outperform end-to-end models, but\nthe intermediate transcription necessarily discards some potentially useful\nnon-textual information. In addition to textual information, speech can convey\ndetails such as accent, mood, and and emphasis, which should be effectively\ncaptured in the encoded representation. In this paper, we investigate whether\nnon-textual information, which is overlooked by pipeline-based models, can be\nleveraged to improve speech-image matching performance. We thoroughly analyze\nand compare End-to-End models, pipeline models, and our proposed dual-channel\nmodel for robust audio-image retrieval on a variety of datasets. Our approach\nachieves a substantial performance gain over the previous state-of-the-art by\nleveraging strong pretrained models, a prompting mechanism and a bifurcated\ndesign.\n", "date": "2024-08-19", "categories": []}
{"id": "2408.10807", "title": "DisMix: Disentangling Mixtures of Musical Instruments for Source-level\n  Pitch and Timbre Manipulation", "abstract": "  Existing work on pitch and timbre disentanglement has been mostly focused on\nsingle-instrument music audio, excluding the cases where multiple instruments\nare presented. To fill the gap, we propose DisMix, a generative framework in\nwhich the pitch and timbre representations act as modular building blocks for\nconstructing the melody and instrument of a source, and the collection of which\nforms a set of per-instrument latent representations underlying the observed\nmixture. By manipulating the representations, our model samples mixtures with\nnovel combinations of pitch and timbre of the constituent instruments. We can\njointly learn the disentangled pitch-timbre representations and a latent\ndiffusion transformer that reconstructs the mixture conditioned on the set of\nsource-level representations. We evaluate the model using both a simple dataset\nof isolated chords and a realistic four-part chorales in the style of J.S.\nBach, identify the key components for the success of disentanglement, and\ndemonstrate the application of mixture transformation based on source-level\nattribute manipulation.\n", "date": "2024-08-20", "categories": []}
{"id": "2408.10998", "title": "Audio Match Cutting: Finding and Creating Matching Audio Transitions in\n  Movies and Videos", "abstract": "  A \"match cut\" is a common video editing technique where a pair of shots that\nhave a similar composition transition fluidly from one to another. Although\nmatch cuts are often visual, certain match cuts involve the fluid transition of\naudio, where sounds from different sources merge into one indistinguishable\ntransition between two shots. In this paper, we explore the ability to\nautomatically find and create \"audio match cuts\" within videos and movies. We\ncreate a self-supervised audio representation for audio match cutting and\ndevelop a coarse-to-fine audio match pipeline that recommends matching shots\nand creates the blended audio. We further annotate a dataset for the proposed\naudio match cut task and compare the ability of multiple audio representations\nto find audio match cut candidates. Finally, we evaluate multiple methods to\nblend two matching audio candidates with the goal of creating a smooth\ntransition. Project page and examples are available at:\nhttps://denfed.github.io/audiomatchcut/\n", "date": "2024-08-20", "categories": []}
{"id": "2406.14176", "title": "A Multi-Stream Fusion Approach with One-Class Learning for Audio-Visual\n  Deepfake Detection", "abstract": "  This paper addresses the challenge of developing a robust audio-visual\ndeepfake detection model. In practical use cases, new generation algorithms are\ncontinually emerging, and these algorithms are not encountered during the\ndevelopment of detection methods. This calls for the generalization ability of\nthe method. Additionally, to ensure the credibility of detection methods, it is\nbeneficial for the model to interpret which cues from the video indicate it is\nfake. Motivated by these considerations, we then propose a multi-stream fusion\napproach with one-class learning as a representation-level regularization\ntechnique. We study the generalization problem of audio-visual deepfake\ndetection by creating a new benchmark by extending and re-splitting the\nexisting FakeAVCeleb dataset. The benchmark contains four categories of fake\nvideos (Real Audio-Fake Visual, Fake Audio-Fake Visual, Fake Audio-Real Visual,\nand Unsynchronized videos). The experimental results demonstrate that our\napproach surpasses the previous models by a large margin. Furthermore, our\nproposed framework offers interpretability, indicating which modality the model\nidentifies as more likely to be fake. The source code is released at\nhttps://github.com/bok-bok/MSOC.\n", "date": "2024-06-20", "categories": []}
{"id": "2407.21611", "title": "Enhancing Partially Spoofed Audio Localization with Boundary-aware\n  Attention Mechanism", "abstract": "  The task of partially spoofed audio localization aims to accurately determine\naudio authenticity at a frame level. Although some works have achieved\nencouraging results, utilizing boundary information within a single model\nremains an unexplored research topic. In this work, we propose a novel method\ncalled Boundary-aware Attention Mechanism (BAM). Specifically, it consists of\ntwo core modules: Boundary Enhancement and Boundary Frame-wise Attention. The\nformer assembles the intra-frame and inter-frame information to extract\ndiscriminative boundary features that are subsequently used for boundary\nposition detection and authenticity decision, while the latter leverages\nboundary prediction results to explicitly control the feature interaction\nbetween frames, which achieves effective discrimination between real and fake\nframes. Experimental results on PartialSpoof database demonstrate our proposed\nmethod achieves the best performance. The code is available at\nhttps://github.com/media-sec-lab/BAM.\n", "date": "2024-07-31", "categories": []}
{"id": "2408.08673", "title": "MAT-SED: A Masked Audio Transformer with Masked-Reconstruction Based\n  Pre-training for Sound Event Detection", "abstract": "  Sound event detection (SED) methods that leverage a large pre-trained\nTransformer encoder network have shown promising performance in recent DCASE\nchallenges. However, they still rely on an RNN-based context network to model\ntemporal dependencies, largely due to the scarcity of labeled data. In this\nwork, we propose a pure Transformer-based SED model with masked-reconstruction\nbased pre-training, termed MAT-SED. Specifically, a Transformer with relative\npositional encoding is first designed as the context network, pre-trained by\nthe masked-reconstruction task on all available target data in a\nself-supervised way. Both the encoder and the context network are jointly\nfine-tuned in a semi-supervised manner. Furthermore, a global-local feature\nfusion strategy is proposed to enhance the localization capability. Evaluation\nof MAT-SED on DCASE2023 task4 surpasses state-of-the-art performance, achieving\n0.587/0.896 PSDS1/PSDS2 respectively.\n", "date": "2024-08-16", "categories": []}
{"id": "2408.09792", "title": "Unsupervised Composable Representations for Audio", "abstract": "  Current generative models are able to generate high-quality artefacts but\nhave been shown to struggle with compositional reasoning, which can be defined\nas the ability to generate complex structures from simpler elements. In this\npaper, we focus on the problem of compositional representation learning for\nmusic data, specifically targeting the fully-unsupervised setting. We propose a\nsimple and extensible framework that leverages an explicit compositional\ninductive bias, defined by a flexible auto-encoding objective that can leverage\nany of the current state-of-art generative models. We demonstrate that our\nframework, used with diffusion models, naturally addresses the task of\nunsupervised audio source separation, showing that our model is able to perform\nhigh-quality separation. Our findings reveal that our proposal achieves\ncomparable or superior performance with respect to other blind source\nseparation methods and, furthermore, it even surpasses current state-of-art\nsupervised baselines on signal-to-interference ratio metrics. Additionally, by\nlearning an a-posteriori masking diffusion model in the space of composable\nrepresentations, we achieve a system capable of seamlessly performing\nunsupervised source separation, unconditional generation, and variation\ngeneration. Finally, as our proposal works in the latent space of pre-trained\nneural audio codecs, it also provides a lower computational cost with respect\nto other neural baselines.\n", "date": "2024-08-19", "categories": []}
{"id": "2408.09933", "title": "SZU-AFS Antispoofing System for the ASVspoof 5 Challenge", "abstract": "  This paper presents the SZU-AFS anti-spoofing system, designed for Track 1 of\nthe ASVspoof 5 Challenge under open conditions. The system is built with four\nstages: selecting a baseline model, exploring effective data augmentation (DA)\nmethods for fine-tuning, applying a co-enhancement strategy based on gradient\nnorm aware minimization (GAM) for secondary fine-tuning, and fusing logits\nscores from the two best-performing fine-tuned models. The system utilizes the\nWav2Vec2 front-end feature extractor and the AASIST back-end classifier as the\nbaseline model. During model fine-tuning, three distinct DA policies have been\ninvestigated: single-DA, random-DA, and cascade-DA. Moreover, the employed\nGAM-based co-enhancement strategy, designed to fine-tune the augmented model at\nboth data and optimizer levels, helps the Adam optimizer find flatter minima,\nthereby boosting model generalization. Overall, the final fusion system\nachieves a minDCF of 0.115 and an EER of 4.04% on the evaluation set.\n", "date": "2024-08-19", "categories": []}
{"id": "2408.08729", "title": "ConcateNet: Dialogue Separation Using Local And Global Feature\n  Concatenation", "abstract": "  Dialogue separation involves isolating a dialogue signal from a mixture, such\nas a movie or a TV program. This can be a necessary step to enable dialogue\nenhancement for broadcast-related applications. In this paper, ConcateNet for\ndialogue separation is proposed, which is based on a novel approach for\nprocessing local and global features aimed at better generalization for\nout-of-domain signals. ConcateNet is trained using a noise reduction-focused,\npublicly available dataset and evaluated using three datasets: two noise\nreduction-focused datasets (in-domain), which show competitive performance for\nConcateNet, and a broadcast-focused dataset (out-of-domain), which verifies the\nbetter generalization performance for the proposed architecture compared to\nconsidered state-of-the-art noise-reduction methods.\n", "date": "2024-08-16", "categories": []}
{"id": "2211.13189", "title": "ASiT: Local-Global Audio Spectrogram vIsion Transformer for Event\n  Classification", "abstract": "  Transformers, which were originally developed for natural language\nprocessing, have recently generated significant interest in the computer vision\nand audio communities due to their flexibility in learning long-range\nrelationships. Constrained by the data hungry nature of transformers and the\nlimited amount of labelled data, most transformer-based models for audio tasks\nare finetuned from ImageNet pretrained models, despite the huge gap between the\ndomain of natural images and audio. This has motivated the research in\nself-supervised pretraining of audio transformers, which reduces the dependency\non large amounts of labeled data and focuses on extracting concise\nrepresentations of audio spectrograms. In this paper, we propose\n\\textbf{L}ocal-\\textbf{G}lobal \\textbf{A}udio \\textbf{S}pectrogram\nv\\textbf{I}sion \\textbf{T}ransformer, namely ASiT, a novel self-supervised\nlearning framework that captures local and global contextual information by\nemploying group masked model learning and self-distillation. We evaluate our\npretrained models on both audio and speech classification tasks, including\naudio event classification, keyword spotting, and speaker identification. We\nfurther conduct comprehensive ablation studies, including evaluations of\ndifferent pretraining strategies. The proposed ASiT framework significantly\nboosts the performance on all tasks and sets a new state-of-the-art performance\nin five audio and speech classification tasks, outperforming recent methods,\nincluding the approaches that use additional datasets for pretraining.\n", "date": "2022-11-23", "categories": []}
{"id": "2408.02009", "title": "Joint Learning of Emotions in Music and Generalized Sounds", "abstract": "  In this study, we aim to determine if generalized sounds and music can share\na common emotional space, improving predictions of emotion in terms of arousal\nand valence. We propose the use of multiple datasets as a multi-domain learning\ntechnique. Our approach involves creating a common space encompassing features\nthat characterize both generalized sounds and music, as they can evoke emotions\nin a similar manner. To achieve this, we utilized two publicly available\ndatasets, namely IADS-E and PMEmo, following a standardized experimental\nprotocol. We employed a wide variety of features that capture diverse aspects\nof the audio structure including key parameters of spectrum, energy, and\nvoicing. Subsequently, we performed joint learning on the common feature space,\nleveraging heterogeneous model architectures. Interestingly, this synergistic\nscheme outperforms the state-of-the-art in both sound and music emotion\nprediction. The code enabling full replication of the presented experimental\npipeline is available at https://github.com/LIMUNIMI/MusicSoundEmotions.\n", "date": "2024-08-04", "categories": []}
{"id": "2408.07113", "title": "A Theory-Based Explainable Deep Learning Architecture for Music Emotion", "abstract": "  This paper paper develops a theory-based, explainable deep learning\nconvolutional neural network (CNN) classifier to predict the time-varying\nemotional response to music. We design novel CNN filters that leverage the\nfrequency harmonics structure from acoustic physics known to impact the\nperception of musical features. Our theory-based model is more parsimonious,\nbut provides comparable predictive performance to atheoretical deep learning\nmodels, while performing better than models using handcrafted features. Our\nmodel can be complemented with handcrafted features, but the performance\nimprovement is marginal. Importantly, the harmonics-based structure placed on\nthe CNN filters provides better explainability for how the model predicts\nemotional response (valence and arousal), because emotion is closely related to\nconsonance--a perceptual feature defined by the alignment of harmonics.\nFinally, we illustrate the utility of our model with an application involving\ndigital advertising. Motivated by YouTube mid-roll ads, we conduct a lab\nexperiment in which we exogenously insert ads at different times within videos.\nWe find that ads placed in emotionally similar contexts increase ad engagement\n(lower skip rates, higher brand recall rates). Ad insertion based on emotional\nsimilarity metrics predicted by our theory-based, explainable model produces\ncomparable or better engagement relative to atheoretical models.\n", "date": "2024-08-13", "categories": []}
{"id": "2408.06500", "title": "Music2Latent: Consistency Autoencoders for Latent Audio Compression", "abstract": "  Efficient audio representations in a compressed continuous latent space are\ncritical for generative audio modeling and Music Information Retrieval (MIR)\ntasks. However, some existing audio autoencoders have limitations, such as\nmulti-stage training procedures, slow iterative sampling, or low reconstruction\nquality. We introduce Music2Latent, an audio autoencoder that overcomes these\nlimitations by leveraging consistency models. Music2Latent encodes samples into\na compressed continuous latent space in a single end-to-end training process\nwhile enabling high-fidelity single-step reconstruction. Key innovations\ninclude conditioning the consistency model on upsampled encoder outputs at all\nlevels through cross connections, using frequency-wise self-attention to\ncapture long-range frequency dependencies, and employing frequency-wise learned\nscaling to handle varying value distributions across frequencies at different\nnoise levels. We demonstrate that Music2Latent outperforms existing continuous\naudio autoencoders in sound quality and reconstruction accuracy while achieving\ncompetitive performance on downstream MIR tasks using its latent\nrepresentations. To our knowledge, this represents the first successful attempt\nat training an end-to-end consistency autoencoder model.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.06911", "title": "Heterogeneous Space Fusion and Dual-Dimension Attention: A New Paradigm\n  for Speech Enhancement", "abstract": "  Self-supervised learning has demonstrated impressive performance in speech\ntasks, yet there remains ample opportunity for advancement in the realm of\nspeech enhancement research. In addressing speech tasks, confining the\nattention mechanism solely to the temporal dimension poses limitations in\neffectively focusing on critical speech features. Considering the\naforementioned issues, our study introduces a novel speech enhancement\nframework, HFSDA, which skillfully integrates heterogeneous spatial features\nand incorporates a dual-dimension attention mechanism to significantly enhance\nspeech clarity and quality in noisy environments. By leveraging self-supervised\nlearning embeddings in tandem with Short-Time Fourier Transform (STFT)\nspectrogram features, our model excels at capturing both high-level semantic\ninformation and detailed spectral data, enabling a more thorough analysis and\nrefinement of speech signals. Furthermore, we employ the innovative\nOmni-dimensional Dynamic Convolution (ODConv) technology within the spectrogram\ninput branch, enabling enhanced extraction and integration of crucial\ninformation across multiple dimensions. Additionally, we refine the Conformer\nmodel by enhancing its feature extraction capabilities not only in the temporal\ndimension but also across the spectral domain. Extensive experiments on the\nVCTK-DEMAND dataset show that HFSDA is comparable to existing state-of-the-art\nmodels, confirming the validity of our approach.\n", "date": "2024-08-13", "categories": []}
{"id": "2408.06922", "title": "Temporal Variability and Multi-Viewed Self-Supervised Representations to\n  Tackle the ASVspoof5 Deepfake Challenge", "abstract": "  ASVspoof5, the fifth edition of the ASVspoof series, is one of the largest\nglobal audio security challenges. It aims to advance the development of\ncountermeasure (CM) to discriminate bonafide and spoofed speech utterances. In\nthis paper, we focus on addressing the problem of open-domain audio deepfake\ndetection, which corresponds directly to the ASVspoof5 Track1 open condition.\nAt first, we comprehensively investigate various CM on ASVspoof5, including\ndata expansion, data augmentation, and self-supervised learning (SSL) features.\nDue to the high-frequency gaps characteristic of the ASVspoof5 dataset, we\nintroduce Frequency Mask, a data augmentation method that masks specific\nfrequency bands to improve CM robustness. Combining various scale of temporal\ninformation with multiple SSL features, our experiments achieved a minDCF of\n0.0158 and an EER of 0.55% on the ASVspoof 5 Track 1 evaluation progress set.\n", "date": "2024-08-13", "categories": []}
{"id": "2408.07020", "title": "Source Separation of Multi-source Raw Music using a Residual Quantized\n  Variational Autoencoder", "abstract": "  I developed a neural audio codec model based on the residual quantized\nvariational autoencoder architecture. I train the model on the Slakh2100\ndataset, a standard dataset for musical source separation, composed of\nmulti-track audio. The model can separate audio sources, achieving almost SoTA\nresults with much less computing power. The code is publicly available at\ngithub.com/LeonardoBerti00/Source-Separation-of-Multi-source-Music-using-Residual-Quantizad-Variational-Autoencoder\n", "date": "2024-08-12", "categories": []}
{"id": "2408.07050", "title": "PSM: Learning Probabilistic Embeddings for Multi-scale Zero-Shot\n  Soundscape Mapping", "abstract": "  A soundscape is defined by the acoustic environment a person perceives at a\nlocation. In this work, we propose a framework for mapping soundscapes across\nthe Earth. Since soundscapes involve sound distributions that span varying\nspatial scales, we represent locations with multi-scale satellite imagery and\nlearn a joint representation among this imagery, audio, and text. To capture\nthe inherent uncertainty in the soundscape of a location, we design the\nrepresentation space to be probabilistic. We also fuse ubiquitous metadata\n(including geolocation, time, and data source) to enable learning of spatially\nand temporally dynamic representations of soundscapes. We demonstrate the\nutility of our framework by creating large-scale soundscape maps integrating\nboth audio and text with temporal control. To facilitate future research on\nthis task, we also introduce a large-scale dataset, GeoSound, containing over\n$300k$ geotagged audio samples paired with both low- and high-resolution\nsatellite imagery. We demonstrate that our method outperforms the existing\nstate-of-the-art on both GeoSound and the existing SoundingEarth dataset. Our\ndataset and code is available at https://github.com/mvrl/PSM.\n", "date": "2024-08-13", "categories": []}
{"id": "2408.06065", "title": "An Investigation Into Explainable Audio Hate Speech Detection", "abstract": "  Research on hate speech has predominantly revolved around detection and\ninterpretation from textual inputs, leaving verbal content largely unexplored.\nWhile there has been limited exploration into hate speech detection within\nverbal acoustic speech inputs, the aspect of interpretability has been\noverlooked. Therefore, we introduce a new task of explainable audio hate speech\ndetection. Specifically, we aim to identify the precise time intervals,\nreferred to as audio frame-level rationales, which serve as evidence for hate\nspeech classification. Towards this end, we propose two different approaches:\ncascading and End-to-End (E2E). The cascading approach initially converts audio\nto transcripts, identifies hate speech within these transcripts, and\nsubsequently locates the corresponding audio time frames. Conversely, the E2E\napproach processes audio utterances directly, which allows it to pinpoint hate\nspeech within specific time frames. Additionally, due to the lack of\nexplainable audio hate speech datasets that include audio frame-level\nrationales, we curated a synthetic audio dataset to train our models. We\nfurther validated these models on actual human speech utterances and found that\nthe E2E approach outperforms the cascading method in terms of the audio frame\nIntersection over Union (IoU) metric. Furthermore, we observed that including\nframe-level rationales significantly enhances hate speech detection accuracy\nfor the E2E approach.\n  \\textbf{Disclaimer} The reader may encounter content of an offensive or\nhateful nature. However, given the nature of the work, this cannot be avoided.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.06264", "title": "Audio Enhancement for Computer Audition -- An Iterative Training\n  Paradigm Using Sample Importance", "abstract": "  Neural network models for audio tasks, such as automatic speech recognition\n(ASR) and acoustic scene classification (ASC), are susceptible to noise\ncontamination for real-life applications. To improve audio quality, an\nenhancement module, which can be developed independently, is explicitly used at\nthe front-end of the target audio applications. In this paper, we present an\nend-to-end learning solution to jointly optimise the models for audio\nenhancement (AE) and the subsequent applications. To guide the optimisation of\nthe AE module towards a target application, and especially to overcome\ndifficult samples, we make use of the sample-wise performance measure as an\nindication of sample importance. In experiments, we consider four\nrepresentative applications to evaluate our training paradigm, i.e., ASR,\nspeech command recognition (SCR), speech emotion recognition (SER), and ASC.\nThese applications are associated with speech and non-speech tasks concerning\nsemantic and non-semantic features, transient and global information, and the\nexperimental results indicate that our proposed approach can considerably boost\nthe noise robustness of the models, especially at low signal-to-noise ratios\n(SNRs), for a wide range of computer audition tasks in everyday-life noisy\nenvironments.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.04708", "title": "MulliVC: Multi-lingual Voice Conversion With Cycle Consistency", "abstract": "  Voice conversion aims to modify the source speaker's voice to resemble the\ntarget speaker while preserving the original speech content. Despite notable\nadvancements in voice conversion these days, multi-lingual voice conversion\n(including both monolingual and cross-lingual scenarios) has yet to be\nextensively studied. It faces two main challenges: 1) the considerable\nvariability in prosody and articulation habits across languages; and 2) the\nrarity of paired multi-lingual datasets from the same speaker. In this paper,\nwe propose MulliVC, a novel voice conversion system that only converts timbre\nand keeps original content and source language prosody without multi-lingual\npaired data. Specifically, each training step of MulliVC contains three\nsubsteps: In step one the model is trained with monolingual speech data; then,\nsteps two and three take inspiration from back translation, construct a\ncyclical process to disentangle the timbre and other information (content,\nprosody, and other language-related information) in the absence of\nmulti-lingual data from the same speaker. Both objective and subjective results\nindicate that MulliVC significantly surpasses other methods in both monolingual\nand cross-lingual contexts, demonstrating the system's efficacy and the\nviability of the three-step approach with cycle consistency. Audio samples can\nbe found on our demo page (mullivc.github.io).\n", "date": "2024-08-08", "categories": []}
{"id": "2408.05057", "title": "SELD-Mamba: Selective State-Space Model for Sound Event Localization and\n  Detection with Source Distance Estimation", "abstract": "  In the Sound Event Localization and Detection (SELD) task, Transformer-based\nmodels have demonstrated impressive capabilities. However, the quadratic\ncomplexity of the Transformer's self-attention mechanism results in\ncomputational inefficiencies. In this paper, we propose a network architecture\nfor SELD called SELD-Mamba, which utilizes Mamba, a selective state-space\nmodel. We adopt the Event-Independent Network V2 (EINV2) as the foundational\nframework and replace its Conformer blocks with bidirectional Mamba blocks to\ncapture a broader range of contextual information while maintaining\ncomputational efficiency. Additionally, we implement a two-stage training\nmethod, with the first stage focusing on Sound Event Detection (SED) and\nDirection of Arrival (DoA) estimation losses, and the second stage\nreintroducing the Source Distance Estimation (SDE) loss. Our experimental\nresults on the 2024 DCASE Challenge Task3 dataset demonstrate the effectiveness\nof the selective state-space model in SELD and highlight the benefits of the\ntwo-stage training approach in enhancing SELD performance.\n", "date": "2024-08-09", "categories": []}
{"id": "2207.03927", "title": "BAST: Binaural Audio Spectrogram Transformer for Binaural Sound\n  Localization", "abstract": "  Accurate sound localization in a reverberation environment is essential for\nhuman auditory perception. Recently, Convolutional Neural Networks (CNNs) have\nbeen utilized to model the binaural human auditory pathway. However, CNN shows\nbarriers in capturing the global acoustic features. To address this issue, we\npropose a novel end-to-end Binaural Audio Spectrogram Transformer (BAST) model\nto predict the sound azimuth in both anechoic and reverberation environments.\nTwo modes of implementation, i.e. BAST-SP and BAST-NSP corresponding to BAST\nmodel with shared and non-shared parameters respectively, are explored. Our\nmodel with subtraction interaural integration and hybrid loss achieves an\nangular distance of 1.29 degrees and a Mean Square Error of 1e-3 at all\nazimuths, significantly surpassing CNN based model. The exploratory analysis of\nthe BAST's performance on the left-right hemifields and anechoic and\nreverberation environments shows its generalization ability as well as the\nfeasibility of binaural Transformers in sound localization. Furthermore, the\nanalysis of the attention maps is provided to give additional insights on the\ninterpretation of the localization process in a natural reverberant\nenvironment.\n", "date": "2022-07-08", "categories": []}
{"id": "2308.11530", "title": "Leveraging Language Model Capabilities for Sound Event Detection", "abstract": "  Large language models reveal deep comprehension and fluent generation in the\nfield of multi-modality. Although significant advancements have been achieved\nin audio multi-modality, existing methods are rarely leverage language model\nfor sound event detection (SED). In this work, we propose an end-to-end\nframework for understanding audio features while simultaneously generating\nsound event and their temporal location. Specifically, we employ pretrained\nacoustic models to capture discriminative features across different categories\nand language models for autoregressive text generation. Conventional methods\ngenerally struggle to obtain features in pure audio domain for classification.\nIn contrast, our framework utilizes the language model to flexibly understand\nabundant semantic context aligned with the acoustic representation. The\nexperimental results showcase the effectiveness of proposed method in enhancing\ntimestamps precision and event classification.\n", "date": "2023-08-22", "categories": []}
{"id": "2408.02514", "title": "Stem-JEPA: A Joint-Embedding Predictive Architecture for Musical Stem\n  Compatibility Estimation", "abstract": "  This paper explores the automated process of determining stem compatibility\nby identifying audio recordings of single instruments that blend well with a\ngiven musical context. To tackle this challenge, we present Stem-JEPA, a novel\nJoint-Embedding Predictive Architecture (JEPA) trained on a multi-track dataset\nusing a self-supervised learning approach.\n  Our model comprises two networks: an encoder and a predictor, which are\njointly trained to predict the embeddings of compatible stems from the\nembeddings of a given context, typically a mix of several instruments. Training\na model in this manner allows its use in estimating stem compatibility -\nretrieving, aligning, or generating a stem to match a given mix - or for\ndownstream tasks such as genre or key estimation, as the training paradigm\nrequires the model to learn information related to timbre, harmony, and rhythm.\n  We evaluate our model's performance on a retrieval task on the MUSDB18\ndataset, testing its ability to find the missing stem from a mix and through a\nsubjective user study. We also show that the learned embeddings capture\ntemporal alignment information and, finally, evaluate the representations\nlearned by our model on several downstream tasks, highlighting that they\neffectively capture meaningful musical features.\n", "date": "2024-08-05", "categories": []}
{"id": "2310.08753", "title": "CompA: Addressing the Gap in Compositional Reasoning in Audio-Language\n  Models", "abstract": "  A fundamental characteristic of audio is its compositional nature.\nAudio-language models (ALMs) trained using a contrastive approach (e.g., CLAP)\nthat learns a shared representation between audio and language modalities have\nimproved performance in many downstream applications, including zero-shot audio\nclassification, audio retrieval, etc. However, the ability of these models to\neffectively perform compositional reasoning remains largely unexplored and\nnecessitates additional research. In this paper, we propose CompA, a collection\nof two expert-annotated benchmarks with a majority of real-world audio samples,\nto evaluate compositional reasoning in ALMs. Our proposed CompA-order evaluates\nhow well an ALM understands the order or occurrence of acoustic events in\naudio, and CompA-attribute evaluates attribute-binding of acoustic events. An\ninstance from either benchmark consists of two audio-caption pairs, where both\naudios have the same acoustic events but with different compositions. An ALM is\nevaluated on how well it matches the right audio to the right caption. Using\nthis benchmark, we first show that current ALMs perform only marginally better\nthan random chance, thereby struggling with compositional reasoning. Next, we\npropose CompA-CLAP, where we fine-tune CLAP using a novel learning method to\nimprove its compositional reasoning abilities. To train CompA-CLAP, we first\npropose improvements to contrastive training with composition-aware hard\nnegatives, allowing for more focused training. Next, we propose a novel modular\ncontrastive loss that helps the model learn fine-grained compositional\nunderstanding and overcomes the acute scarcity of openly available\ncompositional audios. CompA-CLAP significantly improves over all our baseline\nmodels on the CompA benchmark, indicating its superior compositional reasoning\ncapabilities.\n", "date": "2023-10-12", "categories": []}
{"id": "2407.14358", "title": "Stable Audio Open", "abstract": "  Open generative models are vitally important for the community, allowing for\nfine-tunes and serving as baselines when presenting new models. However, most\ncurrent text-to-audio models are private and not accessible for artists and\nresearchers to build upon. Here we describe the architecture and training\nprocess of a new open-weights text-to-audio model trained with Creative Commons\ndata. Our evaluation shows that the model's performance is competitive with the\nstate-of-the-art across various metrics. Notably, the reported FDopenl3 results\n(measuring the realism of the generations) showcase its potential for\nhigh-quality stereo sound synthesis at 44.1kHz.\n", "date": "2024-07-19", "categories": []}
{"id": "2407.21658", "title": "Beat this! Accurate beat tracking without DBN postprocessing", "abstract": "  We propose a system for tracking beats and downbeats with two objectives:\ngenerality across a diverse music range, and high accuracy. We achieve\ngenerality by training on multiple datasets -- including solo instrument\nrecordings, pieces with time signature changes, and classical music with high\ntempo variations -- and by removing the commonly used Dynamic Bayesian Network\n(DBN) postprocessing, which introduces constraints on the meter and tempo. For\nhigh accuracy, among other improvements, we develop a loss function tolerant to\nsmall time shifts of annotations, and an architecture alternating convolutions\nwith transformers either over frequency or time. Our system surpasses the\ncurrent state of the art in F1 score despite using no DBN. However, it can\nstill fail, especially for difficult and underrepresented genres, and performs\nworse on continuity metrics, so we publish our model, code, and preprocessed\ndatasets, and invite others to beat this.\n", "date": "2024-07-31", "categories": []}
{"id": "2303.03857", "title": "Leveraging Pre-trained AudioLDM for Sound Generation: A Benchmark Study", "abstract": "  Deep neural networks have recently achieved breakthroughs in sound\ngeneration. Despite the outstanding sample quality, current sound generation\nmodels face issues on small-scale datasets (e.g., overfitting), significantly\nlimiting performance. In this paper, we make the first attempt to investigate\nthe benefits of pre-training on sound generation with AudioLDM, the\ncutting-edge model for audio generation, as the backbone. Our study\ndemonstrates the advantages of the pre-trained AudioLDM, especially in\ndata-scarcity scenarios. In addition, the baselines and evaluation protocol for\nsound generation systems are not consistent enough to compare different studies\ndirectly. Aiming to facilitate further study on sound generation tasks, we\nbenchmark the sound generation task on various frequently-used datasets. We\nhope our results on transfer learning and benchmarks can provide references for\nfurther research on conditional sound generation.\n", "date": "2023-03-07", "categories": []}
{"id": "2407.19823", "title": "Analyzing and reducing the synthetic-to-real transfer gap in Music\n  Information Retrieval: the task of automatic drum transcription", "abstract": "  Automatic drum transcription is a critical tool in Music Information\nRetrieval for extracting and analyzing the rhythm of a music track, but it is\nlimited by the size of the datasets available for training. A popular method\nused to increase the amount of data is by generating them synthetically from\nmusic scores rendered with virtual instruments. This method can produce a\nvirtually infinite quantity of tracks, but empirical evidence shows that models\ntrained on previously created synthetic datasets do not transfer well to real\ntracks. In this work, besides increasing the amount of data, we identify and\nevaluate three more strategies that practitioners can use to improve the\nrealism of the generated data and, thus, narrow the synthetic-to-real transfer\ngap. To explore their efficacy, we used them to build a new synthetic dataset\nand then we measured how the performance of a model scales and, specifically,\nat what value it will stagnate when increasing the number of training tracks\nfor different datasets. By doing this, we were able to prove that the\naforementioned strategies contribute to make our dataset the one with the most\nrealistic data distribution and the lowest synthetic-to-real transfer gap among\nthe synthetic datasets we evaluated. We conclude by highlighting the limits of\ntraining with infinite data in drum transcription and we show how they can be\novercome.\n", "date": "2024-07-29", "categories": []}
{"id": "2402.07729", "title": "AIR-Bench: Benchmarking Large Audio-Language Models via Generative\n  Comprehension", "abstract": "  Recently, instruction-following audio-language models have received broad\nattention for human-audio interaction. However, the absence of benchmarks\ncapable of evaluating audio-centric interaction capabilities has impeded\nadvancements in this field. Previous models primarily focus on assessing\ndifferent fundamental tasks, such as Automatic Speech Recognition (ASR), and\nlack an assessment of the open-ended generative capabilities centered around\naudio. Thus, it is challenging to track the progression in the Large\nAudio-Language Models (LALMs) domain and to provide guidance for future\nimprovement. In this paper, we introduce AIR-Bench (\\textbf{A}udio\n\\textbf{I}nst\\textbf{R}uction \\textbf{Bench}mark), the first benchmark designed\nto evaluate the ability of LALMs to understand various types of audio signals\n(including human speech, natural sounds, and music), and furthermore, to\ninteract with humans in the textual format. AIR-Bench encompasses two\ndimensions: \\textit{foundation} and \\textit{chat} benchmarks. The former\nconsists of 19 tasks with approximately 19k single-choice questions, intending\nto inspect the basic single-task ability of LALMs. The latter one contains 2k\ninstances of open-ended question-and-answer data, directly assessing the\ncomprehension of the model on complex audio and its capacity to follow\ninstructions. Both benchmarks require the model to generate hypotheses\ndirectly. We design a unified framework that leverages advanced language\nmodels, such as GPT-4, to evaluate the scores of generated hypotheses given the\nmeta-information of the audio. Experimental results demonstrate a high level of\nconsistency between GPT-4-based evaluation and human evaluation. By revealing\nthe limitations of existing LALMs through evaluation results, AIR-Bench can\nprovide insights into the direction of future research.\n", "date": "2024-02-12", "categories": []}
{"id": "2407.00465", "title": "Characterizing Continual Learning Scenarios and Strategies for Audio\n  Analysis", "abstract": "  Audio analysis is useful in many application scenarios. The state-of-the-art\naudio analysis approaches assume the data distribution at training and\ndeployment time will be the same. However, due to various real-life challenges,\nthe data may encounter drift in its distribution or can encounter new classes\nin the late future. Thus, a one-time trained model might not perform\nadequately. Continual learning (CL) approaches are devised to handle such\nchanges in data distribution. There have been a few attempts to use CL\napproaches for audio analysis. Yet, there is a lack of a systematic evaluation\nframework. In this paper, we create a comprehensive CL dataset and characterize\nCL approaches for audio-based monitoring tasks. We have investigated the\nfollowing CL and non-CL approaches: EWC, LwF, SI, GEM, A-GEM, GDumb, Replay,\nNaive, Cumulative, and Joint training. The study is very beneficial for\nresearchers and practitioners working in the area of audio analysis for\ndeveloping adaptive models. We observed that Replay achieved better results\nthan other methods in the DCASE challenge data. It achieved an accuracy of\n70.12% for the domain incremental scenario and an accuracy of 96.98% for the\nclass incremental scenario.\n", "date": "2024-06-29", "categories": []}
{"id": "2407.18517", "title": "SLIM: Style-Linguistics Mismatch Model for Generalized Audio Deepfake\n  Detection", "abstract": "  Audio deepfake detection (ADD) is crucial to combat the misuse of speech\nsynthesized from generative AI models. Existing ADD models suffer from\ngeneralization issues, with a large performance discrepancy between in-domain\nand out-of-domain data. Moreover, the black-box nature of existing models\nlimits their use in real-world scenarios, where explanations are required for\nmodel decisions. To alleviate these issues, we introduce a new ADD model that\nexplicitly uses the StyleLInguistics Mismatch (SLIM) in fake speech to separate\nthem from real speech. SLIM first employs self-supervised pretraining on only\nreal samples to learn the style-linguistics dependency in the real class. The\nlearned features are then used in complement with standard pretrained acoustic\nfeatures (e.g., Wav2vec) to learn a classifier on the real and fake classes.\nWhen the feature encoders are frozen, SLIM outperforms benchmark methods on\nout-of-domain datasets while achieving competitive results on in-domain data.\nThe features learned by SLIM allow us to quantify the (mis)match between style\nand linguistic content in a sample, hence facilitating an explanation of the\nmodel decision.\n", "date": "2024-07-26", "categories": []}
{"id": "2401.07575", "title": "Cascaded Cross-Modal Transformer for Audio-Textual Classification", "abstract": "  Speech classification tasks often require powerful language understanding\nmodels to grasp useful features, which becomes problematic when limited\ntraining data is available. To attain superior classification performance, we\npropose to harness the inherent value of multimodal representations by\ntranscribing speech using automatic speech recognition (ASR) models and\ntranslating the transcripts into different languages via pretrained translation\nmodels. We thus obtain an audio-textual (multimodal) representation for each\ndata sample. Subsequently, we combine language-specific Bidirectional Encoder\nRepresentations from Transformers (BERT) with Wav2Vec2.0 audio features via a\nnovel cascaded cross-modal transformer (CCMT). Our model is based on two\ncascaded transformer blocks. The first one combines text-specific features from\ndistinct languages, while the second one combines acoustic features with\nmultilingual features previously learned by the first transformer block. We\nemployed our system in the Requests Sub-Challenge of the ACM Multimedia 2023\nComputational Paralinguistics Challenge. CCMT was declared the winning\nsolution, obtaining an unweighted average recall (UAR) of 65.41% and 85.87% for\ncomplaint and request detection, respectively. Moreover, we applied our\nframework on the Speech Commands v2 and HarperValleyBank dialog data sets,\nsurpassing previous studies reporting results on these benchmarks. Our code is\nfreely available for download at: https://github.com/ristea/ccmt.\n", "date": "2024-01-15", "categories": []}
{"id": "2406.09272", "title": "Action2Sound: Ambient-Aware Generation of Action Sounds from Egocentric\n  Videos", "abstract": "  Generating realistic audio for human actions is important for many\napplications, such as creating sound effects for films or virtual reality\ngames. Existing approaches implicitly assume total correspondence between the\nvideo and audio during training, yet many sounds happen off-screen and have\nweak to no correspondence with the visuals -- resulting in uncontrolled ambient\nsounds or hallucinations at test time. We propose a novel ambient-aware audio\ngeneration model, AV-LDM. We devise a novel audio-conditioning mechanism to\nlearn to disentangle foreground action sounds from the ambient background\nsounds in in-the-wild training videos. Given a novel silent video, our model\nuses retrieval-augmented generation to create audio that matches the visual\ncontent both semantically and temporally. We train and evaluate our model on\ntwo in-the-wild egocentric video datasets, Ego4D and EPIC-KITCHENS, and we\nintroduce Ego4D-Sounds -- 1.2M curated clips with action-audio correspondence.\nOur model outperforms an array of existing methods, allows controllable\ngeneration of the ambient sound, and even shows promise for generalizing to\ncomputer graphics game clips. Overall, our approach is the first to focus\nvideo-to-audio generation faithfully on the observed visual content despite\ntraining from uncurated clips with natural background sounds.\n", "date": "2024-06-13", "categories": []}
{"id": "2407.18058", "title": "I can listen but cannot read: An evaluation of two-tower multimodal\n  systems for instrument recognition", "abstract": "  Music two-tower multimodal systems integrate audio and text modalities into a\njoint audio-text space, enabling direct comparison between songs and their\ncorresponding labels. These systems enable new approaches for classification\nand retrieval, leveraging both modalities. Despite the promising results they\nhave shown for zero-shot classification and retrieval tasks, closer inspection\nof the embeddings is needed. This paper evaluates the inherent zero-shot\nproperties of joint audio-text spaces for the case-study of instrument\nrecognition. We present an evaluation and analysis of two-tower systems for\nzero-shot instrument recognition and a detailed analysis of the properties of\nthe pre-joint and joint embeddings spaces. Our findings suggest that audio\nencoders alone demonstrate good quality, while challenges remain within the\ntext encoder or joint space projection. Specifically, two-tower systems exhibit\nsensitivity towards specific words, favoring generic prompts over musically\ninformed ones. Despite the large size of textual encoders, they do not yet\nleverage additional textual context or infer instruments accurately from their\ndescriptions. Lastly, a novel approach for quantifying the semantic\nmeaningfulness of the textual space leveraging an instrument ontology is\nproposed. This method reveals deficiencies in the systems' understanding of\ninstruments and provides evidence of the need for fine-tuning text encoders on\nmusical data.\n", "date": "2024-07-25", "categories": []}
{"id": "2407.16564", "title": "Audio Prompt Adapter: Unleashing Music Editing Abilities for\n  Text-to-Music with Lightweight Finetuning", "abstract": "  Text-to-music models allow users to generate nearly realistic musical audio\nwith textual commands. However, editing music audios remains challenging due to\nthe conflicting desiderata of performing fine-grained alterations on the audio\nwhile maintaining a simple user interface. To address this challenge, we\npropose Audio Prompt Adapter (or AP-Adapter), a lightweight addition to\npretrained text-to-music models. We utilize AudioMAE to extract features from\nthe input audio, and construct attention-based adapters to feedthese features\ninto the internal layers of AudioLDM2, a diffusion-based text-to-music model.\nWith 22M trainable parameters, AP-Adapter empowers users to harness both global\n(e.g., genre and timbre) and local (e.g., melody) aspects of music, using the\noriginal audio and a short text as inputs. Through objective and subjective\nstudies, we evaluate AP-Adapter on three tasks: timbre transfer, genre\ntransfer, and accompaniment generation. Additionally, we demonstrate its\neffectiveness on out-of-domain audios containing unseen instruments during\ntraining.\n", "date": "2024-07-23", "categories": []}
{"id": "2407.16554", "title": "Coarse-to-Fine Proposal Refinement Framework for Audio Temporal Forgery\n  Detection and Localization", "abstract": "  Recently, a novel form of audio partial forgery has posed challenges to its\nforensics, requiring advanced countermeasures to detect subtle forgery\nmanipulations within long-duration audio. However, existing countermeasures\nstill serve a classification purpose and fail to perform meaningful analysis of\nthe start and end timestamps of partial forgery segments. To address this\nchallenge, we introduce a novel coarse-to-fine proposal refinement framework\n(CFPRF) that incorporates a frame-level detection network (FDN) and a proposal\nrefinement network (PRN) for audio temporal forgery detection and localization.\nSpecifically, the FDN aims to mine informative inconsistency cues between real\nand fake frames to obtain discriminative features that are beneficial for\nroughly indicating forgery regions. The PRN is responsible for predicting\nconfidence scores and regression offsets to refine the coarse-grained proposals\nderived from the FDN. To learn robust discriminative features, we devise a\ndifference-aware feature learning (DAFL) module guided by contrastive\nrepresentation learning to enlarge the sensitive differences between different\nframes induced by minor manipulations. We further design a boundary-aware\nfeature enhancement (BAFE) module to capture the contextual information of\nmultiple transition boundaries and guide the interaction between boundary\ninformation and temporal features via a cross-attention mechanism. Extensive\nexperiments show that our CFPRF achieves state-of-the-art performance on\nvarious datasets, including LAV-DF, ASVS2019PS, and HAD.\n", "date": "2024-07-23", "categories": []}
{"id": "2407.16643", "title": "Synthesizer Sound Matching Using Audio Spectrogram Transformers", "abstract": "  Systems for synthesizer sound matching, which automatically set the\nparameters of a synthesizer to emulate an input sound, have the potential to\nmake the process of synthesizer programming faster and easier for novice and\nexperienced musicians alike, whilst also affording new means of interaction\nwith synthesizers. Considering the enormous variety of synthesizers in the\nmarketplace, and the complexity of many of them, general-purpose sound matching\nsystems that function with minimal knowledge or prior assumptions about the\nunderlying synthesis architecture are particularly desirable. With this in\nmind, we introduce a synthesizer sound matching model based on the Audio\nSpectrogram Transformer. We demonstrate the viability of this model by training\non a large synthetic dataset of randomly generated samples from the popular\nMassive synthesizer. We show that this model can reconstruct parameters of\nsamples generated from a set of 16 parameters, highlighting its improved\nfidelity relative to multi-layer perceptron and convolutional neural network\nbaselines. We also provide audio examples demonstrating the out-of-domain model\nperformance in emulating vocal imitations, and sounds from other synthesizers\nand musical instruments.\n", "date": "2024-07-23", "categories": []}
{"id": "2407.16691", "title": "Automatic Equalization for Individual Instrument Tracks Using\n  Convolutional Neural Networks", "abstract": "  We propose a novel approach for the automatic equalization of individual\nmusical instrument tracks. Our method begins by identifying the instrument\npresent within a source recording in order to choose its corresponding ideal\nspectrum as a target. Next, the spectral difference between the recording and\nthe target is calculated, and accordingly, an equalizer matching model is used\nto predict settings for a parametric equalizer. To this end, we build upon a\ndifferentiable parametric equalizer matching neural network, demonstrating\nimprovements relative to previously established state-of-the-art. Unlike past\napproaches, we show how our system naturally allows real-world audio data to be\nleveraged during the training of our matching model, effectively generating\nsuitably produced training targets in an automated manner mirroring conditions\nat inference time. Consequently, we illustrate how fine-tuning our matching\nmodel on such examples considerably improves parametric equalizer matching\nperformance in real-world scenarios, decreasing mean absolute error by 24%\nrelative to methods relying solely on random parameter sampling techniques as a\nself-supervised learning strategy. We perform listening tests, and demonstrate\nthat our proposed automatic equalization solution subjectively enhances the\ntonal characteristics for recordings of common instrument types.\n", "date": "2024-07-23", "categories": []}
{"id": "2312.07059", "title": "LSTM-CNN Network for Audio Signature Analysis in Noisy Environments", "abstract": "  There are multiple applications to automatically count people and specify\ntheir gender at work, exhibitions, malls, sales, and industrial usage. Although\ncurrent speech detection methods are supposed to operate well, in most\nsituations, in addition to genders, the number of current speakers is unknown\nand the classification methods are not suitable due to many possible classes.\nIn this study, we focus on a long-short-term memory convolutional neural\nnetwork (LSTM-CNN) to extract time and / or frequency-dependent features of the\nsound data to estimate the number / gender of simultaneous active speakers at\neach frame in noisy environments. Considering the maximum number of speakers as\n10, we have utilized 19000 audio samples with diverse combinations of males,\nfemales, and background noise in public cities, industrial situations, malls,\nexhibitions, workplaces, and nature for learning purposes. This proof of\nconcept shows promising performance with training/validation MSE values of\nabout 0.019/0.017 in detecting count and gender.\n", "date": "2023-12-12", "categories": []}
{"id": "2303.17395", "title": "WavCaps: A ChatGPT-Assisted Weakly-Labelled Audio Captioning Dataset for\n  Audio-Language Multimodal Research", "abstract": "  The advancement of audio-language (AL) multimodal learning tasks has been\nsignificant in recent years. However, researchers face challenges due to the\ncostly and time-consuming collection process of existing audio-language\ndatasets, which are limited in size. To address this data scarcity issue, we\nintroduce WavCaps, the first large-scale weakly-labelled audio captioning\ndataset, comprising approximately 400k audio clips with paired captions. We\nsourced audio clips and their raw descriptions from web sources and a sound\nevent detection dataset. However, the online-harvested raw descriptions are\nhighly noisy and unsuitable for direct use in tasks such as automated audio\ncaptioning. To overcome this issue, we propose a three-stage processing\npipeline for filtering noisy data and generating high-quality captions, where\nChatGPT, a large language model, is leveraged to filter and transform raw\ndescriptions automatically. We conduct a comprehensive analysis of the\ncharacteristics of WavCaps dataset and evaluate it on multiple downstream\naudio-language multimodal learning tasks. The systems trained on WavCaps\noutperform previous state-of-the-art (SOTA) models by a significant margin. Our\naspiration is for the WavCaps dataset we have proposed to facilitate research\nin audio-language multimodal learning and demonstrate the potential of\nutilizing ChatGPT to enhance academic research. Our dataset and codes are\navailable at https://github.com/XinhaoMei/WavCaps.\n", "date": "2023-03-30", "categories": []}
{"id": "2407.13676", "title": "Aligning Sight and Sound: Advanced Sound Source Localization Through\n  Audio-Visual Alignment", "abstract": "  Recent studies on learning-based sound source localization have mainly\nfocused on the localization performance perspective. However, prior work and\nexisting benchmarks overlook a crucial aspect: cross-modal interaction, which\nis essential for interactive sound source localization. Cross-modal interaction\nis vital for understanding semantically matched or mismatched audio-visual\nevents, such as silent objects or off-screen sounds. In this paper, we first\ncomprehensively examine the cross-modal interaction of existing methods,\nbenchmarks, evaluation metrics, and cross-modal understanding tasks. Then, we\nidentify the limitations of previous studies and make several contributions to\novercome the limitations. First, we introduce a new synthetic benchmark for\ninteractive sound source localization. Second, we introduce new evaluation\nmetrics to rigorously assess sound source localization methods, focusing on\naccurately evaluating both localization performance and cross-modal interaction\nability. Third, we propose a learning framework with a cross-modal alignment\nstrategy to enhance cross-modal interaction. Lastly, we evaluate both\ninteractive sound source localization and auxiliary cross-modal retrieval tasks\ntogether to thoroughly assess cross-modal interaction capabilities and\nbenchmark competing methods. Our new benchmarks and evaluation metrics reveal\npreviously overlooked issues in sound source localization studies. Our proposed\nnovel method, with enhanced cross-modal alignment, shows superior sound source\nlocalization performance. This work provides the most comprehensive analysis of\nsound source localization to date, with extensive validation of competing\nmethods on both existing and new benchmarks using new and standard evaluation\nmetrics.\n", "date": "2024-07-18", "categories": []}
{"id": "2405.06134", "title": "Muting Whisper: A Universal Acoustic Adversarial Attack on Speech\n  Foundation Models", "abstract": "  Recent developments in large speech foundation models like Whisper have led\nto their widespread use in many automatic speech recognition (ASR)\napplications. These systems incorporate `special tokens' in their vocabulary,\nsuch as $\\texttt{<|endoftext|>}$, to guide their language generation process.\nHowever, we demonstrate that these tokens can be exploited by adversarial\nattacks to manipulate the model's behavior. We propose a simple yet effective\nmethod to learn a universal acoustic realization of Whisper's\n$\\texttt{<|endoftext|>}$ token, which, when prepended to any speech signal,\nencourages the model to ignore the speech and only transcribe the special\ntoken, effectively `muting' the model. Our experiments demonstrate that the\nsame, universal 0.64-second adversarial audio segment can successfully mute a\ntarget Whisper ASR model for over 97\\% of speech samples. Moreover, we find\nthat this universal adversarial audio segment often transfers to new datasets\nand tasks. Overall this work demonstrates the vulnerability of Whisper models\nto `muting' adversarial attacks, where such attacks can pose both risks and\npotential benefits in real-world settings: for example the attack can be used\nto bypass speech moderation systems, or conversely the attack can also be used\nto protect private speech data.\n", "date": "2024-05-09", "categories": []}
{"id": "2407.11650", "title": "Statistics-aware Audio-visual Deepfake Detector", "abstract": "  In this paper, we propose an enhanced audio-visual deep detection method.\nRecent methods in audio-visual deepfake detection mostly assess the\nsynchronization between audio and visual features. Although they have shown\npromising results, they are based on the maximization/minimization of isolated\nfeature distances without considering feature statistics. Moreover, they rely\non cumbersome deep learning architectures and are heavily dependent on\nempirically fixed hyperparameters. Herein, to overcome these limitations, we\npropose: (1) a statistical feature loss to enhance the discrimination\ncapability of the model, instead of relying solely on feature distances; (2)\nusing the waveform for describing the audio as a replacement of frequency-based\nrepresentations; (3) a post-processing normalization of the fakeness score; (4)\nthe use of shallower network for reducing the computational complexity.\nExperiments on the DFDC and FakeAVCeleb datasets demonstrate the relevance of\nthe proposed method.\n", "date": "2024-07-16", "categories": []}
{"id": "2407.10373", "title": "Mutual Learning for Acoustic Matching and Dereverberation via Visual\n  Scene-driven Diffusion", "abstract": "  Visual acoustic matching (VAM) is pivotal for enhancing the immersive\nexperience, and the task of dereverberation is effective in improving audio\nintelligibility. Existing methods treat each task independently, overlooking\nthe inherent reciprocity between them. Moreover, these methods depend on paired\ntraining data, which is challenging to acquire, impeding the utilization of\nextensive unpaired data. In this paper, we introduce MVSD, a mutual learning\nframework based on diffusion models. MVSD considers the two tasks\nsymmetrically, exploiting the reciprocal relationship to facilitate learning\nfrom inverse tasks and overcome data scarcity. Furthermore, we employ the\ndiffusion model as foundational conditional converters to circumvent the\ntraining instability and over-smoothing drawbacks of conventional GAN\narchitectures. Specifically, MVSD employs two converters: one for VAM called\nreverberator and one for dereverberation called dereverberator. The\ndereverberator judges whether the reverberation audio generated by reverberator\nsounds like being in the conditional visual scenario, and vice versa. By\nforming a closed loop, these two converters can generate informative feedback\nsignals to optimize the inverse tasks, even with easily acquired one-way\nunpaired data. Extensive experiments on two standard benchmarks, i.e.,\nSoundSpaces-Speech and Acoustic AVSpeech, exhibit that our framework can\nimprove the performance of the reverberator and dereverberator and better match\nspecified visual scenarios.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.10387", "title": "Masked Generative Video-to-Audio Transformers with Enhanced\n  Synchronicity", "abstract": "  Video-to-audio (V2A) generation leverages visual-only video features to\nrender plausible sounds that match the scene. Importantly, the generated sound\nonsets should match the visual actions that are aligned with them, otherwise\nunnatural synchronization artifacts arise. Recent works have explored the\nprogression of conditioning sound generators on still images and then video\nfeatures, focusing on quality and semantic matching while ignoring\nsynchronization, or by sacrificing some amount of quality to focus on improving\nsynchronization only. In this work, we propose a V2A generative model, named\nMaskVAT, that interconnects a full-band high-quality general audio codec with a\nsequence-to-sequence masked generative model. This combination allows modeling\nboth high audio quality, semantic matching, and temporal synchronicity at the\nsame time. Our results show that, by combining a high-quality codec with the\nproper pre-trained audio-visual features and a sequence-to-sequence parallel\nstructure, we are able to yield highly synchronized results on one hand, whilst\nbeing competitive with the state of the art of non-codec generative audio\nmodels. Sample videos and generated audios are available at\nhttps://maskvat.github.io .\n", "date": "2024-07-15", "categories": []}
{"id": "2407.10468", "title": "LiteFocus: Accelerated Diffusion Inference for Long Audio Synthesis", "abstract": "  Latent diffusion models have shown promising results in audio generation,\nmaking notable advancements over traditional methods. However, their\nperformance, while impressive with short audio clips, faces challenges when\nextended to longer audio sequences. These challenges are due to model's\nself-attention mechanism and training predominantly on 10-second clips, which\ncomplicates the extension to longer audio without adaptation. In response to\nthese issues, we introduce a novel approach, LiteFocus that enhances the\ninference of existing audio latent diffusion models in long audio synthesis.\nObserved the attention pattern in self-attention, we employ a dual sparse form\nfor attention calculation, designated as same-frequency focus and\ncross-frequency compensation, which curtails the attention computation under\nsame-frequency constraints, while enhancing audio quality through\ncross-frequency refillment. LiteFocus demonstrates substantial reduction on\ninference time with diffusion-based TTA model by 1.99x in synthesizing\n80-second audio clips while also obtaining improved audio quality.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.10759", "title": "Qwen2-Audio Technical Report", "abstract": "  We introduce the latest progress of Qwen-Audio, a large-scale audio-language\nmodel called Qwen2-Audio, which is capable of accepting various audio signal\ninputs and performing audio analysis or direct textual responses with regard to\nspeech instructions. In contrast to complex hierarchical tags, we have\nsimplified the pre-training process by utilizing natural language prompts for\ndifferent data and tasks, and have further expanded the data volume. We have\nboosted the instruction-following capability of Qwen2-Audio and implemented two\ndistinct audio interaction modes for voice chat and audio analysis. In the\nvoice chat mode, users can freely engage in voice interactions with Qwen2-Audio\nwithout text input. In the audio analysis mode, users could provide audio and\ntext instructions for analysis during the interaction. Note that we do not use\nany system prompts to switch between voice chat and audio analysis modes.\nQwen2-Audio is capable of intelligently comprehending the content within audio\nand following voice commands to respond appropriately. For instance, in an\naudio segment that simultaneously contains sounds, multi-speaker conversations,\nand a voice command, Qwen2-Audio can directly understand the command and\nprovide an interpretation and response to the audio. Additionally, DPO has\noptimized the model's performance in terms of factuality and adherence to\ndesired behavior. According to the evaluation results from AIR-Bench,\nQwen2-Audio outperformed previous SOTAs, such as Gemini-1.5-pro, in tests\nfocused on audio-centric instruction-following capabilities. Qwen2-Audio is\nopen-sourced with the aim of fostering the advancement of the multi-modal\nlanguage community.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.10828", "title": "Towards Enhanced Classification of Abnormal Lung sound in Multi-breath:\n  A Light Weight Multi-label and Multi-head Attention Classification Method", "abstract": "  This study aims to develop an auxiliary diagnostic system for classifying\nabnormal lung respiratory sounds, enhancing the accuracy of automatic abnormal\nbreath sound classification through an innovative multi-label learning approach\nand multi-head attention mechanism. Addressing the issue of class imbalance and\nlack of diversity in existing respiratory sound datasets, our study employs a\nlightweight and highly accurate model, using a two-dimensional label set to\nrepresent multiple respiratory sound characteristics. Our method achieved a\n59.2% ICBHI score in the four-category task on the ICBHI2017 dataset,\ndemonstrating its advantages in terms of lightweight and high accuracy. This\nstudy not only improves the accuracy of automatic diagnosis of lung respiratory\nsound abnormalities but also opens new possibilities for clinical applications.\n", "date": "2024-07-15", "categories": []}
{"id": "2406.16990", "title": "AND: Audio Network Dissection for Interpreting Deep Acoustic Models", "abstract": "  Neuron-level interpretations aim to explain network behaviors and properties\nby investigating neurons responsive to specific perceptual or structural input\npatterns. Although there is emerging work in the vision and language domains,\nnone is explored for acoustic models. To bridge the gap, we introduce\n$\\textit{AND}$, the first $\\textbf{A}$udio $\\textbf{N}$etwork\n$\\textbf{D}$issection framework that automatically establishes natural language\nexplanations of acoustic neurons based on highly-responsive audio.\n$\\textit{AND}$ features the use of LLMs to summarize mutual acoustic features\nand identities among audio. Extensive experiments are conducted to verify\n$\\textit{AND}$'s precise and informative descriptions. In addition, we\ndemonstrate a potential use of $\\textit{AND}$ for audio machine unlearning by\nconducting concept-specific pruning based on the generated descriptions.\nFinally, we highlight two acoustic model behaviors with analysis by\n$\\textit{AND}$: (i) models discriminate audio with a combination of basic\nacoustic features rather than high-level abstract concepts; (ii) training\nstrategies affect model behaviors and neuron interpretability -- supervised\ntraining guides neurons to gradually narrow their attention, while\nself-supervised learning encourages neurons to be polysemantic for exploring\nhigh-level features.\n", "date": "2024-06-24", "categories": []}
{"id": "2407.08239", "title": "An Unsupervised Domain Adaptation Method for Locating Manipulated Region\n  in partially fake Audio", "abstract": "  When the task of locating manipulation regions in partially-fake audio (PFA)\ninvolves cross-domain datasets, the performance of deep learning models drops\nsignificantly due to the shift between the source and target domains. To\naddress this issue, existing approaches often employ data augmentation before\ntraining. However, they overlook the characteristics in target domain that are\nabsent in source domain. Inspired by the mixture-of-experts model, we propose\nan unsupervised method named Samples mining with Diversity and Entropy (SDE).\nOur method first learns from a collection of diverse experts that achieve great\nperformance from different perspectives in the source domain, but with\nambiguity on target samples. We leverage these diverse experts to select the\nmost informative samples by calculating their entropy. Furthermore, we\nintroduced a label generation method tailored for these selected samples that\nare incorporated in the training process in source domain integrating the\ntarget domain information. We applied our method to a cross-domain partially\nfake audio detection dataset, ADD2023Track2. By introducing 10% of unknown\nsamples from the target domain, we achieved an F1 score of 43.84%, which\nrepresents a relative increase of 77.2% compared to the second-best method.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.08647", "title": "From Real to Cloned Singer Identification", "abstract": "  Cloned voices of popular singers sound increasingly realistic and have gained\npopularity over the past few years. They however pose a threat to the industry\ndue to personality rights concerns. As such, methods to identify the original\nsinger in synthetic voices are needed. In this paper, we investigate how singer\nidentification methods could be used for such a task. We present three\nembedding models that are trained using a singer-level contrastive learning\nscheme, where positive pairs consist of segments with vocals from the same\nsingers. These segments can be mixtures for the first model, vocals for the\nsecond, and both for the third. We demonstrate that all three models are highly\ncapable of identifying real singers. However, their performance deteriorates\nwhen classifying cloned versions of singers in our evaluation set. This is\nespecially true for models that use mixtures as an input. These findings\nhighlight the need to understand the biases that exist within singer\nidentification systems, and how they can influence the identification of voice\ndeepfakes in music.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.08691", "title": "ElasticAST: An Audio Spectrogram Transformer for All Length and\n  Resolutions", "abstract": "  Transformers have rapidly overtaken CNN-based architectures as the new\nstandard in audio classification. Transformer-based models, such as the Audio\nSpectrogram Transformers (AST), also inherit the fixed-size input paradigm from\nCNNs. However, this leads to performance degradation for ASTs in the inference\nwhen input lengths vary from the training. This paper introduces an approach\nthat enables the use of variable-length audio inputs with AST models during\nboth training and inference. By employing sequence packing, our method\nElasticAST, accommodates any audio length during training, thereby offering\nflexibility across all lengths and resolutions at the inference. This\nflexibility allows ElasticAST to maintain evaluation capabilities at various\nlengths or resolutions and achieve similar performance to standard ASTs trained\nat specific lengths or resolutions. Moreover, experiments demonstrate\nElasticAST's better performance when trained and evaluated on native-length\naudio datasets.\n", "date": "2024-07-11", "categories": []}
{"id": "2402.01413", "title": "Objective and subjective evaluation of speech enhancement methods in the\n  UDASE task of the 7th CHiME challenge", "abstract": "  Supervised models for speech enhancement are trained using artificially\ngenerated mixtures of clean speech and noise signals. However, the synthetic\ntraining conditions may not accurately reflect real-world conditions\nencountered during testing. This discrepancy can result in poor performance\nwhen the test domain significantly differs from the synthetic training domain.\nTo tackle this issue, the UDASE task of the 7th CHiME challenge aimed to\nleverage real-world noisy speech recordings from the test domain for\nunsupervised domain adaptation of speech enhancement models. Specifically, this\ntest domain corresponds to the CHiME-5 dataset, characterized by real\nmulti-speaker and conversational speech recordings made in noisy and\nreverberant domestic environments, for which ground-truth clean speech signals\nare not available. In this paper, we present the objective and subjective\nevaluations of the systems that were submitted to the CHiME-7 UDASE task, and\nwe provide an analysis of the results. This analysis reveals a limited\ncorrelation between subjective ratings and several supervised nonintrusive\nperformance metrics recently proposed for speech enhancement. Conversely, the\nresults suggest that more traditional intrusive objective metrics can be used\nfor in-domain performance evaluation using the reverberant LibriCHiME-5 dataset\ndeveloped for the challenge. The subjective evaluation indicates that all\nsystems successfully reduced the background noise, but always at the expense of\nincreased distortion. Out of the four speech enhancement methods evaluated\nsubjectively, only one demonstrated an improvement in overall quality compared\nto the unprocessed noisy speech, highlighting the difficulty of the task. The\ntools and audio material created for the CHiME-7 UDASE task are shared with the\ncommunity.\n", "date": "2024-02-02", "categories": []}
{"id": "2407.07598", "title": "Targeted Augmented Data for Audio Deepfake Detection", "abstract": "  The availability of highly convincing audio deepfake generators highlights\nthe need for designing robust audio deepfake detectors. Existing works often\nrely solely on real and fake data available in the training set, which may lead\nto overfitting, thereby reducing the robustness to unseen manipulations. To\nenhance the generalization capabilities of audio deepfake detectors, we propose\na novel augmentation method for generating audio pseudo-fakes targeting the\ndecision boundary of the model. Inspired by adversarial attacks, we perturb\noriginal real data to synthesize pseudo-fakes with ambiguous prediction\nprobabilities. Comprehensive experiments on two well-known architectures\ndemonstrate that the proposed augmentation contributes to improving the\ngeneralization capabilities of these architectures.\n", "date": "2024-07-10", "categories": []}
{"id": "2407.07825", "title": "RT-LA-VocE: Real-Time Low-SNR Audio-Visual Speech Enhancement", "abstract": "  In this paper, we aim to generate clean speech frame by frame from a live\nvideo stream and a noisy audio stream without relying on future inputs. To this\nend, we propose RT-LA-VocE, which completely re-designs every component of\nLA-VocE, a state-of-the-art non-causal audio-visual speech enhancement model,\nto perform causal real-time inference with a 40ms input frame. We do so by\ndevising new visual and audio encoders that rely solely on past frames,\nreplacing the Transformer encoder with the Emformer, and designing a new causal\nneural vocoder C-HiFi-GAN. On the popular AVSpeech dataset, we show that our\nalgorithm achieves state-of-the-art results in all real-time scenarios. More\nimportantly, each component is carefully tuned to minimize the algorithm\nlatency to the theoretical minimum (40ms) while maintaining a low end-to-end\nprocessing latency of 28.15ms per frame, enabling real-time frame-by-frame\nenhancement with minimal delay.\n", "date": "2024-07-10", "categories": []}
{"id": "2206.04805", "title": "Motif Mining and Unsupervised Representation Learning for BirdCLEF 2022", "abstract": "  We build a classification model for the BirdCLEF 2022 challenge using\nunsupervised methods. We implement an unsupervised representation of the\ntraining dataset using a triplet loss on spectrogram representation of audio\nmotifs. Our best model performs with a score of 0.48 on the public leaderboard.\n", "date": "2022-06-08", "categories": []}
{"id": "2306.16760", "title": "Transfer Learning with Semi-Supervised Dataset Annotation for Birdcall\n  Classification", "abstract": "  We present working notes on transfer learning with semi-supervised dataset\nannotation for the BirdCLEF 2023 competition, focused on identifying African\nbird species in recorded soundscapes. Our approach utilizes existing\noff-the-shelf models, BirdNET and MixIT, to address representation and labeling\nchallenges in the competition. We explore the embedding space learned by\nBirdNET and propose a process to derive an annotated dataset for supervised\nlearning. Our experiments involve various models and feature engineering\napproaches to maximize performance on the competition leaderboard. The results\ndemonstrate the effectiveness of our approach in classifying bird species and\nhighlight the potential of transfer learning and semi-supervised dataset\nannotation in similar tasks.\n", "date": "2023-06-29", "categories": []}
{"id": "2406.15487", "title": "Improving Text-To-Audio Models with Synthetic Captions", "abstract": "  It is an open challenge to obtain high quality training data, especially\ncaptions, for text-to-audio models. Although prior methods have leveraged\n\\textit{text-only language models} to augment and improve captions, such\nmethods have limitations related to scale and coherence between audio and\ncaptions. In this work, we propose an audio captioning pipeline that uses an\n\\textit{audio language model} to synthesize accurate and diverse captions for\naudio at scale. We leverage this pipeline to produce a dataset of synthetic\ncaptions for AudioSet, named \\texttt{AF-AudioSet}, and then evaluate the\nbenefit of pre-training text-to-audio models on these synthetic captions.\nThrough systematic evaluations on AudioCaps and MusicCaps, we find leveraging\nour pipeline and synthetic captions leads to significant improvements on audio\ngeneration quality, achieving a new \\textit{state-of-the-art}.\n", "date": "2024-06-18", "categories": []}
{"id": "2307.13821", "title": "Fitting Auditory Filterbanks with Multiresolution Neural Networks", "abstract": "  Waveform-based deep learning faces a dilemma between nonparametric and\nparametric approaches. On one hand, convolutional neural networks (convnets)\nmay approximate any linear time-invariant system; yet, in practice, their\nfrequency responses become more irregular as their receptive fields grow. On\nthe other hand, a parametric model such as LEAF is guaranteed to yield Gabor\nfilters, hence an optimal time-frequency localization; yet, this strong\ninductive bias comes at the detriment of representational capacity. In this\npaper, we aim to overcome this dilemma by introducing a neural audio model,\nnamed multiresolution neural network (MuReNN). The key idea behind MuReNN is to\ntrain separate convolutional operators over the octave subbands of a discrete\nwavelet transform (DWT). Since the scale of DWT atoms grows exponentially\nbetween octaves, the receptive fields of the subsequent learnable convolutions\nin MuReNN are dilated accordingly. For a given real-world dataset, we fit the\nmagnitude response of MuReNN to that of a well-established auditory filterbank:\nGammatone for speech, CQT for music, and third-octave for urban sounds,\nrespectively. This is a form of knowledge distillation (KD), in which the\nfilterbank ''teacher'' is engineered by domain knowledge while the neural\nnetwork ''student'' is optimized from data. We compare MuReNN to the state of\nthe art in terms of goodness of fit after KD on a hold-out set and in terms of\nHeisenberg time-frequency localization. Compared to convnets and Gabor\nconvolutions, we find that MuReNN reaches state-of-the-art performance on all\nthree optimization problems.\n", "date": "2023-07-25", "categories": []}
{"id": "2407.05368", "title": "Music Era Recognition Using Supervised Contrastive Learning and Artist\n  Information", "abstract": "  Does popular music from the 60s sound different than that of the 90s? Prior\nstudy has shown that there would exist some variations of patterns and\nregularities related to instrumentation changes and growing loudness across\nmulti-decadal trends. This indicates that perceiving the era of a song from\nmusical features such as audio and artist information is possible. Music era\ninformation can be an important feature for playlist generation and\nrecommendation. However, the release year of a song can be inaccessible in many\ncircumstances. This paper addresses a novel task of music era recognition. We\nformulate the task as a music classification problem and propose solutions\nbased on supervised contrastive learning. An audio-based model is developed to\npredict the era from audio. For the case where the artist information is\navailable, we extend the audio-based model to take multimodal inputs and\ndevelop a framework, called MultiModal Contrastive (MMC) learning, to enhance\nthe training. Experimental result on Million Song Dataset demonstrates that the\naudio-based model achieves 54% in accuracy with a tolerance of 3-years range;\nincorporating the artist information with the MMC framework for training leads\nto 9% improvement further.\n", "date": "2024-07-07", "categories": []}
{"id": "2312.09665", "title": "FlowMur: A Stealthy and Practical Audio Backdoor Attack with Limited\n  Knowledge", "abstract": "  Speech recognition systems driven by DNNs have revolutionized human-computer\ninteraction through voice interfaces, which significantly facilitate our daily\nlives. However, the growing popularity of these systems also raises special\nconcerns on their security, particularly regarding backdoor attacks. A backdoor\nattack inserts one or more hidden backdoors into a DNN model during its\ntraining process, such that it does not affect the model's performance on\nbenign inputs, but forces the model to produce an adversary-desired output if a\nspecific trigger is present in the model input. Despite the initial success of\ncurrent audio backdoor attacks, they suffer from the following limitations: (i)\nMost of them require sufficient knowledge, which limits their widespread\nadoption. (ii) They are not stealthy enough, thus easy to be detected by\nhumans. (iii) Most of them cannot attack live speech, reducing their\npracticality. To address these problems, in this paper, we propose FlowMur, a\nstealthy and practical audio backdoor attack that can be launched with limited\nknowledge. FlowMur constructs an auxiliary dataset and a surrogate model to\naugment adversary knowledge. To achieve dynamicity, it formulates trigger\ngeneration as an optimization problem and optimizes the trigger over different\nattachment positions. To enhance stealthiness, we propose an adaptive data\npoisoning method according to Signal-to-Noise Ratio (SNR). Furthermore, ambient\nnoise is incorporated into the process of trigger generation and data poisoning\nto make FlowMur robust to ambient noise and improve its practicality. Extensive\nexperiments conducted on two datasets demonstrate that FlowMur achieves high\nattack performance in both digital and physical settings while remaining\nresilient to state-of-the-art defenses. In particular, a human study confirms\nthat triggers generated by FlowMur are not easily detected by participants.\n", "date": "2023-12-15", "categories": []}
{"id": "2406.12434", "title": "Towards Audio Codec-based Speech Separation", "abstract": "  Recent improvements in neural audio codec (NAC) models have generated\ninterest in adopting pre-trained codecs for a variety of speech processing\napplications to take advantage of the efficiencies gained from high\ncompression, but these have yet been applied to the speech separation (SS)\ntask. SS can benefit from high compression because the compute required for\ntraditional SS models makes them impractical for many edge computing use cases.\nHowever, SS is a waveform-masking task where compression tends to introduce\ndistortions that severely impact performance. Here we propose a novel task of\nAudio Codec-based SS, where SS is performed within the embedding space of a\nNAC, and propose a new model, Codecformer, to address this task. At inference,\nCodecformer achieves a 52x reduction in MAC while producing separation\nperformance comparable to a cloud deployment of Sepformer. This method charts a\nnew direction for performing efficient SS in practical scenarios.\n", "date": "2024-06-18", "categories": []}
{"id": "2407.00188", "title": "A Novel Labeled Human Voice Signal Dataset for Misbehavior Detection", "abstract": "  Voice signal classification based on human behaviours involves analyzing\nvarious aspects of speech patterns and delivery styles. In this study, a\nreal-time dataset collection is performed where participants are instructed to\nspeak twelve psychology questions in two distinct manners: first, in a harsh\nvoice, which is categorized as \"misbehaved\"; and second, in a polite manner,\ncategorized as \"normal\". These classifications are crucial in understanding how\ndifferent vocal behaviours affect the interpretation and classification of\nvoice signals. This research highlights the significance of voice tone and\ndelivery in automated machine-learning systems for voice analysis and\nrecognition. This research contributes to the broader field of voice signal\nanalysis by elucidating the impact of human behaviour on the perception and\ncategorization of voice signals, thereby enhancing the development of more\naccurate and context-aware voice recognition technologies.\n", "date": "2024-06-28", "categories": []}
{"id": "2407.03440", "title": "Advanced Framework for Animal Sound Classification With Features\n  Optimization", "abstract": "  The automatic classification of animal sounds presents an enduring challenge\nin bioacoustics, owing to the diverse statistical properties of sound signals,\nvariations in recording equipment, and prevalent low Signal-to-Noise Ratio\n(SNR) conditions. Deep learning models like Convolutional Neural Networks (CNN)\nand Long Short-Term Memory (LSTM) have excelled in human speech recognition but\nhave not been effectively tailored to the intricate nature of animal sounds,\nwhich exhibit substantial diversity even within the same domain. We propose an\nautomated classification framework applicable to general animal sound\nclassification. Our approach first optimizes audio features from Mel-frequency\ncepstral coefficients (MFCC) including feature rearrangement and feature\nreduction. It then uses the optimized features for the deep learning model,\ni.e., an attention-based Bidirectional LSTM (Bi-LSTM), to extract deep semantic\nfeatures for sound classification. We also contribute an animal sound benchmark\ndataset encompassing oceanic animals and birds1. Extensive experimentation with\nreal-world datasets demonstrates that our approach consistently outperforms\nbaseline methods by over 25% in precision, recall, and accuracy, promising\nadvancements in animal sound classification.\n", "date": "2024-07-03", "categories": []}
{"id": "2407.03514", "title": "Towards Attention-based Contrastive Learning for Audio Spoof Detection", "abstract": "  Vision transformers (ViT) have made substantial progress for classification\ntasks in computer vision. Recently, Gong et. al. '21, introduced\nattention-based modeling for several audio tasks. However, relatively\nunexplored is the use of a ViT for audio spoof detection task. We bridge this\ngap and introduce ViTs for this task. A vanilla baseline built on fine-tuning\nthe SSAST (Gong et. al. '22) audio ViT model achieves sub-optimal equal error\nrates (EERs). To improve performance, we propose a novel attention-based\ncontrastive learning framework (SSAST-CL) that uses cross-attention to aid the\nrepresentation learning. Experiments show that our framework successfully\ndisentangles the bonafide and spoof classes and helps learn better classifiers\nfor the task. With appropriate data augmentations policy, a model trained on\nour framework achieves competitive performance on the ASVSpoof 2021 challenge.\nWe provide comparisons and ablation studies to justify our claim.\n", "date": "2024-07-03", "categories": []}
{"id": "2407.03736", "title": "Semantic Grouping Network for Audio Source Separation", "abstract": "  Recently, audio-visual separation approaches have taken advantage of the\nnatural synchronization between the two modalities to boost audio source\nseparation performance. They extracted high-level semantics from visual inputs\nas the guidance to help disentangle sound representation for individual\nsources. Can we directly learn to disentangle the individual semantics from the\nsound itself? The dilemma is that multiple sound sources are mixed together in\nthe original space. To tackle the difficulty, in this paper, we present a novel\nSemantic Grouping Network, termed as SGN, that can directly disentangle sound\nrepresentations and extract high-level semantic information for each source\nfrom input audio mixture. Specifically, SGN aggregates category-wise source\nfeatures through learnable class tokens of sounds. Then, the aggregated\nsemantic features can be used as the guidance to separate the corresponding\naudio sources from the mixture. We conducted extensive experiments on\nmusic-only and universal sound separation benchmarks: MUSIC, FUSS, MUSDB18, and\nVGG-Sound. The results demonstrate that our SGN significantly outperforms\nprevious audio-only methods and audio-visual models without utilizing\nadditional visual cues.\n", "date": "2024-07-04", "categories": []}
{"id": "2407.04547", "title": "Real-time Timbre Remapping with Differentiable DSP", "abstract": "  Timbre is a primary mode of expression in diverse musical contexts. However,\nprevalent audio-driven synthesis methods predominantly rely on pitch and\nloudness envelopes, effectively flattening timbral expression from the input.\nOur approach draws on the concept of timbre analogies and investigates how\ntimbral expression from an input signal can be mapped onto controls for a\nsynthesizer. Leveraging differentiable digital signal processing, our method\nfacilitates direct optimization of synthesizer parameters through a novel\nfeature difference loss. This loss function, designed to learn relative timbral\ndifferences between musical events, prioritizes the subtleties of graded timbre\nmodulations within phrases, allowing for meaningful translations in a timbre\nspace. Using snare drum performances as a case study, where timbral expression\nis central, we demonstrate real-time timbre remapping from acoustic snare drums\nto a differentiable synthesizer modeled after the Roland TR-808.\n", "date": "2024-07-05", "categories": []}
{"id": "2312.10402", "title": "Annotation-free Automatic Music Transcription with Scalable Synthetic\n  Data and Adversarial Domain Confusion", "abstract": "  Automatic Music Transcription (AMT) is a vital technology in the field of\nmusic information processing. Despite recent enhancements in performance due to\nmachine learning techniques, current methods typically attain high accuracy in\ndomains where abundant annotated data is available. Addressing domains with low\nor no resources continues to be an unresolved challenge. To tackle this issue,\nwe propose a transcription model that does not require any MIDI-audio paired\ndata through the utilization of scalable synthetic audio for pre-training and\nadversarial domain confusion using unannotated real audio. In experiments, we\nevaluate methods under the real-world application scenario where training\ndatasets do not include the MIDI annotation of audio in the target data domain.\nOur proposed method achieved competitive performance relative to established\nbaseline methods, despite not utilizing any real datasets of paired MIDI-audio.\nAdditionally, ablation studies have provided insights into the scalability of\nthis approach and the forthcoming challenges in the field of AMT research.\n", "date": "2023-12-16", "categories": []}
{"id": "2406.15897", "title": "Fusing Audio and Metadata Embeddings Improves Language-based Audio\n  Retrieval", "abstract": "  Matching raw audio signals with textual descriptions requires understanding\nthe audio's content and the description's semantics and then drawing\nconnections between the two modalities. This paper investigates a hybrid\nretrieval system that utilizes audio metadata as an additional clue to\nunderstand the content of audio signals before matching them with textual\nqueries. We experimented with metadata often attached to audio recordings, such\nas keywords and natural-language descriptions, and we investigated late and\nmid-level fusion strategies to merge audio and metadata. Our hybrid approach\nwith keyword metadata and late fusion improved the retrieval performance over a\ncontent-based baseline by 2.36 and 3.69 pp. mAP@10 on the ClothoV2 and\nAudioCaps benchmarks, respectively.\n", "date": "2024-06-22", "categories": []}
{"id": "2407.01777", "title": "Deepfake Audio Detection Using Spectrogram-based Feature and Ensemble of\n  Deep Learning Models", "abstract": "  In this paper, we propose a deep learning based system for the task of\ndeepfake audio detection. In particular, the draw input audio is first\ntransformed into various spectrograms using three transformation methods of\nShort-time Fourier Transform (STFT), Constant-Q Transform (CQT), Wavelet\nTransform (WT) combined with different auditory-based filters of Mel,\nGammatone, linear filters (LF), and discrete cosine transform (DCT). Given the\nspectrograms, we evaluate a wide range of classification models based on three\ndeep learning approaches. The first approach is to train directly the\nspectrograms using our proposed baseline models of CNN-based model\n(CNN-baseline), RNN-based model (RNN-baseline), C-RNN model (C-RNN baseline).\nMeanwhile, the second approach is transfer learning from computer vision models\nsuch as ResNet-18, MobileNet-V3, EfficientNet-B0, DenseNet-121, SuffleNet-V2,\nSwint, Convnext-Tiny, GoogLeNet, MNASsnet, RegNet. In the third approach, we\nleverage the state-of-the-art audio pre-trained models of Whisper, Seamless,\nSpeechbrain, and Pyannote to extract audio embeddings from the input\nspectrograms. Then, the audio embeddings are explored by a Multilayer\nperceptron (MLP) model to detect the fake or real audio samples. Finally,\nhigh-performance deep learning models from these approaches are fused to\nachieve the best performance. We evaluated our proposed models on ASVspoof 2019\nbenchmark dataset. Our best ensemble model achieved an Equal Error Rate (EER)\nof 0.03, which is highly competitive to top-performing systems in the\nASVspoofing 2019 challenge. Experimental results also highlight the potential\nof selective spectrograms and deep learning approaches to enhance the task of\naudio deepfake detection.\n", "date": "2024-07-01", "categories": []}
{"id": "2405.02179", "title": "Training-Free Deepfake Voice Recognition by Leveraging Large-Scale\n  Pre-Trained Models", "abstract": "  Generalization is a main issue for current audio deepfake detectors, which\nstruggle to provide reliable results on out-of-distribution data. Given the\nspeed at which more and more accurate synthesis methods are developed, it is\nvery important to design techniques that work well also on data they were not\ntrained for. In this paper we study the potential of large-scale pre-trained\nmodels for audio deepfake detection, with special focus on generalization\nability. To this end, the detection problem is reformulated in a speaker\nverification framework and fake audios are exposed by the mismatch between the\nvoice sample under test and the voice of the claimed identity. With this\nparadigm, no fake speech sample is necessary in training, cutting off any link\nwith the generation method at the root, and ensuring full generalization\nability. Features are extracted by general-purpose large pre-trained models,\nwith no need for training or fine-tuning on specific fake detection or speaker\nverification datasets. At detection time only a limited set of voice fragments\nof the identity under test is required. Experiments on several datasets\nwidespread in the community show that detectors based on pre-trained models\nachieve excellent performance and show strong generalization ability, rivaling\nsupervised methods on in-distribution data and largely overcoming them on\nout-of-distribution data.\n", "date": "2024-05-03", "categories": []}
{"id": "2407.00657", "title": "Improving Real-Time Music Accompaniment Separation with MMDenseNet", "abstract": "  Music source separation aims to separate polyphonic music into different\ntypes of sources. Most existing methods focus on enhancing the quality of\nseparated results by using a larger model structure, rendering them unsuitable\nfor deployment on edge devices. Moreover, these methods may produce low-quality\noutput when the input duration is short, making them impractical for real-time\napplications. Therefore, the goal of this paper is to enhance a lightweight\nmodel, MMDenstNet, to strike a balance between separation quality and latency\nfor real-time applications. Different directions of improvement are explored or\nproposed in this paper, including complex ideal ratio mask, self-attention,\nband-merge-split method, and feature look back. Source-to-distortion ratio,\nreal-time factor, and optimal latency are employed to evaluate the performance.\nTo align with our application requirements, the evaluation process in this\npaper focuses on the separation performance of the accompaniment part.\nExperimental results demonstrate that our improvement achieves low real-time\nfactor and optimal latency while maintaining acceptable separation quality.\n", "date": "2024-06-30", "categories": []}
{"id": "2407.01494", "title": "FoleyCrafter: Bring Silent Videos to Life with Lifelike and Synchronized\n  Sounds", "abstract": "  We study Neural Foley, the automatic generation of high-quality sound effects\nsynchronizing with videos, enabling an immersive audio-visual experience.\nDespite its wide range of applications, existing approaches encounter\nlimitations when it comes to simultaneously synthesizing high-quality and\nvideo-aligned (i.e.,, semantic relevant and temporal synchronized) sounds. To\novercome these limitations, we propose FoleyCrafter, a novel framework that\nleverages a pre-trained text-to-audio model to ensure high-quality audio\ngeneration. FoleyCrafter comprises two key components: the semantic adapter for\nsemantic alignment and the temporal controller for precise audio-video\nsynchronization. The semantic adapter utilizes parallel cross-attention layers\nto condition audio generation on video features, producing realistic sound\neffects that are semantically relevant to the visual content. Meanwhile, the\ntemporal controller incorporates an onset detector and a timestampbased adapter\nto achieve precise audio-video alignment. One notable advantage of FoleyCrafter\nis its compatibility with text prompts, enabling the use of text descriptions\nto achieve controllable and diverse video-to-audio generation according to user\nintents. We conduct extensive quantitative and qualitative experiments on\nstandard benchmarks to verify the effectiveness of FoleyCrafter. Models and\ncodes are available at https://github.com/open-mmlab/FoleyCrafter.\n", "date": "2024-07-01", "categories": []}
{"id": "2401.05060", "title": "MuTox: Universal MUltilingual Audio-based TOXicity Dataset and Zero-shot\n  Detector", "abstract": "  Research in toxicity detection in natural language processing for the speech\nmodality (audio-based) is quite limited, particularly for languages other than\nEnglish. To address these limitations and lay the groundwork for truly\nmultilingual audio-based toxicity detection, we introduce MuTox, the first\nhighly multilingual audio-based dataset with toxicity labels. The dataset\ncomprises 20,000 audio utterances for English and Spanish, and 4,000 for the\nother 19 languages. To demonstrate the quality of this dataset, we trained the\nMuTox audio-based toxicity classifier, which enables zero-shot toxicity\ndetection across a wide range of languages. This classifier outperforms\nexisting text-based trainable classifiers by more than 1% AUC, while expanding\nthe language coverage more than tenfold. When compared to a wordlist-based\nclassifier that covers a similar number of languages, MuTox improves precision\nand recall by approximately 2.5 times. This significant improvement underscores\nthe potential of MuTox in advancing the field of audio-based toxicity\ndetection.\n", "date": "2024-01-10", "categories": []}
{"id": "2401.12440", "title": "Post-Training Embedding Alignment for Decoupling Enrollment and Runtime\n  Speaker Recognition Models", "abstract": "  Automated speaker identification (SID) is a crucial step for the\npersonalization of a wide range of speech-enabled services. Typical SID systems\nuse a symmetric enrollment-verification framework with a single model to derive\nembeddings both offline for voice profiles extracted from enrollment\nutterances, and online from runtime utterances. Due to the distinct\ncircumstances of enrollment and runtime, such as different computation and\nlatency constraints, several applications would benefit from an asymmetric\nenrollment-verification framework that uses different models for enrollment and\nruntime embedding generation. To support this asymmetric SID where each of the\ntwo models can be updated independently, we propose using a lightweight neural\nnetwork to map the embeddings from the two independent models to a shared\nspeaker embedding space. Our results show that this approach significantly\noutperforms cosine scoring in a shared speaker logit space for models that were\ntrained with a contrastive loss on large datasets with many speaker identities.\nThis proposed Neural Embedding Speaker Space Alignment (NESSA) combined with an\nasymmetric update of only one of the models delivers at least 60% of the\nperformance gain achieved by updating both models in the standard symmetric SID\napproach.\n", "date": "2024-01-23", "categories": []}
{"id": "2406.18679", "title": "Speakers Unembedded: Embedding-free Approach to Long-form Neural\n  Diarization", "abstract": "  End-to-end neural diarization (EEND) models offer significant improvements\nover traditional embedding-based Speaker Diarization (SD) approaches but falls\nshort on generalizing to long-form audio with large number of speakers.\nEEND-vector-clustering method mitigates this by combining local EEND with\nglobal clustering of speaker embeddings from local windows, but this requires\nan additional speaker embedding framework alongside the EEND module. In this\npaper, we propose a novel framework applying EEND both locally and globally for\nlong-form audio without separate speaker embeddings. This approach achieves\nsignificant relative DER reduction of 13% and 10% over the conventional 1-pass\nEEND on Callhome American English and RT03-CTS datasets respectively and\nmarginal improvements over EEND-vector-clustering without the need for\nadditional speaker embeddings. Furthermore, we discuss the computational\ncomplexity of our proposed framework and explore strategies for reducing\nprocessing times.\n", "date": "2024-06-26", "categories": []}
{"id": "2406.19243", "title": "Application of ASV for Voice Identification after VC and Duration\n  Predictor Improvement in TTS Models", "abstract": "  One of the most crucial components in the field of biometric security is the\nautomatic speaker verification system, which is based on the speaker's voice.\nIt is possible to utilise ASVs in isolation or in conjunction with other AI\nmodels. In the contemporary era, the quality and quantity of neural networks\nare increasing exponentially. Concurrently, there is a growing number of\nsystems that aim to manipulate data through the use of voice conversion and\ntext-to-speech models. The field of voice biometrics forgery is aided by a\nnumber of challenges, including SSTC, ASVSpoof, and SingFake.\n  This paper presents a system for automatic speaker verification. The primary\nobjective of our model is the extraction of embeddings from the target\nspeaker's audio in order to obtain information about important characteristics\nof his voice, such as pitch, energy, and the duration of phonemes. This\ninformation is used in our multivoice TTS pipeline, which is currently under\ndevelopment. However, this model was employed within the SSTC challenge to\nverify users whose voice had undergone voice conversion, where it demonstrated\nan EER of 20.669.\n", "date": "2024-06-27", "categories": []}
{"id": "2402.17775", "title": "WhaleNet: a Novel Deep Learning Architecture for Marine Mammals\n  Vocalizations on Watkins Marine Mammal Sound Database", "abstract": "  Marine mammal communication is a complex field, hindered by the diversity of\nvocalizations and environmental factors. The Watkins Marine Mammal Sound\nDatabase (WMMD) constitutes a comprehensive labeled dataset employed in machine\nlearning applications. Nevertheless, the methodologies for data preparation,\npreprocessing, and classification documented in the literature exhibit\nconsiderable variability and are typically not applied to the dataset in its\nentirety. This study initially undertakes a concise review of the\nstate-of-the-art benchmarks pertaining to the dataset, with a particular focus\non clarifying data preparation and preprocessing techniques. Subsequently, we\nexplore the utilization of the Wavelet Scattering Transform (WST) and Mel\nspectrogram as preprocessing mechanisms for feature extraction. In this paper,\nwe introduce \\textbf{WhaleNet} (Wavelet Highly Adaptive Learning Ensemble\nNetwork), a sophisticated deep ensemble architecture for the classification of\nmarine mammal vocalizations, leveraging both WST and Mel spectrogram for\nenhanced feature discrimination. By integrating the insights derived from WST\nand Mel representations, we achieved an improvement in classification accuracy\nby $8-10\\%$ over existing architectures, corresponding to a classification\naccuracy of $97.61\\%$.\n", "date": "2024-02-20", "categories": []}
{"id": "2406.13275", "title": "Enhancing Automated Audio Captioning via Large Language Models with\n  Optimized Audio Encoding", "abstract": "  Automated audio captioning (AAC) is an audio-to-text task to describe audio\ncontents in natural language. Recently, the advancements in large language\nmodels (LLMs), with improvements in training approaches for audio encoders,\nhave opened up possibilities for improving AAC. Thus, we explore enhancing AAC\nfrom three aspects: 1) a pre-trained audio encoder via consistent ensemble\ndistillation (CED) is used to improve the effectivity of acoustic tokens, with\na querying transformer (Q-Former) bridging the modality gap to LLM and compress\nacoustic tokens; 2) we investigate the advantages of using a Llama 2 with 7B\nparameters as the decoder; 3) another pre-trained LLM corrects text errors\ncaused by insufficient training data and annotation ambiguities. Both the audio\nencoder and text decoder are optimized by low-rank adaptation (LoRA).\nExperiments show that each of these enhancements is effective. Our method\nobtains a 33.0 SPIDEr-FL score, outperforming the winner of DCASE 2023 Task 6A.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.17190", "title": "Sound Tagging in Infant-centric Home Soundscapes", "abstract": "  Certain environmental noises have been associated with negative developmental\noutcomes for infants and young children. Though classifying or tagging sound\nevents in a domestic environment is an active research area, previous studies\nfocused on data collected from a non-stationary microphone placed in the\nenvironment or from the perspective of adults. Further, many of these works\nignore infants or young children in the environment or have data collected from\nonly a single family where noise from the fixed sound source can be moderate at\nthe infant's position or vice versa. Thus, despite the recent success of large\npre-trained models for noise event detection, the performance of these models\non infant-centric noise soundscapes in the home is yet to be explored. To\nbridge this gap, we have collected and labeled noises in home soundscapes from\n22 families in an unobtrusive manner, where the data are collected through an\ninfant-worn recording device. In this paper, we explore the performance of a\nlarge pre-trained model (Audio Spectrogram Transformer [AST]) on our\nnoise-conditioned infant-centric environmental data as well as publicly\navailable home environmental datasets. Utilizing different training strategies\nsuch as resampling, utilizing public datasets, mixing public and infant-centric\ntraining sets, and data augmentation using noise and masking, we evaluate the\nperformance of a large pre-trained model on sparse and imbalanced\ninfant-centric data. Our results show that fine-tuning the large pre-trained\nmodel by combining our collected dataset with public datasets increases the\nF1-score from 0.11 (public datasets) and 0.76 (collected datasets) to 0.84\n(combined datasets) and Cohen's Kappa from 0.013 (public datasets) and 0.77\n(collected datasets) to 0.83 (combined datasets) compared to only training with\npublic or collected datasets, respectively.\n", "date": "2024-06-25", "categories": []}
{"id": "2406.17732", "title": "EMVD dataset: a dataset of extreme vocal distortion techniques used in\n  heavy metal", "abstract": "  In this paper, we introduce the Extreme Metal Vocals Dataset, which comprises\na collection of recordings of extreme vocal techniques performed within the\nrealm of heavy metal music. The dataset consists of 760 audio excerpts of 1\nsecond to 30 seconds long, totaling about 100 min of audio material, roughly\ncomposed of 60 minutes of distorted voices and 40 minutes of clear voice\nrecordings. These vocal recordings are from 27 different singers and are\nprovided without accompanying musical instruments or post-processing effects.\nThe distortion taxonomy within this dataset encompasses four distinct\ndistortion techniques and three vocal effects, all performed in different pitch\nranges. Performance of a state-of-the-art deep learning model is evaluated for\ntwo different classification tasks related to vocal techniques, demonstrating\nthe potential of this resource for the audio processing community.\n", "date": "2024-06-24", "categories": []}
{"id": "2403.09598", "title": "Mixture of Mixups for Multi-label Classification of Rare Anuran Sounds", "abstract": "  Multi-label imbalanced classification poses a significant challenge in\nmachine learning, particularly evident in bioacoustics where animal sounds\noften co-occur, and certain sounds are much less frequent than others. This\npaper focuses on the specific case of classifying anuran species sounds using\nthe dataset AnuraSet, that contains both class imbalance and multi-label\nexamples. To address these challenges, we introduce Mixture of Mixups (Mix2), a\nframework that leverages mixing regularization methods Mixup, Manifold Mixup,\nand MultiMix. Experimental results show that these methods, individually, may\nlead to suboptimal results; however, when applied randomly, with one selected\nat each training iteration, they prove effective in addressing the mentioned\nchallenges, particularly for rare classes with few occurrences. Further\nanalysis reveals that Mix2 is also proficient in classifying sounds across\nvarious levels of class co-occurrences.\n", "date": "2024-03-14", "categories": []}
{"id": "2406.13431", "title": "Children's Speech Recognition through Discrete Token Enhancement", "abstract": "  Children's speech recognition is considered a low-resource task mainly due to\nthe lack of publicly available data. There are several reasons for such data\nscarcity, including expensive data collection and annotation processes, and\ndata privacy, among others. Transforming speech signals into discrete tokens\nthat do not carry sensitive information but capture both linguistic and\nacoustic information could be a solution for privacy concerns. In this study,\nwe investigate the integration of discrete speech tokens into children's speech\nrecognition systems as input without significantly degrading the ASR\nperformance. Additionally, we explored single-view and multi-view strategies\nfor creating these discrete labels. Furthermore, we tested the models for\ngeneralization capabilities with unseen domain and nativity dataset. Results\nreveal that the discrete token ASR for children achieves nearly equivalent\nperformance with an approximate 83% reduction in parameters.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.16315", "title": "Song Data Cleansing for End-to-End Neural Singer Diarization Using\n  Neural Analysis and Synthesis Framework", "abstract": "  We propose a data cleansing method that utilizes a neural analysis and\nsynthesis (NANSY++) framework to train an end-to-end neural diarization model\n(EEND) for singer diarization. Our proposed model converts song data with\nchoral singing which is commonly contained in popular music and unsuitable for\ngenerating a simulated dataset to the solo singing data. This cleansing is\nbased on NANSY++, which is a framework trained to reconstruct an input\nnon-overlapped audio signal. We exploit the pre-trained NANSY++ to convert\nchoral singing into clean, non-overlapped audio. This cleansing process\nmitigates the mislabeling of choral singing to solo singing and helps the\neffective training of EEND models even when the majority of available song data\ncontains choral singing sections. We experimentally evaluated the EEND model\ntrained with a dataset using our proposed method using annotated popular duet\nsongs. As a result, our proposed method improved 14.8 points in diarization\nerror rate.\n", "date": "2024-06-24", "categories": []}
{"id": "2309.04505", "title": "COVID-19 Detection System: A Comparative Analysis of System Performance\n  Based on Acoustic Features of Cough Audio Signals", "abstract": "  A wide range of respiratory diseases, such as cold and flu, asthma, and\nCOVID-19, affect people's daily lives worldwide. In medical practice,\nrespiratory sounds are widely used in medical services to diagnose various\nrespiratory illnesses and lung disorders. The traditional diagnosis of such\nsounds requires specialized knowledge, which can be costly and reliant on human\nexpertise. Despite this, recent advancements, such as cough audio recordings,\nhave emerged as a means to automate the detection of respiratory conditions.\nTherefore, this research aims to explore various acoustic features that enhance\nthe performance of machine learning (ML) models in detecting COVID-19 from\ncough signals. It investigates the efficacy of three feature extraction\ntechniques, including Mel Frequency Cepstral Coefficients (MFCC), Chroma, and\nSpectral Contrast features, when applied to two machine learning algorithms,\nSupport Vector Machine (SVM) and Multilayer Perceptron (MLP), and therefore\nproposes an efficient CovCepNet detection system. The proposed system provides\na practical solution and demonstrates state-of-the-art classification\nperformance, with an AUC of 0.843 on the COUGHVID dataset and 0.953 on the\nVirufy dataset for COVID-19 detection from cough audio signals.\n", "date": "2023-09-08", "categories": []}
{"id": "2403.13086", "title": "Listenable Maps for Audio Classifiers", "abstract": "  Despite the impressive performance of deep learning models across diverse\ntasks, their complexity poses challenges for interpretation. This challenge is\nparticularly evident for audio signals, where conveying interpretations becomes\ninherently difficult. To address this issue, we introduce Listenable Maps for\nAudio Classifiers (L-MAC), a posthoc interpretation method that generates\nfaithful and listenable interpretations. L-MAC utilizes a decoder on top of a\npretrained classifier to generate binary masks that highlight relevant portions\nof the input audio. We train the decoder with a loss function that maximizes\nthe confidence of the classifier decision on the masked-in portion of the audio\nwhile minimizing the probability of model output for the masked-out portion.\nQuantitative evaluations on both in-domain and out-of-domain data demonstrate\nthat L-MAC consistently produces more faithful interpretations than several\ngradient and masking-based methodologies. Furthermore, a user study confirms\nthat, on average, users prefer the interpretations generated by the proposed\ntechnique.\n", "date": "2024-03-19", "categories": []}
{"id": "2406.13385", "title": "Explainable by-design Audio Segmentation through Non-Negative Matrix\n  Factorization and Probing", "abstract": "  Audio segmentation is a key task for many speech technologies, most of which\nare based on neural networks, usually considered as black boxes, with\nhigh-level performances. However, in many domains, among which health or\nforensics, there is not only a need for good performance but also for\nexplanations about the output decision. Explanations derived directly from\nlatent representations need to satisfy \"good\" properties, such as\ninformativeness, compactness, or modularity, to be interpretable. In this\narticle, we propose an explainable-by-design audio segmentation model based on\nnon-negative matrix factorization (NMF) which is a good candidate for the\ndesign of interpretable representations. This paper shows that our model\nreaches good segmentation performances, and presents deep analyses of the\nlatent representation extracted from the non-negative matrix. The proposed\napproach opens new perspectives toward the evaluation of interpretable\nrepresentations according to \"good\" properties.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.13579", "title": "Automated Bioacoustic Monitoring for South African Bird Species on\n  Unlabeled Data", "abstract": "  Analyses for biodiversity monitoring based on passive acoustic monitoring\n(PAM) recordings is time-consuming and challenged by the presence of background\nnoise in recordings. Existing models for sound event detection (SED) worked\nonly on certain avian species and the development of further models required\nlabeled data. The developed framework automatically extracted labeled data from\navailable platforms for selected avian species. The labeled data were embedded\ninto recordings, including environmental sounds and noise, and were used to\ntrain convolutional recurrent neural network (CRNN) models. The models were\nevaluated on unprocessed real world data recorded in urban KwaZulu-Natal\nhabitats. The Adapted SED-CRNN model reached a F1 score of 0.73, demonstrating\nits efficiency under noisy, real-world conditions. The proposed approach to\nautomatically extract labeled data for chosen avian species enables an easy\nadaption of PAM to other species and habitats for future conservation projects.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.13935", "title": "CONMOD: Controllable Neural Frame-based Modulation Effects", "abstract": "  Deep learning models have seen widespread use in modelling LFO-driven audio\neffects, such as phaser and flanger. Although existing neural architectures\nexhibit high-quality emulation of individual effects, they do not possess the\ncapability to manipulate the output via control parameters. To address this\nissue, we introduce Controllable Neural Frame-based Modulation Effects\n(CONMOD), a single black-box model which emulates various LFO-driven effects in\na frame-wise manner, offering control over LFO frequency and feedback\nparameters. Additionally, the model is capable of learning the continuous\nembedding space of two distinct phaser effects, enabling us to steer between\neffects and achieve creative outputs. Our model outperforms previous work while\npossessing both controllability and universality, presenting opportunities to\nenhance creativity in modern LFO-driven audio effects.\n", "date": "2024-06-20", "categories": []}
{"id": "2303.07538", "title": "HiSSNet: Sound Event Detection and Speaker Identification via\n  Hierarchical Prototypical Networks for Low-Resource Headphones", "abstract": "  Modern noise-cancelling headphones have significantly improved users'\nauditory experiences by removing unwanted background noise, but they can also\nblock out sounds that matter to users. Machine learning (ML) models for sound\nevent detection (SED) and speaker identification (SID) can enable headphones to\nselectively pass through important sounds; however, implementing these models\nfor a user-centric experience presents several unique challenges. First, most\npeople spend limited time customizing their headphones, so the sound detection\nshould work reasonably well out of the box. Second, the models should be able\nto learn over time the specific sounds that are important to users based on\ntheir implicit and explicit interactions. Finally, such models should have a\nsmall memory footprint to run on low-power headphones with limited on-chip\nmemory. In this paper, we propose addressing these challenges using HiSSNet\n(Hierarchical SED and SID Network). HiSSNet is an SEID (SED and SID) model that\nuses a hierarchical prototypical network to detect both general and specific\nsounds of interest and characterize both alarm-like and speech sounds. We show\nthat HiSSNet outperforms an SEID model trained using non-hierarchical\nprototypical networks by 6.9 - 8.6 percent. When compared to state-of-the-art\n(SOTA) models trained specifically for SED or SID alone, HiSSNet achieves\nsimilar or better performance while reducing the memory footprint required to\nsupport multiple capabilities on-device.\n", "date": "2023-03-13", "categories": []}
{"id": "2310.15672", "title": "How Much Context Does My Attention-Based ASR System Need?", "abstract": "  For the task of speech recognition, the use of more than 30 seconds of\nacoustic context during training is uncommon and under-investigated in\nliterature. In this work, we conduct an empirical study on the effect of\nscaling the sequence length used to train/evaluate (dense-attention-based)\nacoustic models on speech recognition performance. For these experiments, a\ndataset of roughly 100,000 pseudo-labelled Spotify podcasts is used, with\ncontext lengths of 5 seconds to 1 hour being explored. Zero-shot evaluations\nare presented on the long-format datasets: Earnings-22, Tedlium and Rev16.\nResults demonstrate a benefit from training with up to 21.8 minutes of acoustic\ncontext, showing up to a 14.5\\% relative improvement from a baseline trained\nwith 10 seconds of context. We find that the model's width/depth, positional\nencoding scheme and number of attention heads impact its ability to use longer\ncontexts.\n", "date": "2023-10-24", "categories": []}
{"id": "2406.07012", "title": "Bridging Language Gaps in Audio-Text Retrieval", "abstract": "  Audio-text retrieval is a challenging task, requiring the search for an audio\nclip or a text caption within a database. The predominant focus of existing\nresearch on English descriptions poses a limitation on the applicability of\nsuch models, given the abundance of non-English content in real-world data. To\naddress these linguistic disparities, we propose a language enhancement (LE),\nusing a multilingual text encoder (SONAR) to encode the text data with\nlanguage-specific information. Additionally, we optimize the audio encoder\nthrough the application of consistent ensemble distillation (CED), enhancing\nsupport for variable-length audio-text retrieval. Our methodology excels in\nEnglish audio-text retrieval, demonstrating state-of-the-art (SOTA) performance\non commonly used datasets such as AudioCaps and Clotho. Simultaneously, the\napproach exhibits proficiency in retrieving content in seven other languages\nwith only 10% of additional language-enhanced training data, yielding promising\nresults. The source code is publicly available\nhttps://github.com/zyyan4/ml-clap.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.10507", "title": "Benchmarking Children's ASR with Supervised and Self-supervised Speech\n  Foundation Models", "abstract": "  Speech foundation models (SFMs) have achieved state-of-the-art results for\nvarious speech tasks in supervised (e.g. Whisper) or self-supervised systems\n(e.g. WavLM). However, the performance of SFMs for child ASR has not been\nsystematically studied. In addition, there is no benchmark for child ASR with\nstandard evaluations, making the comparisons of novel ideas difficult. In this\npaper, we initiate and present a comprehensive benchmark on several child\nspeech databases based on various SFMs (Whisper, Wav2vec2.0, HuBERT, and\nWavLM). Moreover, we investigate finetuning strategies by comparing various\ndata augmentation and parameter-efficient finetuning (PEFT) methods. We observe\nthat the behaviors of these methods are different when the model size\nincreases. For example, PEFT matches the performance of full finetuning for\nlarge models but worse for small models. To stabilize finetuning using\naugmented data, we propose a perturbation invariant finetuning (PIF) loss as a\nregularization.\n", "date": "2024-06-15", "categories": []}
{"id": "2406.10591", "title": "MINT: a Multi-modal Image and Narrative Text Dubbing Dataset for Foley\n  Audio Content Planning and Generation", "abstract": "  Foley audio, critical for enhancing the immersive experience in multimedia\ncontent, faces significant challenges in the AI-generated content (AIGC)\nlandscape. Despite advancements in AIGC technologies for text and image\ngeneration, the foley audio dubbing remains rudimentary due to difficulties in\ncross-modal scene matching and content correlation. Current text-to-audio\ntechnology, which relies on detailed and acoustically relevant textual\ndescriptions, falls short in practical video dubbing applications. Existing\ndatasets like AudioSet, AudioCaps, Clotho, Sound-of-Story, and WavCaps do not\nfully meet the requirements for real-world foley audio dubbing task. To address\nthis, we introduce the Multi-modal Image and Narrative Text Dubbing Dataset\n(MINT), designed to enhance mainstream dubbing tasks such as literary story\naudiobooks dubbing, image/silent video dubbing. Besides, to address the\nlimitations of existing TTA technology in understanding and planning complex\nprompts, a Foley Audio Content Planning, Generation, and Alignment (CPGA)\nframework is proposed, which includes a content planning module leveraging\nlarge language models for complex multi-modal prompts comprehension.\nAdditionally, the training process is optimized using Proximal Policy\nOptimization based reinforcement learning, significantly improving the\nalignment and auditory realism of generated foley audio. Experimental results\ndemonstrate that our approach significantly advances the field of foley audio\ndubbing, providing robust solutions for the challenges of multi-modal dubbing.\nEven when utilizing the relatively lightweight GPT-2 model, our framework\noutperforms open-source multimodal large models such as LLaVA, DeepSeek-VL, and\nMoondream2. The dataset is available at https://github.com/borisfrb/MINT .\n", "date": "2024-06-15", "categories": []}
{"id": "2406.10735", "title": "How Should We Extract Discrete Audio Tokens from Self-Supervised Models?", "abstract": "  Discrete audio tokens have recently gained attention for their potential to\nbridge the gap between audio and language processing. Ideal audio tokens must\npreserve content, paralinguistic elements, speaker identity, and many other\naudio details. Current audio tokenization methods fall into two categories:\nSemantic tokens, acquired through quantization of Self-Supervised Learning\n(SSL) models, and Neural compression-based tokens (codecs). Although previous\nstudies have benchmarked codec models to identify optimal configurations, the\nideal setup for quantizing pretrained SSL models remains unclear. This paper\nexplores the optimal configuration of semantic tokens across discriminative and\ngenerative tasks. We propose a scalable solution to train a universal vocoder\nacross multiple SSL layers. Furthermore, an attention mechanism is employed to\nidentify task-specific influential layers, enhancing the adaptability and\nperformance of semantic tokens in diverse audio applications.\n", "date": "2024-06-15", "categories": []}
{"id": "2406.10741", "title": "Speech Emotion Recognition Using CNN and Its Use Case in Digital\n  Healthcare", "abstract": "  The process of identifying human emotion and affective states from speech is\nknown as speech emotion recognition (SER). This is based on the observation\nthat tone and pitch in the voice frequently convey underlying emotion. Speech\nrecognition includes the ability to recognize emotions, which is becoming\nincreasingly popular and in high demand. With the help of appropriate factors\n(such modalities, emotions, intensities, repetitions, etc.) found in the data,\nmy research seeks to use the Convolutional Neural Network (CNN) to distinguish\nemotions from audio recordings and label them in accordance with the range of\ndifferent emotions. I have developed a machine learning model to identify\nemotions from supplied audio files with the aid of machine learning methods.\nThe evaluation is mostly focused on precision, recall, and F1 score, which are\ncommon machine learning metrics. To properly set up and train the machine\nlearning framework, the main objective is to investigate the influence and\ncross-relation of all input and output parameters. To improve the ability to\nrecognize intentions, a key condition for communication, I have evaluated\nemotions using my specialized machine learning algorithm via voice that would\naddress the emotional state from voice with the help of digital healthcare,\nbridging the gap between human and artificial intelligence (AI).\n", "date": "2024-06-15", "categories": []}
{"id": "2406.10873", "title": "Optimizing Automatic Speech Assessment: W-RankSim Regularization and\n  Hybrid Feature Fusion Strategies", "abstract": "  Automatic Speech Assessment (ASA) has seen notable advancements with the\nutilization of self-supervised features (SSL) in recent research. However, a\nkey challenge in ASA lies in the imbalanced distribution of data, particularly\nevident in English test datasets. To address this challenge, we approach ASA as\nan ordinal classification task, introducing Weighted Vectors Ranking Similarity\n(W-RankSim) as a novel regularization technique. W-RankSim encourages closer\nproximity of weighted vectors in the output layer for similar classes, implying\nthat feature vectors with similar labels would be gradually nudged closer to\neach other as they converge towards corresponding weighted vectors. Extensive\nexperimental evaluations confirm the effectiveness of our approach in improving\nordinal classification performance for ASA. Furthermore, we propose a hybrid\nmodel that combines SSL and handcrafted features, showcasing how the inclusion\nof handcrafted features enhances performance in an ASA system.\n", "date": "2024-06-16", "categories": []}
{"id": "2406.11006", "title": "SPEAR: Receiver-to-Receiver Acoustic Neural Warping Field", "abstract": "  We present SPEAR, a continuous receiver-to-receiver acoustic neural warping\nfield for spatial acoustic effects prediction in an acoustic 3D space with a\nsingle stationary audio source. Unlike traditional source-to-receiver modelling\nmethods that require prior space acoustic properties knowledge to rigorously\nmodel audio propagation from source to receiver, we propose to predict by\nwarping the spatial acoustic effects from one reference receiver position to\nanother target receiver position, so that the warped audio essentially\naccommodates all spatial acoustic effects belonging to the target position.\nSPEAR can be trained in a data much more readily accessible manner, in which we\nsimply ask two robots to independently record spatial audio at different\npositions. We further theoretically prove the universal existence of the\nwarping field if and only if one audio source presents. Three physical\nprinciples are incorporated to guide SPEAR network design, leading to the\nlearned warping field physically meaningful. We demonstrate SPEAR superiority\non both synthetic, photo-realistic and real-world dataset, showing the huge\npotential of SPEAR to various down-stream robotic tasks.\n", "date": "2024-06-16", "categories": []}
{"id": "2406.11619", "title": "AV-CrossNet: an Audiovisual Complex Spectral Mapping Network for Speech\n  Separation By Leveraging Narrow- and Cross-Band Modeling", "abstract": "  Adding visual cues to audio-based speech separation can improve separation\nperformance. This paper introduces AV-CrossNet, an audiovisual (AV) system for\nspeech enhancement, target speaker extraction, and multi-talker speaker\nseparation. AV-CrossNet is extended from the CrossNet architecture, which is a\nrecently proposed network that performs complex spectral mapping for speech\nseparation by leveraging global attention and positional encoding. To\neffectively utilize visual cues, the proposed system incorporates pre-extracted\nvisual embeddings and employs a visual encoder comprising temporal\nconvolutional layers. Audio and visual features are fused in an early fusion\nlayer before feeding to AV-CrossNet blocks. We evaluate AV-CrossNet on multiple\ndatasets, including LRS, VoxCeleb, and COG-MHEAR challenge. Evaluation results\ndemonstrate that AV-CrossNet advances the state-of-the-art performance in all\naudiovisual tasks, even on untrained and mismatched datasets.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.11768", "title": "GAMA: A Large Audio-Language Model with Advanced Audio Understanding and\n  Complex Reasoning Abilities", "abstract": "  Perceiving and understanding non-speech sounds and non-verbal speech is\nessential to making decisions that help us interact with our surroundings. In\nthis paper, we propose GAMA, a novel General-purpose Large Audio-Language Model\n(LALM) with Advanced Audio Understanding and Complex Reasoning Abilities. We\nbuild GAMA by integrating an LLM with multiple types of audio representations,\nincluding features from a custom Audio Q-Former, a multi-layer aggregator that\naggregates features from multiple layers of an audio encoder. We fine-tune GAMA\non a large-scale audio-language dataset, which augments it with audio\nunderstanding capabilities. Next, we propose CompA-R (Instruction-Tuning for\nComplex Audio Reasoning), a synthetically generated instruction-tuning (IT)\ndataset with instructions that require the model to perform complex reasoning\non the input audio. We instruction-tune GAMA with CompA-R to endow it with\ncomplex reasoning abilities, where we further add a soft prompt as input with\nhigh-level semantic evidence by leveraging event tags of the input audio.\nFinally, we also propose CompA-R-test, a human-labeled evaluation dataset for\nevaluating the capabilities of LALMs on open-ended audio question-answering\nthat requires complex reasoning. Through automated and expert human\nevaluations, we show that GAMA outperforms all other LALMs in literature on\ndiverse audio understanding tasks by margins of 1%-84%. Further, GAMA IT-ed on\nCompA-R proves to be superior in its complex reasoning and instruction\nfollowing capabilities.\n", "date": "2024-06-17", "categories": []}
{"id": "2309.09329", "title": "A Few-Shot Approach to Dysarthric Speech Intelligibility Level\n  Classification Using Transformers", "abstract": "  Dysarthria is a speech disorder that hinders communication due to\ndifficulties in articulating words. Detection of dysarthria is important for\nseveral reasons as it can be used to develop a treatment plan and help improve\na person's quality of life and ability to communicate effectively. Much of the\nliterature focused on improving ASR systems for dysarthric speech. The\nobjective of the current work is to develop models that can accurately classify\nthe presence of dysarthria and also give information about the intelligibility\nlevel using limited data by employing a few-shot approach using a transformer\nmodel. This work also aims to tackle the data leakage that is present in\nprevious studies. Our whisper-large-v2 transformer model trained on a subset of\nthe UASpeech dataset containing medium intelligibility level patients achieved\nan accuracy of 85%, precision of 0.92, recall of 0.8 F1-score of 0.85, and\nspecificity of 0.91. Experimental results also demonstrate that the model\ntrained using the 'words' dataset performed better compared to the model\ntrained on the 'letters' and 'digits' dataset. Moreover, the multiclass model\nachieved an accuracy of 67%.\n", "date": "2023-09-17", "categories": []}
{"id": "2403.16973", "title": "VoiceCraft: Zero-Shot Speech Editing and Text-to-Speech in the Wild", "abstract": "  We introduce VoiceCraft, a token infilling neural codec language model, that\nachieves state-of-the-art performance on both speech editing and zero-shot\ntext-to-speech (TTS) on audiobooks, internet videos, and podcasts. VoiceCraft\nemploys a Transformer decoder architecture and introduces a token rearrangement\nprocedure that combines causal masking and delayed stacking to enable\ngeneration within an existing sequence. On speech editing tasks, VoiceCraft\nproduces edited speech that is nearly indistinguishable from unedited\nrecordings in terms of naturalness, as evaluated by humans; for zero-shot TTS,\nour model outperforms prior SotA models including VALLE and the popular\ncommercial model XTTS-v2. Crucially, the models are evaluated on challenging\nand realistic datasets, that consist of diverse accents, speaking styles,\nrecording conditions, and background noise and music, and our model performs\nconsistently well compared to other models and real recordings. In particular,\nfor speech editing evaluation, we introduce a high quality, challenging, and\nrealistic dataset named RealEdit. We encourage readers to listen to the demos\nat https://jasonppy.github.io/VoiceCraft_web.\n", "date": "2024-03-25", "categories": []}
{"id": "2406.06786", "title": "BTS: Bridging Text and Sound Modalities for Metadata-Aided Respiratory\n  Sound Classification", "abstract": "  Respiratory sound classification (RSC) is challenging due to varied acoustic\nsignatures, primarily influenced by patient demographics and recording\nenvironments. To address this issue, we introduce a text-audio multimodal model\nthat utilizes metadata of respiratory sounds, which provides useful\ncomplementary information for RSC. Specifically, we fine-tune a pretrained\ntext-audio multimodal model using free-text descriptions derived from the sound\nsamples' metadata which includes the gender and age of patients, type of\nrecording devices, and recording location on the patient's body. Our method\nachieves state-of-the-art performance on the ICBHI dataset, surpassing the\nprevious best result by a notable margin of 1.17%. This result validates the\neffectiveness of leveraging metadata and respiratory sound samples in enhancing\nRSC performance. Additionally, we investigate the model performance in the case\nwhere metadata is partially unavailable, which may occur in real-world clinical\nsetting.\n", "date": "2024-06-10", "categories": []}
{"id": "2309.09548", "title": "Non-Intrusive Speech Intelligibility Prediction for Hearing Aids using\n  Whisper and Metadata", "abstract": "  Automated speech intelligibility assessment is pivotal for hearing aid (HA)\ndevelopment. In this paper, we present three novel methods to improve\nintelligibility prediction accuracy and introduce MBI-Net+, an enhanced version\nof MBI-Net, the top-performing system in the 1st Clarity Prediction Challenge.\nMBI-Net+ leverages Whisper's embeddings to create cross-domain acoustic\nfeatures and includes metadata from speech signals by using a classifier that\ndistinguishes different enhancement methods. Furthermore, MBI-Net+ integrates\nthe hearing-aid speech perception index (HASPI) as a supplementary metric into\nthe objective function to further boost prediction performance. Experimental\nresults demonstrate that MBI-Net+ surpasses several intrusive baseline systems\nand MBI-Net on the Clarity Prediction Challenge 2023 dataset, validating the\neffectiveness of incorporating Whisper embeddings, speech metadata, and related\ncomplementary metrics to improve prediction performance for HA.\n", "date": "2023-09-18", "categories": []}
{"id": "2406.09154", "title": "Diffusion Gaussian Mixture Audio Denoise", "abstract": "  Recent diffusion models have achieved promising performances in\naudio-denoising tasks. The unique property of the reverse process could recover\nclean signals. However, the distribution of real-world noises does not comply\nwith a single Gaussian distribution and is even unknown. The sampling of\nGaussian noise conditions limits its application scenarios. To overcome these\nchallenges, we propose a DiffGMM model, a denoising model based on the\ndiffusion and Gaussian mixture models. We employ the reverse process to\nestimate parameters for the Gaussian mixture model. Given a noisy audio signal,\nwe first apply a 1D-U-Net to extract features and train linear layers to\nestimate parameters for the Gaussian mixture model, and we approximate the real\nnoise distributions. The noisy signal is continuously subtracted from the\nestimated noise to output clean audio signals. Extensive experimental results\ndemonstrate that the proposed DiffGMM model achieves state-of-the-art\nperformance.\n", "date": "2024-06-13", "categories": []}
{"id": "2403.11827", "title": "Sound Event Detection and Localization with Distance Estimation", "abstract": "  Sound Event Detection and Localization (SELD) is a combined task of\nidentifying sound events and their corresponding direction-of-arrival (DOA).\nWhile this task has numerous applications and has been extensively researched\nin recent years, it fails to provide full information about the sound source\nposition. In this paper, we overcome this problem by extending the task to\nSound Event Detection, Localization with Distance Estimation (3D SELD). We\nstudy two ways of integrating distance estimation within the SELD core - a\nmulti-task approach, in which the problem is tackled by a separate model\noutput, and a single-task approach obtained by extending the multi-ACCDOA\nmethod to include distance information. We investigate both methods for the\nAmbisonic and binaural versions of STARSS23: Sony-TAU Realistic Spatial\nSoundscapes 2023. Moreover, our study involves experiments on the loss function\nrelated to the distance estimation part. Our results show that it is possible\nto perform 3D SELD without any degradation of performance in sound event\ndetection and DOA estimation.\n", "date": "2024-03-18", "categories": []}
{"id": "2406.03512", "title": "Harder or Different? Understanding Generalization of Audio Deepfake\n  Detection", "abstract": "  Recent research has highlighted a key issue in speech deepfake detection:\nmodels trained on one set of deepfakes perform poorly on others. The question\narises: is this due to the continuously improving quality of Text-to-Speech\n(TTS) models, i.e., are newer DeepFakes just 'harder' to detect? Or, is it\nbecause deepfakes generated with one model are fundamentally different to those\ngenerated using another model? We answer this question by decomposing the\nperformance gap between in-domain and out-of-domain test data into 'hardness'\nand 'difference' components. Experiments performed using ASVspoof databases\nindicate that the hardness component is practically negligible, with the\nperformance gap being attributed primarily to the difference component. This\nhas direct implications for real-world deepfake detection, highlighting that\nmerely increasing model capacity, the currently-dominant research trend, may\nnot effectively address the generalization challenge.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.07676", "title": "FastAST: Accelerating Audio Spectrogram Transformer via Token Merging\n  and Cross-Model Knowledge Distillation", "abstract": "  Audio classification models, particularly the Audio Spectrogram Transformer\n(AST), play a crucial role in efficient audio analysis. However, optimizing\ntheir efficiency without compromising accuracy remains a challenge. In this\npaper, we introduce FastAST, a framework that integrates Token Merging (ToMe)\ninto the AST framework. FastAST enhances inference speed without requiring\nextensive retraining by merging similar tokens in audio spectrograms.\nFurthermore, during training, FastAST brings about significant speed\nimprovements. The experiments indicate that FastAST can increase audio\nclassification throughput with minimal impact on accuracy. To mitigate the\naccuracy impact, we integrate Cross-Model Knowledge Distillation (CMKD) into\nthe FastAST framework. Integrating ToMe and CMKD into AST results in improved\naccuracy compared to AST while maintaining faster inference speeds. FastAST\nrepresents a step towards real-time, resource-efficient audio analysis.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.07816", "title": "Spoof Diarization: \"What Spoofed When\" in Partially Spoofed Audio", "abstract": "  This paper defines Spoof Diarization as a novel task in the Partial Spoof\n(PS) scenario. It aims to determine what spoofed when, which includes not only\nlocating spoof regions but also clustering them according to different spoofing\nmethods. As a pioneering study in spoof diarization, we focus on defining the\ntask, establishing evaluation metrics, and proposing a benchmark model, namely\nthe Countermeasure-Condition Clustering (3C) model. Utilizing this model, we\nfirst explore how to effectively train countermeasures to support spoof\ndiarization using three labeling schemes. We then utilize spoof localization\npredictions to enhance the diarization performance. This first study reveals\nthe high complexity of the task, even in restricted scenarios where only a\nsingle speaker per audio file and an oracle number of spoofing methods are\nconsidered. Our code is available at\nhttps://github.com/nii-yamagishilab/PartialSpoof.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.08112", "title": "Codecfake: An Initial Dataset for Detecting LLM-based Deepfake Audio", "abstract": "  With the proliferation of Large Language Model (LLM) based deepfake audio,\nthere is an urgent need for effective detection methods. Previous deepfake\naudio generation methods typically involve a multi-step generation process,\nwith the final step using a vocoder to predict the waveform from handcrafted\nfeatures. However, LLM-based audio is directly generated from discrete neural\ncodecs in an end-to-end generation process, skipping the final step of vocoder\nprocessing. This poses a significant challenge for current audio deepfake\ndetection (ADD) models based on vocoder artifacts. To effectively detect\nLLM-based deepfake audio, we focus on the core of the generation process, the\nconversion from neural codec to waveform. We propose Codecfake dataset, which\nis generated by seven representative neural codec methods. Experiment results\nshow that codec-trained ADD models exhibit a 41.406% reduction in average equal\nerror rate compared to vocoder-trained ADD models on the Codecfake test set.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.08402", "title": "Understanding Sounds, Missing the Questions: The Challenge of Object\n  Hallucination in Large Audio-Language Models", "abstract": "  Large audio-language models (LALMs) enhance traditional large language models\nby integrating audio perception capabilities, allowing them to tackle\naudio-related tasks. Previous research has primarily focused on assessing the\nperformance of LALMs across various tasks, yet overlooking their reliability,\nparticularly concerning issues like object hallucination. In our study, we\nintroduce methods to assess the extent of object hallucination of publicly\navailable LALMs. Our findings reveal that LALMs are comparable to specialized\naudio captioning models in their understanding of audio content, but struggle\nto answer discriminative questions, specifically those requiring the\nidentification of the presence of particular object sounds within an audio\nclip. This limitation highlights a critical weakness in current LALMs: their\ninadequate understanding of discriminative queries. Moreover, we explore the\npotential of prompt engineering to enhance LALMs' performance on discriminative\nquestions.\n", "date": "2024-06-12", "categories": []}
{"id": "2309.08398", "title": "Exploring Meta Information for Audio-based Zero-shot Bird Classification", "abstract": "  Advances in passive acoustic monitoring and machine learning have led to the\nprocurement of vast datasets for computational bioacoustic research.\nNevertheless, data scarcity is still an issue for rare and underrepresented\nspecies. This study investigates how meta-information can improve zero-shot\naudio classification, utilising bird species as an example case study due to\nthe availability of rich and diverse meta-data. We investigate three different\nsources of metadata: textual bird sound descriptions encoded via (S)BERT,\nfunctional traits (AVONET), and bird life-history (BLH) characteristics. As\naudio features, we extract audio spectrogram transformer (AST) embeddings and\nproject them to the dimension of the auxiliary information by adopting a single\nlinear layer. Then, we employ the dot product as compatibility function and a\nstandard zero-shot learning ranking hinge loss to determine the correct class.\nThe best results are achieved by concatenating the AVONET and BLH features\nattaining a mean unweighted F1-score of .233 over five different test sets with\n8 to 10 classes.\n", "date": "2023-09-15", "categories": []}
{"id": "2401.02099", "title": "Oceanship: A Large-Scale Dataset for Underwater Audio Target Recognition", "abstract": "  The recognition of underwater audio plays a significant role in identifying a\nvessel while it is in motion. Underwater target recognition tasks have a wide\nrange of applications in areas such as marine environmental protection,\ndetection of ship radiated noise, underwater noise control, and coastal vessel\ndispatch. The traditional UATR task involves training a network to extract\nfeatures from audio data and predict the vessel type. The current UATR dataset\nexhibits shortcomings in both duration and sample quantity. In this paper, we\npropose Oceanship, a large-scale and diverse underwater audio dataset. This\ndataset comprises 15 categories, spans a total duration of 121 hours, and\nincludes comprehensive annotation information such as coordinates, velocity,\nvessel types, and timestamps. We compiled the dataset by crawling and\norganizing original communication data from the Ocean Communication Network\n(ONC) database between 2021 and 2022. While audio retrieval tasks are\nwell-established in general audio classification, they have not been explored\nin the context of underwater audio recognition. Leveraging the Oceanship\ndataset, we introduce a baseline model named Oceannet for underwater audio\nretrieval. This model achieves a recall at 1 (R@1) accuracy of 67.11% and a\nrecall at 5 (R@5) accuracy of 99.13% on the Deepship dataset.\n", "date": "2024-01-04", "categories": []}
{"id": "2402.04866", "title": "Room Transfer Function Reconstruction Using Complex-valued Neural\n  Networks and Irregularly Distributed Microphones", "abstract": "  Reconstructing the room transfer functions needed to calculate the complex\nsound field in a room has several important real-world applications. However,\nan unpractical number of microphones is often required. Recently, in addition\nto classical signal processing methods, deep learning techniques have been\napplied to reconstruct the room transfer function starting from a very limited\nset of measurements at scattered points in the room. In this paper, we employ\ncomplex-valued neural networks to estimate room transfer functions in the\nfrequency range of the first room resonances, using a few irregularly\ndistributed microphones. To the best of our knowledge, this is the first time\nthat complex-valued neural networks are used to estimate room transfer\nfunctions. To analyze the benefits of applying complex-valued optimization to\nthe considered task, we compare the proposed technique with a state-of-the-art\nkernel-based signal processing approach for sound field reconstruction, showing\nthat the proposed technique exhibits relevant advantages in terms of phase\naccuracy and overall quality of the reconstructed sound field. For informative\npurposes, we also compare the model with a similarly-structured data-driven\napproach that, however, applies a real-valued neural network to reconstruct\nonly the magnitude of the sound field.\n", "date": "2024-02-01", "categories": []}
{"id": "2406.07250", "title": "Description and Discussion on DCASE 2024 Challenge Task 2: First-Shot\n  Unsupervised Anomalous Sound Detection for Machine Condition Monitoring", "abstract": "  We present the task description of the Detection and Classification of\nAcoustic Scenes and Events (DCASE) 2024 Challenge Task 2: First-shot\nunsupervised anomalous sound detection (ASD) for machine condition monitoring.\nContinuing from last year's DCASE 2023 Challenge Task 2, we organize the task\nas a first-shot problem under domain generalization required settings. The main\ngoal of the first-shot problem is to enable rapid deployment of ASD systems for\nnew kinds of machines without the need for machine-specific hyperparameter\ntunings. This problem setting was realized by (1) giving only one section for\neach machine type and (2) having completely different machine types for the\ndevelopment and evaluation datasets. For the DCASE 2024 Challenge Task 2, data\nof completely new machine types were newly collected and provided as the\nevaluation dataset. In addition, attribute information such as the machine\noperation conditions were concealed for several machine types to mimic\nsituations where such information are unavailable. We will add challenge\nresults and analysis of the submissions after the challenge submission\ndeadline.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.02178", "title": "Audio Mamba: Selective State Spaces for Self-Supervised Audio\n  Representations", "abstract": "  Despite its widespread adoption as the prominent neural architecture, the\nTransformer has spurred several independent lines of work to address its\nlimitations. One such approach is selective state space models, which have\ndemonstrated promising results for language modelling. However, their\nfeasibility for learning self-supervised, general-purpose audio representations\nis yet to be investigated. This work proposes Audio Mamba, a selective state\nspace model for learning general-purpose audio representations from randomly\nmasked spectrogram patches through self-supervision. Empirical results on ten\ndiverse audio recognition downstream tasks show that the proposed models,\npretrained on the AudioSet dataset, consistently outperform comparable\nself-supervised audio spectrogram transformer (SSAST) baselines by a\nconsiderable margin and demonstrate better performance in dataset size,\nsequence length and model size comparisons.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.03240", "title": "Generalized Source Tracing: Detecting Novel Audio Deepfake Algorithm\n  with Real Emphasis and Fake Dispersion Strategy", "abstract": "  With the proliferation of deepfake audio, there is an urgent need to\ninvestigate their attribution. Current source tracing methods can effectively\ndistinguish in-distribution (ID) categories. However, the rapid evolution of\ndeepfake algorithms poses a critical challenge in the accurate identification\nof out-of-distribution (OOD) novel deepfake algorithms. In this paper, we\npropose Real Emphasis and Fake Dispersion (REFD) strategy for audio deepfake\nalgorithm recognition, demonstrating its effectiveness in discriminating ID\nsamples while identifying OOD samples. For effective OOD detection, we first\nexplore current post-hoc OOD methods and propose NSD, a novel OOD approach in\nidentifying novel deepfake algorithms through the similarity consideration of\nboth feature and logits scores. REFD achieves 86.83% F1-score as a single\nsystem in Audio Deepfake Detection Challenge 2023 Track3, showcasing its\nstate-of-the-art performance.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.05547", "title": "Exploring the Benefits of Tokenization of Discrete Acoustic Units", "abstract": "  Tokenization algorithms that merge the units of a base vocabulary into\nlarger, variable-rate units have become standard in natural language processing\ntasks. This idea, however, has been mostly overlooked when the vocabulary\nconsists of phonemes or Discrete Acoustic Units (DAUs), an audio-based\nrepresentation that is playing an increasingly important role due to the\nsuccess of discrete language-modeling techniques. In this paper, we showcase\nthe advantages of tokenization of phonetic units and of DAUs on three\nprediction tasks: grapheme-to-phoneme, grapheme-to-DAUs, and unsupervised\nspeech generation using DAU language modeling. We demonstrate that tokenization\nyields significant improvements in terms of performance, as well as training\nand inference speed, across all three tasks. We also offer theoretical insights\nto provide some explanation for the superior performance observed.\n", "date": "2024-06-08", "categories": []}
{"id": "2404.04947", "title": "Gull: A Generative Multifunctional Audio Codec", "abstract": "  We introduce Gull, a generative multifunctional audio codec. Gull is a\ngeneral purpose neural audio compression and decompression model which can be\napplied to a wide range of tasks and applications such as real-time\ncommunication, audio super-resolution, and codec language models. The key\ncomponents of Gull include (1) universal-sample-rate modeling via subband\nmodeling schemes motivated by recent progress in audio source separation, (2)\ngain-shape representations motivated by traditional audio codecs, (3) improved\nresidual vector quantization modules, (4) elastic decoder network that enables\nuser-defined model size and complexity during inference time, (5) built-in\nability for audio super-resolution without the increase of bitrate. We compare\nGull with existing traditional and neural audio codecs and show that Gull is\nable to achieve on par or better performance across various sample rates,\nbitrates and model complexities in both subjective and objective evaluation\nmetrics.\n", "date": "2024-04-07", "categories": []}
{"id": "2405.07442", "title": "Rene: A Pre-trained Multi-modal Architecture for Auscultation of\n  Respiratory Diseases", "abstract": "  Compared with invasive examinations that require tissue sampling, respiratory\nsound testing is a non-invasive examination method that is safer and easier for\npatients to accept. In this study, we introduce Rene, a pioneering large-scale\nmodel tailored for respiratory sound recognition. Rene has been rigorously\nfine-tuned with an extensive dataset featuring a broad array of respiratory\naudio samples, targeting disease detection, sound pattern classification, and\nevent identification. Our innovative approach applies a pre-trained speech\nrecognition model to process respiratory sounds, augmented with patient medical\nrecords. The resulting multi-modal deep-learning framework addresses\ninterpretability and real-time diagnostic challenges that have hindered\nprevious respiratory-focused models. Benchmark comparisons reveal that Rene\nsignificantly outperforms existing models, achieving improvements of 10.27%,\n16.15%, 15.29%, and 18.90% in respiratory event detection and audio\nclassification on the SPRSound database. Disease prediction accuracy on the\nICBHI database improved by 23% over the baseline in both mean average and\nharmonic scores. Moreover, we have developed a real-time respiratory sound\ndiscrimination system utilizing the Rene architecture. Employing\nstate-of-the-art Edge AI technology, this system enables rapid and accurate\nresponses for respiratory sound\nauscultation(https://github.com/zpforlove/Rene).\n", "date": "2024-05-13", "categories": []}
{"id": "2405.11093", "title": "AudioSetMix: Enhancing Audio-Language Datasets with LLM-Assisted\n  Augmentations", "abstract": "  Multi-modal learning in the audio-language domain has seen significant\nadvancements in recent years. However, audio-language learning faces challenges\ndue to limited and lower-quality data compared to image-language tasks.\nExisting audio-language datasets are notably smaller, and manual labeling is\nhindered by the need to listen to entire audio clips for accurate labeling.\n  Our method systematically generates audio-caption pairs by augmenting audio\nclips with natural language labels and corresponding audio signal processing\noperations. Leveraging a Large Language Model, we generate descriptions of\naugmented audio clips with a prompt template. This scalable method produces\nAudioSetMix, a high-quality training dataset for text-and-audio related models.\n  Integration of our dataset improves models performance on benchmarks by\nproviding diversified and better-aligned examples. Notably, our dataset\naddresses the absence of modifiers (adjectives and adverbs) in existing\ndatasets. By enabling models to learn these concepts, and generating hard\nnegative examples during training, we achieve state-of-the-art performance on\nmultiple benchmarks.\n", "date": "2024-05-17", "categories": []}
{"id": "2309.09836", "title": "RECAP: Retrieval-Augmented Audio Captioning", "abstract": "  We present RECAP (REtrieval-Augmented Audio CAPtioning), a novel and\neffective audio captioning system that generates captions conditioned on an\ninput audio and other captions similar to the audio retrieved from a datastore.\nAdditionally, our proposed method can transfer to any domain without the need\nfor any additional fine-tuning. To generate a caption for an audio sample, we\nleverage an audio-text model CLAP to retrieve captions similar to it from a\nreplaceable datastore, which are then used to construct a prompt. Next, we feed\nthis prompt to a GPT-2 decoder and introduce cross-attention layers between the\nCLAP encoder and GPT-2 to condition the audio for caption generation.\nExperiments on two benchmark datasets, Clotho and AudioCaps, show that RECAP\nachieves competitive performance in in-domain settings and significant\nimprovements in out-of-domain settings. Additionally, due to its capability to\nexploit a large text-captions-only datastore in a training-free fashion, RECAP\nshows unique capabilities of captioning novel audio events never seen during\ntraining and compositional audios with multiple events. To promote research in\nthis space, we also release 150,000+ new weakly labeled captions for AudioSet,\nAudioCaps, and Clotho.\n", "date": "2023-09-18", "categories": []}
{"id": "2401.17264", "title": "Proactive Detection of Voice Cloning with Localized Watermarking", "abstract": "  In the rapidly evolving field of speech generative models, there is a\npressing need to ensure audio authenticity against the risks of voice cloning.\nWe present AudioSeal, the first audio watermarking technique designed\nspecifically for localized detection of AI-generated speech. AudioSeal employs\na generator/detector architecture trained jointly with a localization loss to\nenable localized watermark detection up to the sample level, and a novel\nperceptual loss inspired by auditory masking, that enables AudioSeal to achieve\nbetter imperceptibility. AudioSeal achieves state-of-the-art performance in\nterms of robustness to real life audio manipulations and imperceptibility based\non automatic and human evaluation metrics. Additionally, AudioSeal is designed\nwith a fast, single-pass detector, that significantly surpasses existing models\nin speed - achieving detection up to two orders of magnitude faster, making it\nideal for large-scale and real-time applications.\n", "date": "2024-01-30", "categories": []}
{"id": "2406.03120", "title": "RevRIR: Joint Reverberant Speech and Room Impulse Response Embedding\n  using Contrastive Learning with Application to Room Shape Classification", "abstract": "  This paper focuses on room fingerprinting, a task involving the analysis of\nan audio recording to determine the specific volume and shape of the room in\nwhich it was captured. While it is relatively straightforward to determine the\nbasic room parameters from the Room Impulse Responses (RIR), doing so from a\nspeech signal is a cumbersome task. To address this challenge, we introduce a\ndual-encoder architecture that facilitates the estimation of room parameters\ndirectly from speech utterances. During pre-training, one encoder receives the\nRIR while the other processes the reverberant speech signal. A contrastive loss\nfunction is employed to embed the speech and the acoustic response jointly. In\nthe fine-tuning stage, the specific classification task is trained. In the test\nphase, only the reverberant utterance is available, and its embedding is used\nfor the task of room shape classification. The proposed scheme is extensively\nevaluated using simulated acoustic environments.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.03251", "title": "ASoBO: Attentive Beamformer Selection for Distant Speaker Diarization in\n  Meetings", "abstract": "  Speaker Diarization (SD) aims at grouping speech segments that belong to the\nsame speaker. This task is required in many speech-processing applications,\nsuch as rich meeting transcription. In this context, distant microphone arrays\nusually capture the audio signal. Beamforming, i.e., spatial filtering, is a\ncommon practice to process multi-microphone audio data. However, it often\nrequires an explicit localization of the active source to steer the filter.\nThis paper proposes a self-attention-based algorithm to select the output of a\nbank of fixed spatial filters. This method serves as a feature extractor for\njoint Voice Activity (VAD) and Overlapped Speech Detection (OSD). The speaker\ndiarization is then inferred from the detected segments. The approach shows\nconvincing distant VAD, OSD, and SD performance, e.g. 14.5% DER on the\nAISHELL-4 dataset. The analysis of the self-attention weights demonstrates\ntheir explainability, as they correlate with the speaker's angular locations.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.03344", "title": "Audio Mamba: Bidirectional State Space Model for Audio Representation\n  Learning", "abstract": "  Transformers have rapidly become the preferred choice for audio\nclassification, surpassing methods based on CNNs. However, Audio Spectrogram\nTransformers (ASTs) exhibit quadratic scaling due to self-attention. The\nremoval of this quadratic self-attention cost presents an appealing direction.\nRecently, state space models (SSMs), such as Mamba, have demonstrated potential\nin language and vision tasks in this regard. In this study, we explore whether\nreliance on self-attention is necessary for audio classification tasks. By\nintroducing Audio Mamba (AuM), the first self-attention-free, purely SSM-based\nmodel for audio classification, we aim to address this question. We evaluate\nAuM on various audio datasets - comprising six different benchmarks - where it\nachieves comparable or better performance compared to well-established AST\nmodel.\n", "date": "2024-06-05", "categories": []}
{"id": "2402.00828", "title": "Efficient Fine-tuning of Audio Spectrogram Transformers via Soft Mixture\n  of Adapters", "abstract": "  Mixture of Experts (MoE) architectures have recently started burgeoning due\nto their ability to scale model's capacity while maintaining the computational\ncost affordable. Furthermore, they can be applied to both Transformers and\nState Space Models, the current state-of-the-art models in numerous fields.\nWhile MoE has been mostly investigated for the pre-training stage, its use in\nparameter-efficient transfer learning settings is under-explored. To narrow\nthis gap, this paper attempts to demystify the use of MoE for\nparameter-efficient fine-tuning of Audio Spectrogram Transformers to audio and\nspeech downstream tasks. Specifically, we propose Soft Mixture of Adapters\n(Soft-MoA). It exploits adapters as the experts and, leveraging the recent Soft\nMoE method, it relies on a soft assignment between the input tokens and experts\nto keep the computational time limited. Extensive experiments across 4\nbenchmarks demonstrate that Soft-MoA outperforms the single adapter method and\nperforms on par with the dense MoA counterpart. We finally present ablation\nstudies on key elements of Soft-MoA, showing for example that Soft-MoA achieves\nbetter scaling with more experts, as well as ensuring that all experts\ncontribute to the computation of the output tokens, thus dispensing with the\nexpert imbalance issue.\n", "date": "2024-02-01", "categories": []}
{"id": "2406.00495", "title": "Audio-Visual Talker Localization in Video for Spatial Sound Reproduction", "abstract": "  Object-based audio production requires the positional metadata to be defined\nfor each point-source object, including the key elements in the foreground of\nthe sound scene. In many media production use cases, both cameras and\nmicrophones are employed to make recordings, and the human voice is often a key\nelement. In this research, we detect and locate the active speaker in the\nvideo, facilitating the automatic extraction of the positional metadata of the\ntalker relative to the camera's reference frame. With the integration of the\nvisual modality, this study expands upon our previous investigation focused\nsolely on audio-based active speaker detection and localization. Our\nexperiments compare conventional audio-visual approaches for active speaker\ndetection that leverage monaural audio, our previous audio-only method that\nleverages multichannel recordings from a microphone array, and a novel\naudio-visual approach integrating vision and multichannel audio. We found the\nrole of the two modalities to complement each other. Multichannel audio,\novercoming the problem of visual occlusions, provides a double-digit reduction\nin detection error compared to audio-visual methods with single-channel audio.\nThe combination of multichannel audio and vision further enhances spatial\naccuracy, leading to a four-percentage point increase in F1 score on the Tragic\nTalkers dataset. Future investigations will assess the robustness of the model\nin noisy and highly reverberant environments, as well as tackle the problem of\noff-screen speakers.\n", "date": "2024-06-01", "categories": []}
{"id": "2406.02483", "title": "How Do Neural Spoofing Countermeasures Detect Partially Spoofed Audio?", "abstract": "  Partially manipulating a sentence can greatly change its meaning. Recent work\nshows that countermeasures (CMs) trained on partially spoofed audio can\neffectively detect such spoofing. However, the current understanding of the\ndecision-making process of CMs is limited. We utilize Grad-CAM and introduce a\nquantitative analysis metric to interpret CMs' decisions. We find that CMs\nprioritize the artifacts of transition regions created when concatenating bona\nfide and spoofed audio. This focus differs from that of CMs trained on fully\nspoofed audio, which concentrate on the pattern differences between bona fide\nand spoofed parts. Our further investigation explains the varying nature of\nCMs' focus while making correct or incorrect predictions. These insights\nprovide a basis for the design of CM models and the creation of datasets.\nMoreover, this work lays a foundation of interpretability in the field of\npartial spoofed audio detection that has not been well explored previously.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.00146", "title": "A Survey of Deep Learning Audio Generation Methods", "abstract": "  This article presents a review of typical techniques used in three distinct\naspects of deep learning model development for audio generation. In the first\npart of the article, we provide an explanation of audio representations,\nbeginning with the fundamental audio waveform. We then progress to the\nfrequency domain, with an emphasis on the attributes of human hearing, and\nfinally introduce a relatively recent development. The main part of the article\nfocuses on explaining basic and extended deep learning architecture variants,\nalong with their practical applications in the field of audio generation. The\nfollowing architectures are addressed: 1) Autoencoders 2) Generative\nadversarial networks 3) Normalizing flows 4) Transformer networks 5) Diffusion\nmodels. Lastly, we will examine four distinct evaluation metrics that are\ncommonly employed in audio generation. This article aims to offer novice\nreaders and beginners in the field a comprehensive understanding of the current\nstate of the art in audio generation methods as well as relevant studies that\ncan be explored for future research.\n", "date": "2024-05-31", "categories": []}
{"id": "2210.05148", "title": "DiffRoll: Diffusion-based Generative Music Transcription with\n  Unsupervised Pretraining Capability", "abstract": "  In this paper we propose a novel generative approach, DiffRoll, to tackle\nautomatic music transcription (AMT). Instead of treating AMT as a\ndiscriminative task in which the model is trained to convert spectrograms into\npiano rolls, we think of it as a conditional generative task where we train our\nmodel to generate realistic looking piano rolls from pure Gaussian noise\nconditioned on spectrograms. This new AMT formulation enables DiffRoll to\ntranscribe, generate and even inpaint music. Due to the classifier-free nature,\nDiffRoll is also able to be trained on unpaired datasets where only piano rolls\nare available. Our experiments show that DiffRoll outperforms its\ndiscriminative counterpart by 19 percentage points (ppt.) and our ablation\nstudies also indicate that it outperforms similar existing methods by 4.8 ppt.\n  Source code and demonstration are available https://sony.github.io/DiffRoll/.\n", "date": "2022-10-11", "categories": []}
{"id": "2311.09655", "title": "Multi-View Spectrogram Transformer for Respiratory Sound Classification", "abstract": "  Deep neural networks have been applied to audio spectrograms for respiratory\nsound classification. Existing models often treat the spectrogram as a\nsynthetic image while overlooking its physical characteristics. In this paper,\na Multi-View Spectrogram Transformer (MVST) is proposed to embed different\nviews of time-frequency characteristics into the vision transformer.\nSpecifically, the proposed MVST splits the mel-spectrogram into different sized\npatches, representing the multi-view acoustic elements of a respiratory sound.\nThese patches and positional embeddings are then fed into transformer encoders\nto extract the attentional information among patches through a self-attention\nmechanism. Finally, a gated fusion scheme is designed to automatically weigh\nthe multi-view features to highlight the best one in a specific scenario.\nExperimental results on the ICBHI dataset demonstrate that the proposed MVST\nsignificantly outperforms state-of-the-art methods for classifying respiratory\nsounds.\n", "date": "2023-11-16", "categories": []}
{"id": "2405.06289", "title": "Look Once to Hear: Target Speech Hearing with Noisy Examples", "abstract": "  In crowded settings, the human brain can focus on speech from a target\nspeaker, given prior knowledge of how they sound. We introduce a novel\nintelligent hearable system that achieves this capability, enabling target\nspeech hearing to ignore all interfering speech and noise, but the target\nspeaker. A naive approach is to require a clean speech example to enroll the\ntarget speaker. This is however not well aligned with the hearable application\ndomain since obtaining a clean example is challenging in real world scenarios,\ncreating a unique user interface problem. We present the first enrollment\ninterface where the wearer looks at the target speaker for a few seconds to\ncapture a single, short, highly noisy, binaural example of the target speaker.\nThis noisy example is used for enrollment and subsequent speech extraction in\nthe presence of interfering speakers and noise. Our system achieves a signal\nquality improvement of 7.01 dB using less than 5 seconds of noisy enrollment\naudio and can process 8 ms of audio chunks in 6.24 ms on an embedded CPU. Our\nuser studies demonstrate generalization to real-world static and mobile\nspeakers in previously unseen indoor and outdoor multipath environments.\nFinally, our enrollment interface for noisy examples does not cause performance\ndegradation compared to clean examples, while being convenient and\nuser-friendly. Taking a step back, this paper takes an important step towards\nenhancing the human auditory perception with artificial intelligence. We\nprovide code and data at: https://github.com/vb000/LookOnceToHear.\n", "date": "2024-05-10", "categories": []}
{"id": "2405.19497", "title": "Gaussian Flow Bridges for Audio Domain Transfer with Unpaired Data", "abstract": "  Audio domain transfer is the process of modifying audio signals to match\ncharacteristics of a different domain, while retaining the original content.\nThis paper investigates the potential of Gaussian Flow Bridges, an emerging\napproach in generative modeling, for this problem. The presented framework\naddresses the transport problem across different distributions of audio signals\nthrough the implementation of a series of two deterministic probability flows.\nThe proposed framework facilitates manipulation of the target distribution\nproperties through a continuous control variable, which defines a certain\naspect of the target domain. Notably, this approach does not rely on paired\nexamples for training. To address identified challenges on maintaining the\nspeech content consistent, we recommend a training strategy that incorporates\nchunk-based minibatch Optimal Transport couplings of data samples and noise.\nComparing our unsupervised method with established baselines, we find\ncompetitive performance in tasks of reverberation and distortion manipulation.\nDespite encoutering limitations, the intriguing results obtained in this study\nunderscore potential for further exploration.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.20059", "title": "Spectral Mapping of Singing Voices: U-Net-Assisted Vocal Segmentation", "abstract": "  Separating vocal elements from musical tracks is a longstanding challenge in\naudio signal processing. This study tackles the distinct separation of vocal\ncomponents from musical spectrograms. We employ the Short Time Fourier\nTransform (STFT) to extract audio waves into detailed frequency-time\nspectrograms, utilizing the benchmark MUSDB18 dataset for music separation.\nSubsequently, we implement a UNet neural network to segment the spectrogram\nimage, aiming to delineate and extract singing voice components accurately. We\nachieved noteworthy results in audio source separation using of our U-Net-based\nmodels. The combination of frequency-axis normalization with Min/Max scaling\nand the Mean Absolute Error (MAE) loss function achieved the highest\nSource-to-Distortion Ratio (SDR) of 7.1 dB, indicating a high level of accuracy\nin preserving the quality of the original signal during separation. This setup\nalso recorded impressive Source-to-Interference Ratio (SIR) and\nSource-to-Artifact Ratio (SAR) scores of 25.2 dB and 7.2 dB, respectively.\nThese values significantly outperformed other configurations, particularly\nthose using Quantile-based normalization or a Mean Squared Error (MSE) loss\nfunction. Our source code, model weights, and demo material can be found at the\nproject's GitHub repository: https://github.com/mbrotos/SoundSeg\n", "date": "2024-05-30", "categories": []}
{"id": "2306.00814", "title": "Vocos: Closing the gap between time-domain and Fourier-based neural\n  vocoders for high-quality audio synthesis", "abstract": "  Recent advancements in neural vocoding are predominantly driven by Generative\nAdversarial Networks (GANs) operating in the time-domain. While effective, this\napproach neglects the inductive bias offered by time-frequency representations,\nresulting in reduntant and computionally-intensive upsampling operations.\nFourier-based time-frequency representation is an appealing alternative,\naligning more accurately with human auditory perception, and benefitting from\nwell-established fast algorithms for its computation. Nevertheless, direct\nreconstruction of complex-valued spectrograms has been historically\nproblematic, primarily due to phase recovery issues. This study seeks to close\nthis gap by presenting Vocos, a new model that directly generates Fourier\nspectral coefficients. Vocos not only matches the state-of-the-art in audio\nquality, as demonstrated in our evaluations, but it also substantially improves\ncomputational efficiency, achieving an order of magnitude increase in speed\ncompared to prevailing time-domain neural vocoding approaches. The source code\nand model weights have been open-sourced at https://github.com/gemelo-ai/vocos.\n", "date": "2023-06-01", "categories": []}
{"id": "2207.11749", "title": "Simultaneous source separation of unknown numbers of single-channel\n  underwater acoustic signals based on deep neural networks with\n  separator-decoder structure", "abstract": "  The separation of single-channel underwater acoustic signals is a challenging\nproblem with practical significance. Few existing studies focus on the source\nseparation problem with unknown numbers of signals, and how to evaluate the\nperformance of the systems is not yet clear. In this paper, a deep\nlearning-based simultaneous separating solution with a fixed number of output\nchannels equal to the maximum number of possible targets is proposed to address\nthese two problems. This solution avoids the dimensional disaster caused by the\npermutation problem induced by the alignment of outputs to targets.\nSpecifically, we propose a two-step learning-based separation model with a\nseparator-decoder structure. A performance evaluation method with two\nquantitative metrics of the separation system for situations with mute channels\nin the output channels that do not contain target signals is also proposed.\nExperiments conducted on simulated mixtures of radiated ship noise show that\nthe proposed solution can achieve similar separation performance to that\nattained with a known number of signals. The proposed separation model with\nseparator-decoder structure achieved competitive performance as two models\ndeveloped for known numbers of signals, which is highly explainable and\nextensible and gets the state of the art under this framework.\n", "date": "2022-07-24", "categories": []}
{"id": "2402.01831", "title": "Audio Flamingo: A Novel Audio Language Model with Few-Shot Learning and\n  Dialogue Abilities", "abstract": "  Augmenting large language models (LLMs) to understand audio -- including\nnon-speech sounds and non-verbal speech -- is critically important for diverse\nreal-world applications of LLMs. In this paper, we propose Audio Flamingo, a\nnovel audio language model with 1) strong audio understanding abilities, 2) the\nability to quickly adapt to unseen tasks via in-context learning and retrieval,\nand 3) strong multi-turn dialogue abilities. We introduce a series of training\ntechniques, architecture design, and data strategies to enhance our model with\nthese abilities. Extensive evaluations across various audio understanding tasks\nconfirm the efficacy of our method, setting new state-of-the-art benchmarks.\nOur demo website is https://audioflamingo.github.io/ and the code is\nopen-sourced at https://github.com/NVIDIA/audio-flamingo.\n", "date": "2024-02-02", "categories": []}
{"id": "2405.16000", "title": "Carnatic Raga Identification System using Rigorous Time-Delay Neural\n  Network", "abstract": "  Large scale machine learning-based Raga identification continues to be a\nnontrivial issue in the computational aspects behind Carnatic music. Each raga\nconsists of many unique and intrinsic melodic patterns that can be used to\neasily identify them from others. These ragas can also then be used to cluster\nsongs within the same raga, as well as identify songs in other closely related\nragas. In this case, the input sound is analyzed using a combination of steps\nincluding using a Discrete Fourier transformation and using Triangular\nFiltering to create custom bins of possible notes, extracting features from the\npresence of particular notes or lack thereof. Using a combination of Neural\nNetworks including 1D Convolutional Neural Networks conventionally known as\nTime-Delay Neural Networks) and Long Short-Term Memory (LSTM), which are a form\nof Recurrent Neural Networks, the backbone of the classification strategy to\nbuild the model can be created. In addition, to help with variations in shruti,\na long-time attention-based mechanism will be implemented to determine the\nrelative changes in frequency rather than the absolute differences. This will\nprovide a much more meaningful data point when training audio clips in\ndifferent shrutis. To evaluate the accuracy of the classifier, a dataset of 676\nrecordings is used. The songs are distributed across the list of ragas. The\ngoal of this program is to be able to effectively and efficiently label a much\nwider range of audio clips in more shrutis, ragas, and with more background\nnoise.\n", "date": "2024-05-25", "categories": []}
{"id": "2405.16797", "title": "A Real-Time Voice Activity Detection Based On Lightweight Neural", "abstract": "  Voice activity detection (VAD) is the task of detecting speech in an audio\nstream, which is challenging due to numerous unseen noises and low\nsignal-to-noise ratios in real environments. Recently, neural network-based\nVADs have alleviated the degradation of performance to some extent. However,\nthe majority of existing studies have employed excessively large models and\nincorporated future context, while neglecting to evaluate the operational\nefficiency and latency of the models. In this paper, we propose a lightweight\nand real-time neural network called MagicNet, which utilizes casual and depth\nseparable 1-D convolutions and GRU. Without relying on future features as\ninput, our proposed model is compared with two state-of-the-art algorithms on\nsynthesized in-domain and out-domain test datasets. The evaluation results\ndemonstrate that MagicNet can achieve improved performance and robustness with\nfewer parameter costs.\n", "date": "2024-05-27", "categories": []}
{"id": "2405.17413", "title": "Enhancing Music Genre Classification through Multi-Algorithm Analysis\n  and User-Friendly Visualization", "abstract": "  The aim of this study is to teach an algorithm how to recognize different\ntypes of music. Users will submit songs for analysis. Since the algorithm\nhasn't heard these songs before, it needs to figure out what makes each song\nunique. It does this by breaking down the songs into different parts and\nstudying things like rhythm, melody, and tone via supervised learning because\nthe program learns from examples that are already labelled. One important thing\nto consider when classifying music is its genre, which can be quite complex. To\nensure accuracy, we use five different algorithms, each working independently,\nto analyze the songs. This helps us get a more complete understanding of each\nsong's characteristics. Therefore, our goal is to correctly identify the genre\nof each submitted song. Once the analysis is done, the results are presented\nusing a graphing tool, making it easy for users to understand and provide\nfeedback.\n", "date": "2024-05-27", "categories": []}
{"id": "2405.15096", "title": "Music Genre Classification: Training an AI model", "abstract": "  Music genre classification is an area that utilizes machine learning models\nand techniques for the processing of audio signals, in which applications range\nfrom content recommendation systems to music recommendation systems. In this\nresearch I explore various machine learning algorithms for the purpose of music\ngenre classification, using features extracted from audio signals.The systems\nare namely, a Multilayer Perceptron (built from scratch), a k-Nearest\nNeighbours (also built from scratch), a Convolutional Neural Network and lastly\na Random Forest wide model. In order to process the audio signals, feature\nextraction methods such as Short-Time Fourier Transform, and the extraction of\nMel Cepstral Coefficients (MFCCs), is performed. Through this extensive\nresearch, I aim to asses the robustness of machine learning models for genre\nclassification, and to compare their results.\n", "date": "2024-05-23", "categories": []}
{"id": "2209.12573", "title": "Faked Speech Detection with Zero Prior Knowledge", "abstract": "  Audio is one of the most used ways of human communication, but at the same\ntime it can be easily misused to trick people. With the revolution of AI, the\nrelated technologies are now accessible to almost everyone, thus making it\nsimple for the criminals to commit crimes and forgeries. In this work, we\nintroduce a neural network method to develop a classifier that will blindly\nclassify an input audio as real or mimicked; the word 'blindly' refers to the\nability to detect mimicked audio without references or real sources. We propose\na deep neural network following a sequential model that comprises three hidden\nlayers, with alternating dense and drop out layers. The proposed model was\ntrained on a set of 26 important features extracted from a large dataset of\naudios to get a classifier that was tested on the same set of features from\ndifferent audios. The data was extracted from two raw datasets, especially\ncomposed for this work; an all English dataset and a mixed dataset (Arabic plus\nEnglish) (The dataset can be provided, in raw form, by writing an email to the\nfirst author). For the purpose of comparison, the audios were also classified\nthrough human inspection with the subjects being the native speakers. The\nensued results were interesting and exhibited formidable accuracy, as we were\nable to get at least 94% correct classification of the test cases, as against\nthe 85% accuracy in the case of human observers.\n", "date": "2022-09-26", "categories": []}
{"id": "2405.04181", "title": "Detecting music deepfakes is easy but actually hard", "abstract": "  In the face of a new era of generative models, the detection of artificially\ngenerated content has become a matter of utmost importance. The ability to\ncreate credible minute-long music deepfakes in a few seconds on user-friendly\nplatforms poses a real threat of fraud on streaming services and unfair\ncompetition to human artists. This paper demonstrates the possibility (and\nsurprising ease) of training classifiers on datasets comprising real audio and\nfake reconstructions, achieving a convincing accuracy of 99.8%. To our\nknowledge, this marks the first publication of a music deepfake detector, a\ntool that will help in the regulation of music forgery. Nevertheless, informed\nby decades of literature on forgery detection in other fields, we stress that a\ngood test score is not the end of the story. We step back from the\nstraightforward ML framework and expose many facets that could be problematic\nwith such a deployed detector: calibration, robustness to audio manipulation,\ngeneralisation to unseen models, interpretability and possibility for recourse.\nThis second part acts as a position for future research steps in the field and\na caveat to a flourishing market of fake content checkers.\n", "date": "2024-05-07", "categories": []}
{"id": "2405.13527", "title": "End-to-End Real-World Polyphonic Piano Audio-to-Score Transcription with\n  Hierarchical Decoding", "abstract": "  Piano audio-to-score transcription (A2S) is an important yet underexplored\ntask with extensive applications for music composition, practice, and analysis.\nHowever, existing end-to-end piano A2S systems faced difficulties in retrieving\nbar-level information such as key and time signatures, and have been trained\nand evaluated with only synthetic data. To address these limitations, we\npropose a sequence-to-sequence (Seq2Seq) model with a hierarchical decoder that\naligns with the hierarchical structure of musical scores, enabling the\ntranscription of score information at both the bar and note levels by\nmulti-task learning. To bridge the gap between synthetic data and recordings of\nhuman performance, we propose a two-stage training scheme, which involves\npre-training the model using an expressive performance rendering (EPR) system\non synthetic audio, followed by fine-tuning the model using recordings of human\nperformance. To preserve the voicing structure for score reconstruction, we\npropose a pre-processing method for **Kern scores in scenarios with an\nunconstrained number of voices. Experimental results support the effectiveness\nof our proposed approaches, in terms of both transcription performance on\nsynthetic audio data in comparison to the current state-of-the-art, and the\nfirst experiment on human recordings.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.13636", "title": "Audio Mamba: Pretrained Audio State Space Model For Audio Tagging", "abstract": "  Audio tagging is an important task of mapping audio samples to their\ncorresponding categories. Recently endeavours that exploit transformer models\nin this field have achieved great success. However, the quadratic\nself-attention cost limits the scaling of audio transformer models and further\nconstrains the development of more universal audio models. In this paper, we\nattempt to solve this problem by proposing Audio Mamba, a self-attention-free\napproach that captures long audio spectrogram dependency with state space\nmodels. Our experimental results on two audio-tagging datasets demonstrate the\nparameter efficiency of Audio Mamba, it achieves comparable results to SOTA\naudio spectrogram transformers with one third parameters.\n", "date": "2024-05-22", "categories": []}
{"id": "2309.07391", "title": "EnCodecMAE: Leveraging neural codecs for universal audio representation\n  learning", "abstract": "  The goal of universal audio representation learning is to obtain foundational\nmodels that can be used for a variety of downstream tasks involving speech,\nmusic and environmental sounds. To approach this problem, methods inspired by\nworks on self-supervised learning for NLP, like BERT, or computer vision, like\nmasked autoencoders (MAE), are often adapted to the audio domain. In this work,\nwe propose masking representations of the audio signal, and training a MAE to\nreconstruct the masked segments. The reconstruction is done by predicting the\ndiscrete units generated by EnCodec, a neural audio codec, from the unmasked\ninputs. We evaluate this approach, which we call EnCodecMAE, on a wide range of\ntasks involving speech, music and environmental sounds. Our best model\noutperforms various state-of-the-art audio representation models in terms of\nglobal performance. Additionally, we evaluate the resulting representations in\nthe challenging task of automatic speech recognition (ASR), obtaining decent\nresults and paving the way for a universal audio representation.\n", "date": "2023-09-14", "categories": []}
{"id": "2312.09663", "title": "Toward Deep Drum Source Separation", "abstract": "  In the past, the field of drum source separation faced significant challenges\ndue to limited data availability, hindering the adoption of cutting-edge deep\nlearning methods that have found success in other related audio applications.\nIn this manuscript, we introduce StemGMD, a large-scale audio dataset of\nisolated single-instrument drum stems. Each audio clip is synthesized from MIDI\nrecordings of expressive drums performances using ten real-sounding acoustic\ndrum kits. Totaling 1224 hours, StemGMD is the largest audio dataset of drums\nto date and the first to comprise isolated audio clips for every instrument in\na canonical nine-piece drum kit. We leverage StemGMD to develop LarsNet, a\nnovel deep drum source separation model. Through a bank of dedicated U-Nets,\nLarsNet can separate five stems from a stereo drum mixture faster than\nreal-time and is shown to significantly outperform state-of-the-art nonnegative\nspectro-temporal factorization methods.\n", "date": "2023-12-15", "categories": []}
{"id": "2405.11913", "title": "Diff-BGM: A Diffusion Model for Video Background Music Generation", "abstract": "  When editing a video, a piece of attractive background music is\nindispensable. However, video background music generation tasks face several\nchallenges, for example, the lack of suitable training datasets, and the\ndifficulties in flexibly controlling the music generation process and\nsequentially aligning the video and music. In this work, we first propose a\nhigh-quality music-video dataset BGM909 with detailed annotation and shot\ndetection to provide multi-modal information about the video and music. We then\npresent evaluation metrics to assess music quality, including music diversity\nand alignment between music and video with retrieval precision metrics.\nFinally, we propose the Diff-BGM framework to automatically generate the\nbackground music for a given video, which uses different signals to control\ndifferent aspects of the music during the generation process, i.e., uses\ndynamic video features to control music rhythm and semantic features to control\nthe melody and atmosphere. We propose to align the video and music sequentially\nby introducing a segment-aware cross-attention layer. Experiments verify the\neffectiveness of our proposed method. The code and models are available at\nhttps://github.com/sizhelee/Diff-BGM.\n", "date": "2024-05-20", "categories": []}
{"id": "2405.10084", "title": "Revisiting Deep Audio-Text Retrieval Through the Lens of Transportation", "abstract": "  The Learning-to-match (LTM) framework proves to be an effective inverse\noptimal transport approach for learning the underlying ground metric between\ntwo sources of data, facilitating subsequent matching. However, the\nconventional LTM framework faces scalability challenges, necessitating the use\nof the entire dataset each time the parameters of the ground metric are\nupdated. In adapting LTM to the deep learning context, we introduce the\nmini-batch Learning-to-match (m-LTM) framework for audio-text retrieval\nproblems. This framework leverages mini-batch subsampling and\nMahalanobis-enhanced family of ground metrics. Moreover, to cope with\nmisaligned training data in practice, we propose a variant using partial\noptimal transport to mitigate the harm of misaligned data pairs in training\ndata. We conduct extensive experiments on audio-text matching problems using\nthree datasets: AudioCaps, Clotho, and ESC-50. Results demonstrate that our\nproposed method is capable of learning rich and expressive joint embedding\nspace, which achieves SOTA performance. Beyond this, the proposed m-LTM\nframework is able to close the modality gap across audio and text embedding,\nwhich surpasses both triplet and contrastive loss in the zero-shot sound event\ndetection task on the ESC-50 dataset. Notably, our strategy of employing\npartial optimal transport with m-LTM demonstrates greater noise tolerance than\ncontrastive loss, especially under varying noise ratios in training data on the\nAudioCaps dataset. Our code is available at\nhttps://github.com/v-manhlt3/m-LTM-Audio-Text-Retrieval\n", "date": "2024-05-16", "categories": []}
{"id": "2405.08679", "title": "Investigating Design Choices in Joint-Embedding Predictive Architectures\n  for General Audio Representation Learning", "abstract": "  This paper addresses the problem of self-supervised general-purpose audio\nrepresentation learning. We explore the use of Joint-Embedding Predictive\nArchitectures (JEPA) for this task, which consists of splitting an input\nmel-spectrogram into two parts (context and target), computing neural\nrepresentations for each, and training the neural network to predict the target\nrepresentations from the context representations. We investigate several design\nchoices within this framework and study their influence through extensive\nexperiments by evaluating our models on various audio classification\nbenchmarks, including environmental sounds, speech and music downstream tasks.\nWe focus notably on which part of the input data is used as context or target\nand show experimentally that it significantly impacts the model's quality. In\nparticular, we notice that some effective design choices in the image domain\nlead to poor performance on audio, thus highlighting major differences between\nthese two modalities.\n", "date": "2024-05-14", "categories": []}
{"id": "2308.05734", "title": "AudioLDM 2: Learning Holistic Audio Generation with Self-supervised\n  Pretraining", "abstract": "  Although audio generation shares commonalities across different types of\naudio, such as speech, music, and sound effects, designing models for each type\nrequires careful consideration of specific objectives and biases that can\nsignificantly differ from those of other types. To bring us closer to a unified\nperspective of audio generation, this paper proposes a framework that utilizes\nthe same learning method for speech, music, and sound effect generation. Our\nframework introduces a general representation of audio, called \"language of\naudio\" (LOA). Any audio can be translated into LOA based on AudioMAE, a\nself-supervised pre-trained representation learning model. In the generation\nprocess, we translate any modalities into LOA by using a GPT-2 model, and we\nperform self-supervised audio generation learning with a latent diffusion model\nconditioned on LOA. The proposed framework naturally brings advantages such as\nin-context learning abilities and reusable self-supervised pretrained AudioMAE\nand latent diffusion models. Experiments on the major benchmarks of\ntext-to-audio, text-to-music, and text-to-speech demonstrate state-of-the-art\nor competitive performance against previous approaches. Our code, pretrained\nmodel, and demo are available at https://audioldm.github.io/audioldm2.\n", "date": "2023-08-10", "categories": []}
{"id": "2402.04825", "title": "Fast Timing-Conditioned Latent Audio Diffusion", "abstract": "  Generating long-form 44.1kHz stereo audio from text prompts can be\ncomputationally demanding. Further, most previous works do not tackle that\nmusic and sound effects naturally vary in their duration. Our research focuses\non the efficient generation of long-form, variable-length stereo music and\nsounds at 44.1kHz using text prompts with a generative model. Stable Audio is\nbased on latent diffusion, with its latent defined by a fully-convolutional\nvariational autoencoder. It is conditioned on text prompts as well as timing\nembeddings, allowing for fine control over both the content and length of the\ngenerated music and sounds. Stable Audio is capable of rendering stereo signals\nof up to 95 sec at 44.1kHz in 8 sec on an A100 GPU. Despite its compute\nefficiency and fast inference, it is one of the best in two public\ntext-to-music and -audio benchmarks and, differently from state-of-the-art\nmodels, can generate music with structure and stereo sounds.\n", "date": "2024-02-07", "categories": []}
{"id": "2405.06747", "title": "Music Emotion Prediction Using Recurrent Neural Networks", "abstract": "  This study explores the application of recurrent neural networks to recognize\nemotions conveyed in music, aiming to enhance music recommendation systems and\nsupport therapeutic interventions by tailoring music to fit listeners'\nemotional states. We utilize Russell's Emotion Quadrant to categorize music\ninto four distinct emotional regions and develop models capable of accurately\npredicting these categories. Our approach involves extracting a comprehensive\nset of audio features using Librosa and applying various recurrent neural\nnetwork architectures, including standard RNNs, Bidirectional RNNs, and Long\nShort-Term Memory (LSTM) networks. Initial experiments are conducted using a\ndataset of 900 audio clips, labeled according to the emotional quadrants. We\ncompare the performance of our neural network models against a set of baseline\nclassifiers and analyze their effectiveness in capturing the temporal dynamics\ninherent in musical expression. The results indicate that simpler RNN\narchitectures may perform comparably or even superiorly to more complex models,\nparticularly in smaller datasets. We've also applied the following experiments\non larger datasets: one is augmented based on our original dataset, and the\nother is from other sources. This research not only enhances our understanding\nof the emotional impact of music but also demonstrates the potential of neural\nnetworks in creating more personalized and emotionally resonant music\nrecommendation and therapy systems.\n", "date": "2024-05-10", "categories": []}
{"id": "2310.03743", "title": "The Un-Kidnappable Robot: Acoustic Localization of Sneaking People", "abstract": "  How easy is it to sneak up on a robot? We examine whether we can detect\npeople using only the incidental sounds they produce as they move, even when\nthey try to be quiet. We collect a robotic dataset of high-quality 4-channel\naudio paired with 360 degree RGB data of people moving in different indoor\nsettings. We train models that predict if there is a moving person nearby and\ntheir location using only audio. We implement our method on a robot, allowing\nit to track a single person moving quietly with only passive audio sensing. For\ndemonstration videos, see our project page:\nhttps://sites.google.com/view/unkidnappable-robot\n", "date": "2023-10-05", "categories": []}
{"id": "2405.05467", "title": "AFEN: Respiratory Disease Classification using Ensemble Learning", "abstract": "  We present AFEN (Audio Feature Ensemble Learning), a model that leverages\nConvolutional Neural Networks (CNN) and XGBoost in an ensemble learning fashion\nto perform state-of-the-art audio classification for a range of respiratory\ndiseases. We use a meticulously selected mix of audio features which provide\nthe salient attributes of the data and allow for accurate classification. The\nextracted features are then used as an input to two separate model classifiers\n1) a multi-feature CNN classifier and 2) an XGBoost Classifier. The outputs of\nthe two models are then fused with the use of soft voting. Thus, by exploiting\nensemble learning, we achieve increased robustness and accuracy. We evaluate\nthe performance of the model on a database of 920 respiratory sounds, which\nundergoes data augmentation techniques to increase the diversity of the data\nand generalizability of the model. We empirically verify that AFEN sets a new\nstate-of-the-art using Precision and Recall as metrics, while decreasing\ntraining time by 60%.\n", "date": "2024-05-08", "categories": []}
{"id": "2405.05244", "title": "SVDD Challenge 2024: A Singing Voice Deepfake Detection Challenge\n  Evaluation Plan", "abstract": "  The rapid advancement of AI-generated singing voices, which now closely mimic\nnatural human singing and align seamlessly with musical scores, has led to\nheightened concerns for artists and the music industry. Unlike spoken voice,\nsinging voice presents unique challenges due to its musical nature and the\npresence of strong background music, making singing voice deepfake detection\n(SVDD) a specialized field requiring focused attention. To promote SVDD\nresearch, we recently proposed the \"SVDD Challenge,\" the very first research\nchallenge focusing on SVDD for lab-controlled and in-the-wild bonafide and\ndeepfake singing voice recordings. The challenge will be held in conjunction\nwith the 2024 IEEE Spoken Language Technology Workshop (SLT 2024).\n", "date": "2024-05-08", "categories": []}
{"id": "2305.07489", "title": "Benchmarks and leaderboards for sound demixing tasks", "abstract": "  Music demixing is the task of separating different tracks from the given\nsingle audio signal into components, such as drums, bass, and vocals from the\nrest of the accompaniment. Separation of sources is useful for a range of\nareas, including entertainment and hearing aids. In this paper, we introduce\ntwo new benchmarks for the sound source separation tasks and compare popular\nmodels for sound demixing, as well as their ensembles, on these benchmarks. For\nthe models' assessments, we provide the leaderboard at\nhttps://mvsep.com/quality_checker/, giving a comparison for a range of models.\nThe new benchmark datasets are available for download. We also develop a novel\napproach for audio separation, based on the ensembling of different models that\nare suited best for the particular stem. The proposed solution was evaluated in\nthe context of the Music Demixing Challenge 2023 and achieved top results in\ndifferent tracks of the challenge. The code and the approach are open-sourced\non GitHub.\n", "date": "2023-05-12", "categories": []}
{"id": "2404.16436", "title": "Leveraging tropical reef, bird and unrelated sounds for superior\n  transfer learning in marine bioacoustics", "abstract": "  Machine learning has the potential to revolutionize passive acoustic\nmonitoring (PAM) for ecological assessments. However, high annotation and\ncompute costs limit the field's efficacy. Generalizable pretrained networks can\novercome these costs, but high-quality pretraining requires vast annotated\nlibraries, limiting its current applicability primarily to bird taxa. Here, we\nidentify the optimum pretraining strategy for a data-deficient domain using\ncoral reef bioacoustics. We assemble ReefSet, a large annotated library of reef\nsounds, though modest compared to bird libraries at 2% of the sample count.\nThrough testing few-shot transfer learning performance, we observe that\npretraining on bird audio provides notably superior generalizability compared\nto pretraining on ReefSet or unrelated audio alone. However, our key findings\nshow that cross-domain mixing which leverages bird, reef and unrelated audio\nduring pretraining maximizes reef generalizability. SurfPerch, our pretrained\nnetwork, provides a strong foundation for automated analysis of marine PAM data\nwith minimal annotation and compute costs.\n", "date": "2024-04-25", "categories": []}
{"id": "2405.04272", "title": "BUDDy: Single-Channel Blind Unsupervised Dereverberation with Diffusion\n  Models", "abstract": "  In this paper, we present an unsupervised single-channel method for joint\nblind dereverberation and room impulse response estimation, based on posterior\nsampling with diffusion models. We parameterize the reverberation operator\nusing a filter with exponential decay for each frequency subband, and\niteratively estimate the corresponding parameters as the speech utterance gets\nrefined along the reverse diffusion trajectory. A measurement consistency\ncriterion enforces the fidelity of the generated speech with the reverberant\nmeasurement, while an unconditional diffusion model implements a strong prior\nfor clean speech generation. Without any knowledge of the room impulse response\nnor any coupled reverberant-anechoic data, we can successfully perform\ndereverberation in various acoustic scenarios. Our method significantly\noutperforms previous blind unsupervised baselines, and we demonstrate its\nincreased robustness to unseen acoustic conditions in comparison to blind\nsupervised methods. Audio samples and code are available online.\n", "date": "2024-05-07", "categories": []}
{"id": "2209.11112", "title": "CMGAN: Conformer-Based Metric-GAN for Monaural Speech Enhancement", "abstract": "  In this work, we further develop the conformer-based metric generative\nadversarial network (CMGAN) model for speech enhancement (SE) in the\ntime-frequency (TF) domain. This paper builds on our previous work but takes a\nmore in-depth look by conducting extensive ablation studies on model inputs and\narchitectural design choices. We rigorously tested the generalization ability\nof the model to unseen noise types and distortions. We have fortified our\nclaims through DNS-MOS measurements and listening tests. Rather than focusing\nexclusively on the speech denoising task, we extend this work to address the\ndereverberation and super-resolution tasks. This necessitated exploring various\narchitectural changes, specifically metric discriminator scores and masking\ntechniques. It is essential to highlight that this is among the earliest works\nthat attempted complex TF-domain super-resolution. Our findings show that CMGAN\noutperforms existing state-of-the-art methods in the three major speech\nenhancement tasks: denoising, dereverberation, and super-resolution. For\nexample, in the denoising task using the Voice Bank+DEMAND dataset, CMGAN\nnotably exceeded the performance of prior models, attaining a PESQ score of\n3.41 and an SSNR of 11.10 dB. Audio samples and CMGAN implementations are\navailable online.\n", "date": "2022-09-22", "categories": []}
{"id": "2303.16621", "title": "AraSpot: Arabic Spoken Command Spotting", "abstract": "  Spoken keyword spotting (KWS) is the task of identifying a keyword in an\naudio stream and is widely used in smart devices at the edge in order to\nactivate voice assistants and perform hands-free tasks. The task is daunting as\nthere is a need, on the one hand, to achieve high accuracy while at the same\ntime ensuring that such systems continue to run efficiently on low power and\npossibly limited computational capabilities devices. This work presents AraSpot\nfor Arabic keyword spotting trained on 40 Arabic keywords, using different\nonline data augmentation, and introducing ConformerGRU model architecture.\nFinally, we further improve the performance of the model by training a\ntext-to-speech model for synthetic data generation. AraSpot achieved a\nState-of-the-Art SOTA 99.59% result outperforming previous approaches.\n", "date": "2023-03-29", "categories": []}
{"id": "2405.02996", "title": "RepAugment: Input-Agnostic Representation-Level Augmentation for\n  Respiratory Sound Classification", "abstract": "  Recent advancements in AI have democratized its deployment as a healthcare\nassistant. While pretrained models from large-scale visual and audio datasets\nhave demonstrably generalized to this task, surprisingly, no studies have\nexplored pretrained speech models, which, as human-originated sounds,\nintuitively would share closer resemblance to lung sounds. This paper explores\nthe efficacy of pretrained speech models for respiratory sound classification.\nWe find that there is a characterization gap between speech and lung sound\nsamples, and to bridge this gap, data augmentation is essential. However, the\nmost widely used augmentation technique for audio and speech, SpecAugment,\nrequires 2-dimensional spectrogram format and cannot be applied to models\npretrained on speech waveforms. To address this, we propose RepAugment, an\ninput-agnostic representation-level augmentation technique that outperforms\nSpecAugment, but is also suitable for respiratory sound classification with\nwaveform pretrained models. Experimental results show that our approach\noutperforms the SpecAugment, demonstrating a substantial improvement in the\naccuracy of minority disease classes, reaching up to 7.14%.\n", "date": "2024-05-05", "categories": []}
{"id": "2405.03567", "title": "Deep Space Separable Distillation for Lightweight Acoustic Scene\n  Classification", "abstract": "  Acoustic scene classification (ASC) is highly important in the real world.\nRecently, deep learning-based methods have been widely employed for acoustic\nscene classification. However, these methods are currently not lightweight\nenough as well as their performance is not satisfactory. To solve these\nproblems, we propose a deep space separable distillation network. Firstly, the\nnetwork performs high-low frequency decomposition on the log-mel spectrogram,\nsignificantly reducing computational complexity while maintaining model\nperformance. Secondly, we specially design three lightweight operators for ASC,\nincluding Separable Convolution (SC), Orthonormal Separable Convolution (OSC),\nand Separable Partial Convolution (SPC). These operators exhibit highly\nefficient feature extraction capabilities in acoustic scene classification\ntasks. The experimental results demonstrate that the proposed method achieves a\nperformance gain of 9.8% compared to the currently popular deep learning\nmethods, while also having smaller parameter count and computational\ncomplexity.\n", "date": "2024-05-06", "categories": []}
{"id": "2207.14682", "title": "Towards Unconstrained Audio Splicing Detection and Localization with\n  Neural Networks", "abstract": "  Freely available and easy-to-use audio editing tools make it straightforward\nto perform audio splicing. Convincing forgeries can be created by combining\nvarious speech samples from the same person. Detection of such splices is\nimportant both in the public sector when considering misinformation, and in a\nlegal context to verify the integrity of evidence. Unfortunately, most existing\ndetection algorithms for audio splicing use handcrafted features and make\nspecific assumptions. However, criminal investigators are often faced with\naudio samples from unconstrained sources with unknown characteristics, which\nraises the need for more generally applicable methods.\n  With this work, we aim to take a first step towards unconstrained audio\nsplicing detection to address this need. We simulate various attack scenarios\nin the form of post-processing operations that may disguise splicing. We\npropose a Transformer sequence-to-sequence (seq2seq) network for splicing\ndetection and localization. Our extensive evaluation shows that the proposed\nmethod outperforms existing dedicated approaches for splicing detection [3, 10]\nas well as the general-purpose networks EfficientNet [28] and RegNet [25].\n", "date": "2022-07-29", "categories": []}
{"id": "2310.09653", "title": "SelfVC: Voice Conversion With Iterative Refinement using Self\n  Transformations", "abstract": "  We propose SelfVC, a training strategy to iteratively improve a voice\nconversion model with self-synthesized examples. Previous efforts on voice\nconversion focus on factorizing speech into explicitly disentangled\nrepresentations that separately encode speaker characteristics and linguistic\ncontent. However, disentangling speech representations to capture such\nattributes using task-specific loss terms can lead to information loss. In this\nwork, instead of explicitly disentangling attributes with loss terms, we\npresent a framework to train a controllable voice conversion model on entangled\nspeech representations derived from self-supervised learning (SSL) and speaker\nverification models. First, we develop techniques to derive prosodic\ninformation from the audio signal and SSL representations to train predictive\nsubmodules in the synthesis model. Next, we propose a training strategy to\niteratively improve the synthesis model for voice conversion, by creating a\nchallenging training objective using self-synthesized examples. We demonstrate\nthat incorporating such self-synthesized examples during training improves the\nspeaker similarity of generated speech as compared to a baseline voice\nconversion model trained solely on heuristically perturbed inputs. Our\nframework is trained without any text and achieves state-of-the-art results in\nzero-shot voice conversion on metrics evaluating naturalness, speaker\nsimilarity, and intelligibility of synthesized audio.\n", "date": "2023-10-14", "categories": []}
{"id": "2405.01293", "title": "Low-resource speech recognition and dialect identification of Irish in a\n  multi-task framework", "abstract": "  This paper explores the use of Hybrid CTC/Attention encoder-decoder models\ntrained with Intermediate CTC (InterCTC) for Irish (Gaelic) low-resource speech\nrecognition (ASR) and dialect identification (DID). Results are compared to the\ncurrent best performing models trained for ASR (TDNN-HMM) and DID (ECAPA-TDNN).\nAn optimal InterCTC setting is initially established using a Conformer encoder.\nThis setting is then used to train a model with an E-branchformer encoder and\nthe performance of both architectures are compared. A multi-task fine-tuning\napproach is adopted for language model (LM) shallow fusion. The experiments\nyielded an improvement in DID accuracy of 10.8% relative to a baseline\nECAPA-TDNN, and WER performance approaching the TDNN-HMM model. This multi-task\napproach emerges as a promising strategy for Irish low-resource ASR and DID.\n", "date": "2024-05-02", "categories": []}
{"id": "2306.06945", "title": "Underwater Acoustic Target Recognition based on Smoothness-inducing\n  Regularization and Spectrogram-based Data Augmentation", "abstract": "  Underwater acoustic target recognition is a challenging task owing to the\nintricate underwater environments and limited data availability. Insufficient\ndata can hinder the ability of recognition systems to support complex modeling,\nthus impeding their advancement. To improve the generalization capacity of\nrecognition models, techniques such as data augmentation have been employed to\nsimulate underwater signals and diversify data distribution. However, the\ncomplexity of underwater environments can cause the simulated signals to\ndeviate from real scenarios, resulting in biased models that are misguided by\nnon-true data. In this study, we propose two strategies to enhance the\ngeneralization ability of models in the case of limited data while avoiding the\nrisk of performance degradation. First, as an alternative to traditional data\naugmentation, we utilize smoothness-inducing regularization, which only\nincorporates simulated signals in the regularization term. Additionally, we\npropose a specialized spectrogram-based data augmentation strategy, namely\nlocal masking and replicating (LMR), to capture inter-class relationships. Our\nexperiments and visualization analysis demonstrate the superiority of our\nproposed strategies.\n", "date": "2023-06-12", "categories": []}
{"id": "2404.19615", "title": "SemiPL: A Semi-supervised Method for Event Sound Source Localization", "abstract": "  In recent years, Event Sound Source Localization has been widely applied in\nvarious fields. Recent works typically relying on the contrastive learning\nframework show impressive performance. However, all work is based on large\nrelatively simple datasets. It's also crucial to understand and analyze human\nbehaviors (actions and interactions of people), voices, and sounds in chaotic\nevents in many applications, e.g., crowd management, and emergency response\nservices. In this paper, we apply the existing model to a more complex dataset,\nexplore the influence of parameters on the model, and propose a semi-supervised\nimprovement method SemiPL. With the increase in data quantity and the influence\nof label quality, self-supervised learning will be an unstoppable trend. The\nexperiment shows that the parameter adjustment will positively affect the\nexisting model. In particular, SSPL achieved an improvement of 12.2% cIoU and\n0.56% AUC in Chaotic World compared to the results provided. The code is\navailable at: https://github.com/ly245422/SSPL\n", "date": "2024-04-30", "categories": []}
{"id": "2404.17608", "title": "Synthesizing Audio from Silent Video using Sequence to Sequence Modeling", "abstract": "  Generating audio from a video's visual context has multiple practical\napplications in improving how we interact with audio-visual media - for\nexample, enhancing CCTV footage analysis, restoring historical videos (e.g.,\nsilent movies), and improving video generation models. We propose a novel\nmethod to generate audio from video using a sequence-to-sequence model,\nimproving on prior work that used CNNs and WaveNet and faced sound diversity\nand generalization challenges. Our approach employs a 3D Vector Quantized\nVariational Autoencoder (VQ-VAE) to capture the video's spatial and temporal\nstructures, decoding with a custom audio decoder for a broader range of sounds.\nTrained on the Youtube8M dataset segment, focusing on specific domains, our\nmodel aims to enhance applications like CCTV footage analysis, silent movie\nrestoration, and video generation models.\n", "date": "2024-04-25", "categories": []}
{"id": "2404.17806", "title": "T-CLAP: Temporal-Enhanced Contrastive Language-Audio Pretraining", "abstract": "  Contrastive language-audio pretraining~(CLAP) has been developed to align the\nrepresentations of audio and language, achieving remarkable performance in\nretrieval and classification tasks. However, current CLAP struggles to capture\ntemporal information within audio and text features, presenting substantial\nlimitations for tasks such as audio retrieval and generation. To address this\ngap, we introduce T-CLAP, a temporal-enhanced CLAP model. We use Large Language\nModels~(LLMs) and mixed-up strategies to generate temporal-contrastive captions\nfor audio clips from extensive audio-text datasets. Subsequently, a new\ntemporal-focused contrastive loss is designed to fine-tune the CLAP model by\nincorporating these synthetic data. We conduct comprehensive experiments and\nanalysis in multiple downstream tasks. T-CLAP shows improved capability in\ncapturing the temporal relationship of sound events and outperforms\nstate-of-the-art models by a significant margin.\n", "date": "2024-04-27", "categories": []}
{"id": "2404.18739", "title": "Towards Dog Bark Decoding: Leveraging Human Speech Processing for\n  Automated Bark Classification", "abstract": "  Similar to humans, animals make extensive use of verbal and non-verbal forms\nof communication, including a large range of audio signals. In this paper, we\naddress dog vocalizations and explore the use of self-supervised speech\nrepresentation models pre-trained on human speech to address dog bark\nclassification tasks that find parallels in human-centered tasks in speech\nrecognition. We specifically address four tasks: dog recognition, breed\nidentification, gender classification, and context grounding. We show that\nusing speech embedding representations significantly improves over simpler\nclassification baselines. Further, we also find that models pre-trained on\nlarge human speech acoustics can provide additional performance boosts on\nseveral tasks.\n", "date": "2024-04-29", "categories": []}
{"id": "2309.05855", "title": "Instabilities in Convnets for Raw Audio", "abstract": "  What makes waveform-based deep learning so hard? Despite numerous attempts at\ntraining convolutional neural networks (convnets) for filterbank design, they\noften fail to outperform hand-crafted baselines. These baselines are linear\ntime-invariant systems: as such, they can be approximated by convnets with wide\nreceptive fields. Yet, in practice, gradient-based optimization leads to\nsuboptimal approximations. In our article, we approach this phenomenon from the\nperspective of initialization. We present a theory of large deviations for the\nenergy response of FIR filterbanks with random Gaussian weights. We find that\ndeviations worsen for large filters and locally periodic input signals, which\nare both typical for audio signal processing applications. Numerical\nsimulations align with our theory and suggest that the condition number of a\nconvolutional layer follows a logarithmic scaling law between the number and\nlength of the filters, which is reminiscent of discrete wavelet bases.\n", "date": "2023-09-11", "categories": []}
{"id": "2404.17018", "title": "Leveraging AI to Generate Audio for User-generated Content in Video\n  Games", "abstract": "  In video game design, audio (both environmental background music and object\nsound effects) play a critical role. Sounds are typically pre-created assets\ndesigned for specific locations or objects in a game. However, user-generated\ncontent is becoming increasingly popular in modern games (e.g. building custom\nenvironments or crafting unique objects). Since the possibilities are virtually\nlimitless, it is impossible for game creators to pre-create audio for\nuser-generated content. We explore the use of generative artificial\nintelligence to create music and sound effects on-the-fly based on\nuser-generated content. We investigate two avenues for audio generation: 1)\ntext-to-audio: using a text description of user-generated content as input to\nthe audio generator, and 2) image-to-audio: using a rendering of the created\nenvironment or object as input to an image-to-text generator, then piping the\nresulting text description into the audio generator. In this paper we discuss\nethical implications of using generative artificial intelligence for\nuser-generated content and highlight two prototype games where audio is\ngenerated for user-created environments and objects.\n", "date": "2024-04-25", "categories": []}
{"id": "2404.17252", "title": "Comparison of self-supervised in-domain and supervised out-domain\n  transfer learning for bird species recognition", "abstract": "  Transferring the weights of a pre-trained model to assist another task has\nbecome a crucial part of modern deep learning, particularly in data-scarce\nscenarios. Pre-training refers to the initial step of training models outside\nthe current task of interest, typically on another dataset. It can be done via\nsupervised models using human-annotated datasets or self-supervised models\ntrained on unlabeled datasets. In both cases, many pre-trained models are\navailable to fine-tune for the task of interest. Interestingly, research has\nshown that pre-trained models from ImageNet can be helpful for audio tasks\ndespite being trained on image datasets. Hence, it's unclear whether in-domain\nmodels would be advantageous compared to competent out-domain models, such as\nconvolutional neural networks from ImageNet. Our experiments will demonstrate\nthe usefulness of in-domain models and datasets for bird species recognition by\nleveraging VICReg, a recent and powerful self-supervised method.\n", "date": "2024-04-26", "categories": []}
{"id": "2312.00834", "title": "AV-RIR: Audio-Visual Room Impulse Response Estimation", "abstract": "  Accurate estimation of Room Impulse Response (RIR), which captures an\nenvironment's acoustic properties, is important for speech processing and AR/VR\napplications. We propose AV-RIR, a novel multi-modal multi-task learning\napproach to accurately estimate the RIR from a given reverberant speech signal\nand the visual cues of its corresponding environment. AV-RIR builds on a novel\nneural codec-based architecture that effectively captures environment geometry\nand materials properties and solves speech dereverberation as an auxiliary task\nby using multi-task learning. We also propose Geo-Mat features that augment\nmaterial information into visual cues and CRIP that improves late reverberation\ncomponents in the estimated RIR via image-to-RIR retrieval by 86%. Empirical\nresults show that AV-RIR quantitatively outperforms previous audio-only and\nvisual-only approaches by achieving 36% - 63% improvement across various\nacoustic metrics in RIR estimation. Additionally, it also achieves higher\npreference scores in human evaluation. As an auxiliary benefit, dereverbed\nspeech from AV-RIR shows competitive performance with the state-of-the-art in\nvarious spoken language processing tasks and outperforms reverberation time\nerror score in the real-world AVSpeech dataset. Qualitative examples of both\nsynthesized reverberant speech and enhanced speech can be found at\nhttps://www.youtube.com/watch?v=tTsKhviukAE.\n", "date": "2023-11-30", "categories": []}
{"id": "2404.15854", "title": "CLAD: Robust Audio Deepfake Detection Against Manipulation Attacks with\n  Contrastive Learning", "abstract": "  The increasing prevalence of audio deepfakes poses significant security\nthreats, necessitating robust detection methods. While existing detection\nsystems exhibit promise, their robustness against malicious audio manipulations\nremains underexplored. To bridge the gap, we undertake the first comprehensive\nstudy of the susceptibility of the most widely adopted audio deepfake detectors\nto manipulation attacks. Surprisingly, even manipulations like volume control\ncan significantly bypass detection without affecting human perception. To\naddress this, we propose CLAD (Contrastive Learning-based Audio deepfake\nDetector) to enhance the robustness against manipulation attacks. The key idea\nis to incorporate contrastive learning to minimize the variations introduced by\nmanipulations, therefore enhancing detection robustness. Additionally, we\nincorporate a length loss, aiming to improve the detection accuracy by\nclustering real audios more closely in the feature space. We comprehensively\nevaluated the most widely adopted audio deepfake detection models and our\nproposed CLAD against various manipulation attacks. The detection models\nexhibited vulnerabilities, with FAR rising to 36.69%, 31.23%, and 51.28% under\nvolume control, fading, and noise injection, respectively. CLAD enhanced\nrobustness, reducing the FAR to 0.81% under noise injection and consistently\nmaintaining an FAR below 1.63% across all tests. Our source code and\ndocumentation are available in the artifact repository\n(https://github.com/CLAD23/CLAD).\n", "date": "2024-04-24", "categories": []}
{"id": "2404.13789", "title": "Anchor-aware Deep Metric Learning for Audio-visual Retrieval", "abstract": "  Metric learning minimizes the gap between similar (positive) pairs of data\npoints and increases the separation of dissimilar (negative) pairs, aiming at\ncapturing the underlying data structure and enhancing the performance of tasks\nlike audio-visual cross-modal retrieval (AV-CMR). Recent works employ sampling\nmethods to select impactful data points from the embedding space during\ntraining. However, the model training fails to fully explore the space due to\nthe scarcity of training data points, resulting in an incomplete representation\nof the overall positive and negative distributions. In this paper, we propose\nan innovative Anchor-aware Deep Metric Learning (AADML) method to address this\nchallenge by uncovering the underlying correlations among existing data points,\nwhich enhances the quality of the shared embedding space. Specifically, our\nmethod establishes a correlation graph-based manifold structure by considering\nthe dependencies between each sample as the anchor and its semantically similar\nsamples. Through dynamic weighting of the correlations within this underlying\nmanifold structure using an attention-driven mechanism, Anchor Awareness (AA)\nscores are obtained for each anchor. These AA scores serve as data proxies to\ncompute relative distances in metric learning approaches. Extensive experiments\nconducted on two audio-visual benchmark datasets demonstrate the effectiveness\nof our proposed AADML method, significantly surpassing state-of-the-art models.\nFurthermore, we investigate the integration of AA proxies with various metric\nlearning methods, further highlighting the efficacy of our approach.\n", "date": "2024-04-21", "categories": []}
{"id": "2404.13892", "title": "Retrieval-Augmented Audio Deepfake Detection", "abstract": "  With recent advances in speech synthesis including text-to-speech (TTS) and\nvoice conversion (VC) systems enabling the generation of ultra-realistic audio\ndeepfakes, there is growing concern about their potential misuse. However, most\ndeepfake (DF) detection methods rely solely on the fuzzy knowledge learned by a\nsingle model, resulting in performance bottlenecks and transparency issues.\nInspired by retrieval-augmented generation (RAG), we propose a\nretrieval-augmented detection (RAD) framework that augments test samples with\nsimilar retrieved samples for enhanced detection. We also extend the\nmulti-fusion attentive classifier to integrate it with our proposed RAD\nframework. Extensive experiments show the superior performance of the proposed\nRAD framework over baseline methods, achieving state-of-the-art results on the\nASVspoof 2021 DF set and competitive results on the 2019 and 2021 LA sets.\nFurther sample analysis indicates that the retriever consistently retrieves\nsamples mostly from the same speaker with acoustic characteristics highly\nconsistent with the query audio, thereby improving detection performance.\n", "date": "2024-04-22", "categories": []}
{"id": "2404.15168", "title": "Artificial Neural Networks to Recognize Speakers Division from\n  Continuous Bengali Speech", "abstract": "  Voice based applications are ruling over the era of automation because speech\nhas a lot of factors that determine a speakers information as well as speech.\nModern Automatic Speech Recognition (ASR) is a blessing in the field of\nHuman-Computer Interaction (HCI) for efficient communication among humans and\ndevices using Artificial Intelligence technology. Speech is one of the easiest\nmediums of communication because it has a lot of identical features for\ndifferent speakers. Nowadays it is possible to determine speakers and their\nidentity using their speech in terms of speaker recognition. In this paper, we\npresented a method that will provide a speakers geographical identity in a\ncertain region using continuous Bengali speech. We consider eight different\ndivisions of Bangladesh as the geographical region. We applied the Mel\nFrequency Cepstral Coefficient (MFCC) and Delta features on an Artificial\nNeural Network to classify speakers division. We performed some preprocessing\ntasks like noise reduction and 8-10 second segmentation of raw audio before\nfeature extraction. We used our dataset of more than 45 hours of audio data\nfrom 633 individual male and female speakers. We recorded the highest accuracy\nof 85.44%.\n", "date": "2024-04-18", "categories": []}
{"id": "2404.13509", "title": "MFHCA: Enhancing Speech Emotion Recognition Via Multi-Spatial Fusion and\n  Hierarchical Cooperative Attention", "abstract": "  Speech emotion recognition is crucial in human-computer interaction, but\nextracting and using emotional cues from audio poses challenges. This paper\nintroduces MFHCA, a novel method for Speech Emotion Recognition using\nMulti-Spatial Fusion and Hierarchical Cooperative Attention on spectrograms and\nraw audio. We employ the Multi-Spatial Fusion module (MF) to efficiently\nidentify emotion-related spectrogram regions and integrate Hubert features for\nhigher-level acoustic information. Our approach also includes a Hierarchical\nCooperative Attention module (HCA) to merge features from various auditory\nlevels. We evaluate our method on the IEMOCAP dataset and achieve 2.6\\% and\n1.87\\% improvements on the weighted accuracy and unweighted accuracy,\nrespectively. Extensive experiments demonstrate the effectiveness of the\nproposed method.\n", "date": "2024-04-21", "categories": []}
{"id": "2404.14063", "title": "LVNS-RAVE: Diversified audio generation with RAVE and Latent Vector\n  Novelty Search", "abstract": "  Evolutionary Algorithms and Generative Deep Learning have been two of the\nmost powerful tools for sound generation tasks. However, they have limitations:\nEvolutionary Algorithms require complicated designs, posing challenges in\ncontrol and achieving realistic sound generation. Generative Deep Learning\nmodels often copy from the dataset and lack creativity. In this paper, we\npropose LVNS-RAVE, a method to combine Evolutionary Algorithms and Generative\nDeep Learning to produce realistic and novel sounds. We use the RAVE model as\nthe sound generator and the VGGish model as a novelty evaluator in the Latent\nVector Novelty Search (LVNS) algorithm. The reported experiments show that the\nmethod can successfully generate diversified, novel audio samples under\ndifferent mutation setups using different pre-trained RAVE models. The\ncharacteristics of the generation process can be easily controlled with the\nmutation parameters. The proposed algorithm can be a creative tool for sound\nartists and musicians.\n", "date": "2024-04-22", "categories": []}
{"id": "2404.12077", "title": "TIMIT Speaker Profiling: A Comparison of Multi-task learning and\n  Single-task learning Approaches", "abstract": "  This study employs deep learning techniques to explore four speaker profiling\ntasks on the TIMIT dataset, namely gender classification, accent\nclassification, age estimation, and speaker identification, highlighting the\npotential and challenges of multi-task learning versus single-task models. The\nmotivation for this research is twofold: firstly, to empirically assess the\nadvantages and drawbacks of multi-task learning over single-task models in the\ncontext of speaker profiling; secondly, to emphasize the undiminished\nsignificance of skillful feature engineering for speaker recognition tasks. The\nfindings reveal challenges in accent classification, and multi-task learning is\nfound advantageous for tasks of similar complexity. Non-sequential features are\nfavored for speaker recognition, but sequential ones can serve as starting\npoints for complex models. The study underscores the necessity of meticulous\nexperimentation and parameter tuning for deep learning models.\n", "date": "2024-04-18", "categories": []}
{"id": "2404.11116", "title": "Music Enhancement with Deep Filters: A Technical Report for The ICASSP\n  2024 Cadenza Challenge", "abstract": "  In this challenge, we disentangle the deep filters from the original\nDeepfilterNet and incorporate them into our Spec-UNet-based network to further\nimprove a hybrid Demucs (hdemucs) based remixing pipeline. The motivation\nbehind the use of the deep filter component lies at its potential in better\nhandling temporal fine structures. We demonstrate an incremental improvement in\nboth the Signal-to-Distortion Ratio (SDR) and the Hearing Aid Audio Quality\nIndex (HAAQI) metrics when comparing the performance of hdemucs against\ndifferent versions of our model.\n", "date": "2024-04-17", "categories": []}
{"id": "2404.09177", "title": "An Experimental Comparison Of Multi-view Self-supervised Methods For\n  Music Tagging", "abstract": "  Self-supervised learning has emerged as a powerful way to pre-train\ngeneralizable machine learning models on large amounts of unlabeled data. It is\nparticularly compelling in the music domain, where obtaining labeled data is\ntime-consuming, error-prone, and ambiguous. During the self-supervised process,\nmodels are trained on pretext tasks, with the primary objective of acquiring\nrobust and informative features that can later be fine-tuned for specific\ndownstream tasks. The choice of the pretext task is critical as it guides the\nmodel to shape the feature space with meaningful constraints for information\nencoding. In the context of music, most works have relied on contrastive\nlearning or masking techniques. In this study, we expand the scope of pretext\ntasks applied to music by investigating and comparing the performance of new\nself-supervised methods for music tagging. We open-source a simple ResNet model\ntrained on a diverse catalog of millions of tracks. Our results demonstrate\nthat, although most of these pre-training methods result in similar downstream\nresults, contrastive learning consistently results in better downstream\nperformance compared to other self-supervised pre-training methods. This holds\ntrue in a limited-data downstream context.\n", "date": "2024-04-14", "categories": []}
{"id": "2404.07616", "title": "Audio Dialogues: Dialogues dataset for audio and music understanding", "abstract": "  Existing datasets for audio understanding primarily focus on single-turn\ninteractions (i.e. audio captioning, audio question answering) for describing\naudio in natural language, thus limiting understanding audio via interactive\ndialogue. To address this gap, we introduce Audio Dialogues: a multi-turn\ndialogue dataset containing 163.8k samples for general audio sounds and music.\nIn addition to dialogues, Audio Dialogues also has question-answer pairs to\nunderstand and compare multiple input audios together. Audio Dialogues\nleverages a prompting-based approach and caption annotations from existing\ndatasets to generate multi-turn dialogues using a Large Language Model (LLM).\nWe evaluate existing audio-augmented large language models on our proposed\ndataset to demonstrate the complexity and applicability of Audio Dialogues. Our\ncode for generating the dataset will be made publicly available. Detailed\nprompts and generated dialogues can be found on the demo website\nhttps://audiodialogues.github.io/.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.06818", "title": "Towards Efficient and Real-Time Piano Transcription Using Neural\n  Autoregressive Models", "abstract": "  In recent years, advancements in neural network designs and the availability\nof large-scale labeled datasets have led to significant improvements in the\naccuracy of piano transcription models. However, most previous work focused on\nhigh-performance offline transcription, neglecting deliberate consideration of\nmodel size. The goal of this work is to implement real-time inference for piano\ntranscription while ensuring both high performance and lightweight. To this\nend, we propose novel architectures for convolutional recurrent neural\nnetworks, redesigning an existing autoregressive piano transcription model.\nFirst, we extend the acoustic module by adding a frequency-conditioned FiLM\nlayer to the CNN module to adapt the convolutional filters on the frequency\naxis. Second, we improve note-state sequence modeling by using a pitchwise LSTM\nthat focuses on note-state transitions within a note. In addition, we augment\nthe autoregressive connection with an enhanced recursive context. Using these\ncomponents, we propose two types of models; one for high performance and the\nother for high compactness. Through extensive experiments, we show that the\nproposed models are comparable to state-of-the-art models in terms of note\naccuracy on the MAESTRO dataset. We also investigate the effective model size\nand real-time inference latency by gradually streamlining the architecture.\nFinally, we conduct cross-data evaluation on unseen piano datasets and in-depth\nanalysis to elucidate the effect of the proposed components in the view of note\nlength and pitch range.\n", "date": "2024-04-10", "categories": []}
{"id": "2402.10100", "title": "Tuning In: Analysis of Audio Classifier Performance in Clinical Settings\n  with Limited Data", "abstract": "  This study assesses deep learning models for audio classification in a\nclinical setting with the constraint of small datasets reflecting real-world\nprospective data collection. We analyze CNNs, including DenseNet and ConvNeXt,\nalongside transformer models like ViT, SWIN, and AST, and compare them against\npre-trained audio models such as YAMNet and VGGish. Our method highlights the\nbenefits of pre-training on large datasets before fine-tuning on specific\nclinical data. We prospectively collected two first-of-their-kind patient audio\ndatasets from stroke patients. We investigated various preprocessing\ntechniques, finding that RGB and grayscale spectrogram transformations affect\nmodel performance differently based on the priors they learn from pre-training.\nOur findings indicate CNNs can match or exceed transformer models in small\ndataset contexts, with DenseNet-Contrastive and AST models showing notable\nperformance. This study highlights the significance of incremental marginal\ngains through model selection, pre-training, and preprocessing in sound\nclassification; this offers valuable insights for clinical diagnostics that\nrely on audio classification.\n", "date": "2024-02-07", "categories": []}
{"id": "2211.06073", "title": "SceneFake: An Initial Dataset and Benchmarks for Scene Fake Audio\n  Detection", "abstract": "  Many datasets have been designed to further the development of fake audio\ndetection. However, fake utterances in previous datasets are mostly generated\nby altering timbre, prosody, linguistic content or channel noise of original\naudio. These datasets leave out a scenario, in which the acoustic scene of an\noriginal audio is manipulated with a forged one. It will pose a major threat to\nour society if some people misuse the manipulated audio with malicious purpose.\nTherefore, this motivates us to fill in the gap. This paper proposes such a\ndataset for scene fake audio detection named SceneFake, where a manipulated\naudio is generated by only tampering with the acoustic scene of an real\nutterance by using speech enhancement technologies. Some scene fake audio\ndetection benchmark results on the SceneFake dataset are reported in this\npaper. In addition, an analysis of fake attacks with different speech\nenhancement technologies and signal-to-noise ratios are presented in this\npaper. The results indicate that scene fake utterances cannot be reliably\ndetected by baseline models trained on the ASVspoof 2019 dataset. Although\nthese models perform well on the SceneFake training set and seen testing set,\ntheir performance is poor on the unseen test set. The dataset\n(https://zenodo.org/record/7663324#.Y_XKMuPYuUk) and benchmark source codes\n(https://github.com/ADDchallenge/SceneFake) are publicly available.\n", "date": "2022-11-11", "categories": []}
{"id": "2404.01611", "title": "Audio Simulation for Sound Source Localization in Virtual Evironment", "abstract": "  Non-line-of-sight localization in signal-deprived environments is a\nchallenging yet pertinent problem. Acoustic methods in such predominantly\nindoor scenarios encounter difficulty due to the reverberant nature. In this\nstudy, we aim to locate sound sources to specific locations within a virtual\nenvironment by leveraging physically grounded sound propagation simulations and\nmachine learning methods. This process attempts to overcome the issue of data\ninsufficiency to localize sound sources to their location of occurrence\nespecially in post-event localization. We achieve 0.786+/- 0.0136 F1-score\nusing an audio transformer spectrogram approach.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.01740", "title": "Weakly-supervised Audio Separation via Bi-modal Semantic Similarity", "abstract": "  Conditional sound separation in multi-source audio mixtures without having\naccess to single source sound data during training is a long standing\nchallenge. Existing mix-and-separate based methods suffer from significant\nperformance drop with multi-source training mixtures due to the lack of\nsupervision signal for single source separation cases during training. However,\nin the case of language-conditional audio separation, we do have access to\ncorresponding text descriptions for each audio mixture in our training data,\nwhich can be seen as (rough) representations of the audio samples in the\nlanguage modality. To this end, in this paper, we propose a generic bi-modal\nseparation framework which can enhance the existing unsupervised frameworks to\nseparate single-source signals in a target modality (i.e., audio) using the\neasily separable corresponding signals in the conditioning modality (i.e.,\nlanguage), without having access to single-source samples in the target\nmodality during training. We empirically show that this is well within reach if\nwe have access to a pretrained joint embedding model between the two modalities\n(i.e., CLAP). Furthermore, we propose to incorporate our framework into two\nfundamental scenarios to enhance separation performance. First, we show that\nour proposed methodology significantly improves the performance of purely\nunsupervised baselines by reducing the distribution shift between training and\ntest samples. In particular, we show that our framework can achieve 71% boost\nin terms of Signal-to-Distortion Ratio (SDR) over the baseline, reaching 97.5%\nof the supervised learning performance. Second, we show that we can further\nimprove the performance of the supervised learning itself by 17% if we augment\nit by our proposed weakly-supervised framework, that enables a powerful\nsemi-supervised framework for audio separation.\n", "date": "2024-04-02", "categories": []}
{"id": "2311.09117", "title": "R-Spin: Efficient Speaker and Noise-invariant Representation Learning\n  with Acoustic Pieces", "abstract": "  This paper introduces Robust Spin (R-Spin), a data-efficient domain-specific\nself-supervision method for speaker and noise-invariant speech representations\nby learning discrete acoustic units with speaker-invariant clustering (Spin).\nR-Spin resolves Spin's issues and enhances content representations by learning\nto predict acoustic pieces. R-Spin offers a 12X reduction in computational\nresources compared to previous state-of-the-art methods while outperforming\nthem in severely distorted speech scenarios. This paper provides detailed\nanalyses to show how discrete units contribute to speech encoder training and\nimproving robustness in diverse acoustic environments.\n", "date": "2023-11-15", "categories": []}
{"id": "2404.01058", "title": "A Novel Audio Representation for Music Genre Identification in MIR", "abstract": "  For Music Information Retrieval downstream tasks, the most common audio\nrepresentation is time-frequency-based, such as Mel spectrograms. In order to\nidentify musical genres, this study explores the possibilities of a new form of\naudio representation one of the most usual MIR downstream tasks. Therefore, to\ndiscretely encoding music using deep vector quantization; a novel audio\nrepresentation was created for the innovative generative music model i.e.\nJukebox. The effectiveness of Jukebox's audio representation is compared to Mel\nspectrograms using a dataset that is almost equivalent to State-of-the-Art\n(SOTA) and an almost same transformer design. The results of this study imply\nthat, at least when the transformers are pretrained using a very modest dataset\nof 20k tracks, Jukebox's audio representation is not superior to Mel\nspectrograms. This could be explained by the fact that Jukebox's audio\nrepresentation does not sufficiently take into account the peculiarities of\nhuman hearing perception. On the other hand, Mel spectrograms are specifically\ncreated with the human auditory sense in mind.\n", "date": "2024-04-01", "categories": []}
{"id": "2403.20184", "title": "Exploring Pathological Speech Quality Assessment with ASR-Powered\n  Wav2Vec2 in Data-Scarce Context", "abstract": "  Automatic speech quality assessment has raised more attention as an\nalternative or support to traditional perceptual clinical evaluation. However,\nmost research so far only gains good results on simple tasks such as binary\nclassification, largely due to data scarcity. To deal with this challenge,\ncurrent works tend to segment patients' audio files into many samples to\naugment the datasets. Nevertheless, this approach has limitations, as it\nindirectly relates overall audio scores to individual segments. This paper\nintroduces a novel approach where the system learns at the audio level instead\nof segments despite data scarcity. This paper proposes to use the pre-trained\nWav2Vec2 architecture for both SSL, and ASR as feature extractor in speech\nassessment. Carried out on the HNC dataset, our ASR-driven approach established\na new baseline compared with other approaches, obtaining average $MSE=0.73$ and\n$MSE=1.15$ for the prediction of intelligibility and severity scores\nrespectively, using only 95 training samples. It shows that the ASR based\nWav2Vec2 model brings the best results and may indicate a strong correlation\nbetween ASR and speech quality assessment. We also measure its ability on\nvariable segment durations and speech content, exploring factors influencing\nits decision.\n", "date": "2024-03-29", "categories": []}
{"id": "2403.19441", "title": "A Novel Stochastic Transformer-based Approach for Post-Traumatic Stress\n  Disorder Detection using Audio Recording of Clinical Interviews", "abstract": "  Post-traumatic stress disorder (PTSD) is a mental disorder that can be\ndeveloped after witnessing or experiencing extremely traumatic events. PTSD can\naffect anyone, regardless of ethnicity, or culture. An estimated one in every\neleven people will experience PTSD during their lifetime. The\nClinician-Administered PTSD Scale (CAPS) and the PTSD Check List for Civilians\n(PCL-C) interviews are gold standards in the diagnosis of PTSD. These\nquestionnaires can be fooled by the subject's responses. This work proposes a\ndeep learning-based approach that achieves state-of-the-art performances for\nPTSD detection using audio recordings during clinical interviews. Our approach\nis based on MFCC low-level features extracted from audio recordings of clinical\ninterviews, followed by deep high-level learning using a Stochastic\nTransformer. Our proposed approach achieves state-of-the-art performances with\nan RMSE of 2.92 on the eDAIC dataset thanks to the stochastic depth, stochastic\ndeep learning layers, and stochastic activation function.\n", "date": "2024-03-28", "categories": []}
{"id": "2308.12882", "title": "LCANets++: Robust Audio Classification using Multi-layer Neural Networks\n  with Lateral Competition", "abstract": "  Audio classification aims at recognizing audio signals, including speech\ncommands or sound events. However, current audio classifiers are susceptible to\nperturbations and adversarial attacks. In addition, real-world audio\nclassification tasks often suffer from limited labeled data. To help bridge\nthese gaps, previous work developed neuro-inspired convolutional neural\nnetworks (CNNs) with sparse coding via the Locally Competitive Algorithm (LCA)\nin the first layer (i.e., LCANets) for computer vision. LCANets learn in a\ncombination of supervised and unsupervised learning, reducing dependency on\nlabeled samples. Motivated by the fact that auditory cortex is also sparse, we\nextend LCANets to audio recognition tasks and introduce LCANets++, which are\nCNNs that perform sparse coding in multiple layers via LCA. We demonstrate that\nLCANets++ are more robust than standard CNNs and LCANets against perturbations,\ne.g., background noise, as well as black-box and white-box attacks, e.g.,\nevasion and fast gradient sign (FGSM) attacks.\n", "date": "2023-08-23", "categories": []}
{"id": "2403.18821", "title": "Real Acoustic Fields: An Audio-Visual Room Acoustics Dataset and\n  Benchmark", "abstract": "  We present a new dataset called Real Acoustic Fields (RAF) that captures real\nacoustic room data from multiple modalities. The dataset includes high-quality\nand densely captured room impulse response data paired with multi-view images,\nand precise 6DoF pose tracking data for sound emitters and listeners in the\nrooms. We used this dataset to evaluate existing methods for novel-view\nacoustic synthesis and impulse response generation which previously relied on\nsynthetic data. In our evaluation, we thoroughly assessed existing audio and\naudio-visual models against multiple criteria and proposed settings to enhance\ntheir performance on real-world data. We also conducted experiments to\ninvestigate the impact of incorporating visual data (i.e., images and depth)\ninto neural acoustic field models. Additionally, we demonstrated the\neffectiveness of a simple sim2real approach, where a model is pre-trained with\nsimulated data and fine-tuned with sparse real-world data, resulting in\nsignificant improvements in the few-shot learning approach. RAF is the first\ndataset to provide densely captured room acoustic data, making it an ideal\nresource for researchers working on audio and audio-visual neural acoustic\nfield modeling techniques. Demos and datasets are available on our project\npage: https://facebookresearch.github.io/real-acoustic-fields/\n", "date": "2024-03-27", "categories": []}
{"id": "2403.17379", "title": "Exploring and Applying Audio-Based Sentiment Analysis in Music", "abstract": "  Sentiment analysis is a continuously explored area of text processing that\ndeals with the computational analysis of opinions, sentiments, and subjectivity\nof text. However, this idea is not limited to text and speech, in fact, it\ncould be applied to other modalities. In reality, humans do not express\nthemselves in text as deeply as they do in music. The ability of a\ncomputational model to interpret musical emotions is largely unexplored and\ncould have implications and uses in therapy and musical queuing. In this paper,\ntwo individual tasks are addressed. This study seeks to (1) predict the emotion\nof a musical clip over time and (2) determine the next emotion value after the\nmusic in a time series to ensure seamless transitions. Utilizing data from the\nEmotions in Music Database, which contains clips of songs selected from the\nFree Music Archive annotated with levels of valence and arousal as reported on\nRussel's circumplex model of affect by multiple volunteers, models are trained\nfor both tasks. Overall, the performance of these models reflected that they\nwere able to perform the tasks they were designed for effectively and\naccurately.\n", "date": "2024-02-22", "categories": []}
{"id": "2212.10744", "title": "An Audio-Visual Speech Separation Model Inspired by\n  Cortico-Thalamo-Cortical Circuits", "abstract": "  Audio-visual approaches involving visual inputs have laid the foundation for\nrecent progress in speech separation. However, the optimization of the\nconcurrent usage of auditory and visual inputs is still an active research\narea. Inspired by the cortico-thalamo-cortical circuit, in which the sensory\nprocessing mechanisms of different modalities modulate one another via the\nnon-lemniscal sensory thalamus, we propose a novel cortico-thalamo-cortical\nneural network (CTCNet) for audio-visual speech separation (AVSS). First, the\nCTCNet learns hierarchical auditory and visual representations in a bottom-up\nmanner in separate auditory and visual subnetworks, mimicking the functions of\nthe auditory and visual cortical areas. Then, inspired by the large number of\nconnections between cortical regions and the thalamus, the model fuses the\nauditory and visual information in a thalamic subnetwork through top-down\nconnections. Finally, the model transmits this fused information back to the\nauditory and visual subnetworks, and the above process is repeated several\ntimes. The results of experiments on three speech separation benchmark datasets\nshow that CTCNet remarkably outperforms existing AVSS methods with considerably\nfewer parameters. These results suggest that mimicking the anatomical\nconnectome of the mammalian brain has great potential for advancing the\ndevelopment of deep neural networks. Project repo is\nhttps://github.com/JusperLee/CTCNet.\n", "date": "2022-12-21", "categories": []}
{"id": "2309.02836", "title": "BigVSAN: Enhancing GAN-based Neural Vocoders with Slicing Adversarial\n  Network", "abstract": "  Generative adversarial network (GAN)-based vocoders have been intensively\nstudied because they can synthesize high-fidelity audio waveforms faster than\nreal-time. However, it has been reported that most GANs fail to obtain the\noptimal projection for discriminating between real and fake data in the feature\nspace. In the literature, it has been demonstrated that slicing adversarial\nnetwork (SAN), an improved GAN training framework that can find the optimal\nprojection, is effective in the image generation task. In this paper, we\ninvestigate the effectiveness of SAN in the vocoding task. For this purpose, we\npropose a scheme to modify least-squares GAN, which most GAN-based vocoders\nadopt, so that their loss functions satisfy the requirements of SAN. Through\nour experiments, we demonstrate that SAN can improve the performance of\nGAN-based vocoders, including BigVGAN, with small modifications. Our code is\navailable at https://github.com/sony/bigvsan.\n", "date": "2023-09-06", "categories": []}
{"id": "2403.16331", "title": "Modeling Analog Dynamic Range Compressors using Deep Learning and\n  State-space Models", "abstract": "  We describe a novel approach for developing realistic digital models of\ndynamic range compressors for digital audio production by analyzing their\nanalog prototypes. While realistic digital dynamic compressors are potentially\nuseful for many applications, the design process is challenging because the\ncompressors operate nonlinearly over long time scales. Our approach is based on\nthe structured state space sequence model (S4), as implementing the state-space\nmodel (SSM) has proven to be efficient at learning long-range dependencies and\nis promising for modeling dynamic range compressors. We present in this paper a\ndeep learning model with S4 layers to model the Teletronix LA-2A analog dynamic\nrange compressor. The model is causal, executes efficiently in real time, and\nachieves roughly the same quality as previous deep-learning models but with\nfewer parameters.\n", "date": "2024-03-24", "categories": []}
{"id": "2309.17189", "title": "RTFS-Net: Recurrent Time-Frequency Modelling for Efficient Audio-Visual\n  Speech Separation", "abstract": "  Audio-visual speech separation methods aim to integrate different modalities\nto generate high-quality separated speech, thereby enhancing the performance of\ndownstream tasks such as speech recognition. Most existing state-of-the-art\n(SOTA) models operate in the time domain. However, their overly simplistic\napproach to modeling acoustic features often necessitates larger and more\ncomputationally intensive models in order to achieve SOTA performance. In this\npaper, we present a novel time-frequency domain audio-visual speech separation\nmethod: Recurrent Time-Frequency Separation Network (RTFS-Net), which applies\nits algorithms on the complex time-frequency bins yielded by the Short-Time\nFourier Transform. We model and capture the time and frequency dimensions of\nthe audio independently using a multi-layered RNN along each dimension.\nFurthermore, we introduce a unique attention-based fusion technique for the\nefficient integration of audio and visual information, and a new mask\nseparation approach that takes advantage of the intrinsic spectral nature of\nthe acoustic features for a clearer separation. RTFS-Net outperforms the prior\nSOTA method in both inference speed and separation quality while reducing the\nnumber of parameters by 90% and MACs by 83%. This is the first time-frequency\ndomain audio-visual speech separation method to outperform all contemporary\ntime-domain counterparts.\n", "date": "2023-09-29", "categories": []}
{"id": "2403.14048", "title": "The NeurIPS 2023 Machine Learning for Audio Workshop: Affective Audio\n  Benchmarks and Novel Data", "abstract": "  The NeurIPS 2023 Machine Learning for Audio Workshop brings together machine\nlearning (ML) experts from various audio domains. There are several valuable\naudio-driven ML tasks, from speech emotion recognition to audio event\ndetection, but the community is sparse compared to other ML areas, e.g.,\ncomputer vision or natural language processing. A major limitation with audio\nis the available data; with audio being a time-dependent modality, high-quality\ndata collection is time-consuming and costly, making it challenging for\nacademic groups to apply their often state-of-the-art strategies to a larger,\nmore generalizable dataset. In this short white paper, to encourage researchers\nwith limited access to large-datasets, the organizers first outline several\nopen-source datasets that are available to the community, and for the duration\nof the workshop are making several propriety datasets available. Namely, three\nvocal datasets, Hume-Prosody, Hume-VocalBurst, an acted emotional speech\ndataset Modulate-Sonata, and an in-game streamer dataset Modulate-Stream. We\noutline the current baselines on these datasets but encourage researchers from\nacross audio to utilize them outside of the initial baseline tasks.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.14246", "title": "CATSE: A Context-Aware Framework for Causal Target Sound Extraction", "abstract": "  Target Sound Extraction (TSE) focuses on the problem of separating sources of\ninterest, indicated by a user's cue, from the input mixture. Most existing\nsolutions operate in an offline fashion and are not suited to the low-latency\ncausal processing constraints imposed by applications in live-streamed content\nsuch as augmented hearing. We introduce a family of context-aware low-latency\ncausal TSE models suitable for real-time processing. First, we explore the\nutility of context by providing the TSE model with oracle information about\nwhat sound classes make up the input mixture, where the objective of the model\nis to extract one or more sources of interest indicated by the user. Since the\npractical applications of oracle models are limited due to their assumptions,\nwe introduce a composite multi-task training objective involving separation and\nclassification losses. Our evaluation involving single- and multi-source\nextraction shows the benefit of using context information in the model either\nby means of providing full context or via the proposed multi-task training loss\nwithout the need for full context information. Specifically, we show that our\nproposed model outperforms size- and latency-matched Waveformer, a\nstate-of-the-art model for real-time TSE.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.14290", "title": "Exploring Green AI for Audio Deepfake Detection", "abstract": "  The state-of-the-art audio deepfake detectors leveraging deep neural networks\nexhibit impressive recognition performance. Nonetheless, this advantage is\naccompanied by a significant carbon footprint. This is mainly due to the use of\nhigh-performance computing with accelerators and high training time. Studies\nshow that average deep NLP model produces around 626k lbs of\nCO\\textsubscript{2} which is equivalent to five times of average US car\nemission at its lifetime. This is certainly a massive threat to the\nenvironment. To tackle this challenge, this study presents a novel framework\nfor audio deepfake detection that can be seamlessly trained using standard CPU\nresources. Our proposed framework utilizes off-the-shelve self-supervised\nlearning (SSL) based models which are pre-trained and available in public\nrepositories. In contrast to existing methods that fine-tune SSL models and\nemploy additional deep neural networks for downstream tasks, we exploit\nclassical machine learning algorithms such as logistic regression and shallow\nneural networks using the SSL embeddings extracted using the pre-trained model.\nOur approach shows competitive results compared to the commonly used\nhigh-carbon footprint approaches. In experiments with the ASVspoof 2019 LA\ndataset, we achieve a 0.90\\% equal error rate (EER) with less than 1k trainable\nmodel parameters. To encourage further research in this direction and support\nreproducible results, the Python code will be made publicly accessible\nfollowing acceptance. Github: https://github.com/sahasubhajit/Speech-Spoofing-\n", "date": "2024-03-21", "categories": []}
{"id": "2309.10787", "title": "AV-SUPERB: A Multi-Task Evaluation Benchmark for Audio-Visual\n  Representation Models", "abstract": "  Audio-visual representation learning aims to develop systems with human-like\nperception by utilizing correlation between auditory and visual information.\nHowever, current models often focus on a limited set of tasks, and\ngeneralization abilities of learned representations are unclear. To this end,\nwe propose the AV-SUPERB benchmark that enables general-purpose evaluation of\nunimodal audio/visual and bimodal fusion representations on 7 datasets covering\n5 audio-visual tasks in speech and audio processing. We evaluate 5 recent\nself-supervised models and show that none of these models generalize to all\ntasks, emphasizing the need for future study on improving universal model\nperformance. In addition, we show that representations may be improved with\nintermediate-task fine-tuning and audio event classification with AudioSet\nserves as a strong intermediate task. We release our benchmark with evaluation\ncode and a model submission platform to encourage further research in\naudio-visual learning.\n", "date": "2023-09-19", "categories": []}
{"id": "2403.11091", "title": "Multitask frame-level learning for few-shot sound event detection", "abstract": "  This paper focuses on few-shot Sound Event Detection (SED), which aims to\nautomatically recognize and classify sound events with limited samples.\nHowever, prevailing methods methods in few-shot SED predominantly rely on\nsegment-level predictions, which often providing detailed, fine-grained\npredictions, particularly for events of brief duration. Although frame-level\nprediction strategies have been proposed to overcome these limitations, these\nstrategies commonly face difficulties with prediction truncation caused by\nbackground noise. To alleviate this issue, we introduces an innovative\nmultitask frame-level SED framework. In addition, we introduce TimeFilterAug, a\nlinear timing mask for data augmentation, to increase the model's robustness\nand adaptability to diverse acoustic environments. The proposed method achieves\na F-score of 63.8%, securing the 1st rank in the few-shot bioacoustic event\ndetection category of the Detection and Classification of Acoustic Scenes and\nEvents Challenge 2023.\n", "date": "2024-03-17", "categories": []}
{"id": "2302.02257", "title": "Multi-Source Diffusion Models for Simultaneous Music Generation and\n  Separation", "abstract": "  In this work, we define a diffusion-based generative model capable of both\nmusic synthesis and source separation by learning the score of the joint\nprobability density of sources sharing a context. Alongside the classic total\ninference tasks (i.e., generating a mixture, separating the sources), we also\nintroduce and experiment on the partial generation task of source imputation,\nwhere we generate a subset of the sources given the others (e.g., play a piano\ntrack that goes well with the drums). Additionally, we introduce a novel\ninference method for the separation task based on Dirac likelihood functions.\nWe train our model on Slakh2100, a standard dataset for musical source\nseparation, provide qualitative results in the generation settings, and\nshowcase competitive quantitative results in the source separation setting. Our\nmethod is the first example of a single model that can handle both generation\nand separation tasks, thus representing a step toward general audio models.\n", "date": "2023-02-04", "categories": []}
{"id": "2403.10549", "title": "On-Device Domain Learning for Keyword Spotting on Low-Power Extreme Edge\n  Embedded Systems", "abstract": "  Keyword spotting accuracy degrades when neural networks are exposed to noisy\nenvironments. On-site adaptation to previously unseen noise is crucial to\nrecovering accuracy loss, and on-device learning is required to ensure that the\nadaptation process happens entirely on the edge device. In this work, we\npropose a fully on-device domain adaptation system achieving up to 14% accuracy\ngains over already-robust keyword spotting models. We enable on-device learning\nwith less than 10 kB of memory, using only 100 labeled utterances to recover 5%\naccuracy after adapting to the complex speech noise. We demonstrate that domain\nadaptation can be achieved on ultra-low-power microcontrollers with as little\nas 806 mJ in only 14 s on always-on, battery-operated devices.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.11778", "title": "Towards the Development of a Real-Time Deepfake Audio Detection System\n  in Communication Platforms", "abstract": "  Deepfake audio poses a rising threat in communication platforms,\nnecessitating real-time detection for audio stream integrity. Unlike\ntraditional non-real-time approaches, this study assesses the viability of\nemploying static deepfake audio detection models in real-time communication\nplatforms. An executable software is developed for cross-platform\ncompatibility, enabling real-time execution. Two deepfake audio detection\nmodels based on Resnet and LCNN architectures are implemented using the\nASVspoof 2019 dataset, achieving benchmark performances compared to ASVspoof\n2019 challenge baselines. The study proposes strategies and frameworks for\nenhancing these models, paving the way for real-time deepfake audio detection\nin communication platforms. This work contributes to the advancement of audio\nstream security, ensuring robust detection capabilities in dynamic, real-time\ncommunication scenarios.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.10024", "title": "MR-MT3: Memory Retaining Multi-Track Music Transcription to Mitigate\n  Instrument Leakage", "abstract": "  This paper presents enhancements to the MT3 model, a state-of-the-art (SOTA)\ntoken-based multi-instrument automatic music transcription (AMT) model. Despite\nSOTA performance, MT3 has the issue of instrument leakage, where transcriptions\nare fragmented across different instruments. To mitigate this, we propose\nMR-MT3, with enhancements including a memory retention mechanism, prior token\nsampling, and token shuffling are proposed. These methods are evaluated on the\nSlakh2100 dataset, demonstrating improved onset F1 scores and reduced\ninstrument leakage. In addition to the conventional multi-instrument\ntranscription F1 score, new metrics such as the instrument leakage ratio and\nthe instrument detection F1 score are introduced for a more comprehensive\nassessment of transcription quality. The study also explores the issue of\ndomain overfitting by evaluating MT3 on single-instrument monophonic datasets\nsuch as ComMU and NSynth. The findings, along with the source code, are shared\nto facilitate future work aimed at refining token-based multi-instrument AMT\nmodels.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.09579", "title": "uaMix-MAE: Efficient Tuning of Pretrained Audio Transformers with\n  Unsupervised Audio Mixtures", "abstract": "  Masked Autoencoders (MAEs) learn rich low-level representations from\nunlabeled data but require substantial labeled data to effectively adapt to\ndownstream tasks. Conversely, Instance Discrimination (ID) emphasizes\nhigh-level semantics, offering a potential solution to alleviate annotation\nrequirements in MAEs. Although combining these two approaches can address\ndownstream tasks with limited labeled data, naively integrating ID into MAEs\nleads to extended training times and high computational costs. To address this\nchallenge, we introduce uaMix-MAE, an efficient ID tuning strategy that\nleverages unsupervised audio mixtures. Utilizing contrastive tuning, uaMix-MAE\naligns the representations of pretrained MAEs, thereby facilitating effective\nadaptation to task-specific semantics. To optimize the model with small amounts\nof unlabeled data, we propose an audio mixing technique that manipulates audio\nsamples in both input and virtual label spaces. Experiments in low/few-shot\nsettings demonstrate that \\modelname achieves 4-6% accuracy improvements over\nvarious benchmarks when tuned with limited unlabeled data, such as\nAudioSet-20K. Code is available at https://github.com/PLAN-Lab/uamix-MAE\n", "date": "2024-03-14", "categories": []}
{"id": "2403.05380", "title": "Spectrogram-Based Detection of Auto-Tuned Vocals in Music Recordings", "abstract": "  In the domain of music production and audio processing, the implementation of\nautomatic pitch correction of the singing voice, also known as Auto-Tune, has\nsignificantly transformed the landscape of vocal performance. While auto-tuning\ntechnology has offered musicians the ability to tune their vocal pitches and\nachieve a desired level of precision, its use has also sparked debates\nregarding its impact on authenticity and artistic integrity. As a result,\ndetecting and analyzing Auto-Tuned vocals in music recordings has become\nessential for music scholars, producers, and listeners. However, to the best of\nour knowledge, no prior effort has been made in this direction. This study\nintroduces a data-driven approach leveraging triplet networks for the detection\nof Auto-Tuned songs, backed by the creation of a dataset composed of original\nand Auto-Tuned audio clips. The experimental results demonstrate the\nsuperiority of the proposed method in both accuracy and robustness compared to\nRawnet2, an end-to-end model proposed for anti-spoofing and widely used for\nother audio forensic tasks.\n", "date": "2024-03-08", "categories": []}
{"id": "2401.04577", "title": "Masked Audio Generation using a Single Non-Autoregressive Transformer", "abstract": "  We introduce MAGNeT, a masked generative sequence modeling method that\noperates directly over several streams of audio tokens. Unlike prior work,\nMAGNeT is comprised of a single-stage, non-autoregressive transformer. During\ntraining, we predict spans of masked tokens obtained from a masking scheduler,\nwhile during inference we gradually construct the output sequence using several\ndecoding steps. To further enhance the quality of the generated audio, we\nintroduce a novel rescoring method in which, we leverage an external\npre-trained model to rescore and rank predictions from MAGNeT, which will be\nthen used for later decoding steps. Lastly, we explore a hybrid version of\nMAGNeT, in which we fuse between autoregressive and non-autoregressive models\nto generate the first few seconds in an autoregressive manner while the rest of\nthe sequence is being decoded in parallel. We demonstrate the efficiency of\nMAGNeT for the task of text-to-music and text-to-audio generation and conduct\nan extensive empirical evaluation, considering both objective metrics and human\nstudies. The proposed approach is comparable to the evaluated baselines, while\nbeing significantly faster (x7 faster than the autoregressive baseline).\nThrough ablation studies and analysis, we shed light on the importance of each\nof the components comprising MAGNeT, together with pointing to the trade-offs\nbetween autoregressive and non-autoregressive modeling, considering latency,\nthroughput, and generation quality. Samples are available on our demo page\nhttps://pages.cs.huji.ac.il/adiyoss-lab/MAGNeT.\n", "date": "2024-01-09", "categories": []}
{"id": "2403.02522", "title": "HeAR -- Health Acoustic Representations", "abstract": "  Health acoustic sounds such as coughs and breaths are known to contain useful\nhealth signals with significant potential for monitoring health and disease,\nyet are underexplored in the medical machine learning community. The existing\ndeep learning systems for health acoustics are often narrowly trained and\nevaluated on a single task, which is limited by data and may hinder\ngeneralization to other tasks. To mitigate these gaps, we develop HeAR, a\nscalable self-supervised learning-based deep learning system using masked\nautoencoders trained on a large dataset of 313 million two-second long audio\nclips. Through linear probes, we establish HeAR as a state-of-the-art health\naudio embedding model on a benchmark of 33 health acoustic tasks across 6\ndatasets. By introducing this work, we hope to enable and accelerate further\nhealth acoustics research.\n", "date": "2024-03-04", "categories": []}
{"id": "2403.03095", "title": "Cross Pseudo-Labeling for Semi-Supervised Audio-Visual Source\n  Localization", "abstract": "  Audio-Visual Source Localization (AVSL) is the task of identifying specific\nsounding objects in the scene given audio cues. In our work, we focus on\nsemi-supervised AVSL with pseudo-labeling. To address the issues with vanilla\nhard pseudo-labels including bias accumulation, noise sensitivity, and\ninstability, we propose a novel method named Cross Pseudo-Labeling (XPL),\nwherein two models learn from each other with the cross-refine mechanism to\navoid bias accumulation. We equip XPL with two effective components. Firstly,\nthe soft pseudo-labels with sharpening and pseudo-label exponential moving\naverage mechanisms enable models to achieve gradual self-improvement and ensure\nstable training. Secondly, the curriculum data selection module adaptively\nselects pseudo-labels with high quality during training to mitigate potential\nbias. Experimental results demonstrate that XPL significantly outperforms\nexisting methods, achieving state-of-the-art performance while effectively\nmitigating confirmation bias and ensuring training stability.\n", "date": "2024-03-05", "categories": []}
{"id": "2402.18007", "title": "Mixer is more than just a model", "abstract": "  Recently, MLP structures have regained popularity, with MLP-Mixer standing\nout as a prominent example. In the field of computer vision, MLP-Mixer is noted\nfor its ability to extract data information from both channel and token\nperspectives, effectively acting as a fusion of channel and token information.\nIndeed, Mixer represents a paradigm for information extraction that amalgamates\nchannel and token information. The essence of Mixer lies in its ability to\nblend information from diverse perspectives, epitomizing the true concept of\n\"mixing\" in the realm of neural network architectures. Beyond channel and token\nconsiderations, it is possible to create more tailored mixers from various\nperspectives to better suit specific task requirements. This study focuses on\nthe domain of audio recognition, introducing a novel model named Audio\nSpectrogram Mixer with Roll-Time and Hermit FFT (ASM-RH) that incorporates\ninsights from both time and frequency domains. Experimental results demonstrate\nthat ASM-RH is particularly well-suited for audio data and yields promising\noutcomes across multiple classification tasks. The models and optimal weights\nfiles will be published.\n", "date": "2024-02-28", "categories": []}
{"id": "2403.00854", "title": "Speaker-Independent Dysarthria Severity Classification using\n  Self-Supervised Transformers and Multi-Task Learning", "abstract": "  Dysarthria, a condition resulting from impaired control of the speech muscles\ndue to neurological disorders, significantly impacts the communication and\nquality of life of patients. The condition's complexity, human scoring and\nvaried presentations make its assessment and management challenging. This study\npresents a transformer-based framework for automatically assessing dysarthria\nseverity from raw speech data. It can offer an objective, repeatable,\naccessible, standardised and cost-effective and compared to traditional methods\nrequiring human expert assessors. We develop a transformer framework, called\nSpeaker-Agnostic Latent Regularisation (SALR), incorporating a multi-task\nlearning objective and contrastive learning for speaker-independent multi-class\ndysarthria severity classification. The multi-task framework is designed to\nreduce reliance on speaker-specific characteristics and address the intrinsic\nintra-class variability of dysarthric speech. We evaluated on the Universal\nAccess Speech dataset using leave-one-speaker-out cross-validation, our model\ndemonstrated superior performance over traditional machine learning approaches,\nwith an accuracy of $70.48\\%$ and an F1 score of $59.23\\%$. Our SALR model also\nexceeded the previous benchmark for AI-based classification, which used support\nvector machines, by $16.58\\%$. We open the black box of our model by\nvisualising the latent space where we can observe how the model substantially\nreduces speaker-specific cues and amplifies task-specific ones, thereby showing\nits robustness. In conclusion, SALR establishes a new benchmark in\nspeaker-independent multi-class dysarthria severity classification using\ngenerative AI. The potential implications of our findings for broader clinical\napplications in automated dysarthria severity assessments.\n", "date": "2024-02-29", "categories": []}
{"id": "2402.02694", "title": "Description on IEEE ICME 2024 Grand Challenge: Semi-supervised Acoustic\n  Scene Classification under Domain Shift", "abstract": "  Acoustic scene classification (ASC) is a crucial research problem in\ncomputational auditory scene analysis, and it aims to recognize the unique\nacoustic characteristics of an environment. One of the challenges of the ASC\ntask is the domain shift between training and testing data. Since 2018, ASC\nchallenges have focused on the generalization of ASC models across different\nrecording devices. Although this task, in recent years, has achieved\nsubstantial progress in device generalization, the challenge of domain shift\nbetween different geographical regions, involving discrepancies such as time,\nspace, culture, and language, remains insufficiently explored at present. In\naddition, considering the abundance of unlabeled acoustic scene data in the\nreal world, it is important to study the possible ways to utilize these\nunlabelled data. Therefore, we introduce the task Semi-supervised Acoustic\nScene Classification under Domain Shift in the ICME 2024 Grand Challenge. We\nencourage participants to innovate with semi-supervised learning techniques,\naiming to develop more robust ASC models under domain shift.\n", "date": "2024-02-05", "categories": []}
{"id": "2306.08510", "title": "Permutation Invariant Recurrent Neural Networks for Sound Source\n  Tracking Applications", "abstract": "  Many multi-source localization and tracking models based on neural networks\nuse one or several recurrent layers at their final stages to track the movement\nof the sources. Conventional recurrent neural networks (RNNs), such as the long\nshort-term memories (LSTMs) or the gated recurrent units (GRUs), take a vector\nas their input and use another vector to store their state. However, this\napproach results in the information from all the sources being contained in a\nsingle ordered vector, which is not optimal for permutation-invariant problems\nsuch as multi-source tracking. In this paper, we present a new recurrent\narchitecture that uses unordered sets to represent both its input and its state\nand that is invariant to the permutations of the input set and equivariant to\nthe permutations of the state set. Hence, the information of every sound source\nis represented in an individual embedding and the new estimates are assigned to\nthe tracked trajectories regardless of their order.\n", "date": "2023-06-14", "categories": []}
{"id": "2311.03509", "title": "MFAAN: Unveiling Audio Deepfakes with a Multi-Feature Authenticity\n  Network", "abstract": "  In the contemporary digital age, the proliferation of deepfakes presents a\nformidable challenge to the sanctity of information dissemination. Audio\ndeepfakes, in particular, can be deceptively realistic, posing significant\nrisks in misinformation campaigns. To address this threat, we introduce the\nMulti-Feature Audio Authenticity Network (MFAAN), an advanced architecture\ntailored for the detection of fabricated audio content. MFAAN incorporates\nmultiple parallel paths designed to harness the strengths of different audio\nrepresentations, including Mel-frequency cepstral coefficients (MFCC),\nlinear-frequency cepstral coefficients (LFCC), and Chroma Short Time Fourier\nTransform (Chroma-STFT). By synergistically fusing these features, MFAAN\nachieves a nuanced understanding of audio content, facilitating robust\ndifferentiation between genuine and manipulated recordings. Preliminary\nevaluations of MFAAN on two benchmark datasets, 'In-the-Wild' Audio Deepfake\nData and The Fake-or-Real Dataset, demonstrate its superior performance,\nachieving accuracies of 98.93% and 94.47% respectively. Such results not only\nunderscore the efficacy of MFAAN but also highlight its potential as a pivotal\ntool in the ongoing battle against deepfake audio content.\n", "date": "2023-11-06", "categories": []}
{"id": "2402.17701", "title": "Real-time Low-latency Music Source Separation using Hybrid\n  Spectrogram-TasNet", "abstract": "  There have been significant advances in deep learning for music demixing in\nrecent years. However, there has been little attention given to how these\nneural networks can be adapted for real-time low-latency applications, which\ncould be helpful for hearing aids, remixing audio streams and live shows. In\nthis paper, we investigate the various challenges involved in adapting current\ndemixing models in the literature for this use case. Subsequently, inspired by\nthe Hybrid Demucs architecture, we propose the Hybrid Spectrogram Time-domain\nAudio Separation Network HS-TasNet, which utilises the advantages of spectral\nand waveform domains. For a latency of 23 ms, the HS-TasNet obtains an overall\nsignal-to-distortion ratio (SDR) of 4.65 on the MusDB test set, and increases\nto 5.55 with additional training data. These results demonstrate the potential\nof efficient demixing for real-time low-latency music applications.\n", "date": "2024-02-27", "categories": []}
{"id": "2402.15569", "title": "Toward Fully Self-Supervised Multi-Pitch Estimation", "abstract": "  Multi-pitch estimation is a decades-long research problem involving the\ndetection of pitch activity associated with concurrent musical events within\nmulti-instrument mixtures. Supervised learning techniques have demonstrated\nsolid performance on more narrow characterizations of the task, but suffer from\nlimitations concerning the shortage of large-scale and diverse polyphonic music\ndatasets with multi-pitch annotations. We present a suite of self-supervised\nlearning objectives for multi-pitch estimation, which encourage the\nconcentration of support around harmonics, invariance to timbral\ntransformations, and equivariance to geometric transformations. These\nobjectives are sufficient to train an entirely convolutional autoencoder to\nproduce multi-pitch salience-grams directly, without any fine-tuning. Despite\ntraining exclusively on a collection of synthetic single-note audio samples,\nour fully self-supervised framework generalizes to polyphonic music mixtures,\nand achieves performance comparable to supervised models trained on\nconventional multi-pitch datasets.\n", "date": "2024-02-23", "categories": []}
{"id": "2402.16321", "title": "Self-Supervised Speech Quality Estimation and Enhancement Using Only\n  Clean Speech", "abstract": "  Speech quality estimation has recently undergone a paradigm shift from\nhuman-hearing expert designs to machine-learning models. However, current\nmodels rely mainly on supervised learning, which is time-consuming and\nexpensive for label collection. To solve this problem, we propose VQScore, a\nself-supervised metric for evaluating speech based on the quantization error of\na vector-quantized-variational autoencoder (VQ-VAE). The training of VQ-VAE\nrelies on clean speech; hence, large quantization errors can be expected when\nthe speech is distorted. To further improve correlation with real quality\nscores, domain knowledge of speech processing is incorporated into the model\ndesign. We found that the vector quantization mechanism could also be used for\nself-supervised speech enhancement (SE) model training. To improve the\nrobustness of the encoder for SE, a novel self-distillation mechanism combined\nwith adversarial training is introduced. In summary, the proposed speech\nquality estimation method and enhancement models require only clean speech for\ntraining without any label requirements. Experimental results show that the\nproposed VQScore and enhancement model are competitive with supervised\nbaselines. The code will be released after publication.\n", "date": "2024-02-26", "categories": []}
{"id": "2312.11234", "title": "Perceptual Musical Features for Interpretable Audio Tagging", "abstract": "  In the age of music streaming platforms, the task of automatically tagging\nmusic audio has garnered significant attention, driving researchers to devise\nmethods aimed at enhancing performance metrics on standard datasets. Most\nrecent approaches rely on deep neural networks, which, despite their impressive\nperformance, possess opacity, making it challenging to elucidate their output\nfor a given input. While the issue of interpretability has been emphasized in\nother fields like medicine, it has not received attention in music-related\ntasks. In this study, we explored the relevance of interpretability in the\ncontext of automatic music tagging. We constructed a workflow that incorporates\nthree different information extraction techniques: a) leveraging symbolic\nknowledge, b) utilizing auxiliary deep neural networks, and c) employing signal\nprocessing to extract perceptual features from audio files. These features were\nsubsequently used to train an interpretable machine-learning model for tag\nprediction. We conducted experiments on two datasets, namely the MTG-Jamendo\ndataset and the GTZAN dataset. Our method surpassed the performance of baseline\nmodels in both tasks and, in certain instances, demonstrated competitiveness\nwith the current state-of-the-art. We conclude that there are use cases where\nthe deterioration in performance is outweighed by the value of\ninterpretability.\n", "date": "2023-12-18", "categories": []}
{"id": "2402.15258", "title": "High Resolution Guitar Transcription via Domain Adaptation", "abstract": "  Automatic music transcription (AMT) has achieved high accuracy for piano due\nto the availability of large, high-quality datasets such as MAESTRO and MAPS,\nbut comparable datasets are not yet available for other instruments. In recent\nwork, however, it has been demonstrated that aligning scores to transcription\nmodel activations can produce high quality AMT training data for instruments\nother than piano. Focusing on the guitar, we refine this approach to training\non score data using a dataset of commercially available score-audio pairs. We\npropose the use of a high-resolution piano transcription model to train a new\nguitar transcription model. The resulting model obtains state-of-the-art\ntranscription results on GuitarSet in a zero-shot context, improving on\npreviously published methods.\n", "date": "2024-02-23", "categories": []}
{"id": "2309.03451", "title": "Cross-domain Sound Recognition for Efficient Underwater Data Analysis", "abstract": "  This paper presents a novel deep learning approach for analyzing massive\nunderwater acoustic data by leveraging a model trained on a broad spectrum of\nnon-underwater (aerial) sounds. Recognizing the challenge in labeling vast\namounts of underwater data, we propose a two-fold methodology to accelerate\nthis labor-intensive procedure.\n  The first part of our approach involves PCA and UMAP visualization of the\nunderwater data using the feature vectors of an aerial sound recognition model.\nThis enables us to cluster the data in a two dimensional space and listen to\npoints within these clusters to understand their defining characteristics. This\ninnovative method simplifies the process of selecting candidate labels for\nfurther training.\n  In the second part, we train a neural network model using both the selected\nunderwater data and the non-underwater dataset. We conducted a quantitative\nanalysis to measure the precision, recall, and F1 score of our model for\nrecognizing airgun sounds, a common type of underwater sound. The F1 score\nachieved by our model exceeded 84.3%, demonstrating the effectiveness of our\napproach in analyzing underwater acoustic data.\n  The methodology presented in this paper holds significant potential to reduce\nthe amount of labor required in underwater data analysis and opens up new\npossibilities for further research in the field of cross-domain data analysis.\n", "date": "2023-09-07", "categories": []}
{"id": "2402.13208", "title": "How do Hyenas deal with Human Speech? Speech Recognition and Translation\n  with ConfHyena", "abstract": "  The attention mechanism, a cornerstone of state-of-the-art neural models,\nfaces computational hurdles in processing long sequences due to its quadratic\ncomplexity. Consequently, research efforts in the last few years focused on\nfinding more efficient alternatives. Among them, Hyena (Poli et al., 2023)\nstands out for achieving competitive results in both language modeling and\nimage classification, while offering sub-quadratic memory and computational\ncomplexity. Building on these promising results, we propose ConfHyena, a\nConformer whose encoder self-attentions are replaced with an adaptation of\nHyena for speech processing, where the long input sequences cause high\ncomputational costs. Through experiments in automatic speech recognition (for\nEnglish) and translation (from English into 8 target languages), we show that\nour best ConfHyena model significantly reduces the training time by 27%, at the\ncost of minimal quality degradation (~1%), which, in most cases, is not\nstatistically significant.\n", "date": "2024-02-20", "categories": []}
{"id": "2402.10547", "title": "Learning Disentangled Audio Representations through Controlled Synthesis", "abstract": "  This paper tackles the scarcity of benchmarking data in disentangled auditory\nrepresentation learning. We introduce SynTone, a synthetic dataset with\nexplicit ground truth explanatory factors for evaluating disentanglement\ntechniques. Benchmarking state-of-the-art methods on SynTone highlights its\nutility for method evaluation. Our results underscore strengths and limitations\nin audio disentanglement, motivating future research.\n", "date": "2024-02-16", "categories": []}
{"id": "2402.01274", "title": "On the Transferability of Large-Scale Self-Supervision to Few-Shot Audio\n  Classification", "abstract": "  In recent years, self-supervised learning has excelled for its capacity to\nlearn robust feature representations from unlabelled data. Networks pretrained\nthrough self-supervision serve as effective feature extractors for downstream\ntasks, including Few-Shot Learning. While the evaluation of unsupervised\napproaches for few-shot learning is well-established in imagery, it is notably\nabsent in acoustics. This study addresses this gap by assessing large-scale\nself-supervised models' performance in few-shot audio classification.\nAdditionally, we explore the relationship between a model's few-shot learning\ncapability and other downstream task benchmarks. Our findings reveal\nstate-of-the-art performance in some few-shot problems such as\nSpeechCommandsv2, as well as strong correlations between speech-based few-shot\nproblems and various downstream audio tasks.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.09245", "title": "Overview of the L3DAS23 Challenge on Audio-Visual Extended Reality", "abstract": "  The primary goal of the L3DAS23 Signal Processing Grand Challenge at ICASSP\n2023 is to promote and support collaborative research on machine learning for\n3D audio signal processing, with a specific emphasis on 3D speech enhancement\nand 3D Sound Event Localization and Detection in Extended Reality applications.\nAs part of our latest competition, we provide a brand-new dataset, which\nmaintains the same general characteristics of the L3DAS21 and L3DAS22 datasets,\nbut with first-order Ambisonics recordings from multiple reverberant simulated\nenvironments. Moreover, we start exploring an audio-visual scenario by\nproviding images of these environments, as perceived by the different\nmicrophone positions and orientations. We also propose updated baseline models\nfor both tasks that can now support audio-image couples as input and a\nsupporting API to replicate our results. Finally, we present the results of the\nparticipants. Further details about the challenge are available at\nhttps://www.l3das.com/icassp2023.\n", "date": "2024-02-14", "categories": []}
{"id": "2402.09318", "title": "Leveraging Pre-Trained Autoencoders for Interpretable Prototype Learning\n  of Music Audio", "abstract": "  We present PECMAE, an interpretable model for music audio classification\nbased on prototype learning. Our model is based on a previous method, APNet,\nwhich jointly learns an autoencoder and a prototypical network. Instead, we\npropose to decouple both training processes. This enables us to leverage\nexisting self-supervised autoencoders pre-trained on much larger data\n(EnCodecMAE), providing representations with better generalization. APNet\nallows prototypes' reconstruction to waveforms for interpretability relying on\nthe nearest training data samples. In contrast, we explore using a diffusion\ndecoder that allows reconstruction without such dependency. We evaluate our\nmethod on datasets for music instrument classification (Medley-Solos-DB) and\ngenre recognition (GTZAN and a larger in-house dataset), the latter being a\nmore challenging task not addressed with prototypical networks before. We find\nthat the prototype-based models preserve most of the performance achieved with\nthe autoencoder embeddings, while the sonification of prototypes benefits\nunderstanding the behavior of the classifier.\n", "date": "2024-02-14", "categories": []}
{"id": "2402.06683", "title": "Sound Source Separation Using Latent Variational Block-Wise\n  Disentanglement", "abstract": "  While neural network approaches have made significant strides in resolving\nclassical signal processing problems, it is often the case that hybrid\napproaches that draw insight from both signal processing and neural networks\nproduce more complete solutions. In this paper, we present a hybrid classical\ndigital signal processing/deep neural network (DSP/DNN) approach to source\nseparation (SS) highlighting the theoretical link between variational\nautoencoder and classical approaches to SS. We propose a system that transforms\nthe single channel under-determined SS task to an equivalent multichannel\nover-determined SS problem in a properly designed latent space. The separation\ntask in the latent space is treated as finding a variational block-wise\ndisentangled representation of the mixture. We show empirically, that the\ndesign choices and the variational formulation of the task at hand motivated by\nthe classical signal processing theoretical results lead to robustness to\nunseen out-of-distribution data and reduction of the overfitting risk. To\naddress the resulting permutation issue we explicitly incorporate a novel\ndifferentiable permutation loss function and augment the model with a memory\nmechanism to keep track of the statistics of the individual sources.\n", "date": "2024-02-08", "categories": []}
{"id": "2312.15824", "title": "Self-Supervised Learning for Few-Shot Bird Sound Classification", "abstract": "  Self-supervised learning (SSL) in audio holds significant potential across\nvarious domains, particularly in situations where abundant, unlabeled data is\nreadily available at no cost. This is pertinent in bioacoustics, where\nbiologists routinely collect extensive sound datasets from the natural\nenvironment. In this study, we demonstrate that SSL is capable of acquiring\nmeaningful representations of bird sounds from audio recordings without the\nneed for annotations. Our experiments showcase that these learned\nrepresentations exhibit the capacity to generalize to new bird species in\nfew-shot learning (FSL) scenarios. Additionally, we show that selecting windows\nwith high bird activation for self-supervised learning, using a pretrained\naudio neural network, significantly enhances the quality of the learned\nrepresentations.\n", "date": "2023-12-25", "categories": []}
{"id": "2302.02809", "title": "Listen2Scene: Interactive material-aware binaural sound propagation for\n  reconstructed 3D scenes", "abstract": "  We present an end-to-end binaural audio rendering approach (Listen2Scene) for\nvirtual reality (VR) and augmented reality (AR) applications. We propose a\nnovel neural-network-based binaural sound propagation method to generate\nacoustic effects for indoor 3D models of real environments. Any clean audio or\ndry audio can be convolved with the generated acoustic effects to render audio\ncorresponding to the real environment. We propose a graph neural network that\nuses both the material and the topology information of the 3D scenes and\ngenerates a scene latent vector. Moreover, we use a conditional generative\nadversarial network (CGAN) to generate acoustic effects from the scene latent\nvector. Our network can handle holes or other artifacts in the reconstructed 3D\nmesh model. We present an efficient cost function for the generator network to\nincorporate spatial audio effects. Given the source and the listener position,\nour learning-based binaural sound propagation approach can generate an acoustic\neffect in 0.1 milliseconds on an NVIDIA GeForce RTX 2080 Ti GPU. We have\nevaluated the accuracy of our approach with binaural acoustic effects generated\nusing an interactive geometric sound propagation algorithm and captured real\nacoustic effects / real-world recordings. We also performed a perceptual\nevaluation and observed that the audio rendered by our approach is more\nplausible than audio rendered using prior learning-based and geometric-based\nsound propagation algorithms. We quantitatively evaluated the accuracy of our\napproach using statistical acoustic parameters, and energy decay curves. The\ndemo videos, code and dataset are available online\n(https://anton-jeran.github.io/Listen2Scene/).\n", "date": "2023-02-02", "categories": []}
{"id": "2305.18441", "title": "DeCoR: Defy Knowledge Forgetting by Predicting Earlier Audio Codes", "abstract": "  Lifelong audio feature extraction involves learning new sound classes\nincrementally, which is essential for adapting to new data distributions over\ntime. However, optimizing the model only on new data can lead to catastrophic\nforgetting of previously learned tasks, which undermines the model's ability to\nperform well over the long term. This paper introduces a new approach to\ncontinual audio representation learning called DeCoR. Unlike other methods that\nstore previous data, features, or models, DeCoR indirectly distills knowledge\nfrom an earlier model to the latest by predicting quantization indices from a\ndelayed codebook. We demonstrate that DeCoR improves acoustic scene\nclassification accuracy and integrates well with continual self-supervised\nrepresentation learning. Our approach introduces minimal storage and\ncomputation overhead, making it a lightweight and efficient solution for\ncontinual learning.\n", "date": "2023-05-29", "categories": []}
{"id": "2402.01824", "title": "Identification of Cognitive Decline from Spoken Language through Feature\n  Selection and the Bag of Acoustic Words Model", "abstract": "  Memory disorders are a central factor in the decline of functioning and daily\nactivities in elderly individuals. The confirmation of the illness, initiation\nof medication to slow its progression, and the commencement of occupational\ntherapy aimed at maintaining and rehabilitating cognitive abilities require a\nmedical diagnosis. The early identification of symptoms of memory disorders,\nespecially the decline in cognitive abilities, plays a significant role in\nensuring the well-being of populations. Features related to speech production\nare known to connect with the speaker's cognitive ability and changes. The lack\nof standardized speech tests in clinical settings has led to a growing emphasis\non developing automatic machine learning techniques for analyzing naturally\nspoken language. Non-lexical but acoustic properties of spoken language have\nproven useful when fast, cost-effective, and scalable solutions are needed for\nthe rapid diagnosis of a disease. The work presents an approach related to\nfeature selection, allowing for the automatic selection of the essential\nfeatures required for diagnosis from the Geneva minimalistic acoustic parameter\nset and relative speech pauses, intended for automatic paralinguistic and\nclinical speech analysis. These features are refined into word histogram\nfeatures, in which machine learning classifiers are trained to classify control\nsubjects and dementia patients from the Dementia Bank's Pitt audio database.\nThe results show that achieving a 75% average classification accuracy with only\ntwenty-five features with the separate ADReSS 2020 competition test data and\nthe Leave-One-Subject-Out cross-validation of the entire competition data is\npossible. The results rank at the top compared to international research, where\nthe same dataset and only acoustic features have been used to diagnose\npatients.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.02184", "title": "Sentiment analysis in non-fixed length audios using a Fully\n  Convolutional Neural Network", "abstract": "  In this work, a sentiment analysis method that is capable of accepting audio\nof any length, without being fixed a priori, is proposed. Mel spectrogram and\nMel Frequency Cepstral Coefficients are used as audio description methods and a\nFully Convolutional Neural Network architecture is proposed as a classifier.\nThe results have been validated using three well known datasets: EMODB,\nRAVDESS, and TESS. The results obtained were promising, outperforming the\nstate-of-the-art methods. Also, thanks to the fact that the proposed method\nadmits audios of any size, it allows a sentiment analysis to be made in near\nreal time, which is very interesting for a wide range of fields such as call\ncenters, medical consultations, or financial brokers.\n", "date": "2024-02-03", "categories": []}
{"id": "2402.02183", "title": "Detecting Respiratory Pathologies Using Convolutional Neural Networks\n  and Variational Autoencoders for Unbalancing Data", "abstract": "  The aim of this paper was the detection of pathologies through respiratory\nsounds. The ICBHI (International Conference on Biomedical and Health\nInformatics) Benchmark was used. This dataset is composed of 920 sounds of\nwhich 810 are of chronic diseases, 75 of non-chronic diseases and only 35 of\nhealthy individuals. As more than 88% of the samples of the dataset are from\nthe same class (Chronic), the use of a Variational Convolutional Autoencoder\nwas proposed to generate new labeled data and other well known oversampling\ntechniques after determining that the dataset classes are unbalanced. Once the\npreprocessing step was carried out, a Convolutional Neural Network (CNN) was\nused to classify the respiratory sounds into healthy, chronic, and non-chronic\ndisease. In addition, we carried out a more challenging classification trying\nto distinguish between the different types of pathologies or healthy: URTI,\nCOPD, Bronchiectasis, Pneumonia, and Bronchiolitis. We achieved results up to\n0.993 F-Score in the three-label classification and 0.990 F-Score in the more\nchallenging six-class classification.\n", "date": "2024-02-03", "categories": []}
{"id": "2402.02754", "title": "Focal Modulation Networks for Interpretable Sound Classification", "abstract": "  The increasing success of deep neural networks has raised concerns about\ntheir inherent black-box nature, posing challenges related to interpretability\nand trust. While there has been extensive exploration of interpretation\ntechniques in vision and language, interpretability in the audio domain has\nreceived limited attention, primarily focusing on post-hoc explanations. This\npaper addresses the problem of interpretability by-design in the audio domain\nby utilizing the recently proposed attention-free focal modulation networks\n(FocalNets). We apply FocalNets to the task of environmental sound\nclassification for the first time and evaluate their interpretability\nproperties on the popular ESC-50 dataset. Our method outperforms a similarly\nsized vision transformer both in terms of accuracy and interpretability.\nFurthermore, it is competitive against PIQ, a method specifically designed for\npost-hoc interpretation in the audio domain.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.02781", "title": "Dual Knowledge Distillation for Efficient Sound Event Detection", "abstract": "  Sound event detection (SED) is essential for recognizing specific sounds and\ntheir temporal locations within acoustic signals. This becomes challenging\nparticularly for on-device applications, where computational resources are\nlimited. To address this issue, we introduce a novel framework referred to as\ndual knowledge distillation for developing efficient SED systems in this work.\nOur proposed dual knowledge distillation commences with temporal-averaging\nknowledge distillation (TAKD), utilizing a mean student model derived from the\ntemporal averaging of the student model's parameters. This allows the student\nmodel to indirectly learn from a pre-trained teacher model, ensuring a stable\nknowledge distillation. Subsequently, we introduce embedding-enhanced feature\ndistillation (EEFD), which involves incorporating an embedding distillation\nlayer within the student model to bolster contextual learning. On DCASE 2023\nTask 4A public evaluation dataset, our proposed SED system with dual knowledge\ndistillation having merely one-third of the baseline model's parameters,\ndemonstrates superior performance in terms of PSDS1 and PSDS2. This highlights\nthe importance of proposed dual knowledge distillation for compact SED systems,\nwhich can be ideal for edge devices.\n", "date": "2024-02-05", "categories": []}
{"id": "2401.15018", "title": "Enhancement of a Text-Independent Speaker Verification System by using\n  Feature Combination and Parallel-Structure Classifiers", "abstract": "  Speaker Verification (SV) systems involve mainly two individual stages:\nfeature extraction and classification. In this paper, we explore these two\nmodules with the aim of improving the performance of a speaker verification\nsystem under noisy conditions. On the one hand, the choice of the most\nappropriate acoustic features is a crucial factor for performing robust speaker\nverification. The acoustic parameters used in the proposed system are: Mel\nFrequency Cepstral Coefficients (MFCC), their first and second derivatives\n(Deltas and Delta- Deltas), Bark Frequency Cepstral Coefficients (BFCC),\nPerceptual Linear Predictive (PLP), and Relative Spectral Transform -\nPerceptual Linear Predictive (RASTA-PLP). In this paper, a complete comparison\nof different combinations of the previous features is discussed. On the other\nhand, the major weakness of a conventional Support Vector Machine (SVM)\nclassifier is the use of generic traditional kernel functions to compute the\ndistances among data points. However, the kernel function of an SVM has great\ninfluence on its performance. In this work, we propose the combination of two\nSVM-based classifiers with different kernel functions: Linear kernel and\nGaussian Radial Basis Function (RBF) kernel with a Logistic Regression (LR)\nclassifier. The combination is carried out by means of a parallel structure\napproach, in which different voting rules to take the final decision are\nconsidered. Results show that significant improvement in the performance of the\nSV system is achieved by using the combined features with the combined\nclassifiers either with clean speech or in the presence of noise. Finally, to\nenhance the system more in noisy environments, the inclusion of the multiband\nnoise removal technique as a preprocessing stage is proposed.\n", "date": "2024-01-26", "categories": []}
{"id": "2402.01753", "title": "SpecDiff-GAN: A Spectrally-Shaped Noise Diffusion GAN for Speech and\n  Music Synthesis", "abstract": "  Generative adversarial network (GAN) models can synthesize highquality audio\nsignals while ensuring fast sample generation. However, they are difficult to\ntrain and are prone to several issues including mode collapse and divergence.\nIn this paper, we introduce SpecDiff-GAN, a neural vocoder based on HiFi-GAN,\nwhich was initially devised for speech synthesis from mel spectrogram. In our\nmodel, the training stability is enhanced by means of a forward diffusion\nprocess which consists in injecting noise from a Gaussian distribution to both\nreal and fake samples before inputting them to the discriminator. We further\nimprove the model by exploiting a spectrally-shaped noise distribution with the\naim to make the discriminator's task more challenging. We then show the merits\nof our proposed model for speech and music synthesis on several datasets. Our\nexperiments confirm that our model compares favorably in audio quality and\nefficiency compared to several baselines.\n", "date": "2024-01-30", "categories": []}
{"id": "2308.08143", "title": "IIANet: An Intra- and Inter-Modality Attention Network for Audio-Visual\n  Speech Separation", "abstract": "  Recent research has made significant progress in designing fusion modules for\naudio-visual speech separation. However, they predominantly focus on\nmulti-modal fusion at a single temporal scale of auditory and visual features\nwithout employing selective attention mechanisms, which is in sharp contrast\nwith the brain. To address this issue, We propose a novel model called Intra-\nand Inter-Attention Network (IIANet), which leverages the attention mechanism\nfor efficient audio-visual feature fusion. IIANet consists of two types of\nattention blocks: intra-attention (IntraA) and inter-attention (InterA) blocks,\nwhere the InterA blocks are distributed at the top, middle and bottom of\nIIANet. Heavily inspired by the way how human brain selectively focuses on\nrelevant content at various temporal scales, these blocks maintain the ability\nto learn modality-specific features and enable the extraction of different\nsemantics from audio-visual features. Comprehensive experiments on three\nstandard audio-visual separation benchmarks (LRS2, LRS3, and VoxCeleb2)\ndemonstrate the effectiveness of IIANet, outperforming previous\nstate-of-the-art methods while maintaining comparable inference time. In\nparticular, the fast version of IIANet (IIANet-fast) has only 7% of CTCNet's\nMACs and is 40% faster than CTCNet on CPUs while achieving better separation\nquality, showing the great potential of attention mechanism for efficient and\neffective multimodal fusion.\n", "date": "2023-08-16", "categories": []}
{"id": "2402.00892", "title": "EVA-GAN: Enhanced Various Audio Generation via Scalable Generative\n  Adversarial Networks", "abstract": "  The advent of Large Models marks a new era in machine learning, significantly\noutperforming smaller models by leveraging vast datasets to capture and\nsynthesize complex patterns. Despite these advancements, the exploration into\nscaling, especially in the audio generation domain, remains limited, with\nprevious efforts didn't extend into the high-fidelity (HiFi) 44.1kHz domain and\nsuffering from both spectral discontinuities and blurriness in the\nhigh-frequency domain, alongside a lack of robustness against out-of-domain\ndata. These limitations restrict the applicability of models to diverse use\ncases, including music and singing generation. Our work introduces Enhanced\nVarious Audio Generation via Scalable Generative Adversarial Networks\n(EVA-GAN), yields significant improvements over previous state-of-the-art in\nspectral and high-frequency reconstruction and robustness in out-of-domain data\nperformance, enabling the generation of HiFi audios by employing an extensive\ndataset of 36,000 hours of 44.1kHz audio, a context-aware module, a\nHuman-In-The-Loop artifact measurement toolkit, and expands the model to\napproximately 200 million parameters. Demonstrations of our work are available\nat https://double-blind-eva-gan.cc.\n", "date": "2024-01-31", "categories": []}
{"id": "2402.01424", "title": "A Data-Driven Analysis of Robust Automatic Piano Transcription", "abstract": "  Algorithms for automatic piano transcription have improved dramatically in\nrecent years due to new datasets and modeling techniques. Recent developments\nhave focused primarily on adapting new neural network architectures, such as\nthe Transformer and Perceiver, in order to yield more accurate systems. In this\nwork, we study transcription systems from the perspective of their training\ndata. By measuring their performance on out-of-distribution annotated piano\ndata, we show how these models can severely overfit to acoustic properties of\nthe training data. We create a new set of audio for the MAESTRO dataset,\ncaptured automatically in a professional studio recording environment via\nYamaha Disklavier playback. Using various data augmentation techniques when\ntraining with the original and re-performed versions of the MAESTRO dataset, we\nachieve state-of-the-art note-onset accuracy of 88.4 F1-score on the MAPS\ndataset, without seeing any of its training data. We subsequently analyze these\ndata augmentation techniques in a series of ablation studies to better\nunderstand their influence on the resulting models.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.01571", "title": "Spiking Music: Audio Compression with Event Based Auto-encoders", "abstract": "  Neurons in the brain communicate information via punctual events called\nspikes. The timing of spikes is thought to carry rich information, but it is\nnot clear how to leverage this in digital systems. We demonstrate that\nevent-based encoding is efficient for audio compression. To build this\nevent-based representation we use a deep binary auto-encoder, and under high\nsparsity pressure, the model enters a regime where the binary event matrix is\nstored more efficiently with sparse matrix storage algorithms. We test this on\nthe large MAESTRO dataset of piano recordings against vector quantized\nauto-encoders. Not only does our \"Spiking Music compression\" algorithm achieve\na competitive compression/reconstruction trade-off, but selectivity and\nsynchrony between encoded events and piano key strikes emerge without\nsupervision in the sparse regime.\n", "date": "2024-02-02", "categories": []}
{"id": "2401.06203", "title": "Remixing Music for Hearing Aids Using Ensemble of Fine-Tuned Source\n  Separators", "abstract": "  This paper introduces our system submission for the Cadenza ICASSP 2024 Grand\nChallenge, which presents the problem of remixing and enhancing music for\nhearing aid users. Our system placed first in the challenge, achieving the best\naverage Hearing-Aid Audio Quality Index (HAAQI) score on the evaluation data\nset. We describe the system, which uses an ensemble of deep learning music\nsource separators that are fine tuned on the challenge data. We demonstrate the\neffectiveness of our system through the challenge results and analyze the\nimportance of different system aspects through ablation studies.\n", "date": "2024-01-11", "categories": []}
{"id": "2401.17690", "title": "EnCLAP: Combining Neural Audio Codec and Audio-Text Joint Embedding for\n  Automated Audio Captioning", "abstract": "  We propose EnCLAP, a novel framework for automated audio captioning. EnCLAP\nemploys two acoustic representation models, EnCodec and CLAP, along with a\npretrained language model, BART. We also introduce a new training objective\ncalled masked codec modeling that improves acoustic awareness of the pretrained\nlanguage model. Experimental results on AudioCaps and Clotho demonstrate that\nour model surpasses the performance of baseline models. Source code will be\navailable at https://github.com/jaeyeonkim99/EnCLAP . An online demo is\navailable at https://huggingface.co/spaces/enclap-team/enclap .\n", "date": "2024-01-31", "categories": []}
{"id": "2306.01433", "title": "Blind Audio Bandwidth Extension: A Diffusion-Based Zero-Shot Approach", "abstract": "  Audio bandwidth extension involves the realistic reconstruction of\nhigh-frequency spectra from bandlimited observations. In cases where the\nlowpass degradation is unknown, such as in restoring historical audio\nrecordings, this becomes a blind problem. This paper introduces a novel method\ncalled BABE (Blind Audio Bandwidth Extension) that addresses the blind problem\nin a zero-shot setting, leveraging the generative priors of a pre-trained\nunconditional diffusion model. During the inference process, BABE utilizes a\ngeneralized version of diffusion posterior sampling, where the degradation\noperator is unknown but parametrized and inferred iteratively. The performance\nof the proposed method is evaluated using objective and subjective metrics, and\nthe results show that BABE surpasses state-of-the-art blind bandwidth extension\nbaselines and achieves competitive performance compared to informed methods\nwhen tested with synthetic data. Moreover, BABE exhibits robust generalization\ncapabilities when enhancing real historical recordings, effectively\nreconstructing the missing high-frequency content while maintaining coherence\nwith the original recording. Subjective preference tests confirm that BABE\nsignificantly improves the audio quality of historical music recordings.\nExamples of historical recordings restored with the proposed method are\navailable on the companion webpage:\n(http://research.spa.aalto.fi/publications/papers/ieee-taslp-babe/)\n", "date": "2023-06-02", "categories": []}
{"id": "2312.08676", "title": "SEF-VC: Speaker Embedding Free Zero-Shot Voice Conversion with Cross\n  Attention", "abstract": "  Zero-shot voice conversion (VC) aims to transfer the source speaker timbre to\narbitrary unseen target speaker timbre, while keeping the linguistic content\nunchanged. Although the voice of generated speech can be controlled by\nproviding the speaker embedding of the target speaker, the speaker similarity\nstill lags behind the ground truth recordings. In this paper, we propose\nSEF-VC, a speaker embedding free voice conversion model, which is designed to\nlearn and incorporate speaker timbre from reference speech via a powerful\nposition-agnostic cross-attention mechanism, and then reconstruct waveform from\nHuBERT semantic tokens in a non-autoregressive manner. The concise design of\nSEF-VC enhances its training stability and voice conversion performance.\nObjective and subjective evaluations demonstrate the superiority of SEF-VC to\ngenerate high-quality speech with better similarity to target reference than\nstrong zero-shot VC baselines, even for very short reference speeches.\n", "date": "2023-12-14", "categories": []}
{"id": "2401.17129", "title": "Enhanced Sound Event Localization and Detection in Real 360-degree\n  audio-visual soundscapes", "abstract": "  This technical report details our work towards building an enhanced\naudio-visual sound event localization and detection (SELD) network. We build on\ntop of the audio-only SELDnet23 model and adapt it to be audio-visual by\nmerging both audio and video information prior to the gated recurrent unit\n(GRU) of the audio-only network. Our model leverages YOLO and DETIC object\ndetectors. We also build a framework that implements audio-visual data\naugmentation and audio-visual synthetic data generation. We deliver an\naudio-visual SELDnet system that outperforms the existing audio-visual SELD\nbaseline.\n", "date": "2024-01-29", "categories": []}
{"id": "2310.19063", "title": "Feature Aggregation in Joint Sound Classification and Localization\n  Neural Networks", "abstract": "  This study addresses the application of deep learning techniques in joint\nsound signal classification and localization networks. Current state-of-the-art\nsound source localization deep learning networks lack feature aggregation\nwithin their architecture. Feature aggregation enhances model performance by\nenabling the consolidation of information from different feature scales,\nthereby improving feature robustness and invariance. This is particularly\nimportant in SSL networks, which must differentiate direct and indirect\nacoustic signals. To address this gap, we adapt feature aggregation techniques\nfrom computer vision neural networks to signal detection neural networks.\nAdditionally, we propose the Scale Encoding Network (SEN) for feature\naggregation to encode features from various scales, compressing the network for\nmore computationally efficient aggregation. To evaluate the efficacy of feature\naggregation in SSL networks, we integrated the following computer vision\nfeature aggregation sub-architectures into a SSL control architecture: Path\nAggregation Network (PANet), Weighted Bi-directional Feature Pyramid Network\n(BiFPN), and SEN. These sub-architectures were evaluated using two metrics for\nsignal classification and two metrics for direction-of-arrival regression.\nPANet and BiFPN are established aggregators in computer vision models, while\nthe proposed SEN is a more compact aggregator. The results suggest that models\nincorporating feature aggregations outperformed the control model, the Sound\nEvent Localization and Detection network (SELDnet), in both sound signal\nclassification and localization. The feature aggregation techniques enhance the\nperformance of sound detection neural networks, particularly in\ndirection-of-arrival regression.\n", "date": "2023-10-29", "categories": []}
{"id": "2401.15323", "title": "Music Auto-Tagging with Robust Music Representation Learned via Domain\n  Adversarial Training", "abstract": "  Music auto-tagging is crucial for enhancing music discovery and\nrecommendation. Existing models in Music Information Retrieval (MIR) struggle\nwith real-world noise such as environmental and speech sounds in multimedia\ncontent. This study proposes a method inspired by speech-related tasks to\nenhance music auto-tagging performance in noisy settings. The approach\nintegrates Domain Adversarial Training (DAT) into the music domain, enabling\nrobust music representations that withstand noise. Unlike previous research,\nthis approach involves an additional pretraining phase for the domain\nclassifier, to avoid performance degradation in the subsequent phase. Adding\nvarious synthesized noisy music data improves the model's generalization across\ndifferent noise levels. The proposed architecture demonstrates enhanced\nperformance in music auto-tagging by effectively utilizing unlabeled noisy\nmusic data. Additional experiments with supplementary unlabeled data further\nimproves the model's performance, underscoring its robust generalization\ncapabilities and broad applicability.\n", "date": "2024-01-27", "categories": []}
{"id": "2211.01839", "title": "HyperSound: Generating Implicit Neural Representations of Audio Signals\n  with Hypernetworks", "abstract": "  Implicit neural representations (INRs) are a rapidly growing research field,\nwhich provides alternative ways to represent multimedia signals. Recent\napplications of INRs include image super-resolution, compression of\nhigh-dimensional signals, or 3D rendering. However, these solutions usually\nfocus on visual data, and adapting them to the audio domain is not trivial.\nMoreover, it requires a separately trained model for every data sample. To\naddress this limitation, we propose HyperSound, a meta-learning method\nleveraging hypernetworks to produce INRs for audio signals unseen at training\ntime. We show that our approach can reconstruct sound waves with quality\ncomparable to other state-of-the-art models.\n", "date": "2022-11-03", "categories": []}
{"id": "2401.14185", "title": "TDFNet: An Efficient Audio-Visual Speech Separation Model with Top-down\n  Fusion", "abstract": "  Audio-visual speech separation has gained significant traction in recent\nyears due to its potential applications in various fields such as speech\nrecognition, diarization, scene analysis and assistive technologies. Designing\na lightweight audio-visual speech separation network is important for\nlow-latency applications, but existing methods often require higher\ncomputational costs and more parameters to achieve better separation\nperformance. In this paper, we present an audio-visual speech separation model\ncalled Top-Down-Fusion Net (TDFNet), a state-of-the-art (SOTA) model for\naudio-visual speech separation, which builds upon the architecture of TDANet,\nan audio-only speech separation method. TDANet serves as the architectural\nfoundation for the auditory and visual networks within TDFNet, offering an\nefficient model with fewer parameters. On the LRS2-2Mix dataset, TDFNet\nachieves a performance increase of up to 10\\% across all performance metrics\ncompared with the previous SOTA method CTCNet. Remarkably, these results are\nachieved using fewer parameters and only 28\\% of the multiply-accumulate\noperations (MACs) of CTCNet. In essence, our method presents a highly effective\nand efficient solution to the challenges of speech separation within the\naudio-visual domain, making significant strides in harnessing visual\ninformation optimally.\n", "date": "2024-01-25", "categories": []}
{"id": "2309.15717", "title": "Timbre-Trap: A Low-Resource Framework for Instrument-Agnostic Music\n  Transcription", "abstract": "  In recent years, research on music transcription has focused mainly on\narchitecture design and instrument-specific data acquisition. With the lack of\navailability of diverse datasets, progress is often limited to solo-instrument\ntasks such as piano transcription. Several works have explored multi-instrument\ntranscription as a means to bolster the performance of models on low-resource\ntasks, but these methods face the same data availability issues. We propose\nTimbre-Trap, a novel framework which unifies music transcription and audio\nreconstruction by exploiting the strong separability between pitch and timbre.\nWe train a single autoencoder to simultaneously estimate pitch salience and\nreconstruct complex spectral coefficients, selecting between either output\nduring the decoding stage via a simple switch mechanism. In this way, the model\nlearns to produce coefficients corresponding to timbre-less audio, which can be\ninterpreted as pitch salience. We demonstrate that the framework leads to\nperformance comparable to state-of-the-art instrument-agnostic transcription\nmethods, while only requiring a small amount of annotated data.\n", "date": "2023-09-27", "categories": []}
{"id": "2401.13498", "title": "Expressive Acoustic Guitar Sound Synthesis with an Instrument-Specific\n  Input Representation and Diffusion Outpainting", "abstract": "  Synthesizing performing guitar sound is a highly challenging task due to the\npolyphony and high variability in expression. Recently, deep generative models\nhave shown promising results in synthesizing expressive polyphonic instrument\nsounds from music scores, often using a generic MIDI input. In this work, we\npropose an expressive acoustic guitar sound synthesis model with a customized\ninput representation to the instrument, which we call guitarroll. We implement\nthe proposed approach using diffusion-based outpainting which can generate\naudio with long-term consistency. To overcome the lack of MIDI/audio-paired\ndatasets, we used not only an existing guitar dataset but also collected data\nfrom a high quality sample-based guitar synthesizer. Through quantitative and\nqualitative evaluations, we show that our proposed model has higher audio\nquality than the baseline model and generates more realistic timbre sounds than\nthe previous leading work.\n", "date": "2024-01-24", "categories": []}
{"id": "2312.16613", "title": "Self-supervised Pretraining for Robust Personalized Voice Activity\n  Detection in Adverse Conditions", "abstract": "  In this paper, we propose the use of self-supervised pretraining on a large\nunlabelled data set to improve the performance of a personalized voice activity\ndetection (VAD) model in adverse conditions. We pretrain a long short-term\nmemory (LSTM)-encoder using the autoregressive predictive coding (APC)\nframework and fine-tune it for personalized VAD. We also propose a denoising\nvariant of APC, with the goal of improving the robustness of personalized VAD.\nThe trained models are systematically evaluated on both clean speech and speech\ncontaminated by various types of noise at different SNR-levels and compared to\na purely supervised model. Our experiments show that self-supervised\npretraining not only improves performance in clean conditions, but also yields\nmodels which are more robust to adverse conditions compared to purely\nsupervised learning.\n", "date": "2023-12-27", "categories": []}
{"id": "2401.10543", "title": "Multilingual acoustic word embeddings for zero-resource languages", "abstract": "  This research addresses the challenge of developing speech applications for\nzero-resource languages that lack labelled data. It specifically uses acoustic\nword embedding (AWE) -- fixed-dimensional representations of variable-duration\nspeech segments -- employing multilingual transfer, where labelled data from\nseveral well-resourced languages are used for pertaining. The study introduces\na new neural network that outperforms existing AWE models on zero-resource\nlanguages. It explores the impact of the choice of well-resourced languages.\nAWEs are applied to a keyword-spotting system for hate speech detection in\nSwahili radio broadcasts, demonstrating robustness in real-world scenarios.\nAdditionally, novel semantic AWE models improve semantic query-by-example\nsearch.\n", "date": "2024-01-19", "categories": []}
{"id": "2401.12238", "title": "Spatial Scaper: A Library to Simulate and Augment Soundscapes for Sound\n  Event Localization and Detection in Realistic Rooms", "abstract": "  Sound event localization and detection (SELD) is an important task in machine\nlistening. Major advancements rely on simulated data with sound events in\nspecific rooms and strong spatio-temporal labels. SELD data is simulated by\nconvolving spatialy-localized room impulse responses (RIRs) with sound\nwaveforms to place sound events in a soundscape. However, RIRs require manual\ncollection in specific rooms. We present SpatialScaper, a library for SELD data\nsimulation and augmentation. Compared to existing tools, SpatialScaper emulates\nvirtual rooms via parameters such as size and wall absorption. This allows for\nparameterized placement (including movement) of foreground and background sound\nsources. SpatialScaper also includes data augmentation pipelines that can be\napplied to existing SELD data. As a case study, we use SpatialScaper to add\nrooms to the DCASE SELD data. Training a model with our data led to progressive\nperformance improves as a direct function of acoustic diversity. These results\nshow that SpatialScaper is valuable to train robust SELD models.\n", "date": "2024-01-19", "categories": []}
{"id": "2401.12570", "title": "DiffMoog: a Differentiable Modular Synthesizer for Sound Matching", "abstract": "  This paper presents DiffMoog - a differentiable modular synthesizer with a\ncomprehensive set of modules typically found in commercial instruments. Being\ndifferentiable, it allows integration into neural networks, enabling automated\nsound matching, to replicate a given audio input. Notably, DiffMoog facilitates\nmodulation capabilities (FM/AM), low-frequency oscillators (LFOs), filters,\nenvelope shapers, and the ability for users to create custom signal chains. We\nintroduce an open-source platform that comprises DiffMoog and an end-to-end\nsound matching framework. This framework utilizes a novel signal-chain loss and\nan encoder network that self-programs its outputs to predict DiffMoogs\nparameters based on the user-defined modular architecture. Moreover, we provide\ninsights and lessons learned towards sound matching using differentiable\nsynthesis. Combining robust sound capabilities with a holistic platform,\nDiffMoog stands as a premier asset for expediting research in audio synthesis\nand machine learning.\n", "date": "2024-01-23", "categories": []}
{"id": "2401.11199", "title": "Projected Belief Networks With Discriminative Alignment for Acoustic\n  Event Classification: Rivaling State of the Art CNNs", "abstract": "  The projected belief network (PBN) is a generative stochastic network with\ntractable likelihood function based on a feed-forward neural network (FFNN).\nThe generative function operates by \"backing up\" through the FFNN. The PBN is\ntwo networks in one, a FFNN that operates in the forward direction, and a\ngenerative network that operates in the backward direction. Both networks\nco-exist based on the same parameter set, have their own cost functions, and\ncan be separately or jointly trained. The PBN therefore has the potential to\npossess the best qualities of both discriminative and generative classifiers.\nTo realize this potential, a separate PBN is trained on each class, maximizing\nthe generative likelihood function for the given class, while minimizing the\ndiscriminative cost for the FFNN against \"all other classes\". This technique,\ncalled discriminative alignment (PBN-DA), aligns the contours of the likelihood\nfunction to the decision boundaries and attains vastly improved classification\nperformance, rivaling that of state of the art discriminative networks. The\nmethod may be further improved using a hidden Markov model (HMM) as a component\nof the PBN, called PBN-DA-HMM. This paper provides a comprehensive treatment of\nPBN, PBN-DA, and PBN-DA-HMM. In addition, the results of two new classification\nexperiments are provided. The first experiment uses air-acoustic events, and\nthe second uses underwater acoustic data consisting of marine mammal calls. In\nboth experiments, PBN-DA-HMM attains comparable or better performance as a\nstate of the art CNN, and attain a factor of two error reduction when combined\nwith the CNN.\n", "date": "2024-01-20", "categories": []}
{"id": "2401.12068", "title": "Resource-constrained stereo singing voice cancellation", "abstract": "  We study the problem of stereo singing voice cancellation, a subtask of music\nsource separation, whose goal is to estimate an instrumental background from a\nstereo mix. We explore how to achieve performance similar to large\nstate-of-the-art source separation networks starting from a small, efficient\nmodel for real-time speech separation. Such a model is useful when memory and\ncompute are limited and singing voice processing has to run with limited\nlook-ahead. In practice, this is realised by adapting an existing mono model to\nhandle stereo input. Improvements in quality are obtained by tuning model\nparameters and expanding the training set. Moreover, we highlight the benefits\na stereo model brings by introducing a new metric which detects attenuation\ninconsistencies between channels. Our approach is evaluated using objective\noffline metrics and a large-scale MUSHRA trial, confirming the effectiveness of\nour techniques in stringent listening tests.\n", "date": "2024-01-22", "categories": []}
{"id": "2303.03183", "title": "Utilizing synthetic training data for the supervised classification of\n  rat ultrasonic vocalizations", "abstract": "  Murine rodents generate ultrasonic vocalizations (USVs) with frequencies that\nextend to around 120kHz. These calls are important in social behaviour, and so\ntheir analysis can provide insights into the function of vocal communication,\nand its dysfunction. The manual identification of USVs, and subsequent\nclassification into different subcategories is time consuming. Although machine\nlearning approaches for identification and classification can lead to enormous\nefficiency gains, the time and effort required to generate training data can be\nhigh, and the accuracy of current approaches can be problematic. Here we\ncompare the detection and classification performance of a trained human against\ntwo convolutional neural networks (CNNs), DeepSqueak and VocalMat, on audio\ncontaining rat USVs. Furthermore, we test the effect of inserting synthetic\nUSVs into the training data of the VocalMat CNN as a means of reducing the\nworkload associated with generating a training set. Our results indicate that\nVocalMat outperformed the DeepSqueak CNN on measures of call identification,\nand classification. Additionally, we found that the augmentation of training\ndata with synthetic images resulted in a further improvement in accuracy, such\nthat it was sufficiently close to human performance to allow for the use of\nthis software in laboratory conditions.\n", "date": "2023-03-03", "categories": []}
{"id": "2401.10544", "title": "AAT: Adapting Audio Transformer for Various Acoustics Recognition Tasks", "abstract": "  Recently, Transformers have been introduced into the field of acoustics\nrecognition. They are pre-trained on large-scale datasets using methods such as\nsupervised learning and semi-supervised learning, demonstrating robust\ngenerality--It fine-tunes easily to downstream tasks and shows more robust\nperformance. However, the predominant fine-tuning method currently used is\nstill full fine-tuning, which involves updating all parameters during training.\nThis not only incurs significant memory usage and time costs but also\ncompromises the model's generality. Other fine-tuning methods either struggle\nto address this issue or fail to achieve matching performance. Therefore, we\nconducted a comprehensive analysis of existing fine-tuning methods and proposed\nan efficient fine-tuning approach based on Adapter tuning, namely AAT. The core\nidea is to freeze the audio Transformer model and insert extra learnable\nAdapters, efficiently acquiring downstream task knowledge without compromising\nthe model's original generality. Extensive experiments have shown that our\nmethod achieves performance comparable to or even superior to full fine-tuning\nwhile optimizing only 7.118% of the parameters. It also demonstrates\nsuperiority over other fine-tuning methods.\n", "date": "2024-01-19", "categories": []}
{"id": "2401.09880", "title": "Attention-Based Recurrent Neural Network For Automatic Behavior Laying\n  Hen Recognition", "abstract": "  One of the interests of modern poultry farming is the vocalization of laying\nhens which contain very useful information on health behavior. This information\nis used as health and well-being indicators that help breeders better monitor\nlaying hens, which involves early detection of problems for rapid and more\neffective intervention. In this work, we focus on the sound analysis for the\nrecognition of the types of calls of the laying hens in order to propose a\nrobust system of characterization of their behavior for a better monitoring. To\ndo this, we first collected and annotated laying hen call signals, then\ndesigned an optimal acoustic characterization based on the combination of time\nand frequency domain features. We then used these features to build the\nmulti-label classification models based on recurrent neural network to assign a\nsemantic class to the vocalization that characterize the laying hen behavior.\nThe results show an overall performance with our model based on the combination\nof time and frequency domain features that obtained the highest F1-score\n(F1=92.75) with a gain of 17% on the models using the frequency domain features\nand of 8% on the compared approaches from the litterature.\n", "date": "2024-01-18", "categories": []}
{"id": "2309.08971", "title": "Regularized Contrastive Pre-training for Few-shot Bioacoustic Sound\n  Detection", "abstract": "  Bioacoustic sound event detection allows for better understanding of animal\nbehavior and for better monitoring biodiversity using audio. Deep learning\nsystems can help achieve this goal, however it is difficult to acquire\nsufficient annotated data to train these systems from scratch. To address this\nlimitation, the Detection and Classification of Acoustic Scenes and Events\n(DCASE) community has recasted the problem within the framework of few-shot\nlearning and organize an annual challenge for learning to detect animal sounds\nfrom only five annotated examples. In this work, we regularize supervised\ncontrastive pre-training to learn features that can transfer well on new target\ntasks with animal sounds unseen during training, achieving a high F-score of\n61.52%(0.48) when no feature adaptation is applied, and an F-score of\n68.19%(0.75) when we further adapt the learned features for each new target\ntask. This work aims to lower the entry bar to few-shot bioacoustic sound event\ndetection by proposing a simple and yet effective framework for this task, by\nalso providing open-source code.\n", "date": "2023-09-16", "categories": []}
{"id": "2401.08268", "title": "An Explainable Proxy Model for Multiabel Audio Segmentation", "abstract": "  Audio signal segmentation is a key task for automatic audio indexing. It\nconsists of detecting the boundaries of class-homogeneous segments in the\nsignal. In many applications, explainable AI is a vital process for\ntransparency of decision-making with machine learning. In this paper, we\npropose an explainable multilabel segmentation model that solves speech\nactivity (SAD), music (MD), noise (ND), and overlapped speech detection (OSD)\nsimultaneously. This proxy uses the non-negative matrix factorization (NMF) to\nmap the embedding used for the segmentation to the frequency domain.\nExperiments conducted on two datasets show similar performances as the\npre-trained black box model while showing strong explainability features.\nSpecifically, the frequency bins used for the decision can be easily identified\nat both the segment level (local explanations) and global level (class\nprototypes).\n", "date": "2024-01-16", "categories": []}
{"id": "2401.08864", "title": "Binaural Angular Separation Network", "abstract": "  We propose a neural network model that can separate target speech sources\nfrom interfering sources at different angular regions using two microphones.\nThe model is trained with simulated room impulse responses (RIRs) using\nomni-directional microphones without needing to collect real RIRs. By relying\non specific angular regions and multiple room simulations, the model utilizes\nconsistent time difference of arrival (TDOA) cues, or what we call delay\ncontrast, to separate target and interference sources while remaining robust in\nvarious reverberation environments. We demonstrate the model is not only\ngeneralizable to a commercially available device with a slightly different\nmicrophone geometry, but also outperforms our previous work which uses one\nadditional microphone on the same device. The model runs in real-time on-device\nand is suitable for low-latency streaming applications such as telephony and\nvideo conferencing.\n", "date": "2024-01-16", "categories": []}
{"id": "2401.08889", "title": "On the Effect of Data-Augmentation on Local Embedding Properties in the\n  Contrastive Learning of Music Audio Representations", "abstract": "  Audio embeddings are crucial tools in understanding large catalogs of music.\nTypically embeddings are evaluated on the basis of the performance they provide\nin a wide range of downstream tasks, however few studies have investigated the\nlocal properties of the embedding spaces themselves which are important in\nnearest neighbor algorithms, commonly used in music search and recommendation.\nIn this work we show that when learning audio representations on music datasets\nvia contrastive learning, musical properties that are typically homogeneous\nwithin a track (e.g., key and tempo) are reflected in the locality of\nneighborhoods in the resulting embedding space. By applying appropriate data\naugmentation strategies, localisation of such properties can not only be\nreduced but the localisation of other attributes is increased. For example,\nlocality of features such as pitch and tempo that are less relevant to\nnon-expert listeners, may be mitigated while improving the locality of more\nsalient features such as genre and mood, achieving state-of-the-art performance\nin nearest neighbor retrieval accuracy. Similarly, we show that the optimal\nselection of data augmentation strategies for contrastive learning of music\naudio embeddings is dependent on the downstream task, highlighting this as an\nimportant embedding design decision.\n", "date": "2024-01-17", "categories": []}
{"id": "2401.08891", "title": "Tempo estimation as fully self-supervised binary classification", "abstract": "  This paper addresses the problem of global tempo estimation in musical audio.\nGiven that annotating tempo is time-consuming and requires certain musical\nexpertise, few publicly available data sources exist to train machine learning\nmodels for this task. Towards alleviating this issue, we propose a fully\nself-supervised approach that does not rely on any human labeled data. Our\nmethod builds on the fact that generic (music) audio embeddings already encode\na variety of properties, including information about tempo, making them easily\nadaptable for downstream tasks. While recent work in self-supervised tempo\nestimation aimed to learn a tempo specific representation that was subsequently\nused to train a supervised classifier, we reformulate the task into the binary\nclassification problem of predicting whether a target track has the same or a\ndifferent tempo compared to a reference. While the former still requires\nlabeled training data for the final classification model, our approach uses\narbitrary unlabeled music data in combination with time-stretching for model\ntraining as well as a small set of synthetically created reference samples for\npredicting the final tempo. Evaluation of our approach in comparison with the\nstate-of-the-art reveals highly competitive performance when the constraint of\nfinding the precise tempo octave is relaxed.\n", "date": "2024-01-17", "categories": []}
{"id": "2401.08902", "title": "Similar but Faster: Manipulation of Tempo in Music Audio Embeddings for\n  Tempo Prediction and Search", "abstract": "  Audio embeddings enable large scale comparisons of the similarity of audio\nfiles for applications such as search and recommendation. Due to the\nsubjectivity of audio similarity, it can be desirable to design systems that\nanswer not only whether audio is similar, but similar in what way (e.g., wrt.\ntempo, mood or genre). Previous works have proposed disentangled embedding\nspaces where subspaces representing specific, yet possibly correlated,\nattributes can be weighted to emphasize those attributes in downstream tasks.\nHowever, no research has been conducted into the independence of these\nsubspaces, nor their manipulation, in order to retrieve tracks that are similar\nbut different in a specific way. Here, we explore the manipulation of tempo in\nembedding spaces as a case-study towards this goal. We propose tempo\ntranslation functions that allow for efficient manipulation of tempo within a\npre-existing embedding space whilst maintaining other properties such as genre.\nAs this translation is specific to tempo it enables retrieval of tracks that\nare similar but have specifically different tempi. We show that such a function\ncan be used as an efficient data augmentation strategy for both training of\ndownstream tempo predictors, and improved nearest neighbor retrieval of\nproperties largely independent of tempo.\n", "date": "2024-01-17", "categories": []}
{"id": "2401.09200", "title": "A Real-Time Lyrics Alignment System Using Chroma And Phonetic Features\n  For Classical Vocal Performance", "abstract": "  The goal of real-time lyrics alignment is to take live singing audio as input\nand to pinpoint the exact position within given lyrics on the fly. The task can\nbenefit real-world applications such as the automatic subtitling of live\nconcerts or operas. However, designing a real-time model poses a great\nchallenge due to the constraints of only using past input and operating within\na minimal latency. Furthermore, due to the lack of datasets for real-time\nmodels for lyrics alignment, previous studies have mostly evaluated with\nprivate in-house datasets, resulting in a lack of standard evaluation methods.\nThis paper presents a real-time lyrics alignment system for classical vocal\nperformances with two contributions. First, we improve the lyrics alignment\nalgorithm by finding an optimal combination of chromagram and phonetic\nposteriorgram (PPG) that capture melodic and phonetics features of the singing\nvoice, respectively. Second, we recast the Schubert Winterreise Dataset (SWD)\nwhich contains multiple performance renditions of the same pieces as an\nevaluation set for the real-time lyrics alignment.\n", "date": "2024-01-17", "categories": []}
{"id": "2401.09294", "title": "T-FOLEY: A Controllable Waveform-Domain Diffusion Model for\n  Temporal-Event-Guided Foley Sound Synthesis", "abstract": "  Foley sound, audio content inserted synchronously with videos, plays a\ncritical role in the user experience of multimedia content. Recently, there has\nbeen active research in Foley sound synthesis, leveraging the advancements in\ndeep generative models. However, such works mainly focus on replicating a\nsingle sound class or a textual sound description, neglecting temporal\ninformation, which is crucial in the practical applications of Foley sound. We\npresent T-Foley, a Temporal-event-guided waveform generation model for Foley\nsound synthesis. T-Foley generates high-quality audio using two conditions: the\nsound class and temporal event feature. For temporal conditioning, we devise a\ntemporal event feature and a novel conditioning technique named Block-FiLM.\nT-Foley achieves superior performance in both objective and subjective\nevaluation metrics and generates Foley sound well-synchronized with the\ntemporal events. Additionally, we showcase T-Foley's practical applications,\nparticularly in scenarios involving vocal mimicry for temporal event control.\nWe show the demo on our companion website.\n", "date": "2024-01-17", "categories": []}
{"id": "2312.14507", "title": "Unsupervised Harmonic Parameter Estimation Using Differentiable DSP and\n  Spectral Optimal Transport", "abstract": "  In neural audio signal processing, pitch conditioning has been used to\nenhance the performance of synthesizers. However, jointly training pitch\nestimators and synthesizers is a challenge when using standard audio-to-audio\nreconstruction loss, leading to reliance on external pitch trackers. To address\nthis issue, we propose using a spectral loss function inspired by optimal\ntransportation theory that minimizes the displacement of spectral energy. We\nvalidate this approach through an unsupervised autoencoding task that fits a\nharmonic template to harmonic signals. We jointly estimate the fundamental\nfrequency and amplitudes of harmonics using a lightweight encoder and\nreconstruct the signals using a differentiable harmonic synthesizer. The\nproposed approach offers a promising direction for improving unsupervised\nparameter estimation in neural audio applications.\n", "date": "2023-12-22", "categories": []}
{"id": "2401.06913", "title": "Microphone Conversion: Mitigating Device Variability in Sound Event\n  Classification", "abstract": "  In this study, we introduce a new augmentation technique to enhance the\nresilience of sound event classification (SEC) systems against device\nvariability through the use of CycleGAN. We also present a unique dataset to\nevaluate this method. As SEC systems become increasingly common, it is crucial\nthat they work well with audio from diverse recording devices. Our method\naddresses limited device diversity in training data by enabling unpaired\ntraining to transform input spectrograms as if they are recorded on a different\ndevice. Our experiments show that our approach outperforms existing methods in\ngeneralization by 5.2% - 11.5% in weighted f1 score. Additionally, it surpasses\nthe current methods in adaptability across diverse recording devices by\nachieving a 6.5% - 12.8% improvement in weighted f1 score.\n", "date": "2024-01-12", "categories": []}
{"id": "2401.08415", "title": "From Coarse to Fine: Efficient Training for Audio Spectrogram\n  Transformers", "abstract": "  Transformers have become central to recent advances in audio classification.\nHowever, training an audio spectrogram transformer, e.g. AST, from scratch can\nbe resource and time-intensive. Furthermore, the complexity of transformers\nheavily depends on the input audio spectrogram size. In this work, we aim to\noptimize AST training by linking to the resolution in the time-axis. We\nintroduce multi-phase training of audio spectrogram transformers by connecting\nthe seminal idea of coarse-to-fine with transformer models. To achieve this, we\npropose a set of methods for temporal compression. By employing one of these\nmethods, the transformer model learns from lower-resolution (coarse) data in\nthe initial phases, and then is fine-tuned with high-resolution data later in a\ncurriculum learning strategy. Experimental results demonstrate that the\nproposed training mechanism for AST leads to improved (or on-par) performance\nwith faster convergence, i.e. requiring fewer computational resources and less\ntime. This approach is also generalizable to other AST-based methods regardless\nof their learning paradigms.\n", "date": "2024-01-16", "categories": []}
{"id": "2210.01719", "title": "Learning Temporal Resolution in Spectrogram for Audio Classification", "abstract": "  The audio spectrogram is a time-frequency representation that has been widely\nused for audio classification. One of the key attributes of the audio\nspectrogram is the temporal resolution, which depends on the hop size used in\nthe Short-Time Fourier Transform (STFT). Previous works generally assume the\nhop size should be a constant value (e.g., 10 ms). However, a fixed temporal\nresolution is not always optimal for different types of sound. The temporal\nresolution affects not only classification accuracy but also computational\ncost. This paper proposes a novel method, DiffRes, that enables differentiable\ntemporal resolution modeling for audio classification. Given a spectrogram\ncalculated with a fixed hop size, DiffRes merges non-essential time frames\nwhile preserving important frames. DiffRes acts as a \"drop-in\" module between\nan audio spectrogram and a classifier and can be jointly optimized with the\nclassification task. We evaluate DiffRes on five audio classification tasks,\nusing mel-spectrograms as the acoustic features, followed by off-the-shelf\nclassifier backbones. Compared with previous methods using the fixed temporal\nresolution, the DiffRes-based method can achieve the equivalent or better\nclassification accuracy with at least 25% computational cost reduction. We\nfurther show that DiffRes can improve classification accuracy by increasing the\ntemporal resolution of input acoustic features, without adding to the\ncomputational cost.\n", "date": "2022-10-04", "categories": []}
{"id": "2311.15830", "title": "A-JEPA: Joint-Embedding Predictive Architecture Can Listen", "abstract": "  This paper presents that the masked-modeling principle driving the success of\nlarge foundational vision models can be effectively applied to audio by making\npredictions in a latent space. We introduce Audio-based Joint-Embedding\nPredictive Architecture (A-JEPA), a simple extension method for self-supervised\nlearning from the audio spectrum. Following the design of I-JEPA, our A-JEPA\nencodes visible audio spectrogram patches with a curriculum masking strategy\nvia context encoder, and predicts the representations of regions sampled at\nwell-designed locations. The target representations of those regions are\nextracted by the exponential moving average of context encoder, \\emph{i.e.},\ntarget encoder, on the whole spectrogram. We find it beneficial to transfer\nrandom block masking into time-frequency aware masking in a curriculum manner,\nconsidering the complexity of highly correlated in local time and frequency in\naudio spectrograms. To enhance contextual semantic understanding and\nrobustness, we fine-tune the encoder with a regularized masking on target\ndatasets, instead of input dropping or zero. Empirically, when built with\nVision Transformers structure, we find A-JEPA to be highly scalable and sets\nnew state-of-the-art performance on multiple audio and speech classification\ntasks, outperforming other recent models that use externally supervised\npre-training.\n", "date": "2023-11-27", "categories": []}
{"id": "2401.04737", "title": "Music Genre Classification: A Comparative Analysis of CNN and XGBoost\n  Approaches with Mel-frequency cepstral coefficients and Mel Spectrograms", "abstract": "  In recent years, various well-designed algorithms have empowered music\nplatforms to provide content based on one's preferences. Music genres are\ndefined through various aspects, including acoustic features and cultural\nconsiderations. Music genre classification works well with content-based\nfiltering, which recommends content based on music similarity to users. Given a\nconsiderable dataset, one premise is automatic annotation using machine\nlearning or deep learning methods that can effectively classify audio files.\nThe effectiveness of systems largely depends on feature and model selection, as\ndifferent architectures and features can facilitate each other and yield\ndifferent results. In this study, we conduct a comparative study investigating\nthe performances of three models: a proposed convolutional neural network\n(CNN), the VGG16 with fully connected layers (FC), and an eXtreme Gradient\nBoosting (XGBoost) approach on different features: 30-second Mel spectrogram\nand 3-second Mel-frequency cepstral coefficients (MFCCs). The results show that\nthe MFCC XGBoost model outperformed the others. Furthermore, applying data\nsegmentation in the data preprocessing phase can significantly enhance the\nperformance of the CNNs.\n", "date": "2024-01-09", "categories": []}
{"id": "2401.04935", "title": "Learning Audio Concepts from Counterfactual Natural Language", "abstract": "  Conventional audio classification relied on predefined classes, lacking the\nability to learn from free-form text. Recent methods unlock learning joint\naudio-text embeddings from raw audio-text pairs describing audio in natural\nlanguage. Despite recent advancements, there is little exploration of\nsystematic methods to train models for recognizing sound events and sources in\nalternative scenarios, such as distinguishing fireworks from gunshots at\noutdoor events in similar situations. This study introduces causal reasoning\nand counterfactual analysis in the audio domain. We use counterfactual\ninstances and include them in our model across different aspects. Our model\nconsiders acoustic characteristics and sound source information from\nhuman-annotated reference texts. To validate the effectiveness of our model, we\nconducted pre-training utilizing multiple audio captioning datasets. We then\nevaluate with several common downstream tasks, demonstrating the merits of the\nproposed method as one of the first works leveraging counterfactual information\nin audio domain. Specifically, the top-1 accuracy in open-ended language-based\naudio retrieval task increased by more than 43%.\n", "date": "2024-01-10", "categories": []}
{"id": "2310.09388", "title": "CORN: Co-Trained Full- And No-Reference Speech Quality Assessment", "abstract": "  Perceptual evaluation constitutes a crucial aspect of various\naudio-processing tasks. Full reference (FR) or similarity-based metrics rely on\nhigh-quality reference recordings, to which lower-quality or corrupted versions\nof the recording may be compared for evaluation. In contrast, no-reference (NR)\nmetrics evaluate a recording without relying on a reference. Both the FR and NR\napproaches exhibit advantages and drawbacks relative to each other. In this\npaper, we present a novel framework called CORN that amalgamates these dual\napproaches, concurrently training both FR and NR models together. After\ntraining, the models can be applied independently. We evaluate CORN by\npredicting several common objective metrics and across two different\narchitectures. The NR model trained using CORN has access to a reference\nrecording during training, and thus, as one would expect, it consistently\noutperforms baseline NR models trained independently. Perhaps even more\nremarkable is that the CORN FR model also outperforms its baseline counterpart,\neven though it relies on the same training data and the same model\narchitecture. Thus, a single training regime produces two independently useful\nmodels, each outperforming independently trained models\n", "date": "2023-10-13", "categories": []}
{"id": "2401.04511", "title": "Zero Shot Audio to Audio Emotion Transfer With Speaker Disentanglement", "abstract": "  The problem of audio-to-audio (A2A) style transfer involves replacing the\nstyle features of the source audio with those from the target audio while\npreserving the content related attributes of the source audio. In this paper,\nwe propose an efficient approach, termed as Zero-shot Emotion Style Transfer\n(ZEST), that allows the transfer of emotional content present in the given\nsource audio with the one embedded in the target audio while retaining the\nspeaker and speech content from the source. The proposed system builds upon\ndecomposing speech into semantic tokens, speaker representations and emotion\nembeddings. Using these factors, we propose a framework to reconstruct the\npitch contour of the given speech signal and train a decoder that reconstructs\nthe speech signal. The model is trained using a self-supervision based\nreconstruction loss. During conversion, the emotion embedding is alone derived\nfrom the target audio, while rest of the factors are derived from the source\naudio. In our experiments, we show that, even without using parallel training\ndata or labels from the source or target audio, we illustrate zero shot emotion\ntransfer capabilities of the proposed ZEST model using objective and subjective\nquality evaluations.\n", "date": "2024-01-09", "categories": []}
{"id": "2401.04558", "title": "HyperGANStrument: Instrument Sound Synthesis and Editing with\n  Pitch-Invariant Hypernetworks", "abstract": "  GANStrument, exploiting GANs with a pitch-invariant feature extractor and\ninstance conditioning technique, has shown remarkable capabilities in\nsynthesizing realistic instrument sounds. To further improve the reconstruction\nability and pitch accuracy to enhance the editability of user-provided sound,\nwe propose HyperGANStrument, which introduces a pitch-invariant hypernetwork to\nmodulate the weights of a pre-trained GANStrument generator, given a one-shot\nsound as input. The hypernetwork modulation provides feedback for the generator\nin the reconstruction of the input sound. In addition, we take advantage of an\nadversarial fine-tuning scheme for the hypernetwork to improve the\nreconstruction fidelity and generation diversity of the generator. Experimental\nresults show that the proposed model not only enhances the generation\ncapability of GANStrument but also significantly improves the editability of\nsynthesized sounds. Audio examples are available at the online demo page.\n", "date": "2024-01-09", "categories": []}
{"id": "2312.16228", "title": "Deformable Audio Transformer for Audio Event Detection", "abstract": "  Transformers have achieved promising results on a variety of tasks. However,\nthe quadratic complexity in self-attention computation has limited the\napplications, especially in low-resource settings and mobile or edge devices.\nExisting works have proposed to exploit hand-crafted attention patterns to\nreduce computation complexity. However, such hand-crafted patterns are\ndata-agnostic and may not be optimal. Hence, it is likely that relevant keys or\nvalues are being reduced, while less important ones are still preserved. Based\non this key insight, we propose a novel deformable audio Transformer for audio\nrecognition, named DATAR, where a deformable attention equipping with a pyramid\ntransformer backbone is constructed and learnable. Such an architecture has\nbeen proven effective in prediction tasks,~\\textit{e.g.}, event classification.\nMoreover, we identify that the deformable attention map computation may\nover-simplify the input feature, which can be further enhanced. Hence, we\nintroduce a learnable input adaptor to alleviate this issue, and DATAR achieves\nstate-of-the-art performance.\n", "date": "2023-12-24", "categories": []}
{"id": "2401.02996", "title": "An AI-enabled Bias-Free Respiratory Disease Diagnosis Model using Cough\n  Audio: A Case Study for COVID-19", "abstract": "  Cough-based diagnosis for Respiratory Diseases (RDs) using Artificial\nIntelligence (AI) has attracted considerable attention, yet many existing\nstudies overlook confounding variables in their predictive models. These\nvariables can distort the relationship between cough recordings (input data)\nand RD status (output variable), leading to biased associations and unrealistic\nmodel performance. To address this gap, we propose the Bias Free Network\n(RBFNet), an end to end solution that effectively mitigates the impact of\nconfounders in the training data distribution. RBFNet ensures accurate and\nunbiased RD diagnosis features, emphasizing its relevance by incorporating a\nCOVID19 dataset in this study. This approach aims to enhance the reliability of\nAI based RD diagnosis models by navigating the challenges posed by confounding\nvariables. A hybrid of a Convolutional Neural Networks (CNN) and Long-Short\nTerm Memory (LSTM) networks is proposed for the feature encoder module of\nRBFNet. An additional bias predictor is incorporated in the classification\nscheme to formulate a conditional Generative Adversarial Network (cGAN) which\nhelps in decorrelating the impact of confounding variables from RD prediction.\nThe merit of RBFNet is demonstrated by comparing classification performance\nwith State of The Art (SoTA) Deep Learning (DL) model (CNN LSTM) after training\non different unbalanced COVID-19 data sets, created by using a large scale\nproprietary cough data set. RBF-Net proved its robustness against extremely\nbiased training scenarios by achieving test set accuracies of 84.1%, 84.6%, and\n80.5% for the following confounding variables gender, age, and smoking status,\nrespectively. RBF-Net outperforms the CNN-LSTM model test set accuracies by\n5.5%, 7.7%, and 8.2%, respectively\n", "date": "2024-01-04", "categories": []}
{"id": "2401.03497", "title": "EAT: Self-Supervised Pre-Training with Efficient Audio Transformer", "abstract": "  Audio self-supervised learning (SSL) pre-training, which aims to learn good\nrepresentations from unlabeled audio, has made remarkable progress. However,\nthe extensive computational demands during pre-training pose a significant\nbarrier to the potential application and optimization of audio SSL models. In\nthis paper, inspired by the success of data2vec 2.0 in image modality and\nAudio-MAE in audio modality, we introduce Efficient Audio Transformer (EAT) to\nfurther improve the effectiveness and efficiency in audio SSL. The proposed EAT\nadopts the bootstrap self-supervised training paradigm to the audio domain. A\nnovel Utterance-Frame Objective (UFO) is designed to enhance the modeling\ncapability of acoustic events. Furthermore, we reveal that the masking strategy\nis critical in audio SSL pre-training, and superior audio representations can\nbe obtained with large inverse block masks. Experiment results demonstrate that\nEAT achieves state-of-the-art (SOTA) performance on a range of audio-related\ntasks, including AudioSet (AS-2M, AS-20K), ESC-50, and SPC-2, along with a\nsignificant pre-training speedup up to ~15x compared to existing audio SSL\nmodels.\n", "date": "2024-01-07", "categories": []}
{"id": "2310.03937", "title": "Diffusion Models as Masked Audio-Video Learners", "abstract": "  Over the past several years, the synchronization between audio and visual\nsignals has been leveraged to learn richer audio-visual representations. Aided\nby the large availability of unlabeled videos, many unsupervised training\nframeworks have demonstrated impressive results in various downstream audio and\nvideo tasks. Recently, Masked Audio-Video Learners (MAViL) has emerged as a\nstate-of-the-art audio-video pre-training framework. MAViL couples contrastive\nlearning with masked autoencoding to jointly reconstruct audio spectrograms and\nvideo frames by fusing information from both modalities. In this paper, we\nstudy the potential synergy between diffusion models and MAViL, seeking to\nderive mutual benefits from these two frameworks. The incorporation of\ndiffusion into MAViL, combined with various training efficiency methodologies\nthat include the utilization of a masking ratio curriculum and adaptive batch\nsizing, results in a notable 32% reduction in pre-training Floating-Point\nOperations (FLOPS) and an 18% decrease in pre-training wall clock time.\nCrucially, this enhanced efficiency does not compromise the model's performance\nin downstream audio-classification tasks when compared to MAViL's performance.\n", "date": "2023-10-05", "categories": []}
{"id": "2305.00011", "title": "Adversarial Representation Learning for Robust Privacy Preservation in\n  Audio", "abstract": "  Sound event detection systems are widely used in various applications such as\nsurveillance and environmental monitoring where data is automatically\ncollected, processed, and sent to a cloud for sound recognition. However, this\nprocess may inadvertently reveal sensitive information about users or their\nsurroundings, hence raising privacy concerns. In this study, we propose a novel\nadversarial training method for learning representations of audio recordings\nthat effectively prevents the detection of speech activity from the latent\nfeatures of the recordings. The proposed method trains a model to generate\ninvariant latent representations of speech-containing audio recordings that\ncannot be distinguished from non-speech recordings by a speech classifier. The\nnovelty of our work is in the optimization algorithm, where the speech\nclassifier's weights are regularly replaced with the weights of classifiers\ntrained in a supervised manner. This increases the discrimination power of the\nspeech classifier constantly during the adversarial training, motivating the\nmodel to generate latent representations in which speech is not\ndistinguishable, even using new speech classifiers trained outside the\nadversarial training loop. The proposed method is evaluated against a baseline\napproach with no privacy measures and a prior adversarial training method,\ndemonstrating a significant reduction in privacy violations compared to the\nbaseline approach. Additionally, we show that the prior adversarial method is\npractically ineffective for this purpose.\n", "date": "2023-04-29", "categories": []}
{"id": "2308.06382", "title": "Phoneme Hallucinator: One-shot Voice Conversion via Set Expansion", "abstract": "  Voice conversion (VC) aims at altering a person's voice to make it sound\nsimilar to the voice of another person while preserving linguistic content.\nExisting methods suffer from a dilemma between content intelligibility and\nspeaker similarity; i.e., methods with higher intelligibility usually have a\nlower speaker similarity, while methods with higher speaker similarity usually\nrequire plenty of target speaker voice data to achieve high intelligibility. In\nthis work, we propose a novel method \\textit{Phoneme Hallucinator} that\nachieves the best of both worlds. Phoneme Hallucinator is a one-shot VC model;\nit adopts a novel model to hallucinate diversified and high-fidelity target\nspeaker phonemes based just on a short target speaker voice (e.g. 3 seconds).\nThe hallucinated phonemes are then exploited to perform neighbor-based voice\nconversion. Our model is a text-free, any-to-any VC model that requires no text\nannotations and supports conversion to any unseen speaker. Objective and\nsubjective evaluations show that \\textit{Phoneme Hallucinator} outperforms\nexisting VC methods for both intelligibility and speaker similarity.\n", "date": "2023-08-11", "categories": []}
{"id": "2312.15190", "title": "SAIC: Integration of Speech Anonymization and Identity Classification", "abstract": "  Speech anonymization and de-identification have garnered significant\nattention recently, especially in the healthcare area including telehealth\nconsultations, patient voiceprint matching, and patient real-time monitoring.\nSpeaker identity classification tasks, which involve recognizing specific\nspeakers from audio to learn identity features, are crucial for\nde-identification. Since rare studies have effectively combined speech\nanonymization with identity classification, we propose SAIC - an innovative\npipeline for integrating Speech Anonymization and Identity Classification. SAIC\ndemonstrates remarkable performance and reaches state-of-the-art in the speaker\nidentity classification task on the Voxceleb1 dataset, with a top-1 accuracy of\n96.1%. Although SAIC is not trained or evaluated specifically on clinical data,\nthe result strongly proves the model's effectiveness and the possibility to\ngeneralize into the healthcare area, providing insightful guidance for future\nwork.\n", "date": "2023-12-23", "categories": []}
{"id": "2312.15821", "title": "Audiobox: Unified Audio Generation with Natural Language Prompts", "abstract": "  Audio is an essential part of our life, but creating it often requires\nexpertise and is time-consuming. Research communities have made great progress\nover the past year advancing the performance of large scale audio generative\nmodels for a single modality (speech, sound, or music) through adopting more\npowerful generative models and scaling data. However, these models lack\ncontrollability in several aspects: speech generation models cannot synthesize\nnovel styles based on text description and are limited on domain coverage such\nas outdoor environments; sound generation models only provide coarse-grained\ncontrol based on descriptions like \"a person speaking\" and would only generate\nmumbling human voices. This paper presents Audiobox, a unified model based on\nflow-matching that is capable of generating various audio modalities. We design\ndescription-based and example-based prompting to enhance controllability and\nunify speech and sound generation paradigms. We allow transcript, vocal, and\nother audio styles to be controlled independently when generating speech. To\nimprove model generalization with limited labels, we adapt a self-supervised\ninfilling objective to pre-train on large quantities of unlabeled audio.\nAudiobox sets new benchmarks on speech and sound generation (0.745 similarity\non Librispeech for zero-shot TTS; 0.77 FAD on AudioCaps for text-to-sound) and\nunlocks new methods for generating audio with novel vocal and acoustic styles.\nWe further integrate Bespoke Solvers, which speeds up generation by over 25\ntimes compared to the default ODE solver for flow-matching, without loss of\nperformance on several tasks. Our demo is available at\nhttps://audiobox.metademolab.com/\n", "date": "2023-12-25", "categories": []}
{"id": "2311.07919", "title": "Qwen-Audio: Advancing Universal Audio Understanding via Unified\n  Large-Scale Audio-Language Models", "abstract": "  Recently, instruction-following audio-language models have received broad\nattention for audio interaction with humans. However, the absence of\npre-trained audio models capable of handling diverse audio types and tasks has\nhindered progress in this field. Consequently, most existing works have only\nbeen able to support a limited range of interaction capabilities. In this\npaper, we develop the Qwen-Audio model and address this limitation by scaling\nup audio-language pre-training to cover over 30 tasks and various audio types,\nsuch as human speech, natural sounds, music, and songs, to facilitate universal\naudio understanding abilities. However, directly co-training all tasks and\ndatasets can lead to interference issues, as the textual labels associated with\ndifferent datasets exhibit considerable variations due to differences in task\nfocus, language, granularity of annotation, and text structure. To overcome the\none-to-many interference, we carefully design a multi-task training framework\nby conditioning on a sequence of hierarchical tags to the decoder for\nencouraging knowledge sharing and avoiding interference through shared and\nspecified tags respectively. Remarkably, Qwen-Audio achieves impressive\nperformance across diverse benchmark tasks without requiring any task-specific\nfine-tuning, surpassing its counterparts. Building upon the capabilities of\nQwen-Audio, we further develop Qwen-Audio-Chat, which allows for input from\nvarious audios and text inputs, enabling multi-turn dialogues and supporting\nvarious audio-central scenarios.\n", "date": "2023-11-14", "categories": []}
{"id": "2312.14005", "title": "On the choice of the optimal temporal support for audio classification\n  with Pre-trained embeddings", "abstract": "  Current state-of-the-art audio analysis systems rely on pre-trained embedding\nmodels, often used off-the-shelf as (frozen) feature extractors. Choosing the\nbest one for a set of tasks is the subject of many recent publications.\nHowever, one aspect often overlooked in these works is the influence of the\nduration of audio input considered to extract an embedding, which we refer to\nas Temporal Support (TS). In this work, we study the influence of the TS for\nwell-established or emerging pre-trained embeddings, chosen to represent\ndifferent types of architectures and learning paradigms. We conduct this\nevaluation using both musical instrument and environmental sound datasets,\nnamely OpenMIC, TAU Urban Acoustic Scenes 2020 Mobile, and ESC-50. We\nespecially highlight that Audio Spectrogram Transformer-based systems (PaSST\nand BEATs) remain effective with smaller TS, which therefore allows for a\ndrastic reduction in memory and computational cost. Moreover, we show that by\nchoosing the optimal TS we reach competitive results across all tasks. In\nparticular, we improve the state-of-the-art results on OpenMIC, using BEATs and\nPaSST without any fine-tuning.\n", "date": "2023-12-21", "categories": []}
{"id": "2312.14021", "title": "Leveraging Visual Supervision for Array-based Active Speaker Detection\n  and Localization", "abstract": "  Conventional audio-visual approaches for active speaker detection (ASD)\ntypically rely on visually pre-extracted face tracks and the corresponding\nsingle-channel audio to find the speaker in a video. Therefore, they tend to\nfail every time the face of the speaker is not visible. We demonstrate that a\nsimple audio convolutional recurrent neural network (CRNN) trained with spatial\ninput features extracted from multichannel audio can perform simultaneous\nhorizontal active speaker detection and localization (ASDL), independently of\nthe visual modality. To address the time and cost of generating ground truth\nlabels to train such a system, we propose a new self-supervised training\npipeline that embraces a ``student-teacher'' learning approach. A conventional\npre-trained active speaker detector is adopted as a ``teacher'' network to\nprovide the position of the speakers as pseudo-labels. The multichannel audio\n``student'' network is trained to generate the same results. At inference, the\nstudent network can generalize and locate also the occluded speakers that the\nteacher network is not able to detect visually, yielding considerable\nimprovements in recall rate. Experiments on the TragicTalkers dataset show that\nan audio network trained with the proposed self-supervised learning approach\ncan exceed the performance of the typical audio-visual methods and produce\nresults competitive with the costly conventional supervised training. We\ndemonstrate that improvements can be achieved when minimal manual supervision\nis introduced in the learning pipeline. Further gains may be sought with larger\ntraining sets and integrating vision with the multichannel audio system.\n", "date": "2023-12-21", "categories": []}
{"id": "2308.07170", "title": "Human Voice Pitch Estimation: A Convolutional Network with Auto-Labeled\n  and Synthetic Data", "abstract": "  In the domain of music and sound processing, pitch extraction plays a pivotal\nrole. Our research presents a specialized convolutional neural network designed\nfor pitch extraction, particularly from the human singing voice in acapella\nperformances. Notably, our approach combines synthetic data with auto-labeled\nacapella sung audio, creating a robust training environment. Evaluation across\ndatasets comprising synthetic sounds, opera recordings, and time-stretched\nvowels demonstrates its efficacy. This work paves the way for enhanced pitch\nextraction in both music and voice settings.\n", "date": "2023-08-14", "categories": []}
{"id": "2312.10949", "title": "Leveraged Mel spectrograms using Harmonic and Percussive Components in\n  Speech Emotion Recognition", "abstract": "  Speech Emotion Recognition (SER) affective technology enables the intelligent\nembedded devices to interact with sensitivity. Similarly, call centre employees\nrecognise customers' emotions from their pitch, energy, and tone of voice so as\nto modify their speech for a high-quality interaction with customers. This work\nexplores, for the first time, the effects of the harmonic and percussive\ncomponents of Mel spectrograms in SER. We attempt to leverage the Mel\nspectrogram by decomposing distinguishable acoustic features for exploitation\nin our proposed architecture, which includes a novel feature map generator\nalgorithm, a CNN-based network feature extractor and a multi-layer perceptron\n(MLP) classifier. This study specifically focuses on effective data\naugmentation techniques for building an enriched hybrid-based feature map. This\nprocess results in a function that outputs a 2D image so that it can be used as\ninput data for a pre-trained CNN-VGG16 feature extractor. Furthermore, we also\ninvestigate other acoustic features such as MFCCs, chromagram, spectral\ncontrast, and the tonnetz to assess our proposed framework. A test accuracy of\n92.79% on the Berlin EMO-DB database is achieved. Our result is higher than\nprevious works using CNN-VGG16.\n", "date": "2023-12-18", "categories": []}
{"id": "2309.02730", "title": "Stylebook: Content-Dependent Speaking Style Modeling for Any-to-Any\n  Voice Conversion using Only Speech Data", "abstract": "  While many recent any-to-any voice conversion models succeed in transferring\nsome target speech's style information to the converted speech, they still lack\nthe ability to faithfully reproduce the speaking style of the target speaker.\nIn this work, we propose a novel method to extract rich style information from\ntarget utterances and to efficiently transfer it to source speech content\nwithout requiring text transcriptions or speaker labeling. Our proposed\napproach introduces an attention mechanism utilizing a self-supervised learning\n(SSL) model to collect the speaking styles of a target speaker each\ncorresponding to the different phonetic content. The styles are represented\nwith a set of embeddings called stylebook. In the next step, the stylebook is\nattended with the source speech's phonetic content to determine the final\ntarget style for each source content. Finally, content information extracted\nfrom the source speech and content-dependent target style embeddings are fed\ninto a diffusion-based decoder to generate the converted speech\nmel-spectrogram. Experiment results show that our proposed method combined with\na diffusion-based generative model can achieve better speaker similarity in\nany-to-any voice conversion tasks when compared to baseline models, while the\nincrease in computational complexity with longer utterances is suppressed.\n", "date": "2023-09-06", "categories": []}
{"id": "2312.09578", "title": "Self-Supervised Learning for Anomalous Sound Detection", "abstract": "  State-of-the-art anomalous sound detection (ASD) systems are often trained by\nusing an auxiliary classification task to learn an embedding space. Doing so\nenables the system to learn embeddings that are robust to noise and are\nignoring non-target sound events but requires manually annotated meta\ninformation to be used as class labels. However, the less difficult the\nclassification task becomes, the less informative are the embeddings and the\nworse is the resulting ASD performance. A solution to this problem is to\nutilize self-supervised learning (SSL). In this work, feature exchange\n(FeatEx), a simple yet effective SSL approach for ASD, is proposed. In\naddition, FeatEx is compared to and combined with existing SSL approaches. As\nthe main result, a new state-of-the-art performance for the DCASE2023 ASD\ndataset is obtained that outperforms all other published results on this\ndataset by a large margin.\n", "date": "2023-12-15", "categories": []}
{"id": "2312.09603", "title": "Stethoscope-guided Supervised Contrastive Learning for Cross-domain\n  Adaptation on Respiratory Sound Classification", "abstract": "  Despite the remarkable advances in deep learning technology, achieving\nsatisfactory performance in lung sound classification remains a challenge due\nto the scarcity of available data. Moreover, the respiratory sound samples are\ncollected from a variety of electronic stethoscopes, which could potentially\nintroduce biases into the trained models. When a significant distribution shift\noccurs within the test dataset or in a practical scenario, it can substantially\ndecrease the performance. To tackle this issue, we introduce cross-domain\nadaptation techniques, which transfer the knowledge from a source domain to a\ndistinct target domain. In particular, by considering different stethoscope\ntypes as individual domains, we propose a novel stethoscope-guided supervised\ncontrastive learning approach. This method can mitigate any domain-related\ndisparities and thus enables the model to distinguish respiratory sounds of the\nrecording variation of the stethoscope. The experimental results on the ICBHI\ndataset demonstrate that the proposed methods are effective in reducing the\ndomain dependency and achieving the ICBHI Score of 61.71%, which is a\nsignificant improvement of 2.16% over the baseline.\n", "date": "2023-12-15", "categories": []}
{"id": "2312.09651", "title": "What to Remember: Self-Adaptive Continual Learning for Audio Deepfake\n  Detection", "abstract": "  The rapid evolution of speech synthesis and voice conversion has raised\nsubstantial concerns due to the potential misuse of such technology, prompting\na pressing need for effective audio deepfake detection mechanisms. Existing\ndetection models have shown remarkable success in discriminating known deepfake\naudio, but struggle when encountering new attack types. To address this\nchallenge, one of the emergent effective approaches is continual learning. In\nthis paper, we propose a continual learning approach called Radian Weight\nModification (RWM) for audio deepfake detection. The fundamental concept\nunderlying RWM involves categorizing all classes into two groups: those with\ncompact feature distributions across tasks, such as genuine audio, and those\nwith more spread-out distributions, like various types of fake audio. These\ndistinctions are quantified by means of the in-class cosine distance, which\nsubsequently serves as the basis for RWM to introduce a trainable gradient\nmodification direction for distinct data types. Experimental evaluations\nagainst mainstream continual learning methods reveal the superiority of RWM in\nterms of knowledge acquisition and mitigating forgetting in audio deepfake\ndetection. Furthermore, RWM's applicability extends beyond audio deepfake\ndetection, demonstrating its potential significance in diverse machine learning\ndomains such as image recognition.\n", "date": "2023-12-15", "categories": []}
{"id": "2308.09300", "title": "V2A-Mapper: A Lightweight Solution for Vision-to-Audio Generation by\n  Connecting Foundation Models", "abstract": "  Building artificial intelligence (AI) systems on top of a set of foundation\nmodels (FMs) is becoming a new paradigm in AI research. Their representative\nand generative abilities learnt from vast amounts of data can be easily adapted\nand transferred to a wide range of downstream tasks without extra training from\nscratch. However, leveraging FMs in cross-modal generation remains\nunder-researched when audio modality is involved. On the other hand,\nautomatically generating semantically-relevant sound from visual input is an\nimportant problem in cross-modal generation studies. To solve this\nvision-to-audio (V2A) generation problem, existing methods tend to design and\nbuild complex systems from scratch using modestly sized datasets. In this\npaper, we propose a lightweight solution to this problem by leveraging\nfoundation models, specifically CLIP, CLAP, and AudioLDM. We first investigate\nthe domain gap between the latent space of the visual CLIP and the auditory\nCLAP models. Then we propose a simple yet effective mapper mechanism\n(V2A-Mapper) to bridge the domain gap by translating the visual input between\nCLIP and CLAP spaces. Conditioned on the translated CLAP embedding, pretrained\naudio generative FM AudioLDM is adopted to produce high-fidelity and\nvisually-aligned sound. Compared to previous approaches, our method only\nrequires a quick training of the V2A-Mapper. We further analyze and conduct\nextensive experiments on the choice of the V2A-Mapper and show that a\ngenerative mapper is better at fidelity and variability (FD) while a regression\nmapper is slightly better at relevance (CS). Both objective and subjective\nevaluation on two V2A datasets demonstrate the superiority of our proposed\nmethod compared to current state-of-the-art approaches - trained with 86% fewer\nparameters but achieving 53% and 19% improvement in FD and CS, respectively.\n", "date": "2023-08-18", "categories": []}
{"id": "2309.15643", "title": "Why do Angular Margin Losses work well for Semi-Supervised Anomalous\n  Sound Detection?", "abstract": "  State-of-the-art anomalous sound detection systems often utilize angular\nmargin losses to learn suitable representations of acoustic data using an\nauxiliary task, which usually is a supervised or self-supervised classification\ntask. The underlying idea is that, in order to solve this auxiliary task,\nspecific information about normal data needs to be captured in the learned\nrepresentations and that this information is also sufficient to differentiate\nbetween normal and anomalous samples. Especially in noisy conditions,\ndiscriminative models based on angular margin losses tend to significantly\noutperform systems based on generative or one-class models. The goal of this\nwork is to investigate why using angular margin losses with auxiliary tasks\nworks well for detecting anomalous sounds. To this end, it is shown, both\ntheoretically and experimentally, that minimizing angular margin losses also\nminimizes compactness loss while inherently preventing learning trivial\nsolutions. Furthermore, multiple experiments are conducted to show that using a\nrelated classification task as an auxiliary task teaches the model to learn\nrepresentations suitable for detecting anomalous sounds in noisy conditions.\nAmong these experiments are performance evaluations, visualizing the embedding\nspace with t-SNE and visualizing the input representations with respect to the\nanomaly score using randomized input sampling for explanation.\n", "date": "2023-09-27", "categories": []}
{"id": "2309.14405", "title": "Joint Audio and Speech Understanding", "abstract": "  Humans are surrounded by audio signals that include both speech and\nnon-speech sounds. The recognition and understanding of speech and non-speech\naudio events, along with a profound comprehension of the relationship between\nthem, constitute fundamental cognitive capabilities. For the first time, we\nbuild a machine learning model, called LTU-AS, that has a conceptually similar\nuniversal audio perception and advanced reasoning ability. Specifically, by\nintegrating Whisper as a perception module and LLaMA as a reasoning module,\nLTU-AS can simultaneously recognize and jointly understand spoken text, speech\nparalinguistics, and non-speech audio events - almost everything perceivable\nfrom audio signals.\n", "date": "2023-09-25", "categories": []}
{"id": "2312.03455", "title": "Data is Overrated: Perceptual Metrics Can Lead Learning in the Absence\n  of Training Data", "abstract": "  Perceptual metrics are traditionally used to evaluate the quality of natural\nsignals, such as images and audio. They are designed to mimic the perceptual\nbehaviour of human observers and usually reflect structures found in natural\nsignals. This motivates their use as loss functions for training generative\nmodels such that models will learn to capture the structure held in the metric.\nWe take this idea to the extreme in the audio domain by training a compressive\nautoencoder to reconstruct uniform noise, in lieu of natural data. We show that\ntraining with perceptual losses improves the reconstruction of spectrograms and\nre-synthesized audio at test time over models trained with a standard Euclidean\nloss. This demonstrates better generalisation to unseen natural signals when\nusing perceptual metrics.\n", "date": "2023-12-06", "categories": []}
{"id": "2312.03666", "title": "Towards small and accurate convolutional neural networks for acoustic\n  biodiversity monitoring", "abstract": "  Automated classification of animal sounds is a prerequisite for large-scale\nmonitoring of biodiversity. Convolutional Neural Networks (CNNs) are among the\nmost promising algorithms but they are slow, often achieve poor classification\nin the field and typically require large training data sets. Our objective was\nto design CNNs that are fast at inference time and achieve good classification\nperformance while learning from moderate-sized data. Recordings from a\nrainforest ecosystem were used. Start and end-point of sounds from 20 bird\nspecies were manually annotated. Spectrograms from 10 second segments were used\nas CNN input. We designed simple CNNs with a frequency unwrapping layer\n(SIMP-FU models) such that any output unit was connected to all spectrogram\nfrequencies but only to a sub-region of time, the Receptive Field (RF). Our\nmodels allowed experimentation with different RF durations. Models either used\nthe time-indexed labels that encode start and end-point of sounds or simpler\nsegment-level labels. Models learning from time-indexed labels performed\nconsiderably better than their segment-level counterparts. Best classification\nperformances was achieved for models with intermediate RF duration of 1.5\nseconds. The best SIMP-FU models achieved AUCs over 0.95 in 18 of 20 classes on\nthe test set. On compact low-cost hardware the best SIMP-FU models evaluated up\nto seven times faster than real-time data acquisition. RF duration was a major\ndriver of classification performance. The optimum of 1.5 s was in the same\nrange as the duration of the sounds. Our models achieved good classification\nperformance while learning from moderate-sized training data. This is explained\nby the usage of time-indexed labels during training and adequately sized RF.\nResults confirm the feasibility of deploying small CNNs with good\nclassification performance on compact low-cost devices.\n", "date": "2023-12-06", "categories": []}
{"id": "2312.01017", "title": "Unveiling the Power of Audio-Visual Early Fusion Transformers with Dense\n  Interactions through Masked Modeling", "abstract": "  Humans possess a remarkable ability to integrate auditory and visual\ninformation, enabling a deeper understanding of the surrounding environment.\nThis early fusion of audio and visual cues, demonstrated through cognitive\npsychology and neuroscience research, offers promising potential for developing\nmultimodal perception models. However, training early fusion architectures\nposes significant challenges, as the increased model expressivity requires\nrobust learning frameworks to harness their enhanced capabilities. In this\npaper, we address this challenge by leveraging the masked reconstruction\nframework, previously successful in unimodal settings, to train audio-visual\nencoders with early fusion. Additionally, we propose an attention-based fusion\nmodule that captures interactions between local audio and visual\nrepresentations, enhancing the model's ability to capture fine-grained\ninteractions. While effective, this procedure can become computationally\nintractable, as the number of local representations increases. Thus, to address\nthe computational complexity, we propose an alternative procedure that\nfactorizes the local representations before representing audio-visual\ninteractions. Extensive evaluations on a variety of datasets demonstrate the\nsuperiority of our approach in audio-event classification, visual sound\nlocalization, sound separation, and audio-visual segmentation. These\ncontributions enable the efficient training of deeply integrated audio-visual\nmodels and significantly advance the usefulness of early fusion architectures.\n", "date": "2023-12-02", "categories": []}
{"id": "2312.01092", "title": "A Semi-Supervised Deep Learning Approach to Dataset Collection for\n  Query-By-Humming Task", "abstract": "  Query-by-Humming (QbH) is a task that involves finding the most relevant song\nbased on a hummed or sung fragment. Despite recent successful commercial\nsolutions, implementing QbH systems remains challenging due to the lack of\nhigh-quality datasets for training machine learning models. In this paper, we\npropose a deep learning data collection technique and introduce Covers and\nHummings Aligned Dataset (CHAD), a novel dataset that contains 18 hours of\nshort music fragments, paired with time-aligned hummed versions. To expand our\ndataset, we employ a semi-supervised model training pipeline that leverages the\nQbH task as a specialized case of cover song identification (CSI) task.\nStarting with a model trained on the initial dataset, we iteratively collect\ngroups of fragments of cover versions of the same song and retrain the model on\nthe extended data. Using this pipeline, we collect over 308 hours of additional\nmusic fragments, paired with time-aligned cover versions. The final model is\nsuccessfully applied to the QbH task and achieves competitive results on\nbenchmark datasets. Our study shows that the proposed dataset and training\npipeline can effectively facilitate the implementation of QbH systems.\n", "date": "2023-12-02", "categories": []}
{"id": "2304.10001", "title": "Weakly Supervised Detection of Baby Cry", "abstract": "  Detection of baby cries is an important part of baby monitoring and health\ncare. Almost all existing methods use supervised SVM, CNN, or their varieties.\nIn this work, we propose to use weakly supervised anomaly detection to detect a\nbaby cry. In this weak supervision, we only need weak annotation if there is a\ncry in an audio file. We design a data mining technique using the pre-trained\nVGGish feature extractor and an anomaly detection network on long untrimmed\naudio files. The obtained datasets are used to train a simple CNN feature\nnetwork for cry/non-cry classification. This CNN is then used as a feature\nextractor in an anomaly detection framework to achieve better cry detection\nperformance.\n", "date": "2023-04-19", "categories": []}
{"id": "2311.15959", "title": "CheapNET: Improving Light-weight speech enhancement network by projected\n  loss function", "abstract": "  Noise suppression and echo cancellation are critical in speech enhancement\nand essential for smart devices and real-time communication. Deployed in voice\nprocessing front-ends and edge devices, these algorithms must ensure efficient\nreal-time inference with low computational demands. Traditional edge-based\nnoise suppression often uses MSE-based amplitude spectrum mask training, but\nthis approach has limitations. We introduce a novel projection loss function,\ndiverging from MSE, to enhance noise suppression. This method uses projection\ntechniques to isolate key audio components from noise, significantly improving\nmodel performance. For echo cancellation, the function enables direct\npredictions on LAEC pre-processed outputs, substantially enhancing performance.\nOur noise suppression model achieves near state-of-the-art results with only\n3.1M parameters and 0.4GFlops/s computational load. Moreover, our echo\ncancellation model outperforms replicated industry-leading models, introducing\na new perspective in speech enhancement.\n", "date": "2023-11-27", "categories": []}
{"id": "2306.10548", "title": "MARBLE: Music Audio Representation Benchmark for Universal Evaluation", "abstract": "  In the era of extensive intersection between art and Artificial Intelligence\n(AI), such as image generation and fiction co-creation, AI for music remains\nrelatively nascent, particularly in music understanding. This is evident in the\nlimited work on deep music representations, the scarcity of large-scale\ndatasets, and the absence of a universal and community-driven benchmark. To\naddress this issue, we introduce the Music Audio Representation Benchmark for\nuniversaL Evaluation, termed MARBLE. It aims to provide a benchmark for various\nMusic Information Retrieval (MIR) tasks by defining a comprehensive taxonomy\nwith four hierarchy levels, including acoustic, performance, score, and\nhigh-level description. We then establish a unified protocol based on 14 tasks\non 8 public-available datasets, providing a fair and standard assessment of\nrepresentations of all open-sourced pre-trained models developed on music\nrecordings as baselines. Besides, MARBLE offers an easy-to-use, extendable, and\nreproducible suite for the community, with a clear statement on copyright\nissues on datasets. Results suggest recently proposed large-scale pre-trained\nmusical language models perform the best in most tasks, with room for further\nimprovement. The leaderboard and toolkit repository are published at\nhttps://marble-bm.shef.ac.uk to promote future music AI research.\n", "date": "2023-06-18", "categories": []}
{"id": "2307.15064", "title": "Self-Supervised Visual Acoustic Matching", "abstract": "  Acoustic matching aims to re-synthesize an audio clip to sound as if it were\nrecorded in a target acoustic environment. Existing methods assume access to\npaired training data, where the audio is observed in both source and target\nenvironments, but this limits the diversity of training data or requires the\nuse of simulated data or heuristics to create paired samples. We propose a\nself-supervised approach to visual acoustic matching where training samples\ninclude only the target scene image and audio -- without acoustically\nmismatched source audio for reference. Our approach jointly learns to\ndisentangle room acoustics and re-synthesize audio into the target environment,\nvia a conditional GAN framework and a novel metric that quantifies the level of\nresidual acoustic information in the de-biased audio. Training with either\nin-the-wild web data or simulated data, we demonstrate it outperforms the\nstate-of-the-art on multiple challenging datasets and a wide variety of\nreal-world audio and environments.\n", "date": "2023-07-27", "categories": []}
{"id": "2311.14275", "title": "Cooperative Dual Attention for Audio-Visual Speech Enhancement with\n  Facial Cues", "abstract": "  In this work, we focus on leveraging facial cues beyond the lip region for\nrobust Audio-Visual Speech Enhancement (AVSE). The facial region, encompassing\nthe lip region, reflects additional speech-related attributes such as gender,\nskin color, nationality, etc., which contribute to the effectiveness of AVSE.\nHowever, static and dynamic speech-unrelated attributes also exist, causing\nappearance changes during speech. To address these challenges, we propose a\nDual Attention Cooperative Framework, DualAVSE, to ignore speech-unrelated\ninformation, capture speech-related information with facial cues, and\ndynamically integrate it with the audio signal for AVSE. Specifically, we\nintroduce a spatial attention-based visual encoder to capture and enhance\nvisual speech information beyond the lip region, incorporating global facial\ncontext and automatically ignoring speech-unrelated information for robust\nvisual feature extraction. Additionally, a dynamic visual feature fusion\nstrategy is introduced by integrating a temporal-dimensional self-attention\nmodule, enabling the model to robustly handle facial variations. The acoustic\nnoise in the speaking process is variable, impacting audio quality. Therefore,\na dynamic fusion strategy for both audio and visual features is introduced to\naddress this issue. By integrating cooperative dual attention in the visual\nencoder and audio-visual fusion strategy, our model effectively extracts\nbeneficial speech information from both audio and visual cues for AVSE.\nThorough analysis and comparison on different datasets, including normal and\nchallenging cases with unreliable or absent visual information, consistently\nshow our model outperforming existing methods across multiple metrics.\n", "date": "2023-11-24", "categories": []}
{"id": "2309.13972", "title": "Audio classification with Dilated Convolution with Learnable Spacings", "abstract": "  Dilated convolution with learnable spacings (DCLS) is a recent convolution\nmethod in which the positions of the kernel elements are learned throughout\ntraining by backpropagation. Its interest has recently been demonstrated in\ncomputer vision (ImageNet classification and downstream tasks). Here we show\nthat DCLS is also useful for audio tagging using the AudioSet classification\nbenchmark. We took two state-of-the-art convolutional architectures using\ndepthwise separable convolutions (DSC), ConvNeXt and ConvFormer, and a hybrid\none using attention in addition, FastViT, and drop-in replaced all the DSC\nlayers by DCLS ones. This significantly improved the mean average precision\n(mAP) with the three architectures without increasing the number of parameters\nand with only a low cost on the throughput. The method code is based on PyTorch\nand is available at https://github.com/K-H-Ismail/DCLS-Audio\n", "date": "2023-09-25", "categories": []}
{"id": "2308.07121", "title": "Active Bird2Vec: Towards End-to-End Bird Sound Monitoring with\n  Transformers", "abstract": "  We propose a shift towards end-to-end learning in bird sound monitoring by\ncombining self-supervised (SSL) and deep active learning (DAL). Leveraging\ntransformer models, we aim to bypass traditional spectrogram conversions,\nenabling direct raw audio processing. ActiveBird2Vec is set to generate\nhigh-quality bird sound representations through SSL, potentially accelerating\nthe assessment of environmental changes and decision-making processes for wind\nfarms. Additionally, we seek to utilize the wide variety of bird vocalizations\nthrough DAL, reducing the reliance on extensively labeled datasets by human\nexperts. We plan to curate a comprehensive set of tasks through Huggingface\nDatasets, enhancing future comparability and reproducibility of bioacoustic\nresearch. A comparative analysis between various transformer models will be\nconducted to evaluate their proficiency in bird sound recognition tasks. We aim\nto accelerate the progression of avian bioacoustic research and contribute to\nmore effective conservation strategies.\n", "date": "2023-08-14", "categories": []}
{"id": "2309.03036", "title": "An Efficient Temporary Deepfake Location Approach Based Embeddings for\n  Partially Spoofed Audio Detection", "abstract": "  Partially spoofed audio detection is a challenging task, lying in the need to\naccurately locate the authenticity of audio at the frame level. To address this\nissue, we propose a fine-grained partially spoofed audio detection method,\nnamely Temporal Deepfake Location (TDL), which can effectively capture\ninformation of both features and locations. Specifically, our approach involves\ntwo novel parts: embedding similarity module and temporal convolution\noperation. To enhance the identification between the real and fake features,\nthe embedding similarity module is designed to generate an embedding space that\ncan separate the real frames from fake frames. To effectively concentrate on\nthe position information, temporal convolution operation is proposed to\ncalculate the frame-specific similarities among neighboring frames, and\ndynamically select informative neighbors to convolution. Extensive experiments\nshow that our method outperform baseline models in ASVspoof2019 Partial Spoof\ndataset and demonstrate superior performance even in the crossdataset scenario.\n", "date": "2023-09-06", "categories": []}
{"id": "2206.13272", "title": "Wideband Audio Waveform Evaluation Networks: Efficient, Accurate\n  Estimation of Speech Qualities", "abstract": "  Wideband Audio Waveform Evaluation Networks (WAWEnets) are convolutional\nneural networks that operate directly on wideband audio waveforms in order to\nproduce evaluations of those waveforms. In the present work these evaluations\ngive qualities of telecommunications speech (e.g., noisiness, intelligibility,\noverall speech quality). WAWEnets are no-reference networks because they do not\nrequire ``reference'' (original or undistorted) versions of the waveforms they\nevaluate. Our initial WAWEnet publication introduced four WAWEnets and each\nemulated the output of an established full-reference speech quality or\nintelligibility estimation algorithm. We have updated the WAWEnet architecture\nto be more efficient and effective. Here we present a single WAWEnet that\nclosely tracks seven different quality and intelligibility values. We create a\nsecond network that additionally tracks four subjective speech quality\ndimensions. We offer a third network that focuses on just subjective quality\nscores and achieves very high levels of agreement. This work has leveraged 334\nhours of speech in 13 languages, over two million full-reference target values\nand over 93,000 subjective mean opinion scores. We also interpret the operation\nof WAWEnets and identify the key to their operation using the language of\nsignal processing: ReLUs strategically move spectral information from non-DC\ncomponents into the DC component. The DC values of 96 output signals define a\nvector in a 96-D latent space and this vector is then mapped to a quality or\nintelligibility value for the input waveform.\n", "date": "2022-06-27", "categories": []}
{"id": "2311.07780", "title": "Parrot-Trained Adversarial Examples: Pushing the Practicality of\n  Black-Box Audio Attacks against Speaker Recognition Models", "abstract": "  Audio adversarial examples (AEs) have posed significant security challenges\nto real-world speaker recognition systems. Most black-box attacks still require\ncertain information from the speaker recognition model to be effective (e.g.,\nkeeping probing and requiring the knowledge of similarity scores). This work\naims to push the practicality of the black-box attacks by minimizing the\nattacker's knowledge about a target speaker recognition model. Although it is\nnot feasible for an attacker to succeed with completely zero knowledge, we\nassume that the attacker only knows a short (or a few seconds) speech sample of\na target speaker. Without any probing to gain further knowledge about the\ntarget model, we propose a new mechanism, called parrot training, to generate\nAEs against the target model. Motivated by recent advancements in voice\nconversion (VC), we propose to use the one short sentence knowledge to generate\nmore synthetic speech samples that sound like the target speaker, called parrot\nspeech. Then, we use these parrot speech samples to train a parrot-trained(PT)\nsurrogate model for the attacker. Under a joint transferability and perception\nframework, we investigate different ways to generate AEs on the PT model\n(called PT-AEs) to ensure the PT-AEs can be generated with high transferability\nto a black-box target model with good human perceptual quality. Real-world\nexperiments show that the resultant PT-AEs achieve the attack success rates of\n45.8% - 80.8% against the open-source models in the digital-line scenario and\n47.9% - 58.3% against smart devices, including Apple HomePod (Siri), Amazon\nEcho, and Google Home, in the over-the-air scenario.\n", "date": "2023-11-13", "categories": []}
{"id": "2306.09126", "title": "STARSS23: An Audio-Visual Dataset of Spatial Recordings of Real Scenes\n  with Spatiotemporal Annotations of Sound Events", "abstract": "  While direction of arrival (DOA) of sound events is generally estimated from\nmultichannel audio data recorded in a microphone array, sound events usually\nderive from visually perceptible source objects, e.g., sounds of footsteps come\nfrom the feet of a walker. This paper proposes an audio-visual sound event\nlocalization and detection (SELD) task, which uses multichannel audio and video\ninformation to estimate the temporal activation and DOA of target sound events.\nAudio-visual SELD systems can detect and localize sound events using signals\nfrom a microphone array and audio-visual correspondence. We also introduce an\naudio-visual dataset, Sony-TAu Realistic Spatial Soundscapes 2023 (STARSS23),\nwhich consists of multichannel audio data recorded with a microphone array,\nvideo data, and spatiotemporal annotation of sound events. Sound scenes in\nSTARSS23 are recorded with instructions, which guide recording participants to\nensure adequate activity and occurrences of sound events. STARSS23 also serves\nhuman-annotated temporal activation labels and human-confirmed DOA labels,\nwhich are based on tracking results of a motion capture system. Our benchmark\nresults demonstrate the benefits of using visual object positions in\naudio-visual SELD tasks. The data is available at\nhttps://zenodo.org/record/7880637.\n", "date": "2023-06-15", "categories": []}
{"id": "2311.07534", "title": "Unsupervised Musical Object Discovery from Audio", "abstract": "  Current object-centric learning models such as the popular SlotAttention\narchitecture allow for unsupervised visual scene decomposition. Our novel\nMusicSlots method adapts SlotAttention to the audio domain, to achieve\nunsupervised music decomposition. Since concepts of opacity and occlusion in\nvision have no auditory analogues, the softmax normalization of alpha masks in\nthe decoders of visual object-centric models is not well-suited for decomposing\naudio objects. MusicSlots overcomes this problem. We introduce a\nspectrogram-based multi-object music dataset tailored to evaluate\nobject-centric learning on western tonal music. MusicSlots achieves good\nperformance on unsupervised note discovery and outperforms several established\nbaselines on supervised note property prediction tasks.\n", "date": "2023-11-13", "categories": []}
{"id": "2311.07630", "title": "Cross-modal Generative Model for Visual-Guided Binaural Stereo\n  Generation", "abstract": "  Binaural stereo audio is recorded by imitating the way the human ear receives\nsound, which provides people with an immersive listening experience. Existing\napproaches leverage autoencoders and directly exploit visual spatial\ninformation to synthesize binaural stereo, resulting in a limited\nrepresentation of visual guidance. For the first time, we propose a visually\nguided generative adversarial approach for generating binaural stereo audio\nfrom mono audio. Specifically, we develop a Stereo Audio Generation Model\n(SAGM), which utilizes shared spatio-temporal visual information to guide the\ngenerator and the discriminator to work separately. The shared visual\ninformation is updated alternately in the generative adversarial stage,\nallowing the generator and discriminator to deliver their respective guided\nknowledge while visually sharing. The proposed method learns bidirectional\ncomplementary visual information, which facilitates the expression of visual\nguidance in generation. In addition, spatial perception is a crucial attribute\nof binaural stereo audio, and thus the evaluation of stereo spatial perception\nis essential. However, previous metrics failed to measure the spatial\nperception of audio. To this end, a metric to measure the spatial perception of\naudio is proposed for the first time. The proposed metric is capable of\nmeasuring the magnitude and direction of spatial perception in the temporal\ndimension. Further, considering its function, it is feasible to utilize it\ninstead of demanding user studies to some extent. The proposed method achieves\nstate-of-the-art performance on 2 datasets and 5 evaluation metrics.\nQualitative experiments and user studies demonstrate that the method generates\nspace-realistic stereo audio.\n", "date": "2023-11-13", "categories": []}
{"id": "2311.06368", "title": "The AeroSonicDB (YPAD-0523) Dataset for Acoustic Detection and\n  Classification of Aircraft", "abstract": "  The time and expense required to collect and label audio data has been a\nprohibitive factor in the availability of domain specific audio datasets. As\nthe predictive specificity of a classifier depends on the specificity of the\nlabels it is trained on, it follows that finely-labelled datasets are crucial\nfor advances in machine learning. Aiming to stimulate progress in the field of\nmachine listening, this paper introduces AeroSonicDB (YPAD-0523), a dataset of\nlow-flying aircraft sounds for training acoustic detection and classification\nsystems. This paper describes the method of exploiting ADS-B radio\ntransmissions to passively collect and label audio samples. Provides a summary\nof the collated dataset. Presents baseline results from three binary\nclassification models, then discusses the limitations of the current dataset\nand its future potential. The dataset contains 625 aircraft recordings ranging\nin event duration from 18 to 60 seconds, for a total of 8.87 hours of aircraft\naudio. These 625 samples feature 301 unique aircraft, each of which are\nsupplied with 14 supplementary (non-acoustic) labels to describe the aircraft.\nThe dataset also contains 3.52 hours of ambient background audio (\"silence\"),\nas a means to distinguish aircraft noise from other local environmental noises.\nAdditionally, 6 hours of urban soundscape recordings (with aircraft\nannotations) are included as an ancillary method for evaluating model\nperformance, and to provide a testing ground for real-time applications.\n", "date": "2023-11-10", "categories": []}
{"id": "2311.06480", "title": "Adversarial Fine-tuning using Generated Respiratory Sound to Address\n  Class Imbalance", "abstract": "  Deep generative models have emerged as a promising approach in the medical\nimage domain to address data scarcity. However, their use for sequential data\nlike respiratory sounds is less explored. In this work, we propose a\nstraightforward approach to augment imbalanced respiratory sound data using an\naudio diffusion model as a conditional neural vocoder. We also demonstrate a\nsimple yet effective adversarial fine-tuning method to align features between\nthe synthetic and real respiratory sound samples to improve respiratory sound\nclassification performance. Our experimental results on the ICBHI dataset\ndemonstrate that the proposed adversarial fine-tuning is effective, while only\nusing the conventional augmentation method shows performance degradation.\nMoreover, our method outperforms the baseline by 2.24% on the ICBHI Score and\nimproves the accuracy of the minority classes up to 26.58%. For the\nsupplementary material, we provide the code at\nhttps://github.com/kaen2891/adversarial_fine-tuning_using_generated_respiratory_sound.\n", "date": "2023-11-11", "categories": []}
{"id": "2311.05203", "title": "Whisper in Focus: Enhancing Stuttered Speech Classification with Encoder\n  Layer Optimization", "abstract": "  In recent years, advancements in the field of speech processing have led to\ncutting-edge deep learning algorithms with immense potential for real-world\napplications. The automated identification of stuttered speech is one of such\napplications that the researchers are addressing by employing deep learning\ntechniques. Recently, researchers have utilized Wav2vec2.0, a speech\nrecognition model to classify disfluency types in stuttered speech. Although\nWav2vec2.0 has shown commendable results, its ability to generalize across all\ndisfluency types is limited. In addition, since its base model uses 12 encoder\nlayers, it is considered a resource-intensive model. Our study unravels the\ncapabilities of Whisper for the classification of disfluency types in stuttered\nspeech. We have made notable contributions in three pivotal areas: enhancing\nthe quality of SEP28-k benchmark dataset, exploration of Whisper for\nclassification, and introducing an efficient encoder layer freezing strategy.\nThe optimized Whisper model has achieved the average F1-score of 0.81, which\nproffers its abilities. This study also unwinds the significance of deeper\nencoder layers in the identification of disfluency types, as the results\ndemonstrate their greater contribution compared to initial layers. This\nresearch represents substantial contributions, shifting the emphasis towards an\nefficient solution, thereby thriving towards prospective innovation.\n", "date": "2023-11-09", "categories": []}
{"id": "2311.04693", "title": "Diff-HierVC: Diffusion-based Hierarchical Voice Conversion with Robust\n  Pitch Generation and Masked Prior for Zero-shot Speaker Adaptation", "abstract": "  Although voice conversion (VC) systems have shown a remarkable ability to\ntransfer voice style, existing methods still have an inaccurate pitch and low\nspeaker adaptation quality. To address these challenges, we introduce\nDiff-HierVC, a hierarchical VC system based on two diffusion models. We first\nintroduce DiffPitch, which can effectively generate F0 with the target voice\nstyle. Subsequently, the generated F0 is fed to DiffVoice to convert the speech\nwith a target voice style. Furthermore, using the source-filter encoder, we\ndisentangle the speech and use the converted Mel-spectrogram as a data-driven\nprior in DiffVoice to improve the voice style transfer capacity. Finally, by\nusing the masked prior in diffusion models, our model can improve the speaker\nadaptation quality. Experimental results verify the superiority of our model in\npitch generation and voice style transfer performance, and our model also\nachieves a CER of 0.83% and EER of 3.29% in zero-shot VC scenarios.\n", "date": "2023-11-08", "categories": []}
{"id": "2306.04186", "title": "Self-supervised Audio Teacher-Student Transformer for Both Clip-level\n  and Frame-level Tasks", "abstract": "  Self-supervised learning (SSL) has emerged as a popular approach for learning\naudio representations. One goal of audio self-supervised pre-training is to\ntransfer knowledge to downstream audio tasks, generally including clip-level\nand frame-level tasks. While frame-level tasks are important for fine-grained\nacoustic scene/event understanding, prior studies primarily evaluate on\nclip-level downstream tasks. In order to tackle both clip-level and frame-level\ntasks, this paper proposes Audio Teacher-Student Transformer (ATST), with a\nclip-level version (named ATST-Clip) and a frame-level version (named\nATST-Frame), responsible for learning clip-level and frame-level\nrepresentations, respectively. Both methods use a Transformer encoder and a\nteacher-student training scheme. We have carefully designed the view creation\nstrategy for ATST-Clip and ATST-Frame. Specifically, ATST-Clip uses\nsegment-wise data augmentations, and ATST-Frame integrates frame-wise data\naugmentations and masking. Experimental results show that our ATST-Frame model\nobtains state-of-the-art (SOTA) performances on most of the clip-level and\nframe-level downstream tasks. Especially, it outperforms other models by a\nlarge margin on the frame-level sound event detection task. In addition, the\nperformance can be further improved by combining the two models through\nknowledge distillation. Our code is available online.\n", "date": "2023-06-07", "categories": []}
{"id": "2311.02482", "title": "Generalized zero-shot audio-to-intent classification", "abstract": "  Spoken language understanding systems using audio-only data are gaining\npopularity, yet their ability to handle unseen intents remains limited. In this\nstudy, we propose a generalized zero-shot audio-to-intent classification\nframework with only a few sample text sentences per intent. To achieve this, we\nfirst train a supervised audio-to-intent classifier by making use of a\nself-supervised pre-trained model. We then leverage a neural audio synthesizer\nto create audio embeddings for sample text utterances and perform generalized\nzero-shot classification on unseen intents using cosine similarity. We also\npropose a multimodal training strategy that incorporates lexical information\ninto the audio representation to improve zero-shot performance. Our multimodal\ntraining approach improves the accuracy of zero-shot intent classification on\nunseen intents of SLURP by 2.75% and 18.2% for the SLURP and internal\ngoal-oriented dialog datasets, respectively, compared to audio-only training.\n", "date": "2023-11-04", "categories": []}
{"id": "2311.01526", "title": "ATGNN: Audio Tagging Graph Neural Network", "abstract": "  Deep learning models such as CNNs and Transformers have achieved impressive\nperformance for end-to-end audio tagging. Recent works have shown that despite\nstacking multiple layers, the receptive field of CNNs remains severely limited.\nTransformers on the other hand are able to map global context through\nself-attention, but treat the spectrogram as a sequence of patches which is not\nflexible enough to capture irregular audio objects. In this work, we treat the\nspectrogram in a more flexible way by considering it as graph structure and\nprocess it with a novel graph neural architecture called ATGNN. ATGNN not only\ncombines the capability of CNNs with the global information sharing ability of\nGraph Neural Networks, but also maps semantic relationships between learnable\nclass embeddings and corresponding spectrogram regions. We evaluate ATGNN on\ntwo audio tagging tasks, where it achieves 0.585 mAP on the FSD50K dataset and\n0.335 mAP on the AudioSet-balanced dataset, achieving comparable results to\nTransformer based models with significantly lower number of learnable\nparameters.\n", "date": "2023-11-02", "categories": []}
{"id": "2311.01615", "title": "FLAP: Fast Language-Audio Pre-training", "abstract": "  We propose Fast Language-Audio Pre-training (FLAP), a self-supervised\napproach that efficiently and effectively learns aligned audio and language\nrepresentations through masking, contrastive learning and reconstruction. For\nefficiency, FLAP randomly drops audio spectrogram tokens, focusing solely on\nthe remaining ones for self-supervision. Through inter-modal contrastive\nlearning, FLAP learns to align paired audio and text representations in a\nshared latent space. Notably, FLAP leverages multiple augmented views via\nmasking for inter-modal contrast and learns to reconstruct the masked portion\nof audio tokens. Moreover, FLAP leverages large language models (LLMs) to\naugment the text inputs, contributing to improved performance. These approaches\nlead to more robust and informative audio-text representations, enabling FLAP\nto achieve state-of-the-art (SoTA) performance on audio-text retrieval tasks on\nAudioCaps (achieving 53.0% R@1) and Clotho (achieving 25.5% R@1).\n", "date": "2023-11-02", "categories": []}
{"id": "2305.07828", "title": "Description and Discussion on DCASE 2023 Challenge Task 2: First-Shot\n  Unsupervised Anomalous Sound Detection for Machine Condition Monitoring", "abstract": "  We present the task description of the Detection and Classification of\nAcoustic Scenes and Events (DCASE) 2023 Challenge Task 2: ``First-shot\nunsupervised anomalous sound detection (ASD) for machine condition\nmonitoring''. The main goal is to enable rapid deployment of ASD systems for\nnew kinds of machines without the need for hyperparameter tuning. In the past\nASD tasks, developed methods tuned hyperparameters for each machine type, as\nthe development and evaluation datasets had the same machine types. However,\ncollecting normal and anomalous data as the development dataset can be\ninfeasible in practice. In 2023 Task 2, we focus on solving the first-shot\nproblem, which is the challenge of training a model on a completely novel\nmachine type. Specifically, (i) each machine type has only one section (a\nsubset of machine type) and (ii) machine types in the development and\nevaluation datasets are completely different. Analysis of 86 submissions from\n23 teams revealed that the keys to outperform baselines were: 1) sampling\ntechniques for dealing with class imbalances across different domains and\nattributes, 2) generation of synthetic samples for robust detection, and 3) use\nof multiple large pre-trained models to extract meaningful embeddings for the\nanomaly detector.\n", "date": "2023-05-13", "categories": []}
{"id": "2311.00867", "title": "Automatic Disfluency Detection from Untranscribed Speech", "abstract": "  Speech disfluencies, such as filled pauses or repetitions, are disruptions in\nthe typical flow of speech. Stuttering is a speech disorder characterized by a\nhigh rate of disfluencies, but all individuals speak with some disfluencies and\nthe rates of disfluencies may by increased by factors such as cognitive load.\nClinically, automatic disfluency detection may help in treatment planning for\nindividuals who stutter. Outside of the clinic, automatic disfluency detection\nmay serve as a pre-processing step to improve natural language understanding in\ndownstream applications. With this wide range of applications in mind, we\ninvestigate language, acoustic, and multimodal methods for frame-level\nautomatic disfluency detection and categorization. Each of these methods relies\non audio as an input. First, we evaluate several automatic speech recognition\n(ASR) systems in terms of their ability to transcribe disfluencies, measured\nusing disfluency error rates. We then use these ASR transcripts as input to a\nlanguage-based disfluency detection model. We find that disfluency detection\nperformance is largely limited by the quality of transcripts and alignments. We\nfind that an acoustic-based approach that does not require transcription as an\nintermediate step outperforms the ASR language approach. Finally, we present\nmultimodal architectures which we find improve disfluency detection performance\nover the unimodal approaches. Ultimately, this work introduces novel approaches\nfor automatic frame-level disfluency and categorization. In the long term, this\nwill help researchers incorporate automatic disfluency detection into a range\nof applications.\n", "date": "2023-11-01", "categories": []}
{"id": "2311.00895", "title": "In-Context Prompt Editing For Conditional Audio Generation", "abstract": "  Distributional shift is a central challenge in the deployment of machine\nlearning models as they can be ill-equipped for real-world data. This is\nparticularly evident in text-to-audio generation where the encoded\nrepresentations are easily undermined by unseen prompts, which leads to the\ndegradation of generated audio -- the limited set of the text-audio pairs\nremains inadequate for conditional audio generation in the wild as user prompts\nare under-specified. In particular, we observe a consistent audio quality\ndegradation in generated audio samples with user prompts, as opposed to\ntraining set prompts. To this end, we present a retrieval-based in-context\nprompt editing framework that leverages the training captions as demonstrative\nexemplars to revisit the user prompts. We show that the framework enhanced the\naudio quality across the set of collected user prompts, which were edited with\nreference to the training captions as exemplars.\n", "date": "2023-11-01", "categories": []}
{"id": "2311.00897", "title": "On The Open Prompt Challenge In Conditional Audio Generation", "abstract": "  Text-to-audio generation (TTA) produces audio from a text description,\nlearning from pairs of audio samples and hand-annotated text. However,\ncommercializing audio generation is challenging as user-input prompts are often\nunder-specified when compared to text descriptions used to train TTA models. In\nthis work, we treat TTA models as a ``blackbox'' and address the user prompt\nchallenge with two key insights: (1) User prompts are generally\nunder-specified, leading to a large alignment gap between user prompts and\ntraining prompts. (2) There is a distribution of audio descriptions for which\nTTA models are better at generating higher quality audio, which we refer to as\n``audionese''. To this end, we rewrite prompts with instruction-tuned models\nand propose utilizing text-audio alignment as feedback signals via margin\nranking learning for audio improvements. On both objective and subjective human\nevaluations, we observed marked improvements in both text-audio alignment and\nmusic audio quality.\n", "date": "2023-11-01", "categories": []}
{"id": "2306.10608", "title": "STHG: Spatial-Temporal Heterogeneous Graph Learning for Advanced\n  Audio-Visual Diarization", "abstract": "  This report introduces our novel method named STHG for the Audio-Visual\nDiarization task of the Ego4D Challenge 2023. Our key innovation is that we\nmodel all the speakers in a video using a single, unified heterogeneous graph\nlearning framework. Unlike previous approaches that require a separate\ncomponent solely for the camera wearer, STHG can jointly detect the speech\nactivities of all people including the camera wearer. Our final method obtains\n61.1% DER on the test set of Ego4D, which significantly outperforms all the\nbaselines as well as last year's winner. Our submission achieved 1st place in\nthe Ego4D Challenge 2023. We additionally demonstrate that applying the\noff-the-shelf speech recognition system to the diarized speech segments by STHG\nproduces a competitive performance on the Speech Transcription task of this\nchallenge.\n", "date": "2023-06-18", "categories": []}
{"id": "2311.00320", "title": "Semantic Hearing: Programming Acoustic Scenes with Binaural Hearables", "abstract": "  Imagine being able to listen to the birds chirping in a park without hearing\nthe chatter from other hikers, or being able to block out traffic noise on a\nbusy street while still being able to hear emergency sirens and car honks. We\nintroduce semantic hearing, a novel capability for hearable devices that\nenables them to, in real-time, focus on, or ignore, specific sounds from\nreal-world environments, while also preserving the spatial cues. To achieve\nthis, we make two technical contributions: 1) we present the first neural\nnetwork that can achieve binaural target sound extraction in the presence of\ninterfering sounds and background noise, and 2) we design a training\nmethodology that allows our system to generalize to real-world use. Results\nshow that our system can operate with 20 sound classes and that our\ntransformer-based network has a runtime of 6.56 ms on a connected smartphone.\nIn-the-wild evaluation with participants in previously unseen indoor and\noutdoor scenarios shows that our proof-of-concept system can extract the target\nsounds and generalize to preserve the spatial cues in its binaural output.\nProject page with code: https://semantichearing.cs.washington.edu\n", "date": "2023-11-01", "categories": []}
{"id": "2310.20446", "title": "LAVSS: Location-Guided Audio-Visual Spatial Audio Separation", "abstract": "  Existing machine learning research has achieved promising results in monaural\naudio-visual separation (MAVS). However, most MAVS methods purely consider what\nthe sound source is, not where it is located. This can be a problem in VR/AR\nscenarios, where listeners need to be able to distinguish between similar audio\nsources located in different directions. To address this limitation, we have\ngeneralized MAVS to spatial audio separation and proposed LAVSS: a\nlocation-guided audio-visual spatial audio separator. LAVSS is inspired by the\ncorrelation between spatial audio and visual location. We introduce the phase\ndifference carried by binaural audio as spatial cues, and we utilize positional\nrepresentations of sounding objects as additional modality guidance. We also\nleverage multi-level cross-modal attention to perform visual-positional\ncollaboration with audio features. In addition, we adopt a pre-trained monaural\nseparator to transfer knowledge from rich mono sounds to boost spatial audio\nseparation. This exploits the correlation between monaural and binaural\nchannels. Experiments on the FAIR-Play dataset demonstrate the superiority of\nthe proposed LAVSS over existing benchmarks of audio-visual separation. Our\nproject page: https://yyx666660.github.io/LAVSS/.\n", "date": "2023-10-31", "categories": []}
{"id": "2310.13388", "title": "Music Augmentation and Denoising For Peak-Based Audio Fingerprinting", "abstract": "  Audio fingerprinting is a well-established solution for song identification\nfrom short recording excerpts. Popular methods rely on the extraction of sparse\nrepresentations, generally spectral peaks, and have proven to be accurate,\nfast, and scalable to large collections. However, real-world applications of\naudio identification often happen in noisy environments, which can cause these\nsystems to fail. In this work, we tackle this problem by introducing and\nreleasing a new audio augmentation pipeline that adds noise to music snippets\nin a realistic way, by stochastically mimicking real-world scenarios. We then\npropose and release a deep learning model that removes noisy components from\nspectrograms in order to improve peak-based fingerprinting systems' accuracy.\nWe show that the addition of our model improves the identification performance\nof commonly used audio fingerprinting systems, even under noisy conditions.\n", "date": "2023-10-20", "categories": []}
{"id": "2310.19581", "title": "Seeing Through the Conversation: Audio-Visual Speech Separation based on\n  Diffusion Model", "abstract": "  The objective of this work is to extract target speaker's voice from a\nmixture of voices using visual cues. Existing works on audio-visual speech\nseparation have demonstrated their performance with promising intelligibility,\nbut maintaining naturalness remains a challenge. To address this issue, we\npropose AVDiffuSS, an audio-visual speech separation model based on a diffusion\nmechanism known for its capability in generating natural samples. For an\neffective fusion of the two modalities for diffusion, we also propose a\ncross-attention-based feature fusion mechanism. This mechanism is specifically\ntailored for the speech domain to integrate the phonetic information from\naudio-visual correspondence in speech generation. In this way, the fusion\nprocess maintains the high temporal resolution of the features, without\nexcessive computational requirements. We demonstrate that the proposed\nframework achieves state-of-the-art results on two benchmarks, including\nVoxCeleb2 and LRS3, producing speech with notably better naturalness.\n", "date": "2023-10-30", "categories": []}
{"id": "2306.06546", "title": "High-Fidelity Audio Compression with Improved RVQGAN", "abstract": "  Language models have been successfully used to model natural signals, such as\nimages, speech, and music. A key component of these models is a high quality\nneural compression model that can compress high-dimensional natural signals\ninto lower dimensional discrete tokens. To that end, we introduce a\nhigh-fidelity universal neural audio compression algorithm that achieves ~90x\ncompression of 44.1 KHz audio into tokens at just 8kbps bandwidth. We achieve\nthis by combining advances in high-fidelity audio generation with better vector\nquantization techniques from the image domain, along with improved adversarial\nand reconstruction losses. We compress all domains (speech, environment, music,\netc.) with a single universal model, making it widely applicable to generative\nmodeling of all audio. We compare with competing audio compression algorithms,\nand find our method outperforms them significantly. We provide thorough\nablations for every design choice, as well as open-source code and trained\nmodel weights. We hope our work can lay the foundation for the next generation\nof high-fidelity audio modeling.\n", "date": "2023-06-11", "categories": []}
{"id": "2301.07087", "title": "MooseNet: A Trainable Metric for Synthesized Speech with a PLDA Module", "abstract": "  We present MooseNet, a trainable speech metric that predicts the listeners'\nMean Opinion Score (MOS). We propose a novel approach where the Probabilistic\nLinear Discriminative Analysis (PLDA) generative model is used on top of an\nembedding obtained from a self-supervised learning (SSL) neural network (NN)\nmodel. We show that PLDA works well with a non-finetuned SSL model when trained\nonly on 136 utterances (ca. one minute training time) and that PLDA\nconsistently improves various neural MOS prediction models, even\nstate-of-the-art models with task-specific fine-tuning. Our ablation study\nshows PLDA training superiority over SSL model fine-tuning in a low-resource\nscenario. We also improve SSL model fine-tuning using a convenient optimizer\nchoice and additional contrastive and multi-task training objectives. The\nfine-tuned MooseNet NN with the PLDA module achieves the best results,\nsurpassing the SSL baseline on the VoiceMOS Challenge data.\n", "date": "2023-01-17", "categories": []}
{"id": "2310.17448", "title": "Dialect Adaptation and Data Augmentation for Low-Resource ASR: TalTech\n  Systems for the MADASR 2023 Challenge", "abstract": "  This paper describes Tallinn University of Technology (TalTech) systems\ndeveloped for the ASRU MADASR 2023 Challenge. The challenge focuses on\nautomatic speech recognition of dialect-rich Indian languages with limited\ntraining audio and text data. TalTech participated in two tracks of the\nchallenge: Track 1 that allowed using only the provided training data and Track\n3 which allowed using additional audio data. In both tracks, we relied on\nwav2vec2.0 models. Our methodology diverges from the traditional procedure of\nfinetuning pretrained wav2vec2.0 models in two key points: firstly, through the\nimplementation of the aligned data augmentation technique to enhance the\nlinguistic diversity of the training data, and secondly, via the application of\ndeep prefix tuning for dialect adaptation of wav2vec2.0 models. In both tracks,\nour approach yielded significant improvements over the provided baselines,\nachieving the lowest word error rates across all participating teams.\n", "date": "2023-10-26", "categories": []}
{"id": "2301.08730", "title": "Novel-View Acoustic Synthesis", "abstract": "  We introduce the novel-view acoustic synthesis (NVAS) task: given the sight\nand sound observed at a source viewpoint, can we synthesize the sound of that\nscene from an unseen target viewpoint? We propose a neural rendering approach:\nVisually-Guided Acoustic Synthesis (ViGAS) network that learns to synthesize\nthe sound of an arbitrary point in space by analyzing the input audio-visual\ncues. To benchmark this task, we collect two first-of-their-kind large-scale\nmulti-view audio-visual datasets, one synthetic and one real. We show that our\nmodel successfully reasons about the spatial cues and synthesizes faithful\naudio on both datasets. To our knowledge, this work represents the very first\nformulation, dataset, and approach to solve the novel-view acoustic synthesis\ntask, which has exciting potential applications ranging from AR/VR to art and\ndesign. Unlocked by this work, we believe that the future of novel-view\nsynthesis is in multi-modal learning from videos.\n", "date": "2023-01-20", "categories": []}
{"id": "2310.16109", "title": "Complex Image Generation SwinTransformer Network for Audio Denoising", "abstract": "  Achieving high-performance audio denoising is still a challenging task in\nreal-world applications. Existing time-frequency methods often ignore the\nquality of generated frequency domain images. This paper converts the audio\ndenoising problem into an image generation task. We first develop a complex\nimage generation SwinTransformer network to capture more information from the\ncomplex Fourier domain. We then impose structure similarity and detailed loss\nfunctions to generate high-quality images and develop an SDR loss to minimize\nthe difference between denoised and clean audios. Extensive experiments on two\nbenchmark datasets demonstrate that our proposed model is better than\nstate-of-the-art methods.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.15247", "title": "SyncFusion: Multimodal Onset-synchronized Video-to-Audio Foley Synthesis", "abstract": "  Sound design involves creatively selecting, recording, and editing sound\neffects for various media like cinema, video games, and virtual/augmented\nreality. One of the most time-consuming steps when designing sound is\nsynchronizing audio with video. In some cases, environmental recordings from\nvideo shoots are available, which can aid in the process. However, in video\ngames and animations, no reference audio exists, requiring manual annotation of\nevent timings from the video. We propose a system to extract repetitive actions\nonsets from a video, which are then used - in conjunction with audio or textual\nembeddings - to condition a diffusion model trained to generate a new\nsynchronized sound effects audio track. In this way, we leave complete creative\ncontrol to the sound designer while removing the burden of synchronization with\nvideo. Furthermore, editing the onset track or changing the conditioning\nembedding requires much less effort than editing the audio track itself,\nsimplifying the sonification process. We provide sound examples, source code,\nand pretrained models to faciliate reproducibility\n", "date": "2023-10-23", "categories": []}
{"id": "2310.15261", "title": "Modality Dropout for Multimodal Device Directed Speech Detection using\n  Verbal and Non-Verbal Features", "abstract": "  Device-directed speech detection (DDSD) is the binary classification task of\ndistinguishing between queries directed at a voice assistant versus side\nconversation or background speech. State-of-the-art DDSD systems use verbal\ncues, e.g acoustic, text and/or automatic speech recognition system (ASR)\nfeatures, to classify speech as device-directed or otherwise, and often have to\ncontend with one or more of these modalities being unavailable when deployed in\nreal-world settings. In this paper, we investigate fusion schemes for DDSD\nsystems that can be made more robust to missing modalities. Concurrently, we\nstudy the use of non-verbal cues, specifically prosody features, in addition to\nverbal cues for DDSD. We present different approaches to combine scores and\nembeddings from prosody with the corresponding verbal cues, finding that\nprosody improves DDSD performance by upto 8.5% in terms of false acceptance\nrate (FA) at a given fixed operating point via non-linear intermediate fusion,\nwhile our use of modality dropout techniques improves the performance of these\nmodels by 7.4% in terms of FA when evaluated with missing modalities during\ninference time.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.15648", "title": "Dynamic Convolutional Neural Networks as Efficient Pre-trained Audio\n  Models", "abstract": "  The introduction of large-scale audio datasets, such as AudioSet, paved the\nway for Transformers to conquer the audio domain and replace CNNs as the\nstate-of-the-art neural network architecture for many tasks. Audio Spectrogram\nTransformers are excellent at exploiting large datasets, creating powerful\npre-trained models that surpass CNNs when fine-tuned on downstream tasks.\nHowever, current popular Audio Spectrogram Transformers are demanding in terms\nof computational complexity compared to CNNs. Recently, we have shown that, by\nemploying Transformer-to-CNN Knowledge Distillation, efficient CNNs can catch\nup with and even outperform Transformers on large datasets. In this work, we\nextend this line of research and increase the capacity of efficient CNNs by\nintroducing dynamic CNN blocks, constructed of dynamic non-linearities, dynamic\nconvolutions and attention mechanisms. We show that these dynamic CNNs\noutperform traditional efficient CNNs, in terms of the performance-complexity\ntrade-off and parameter efficiency, at the task of audio tagging on the\nlarge-scale AudioSet. Our experiments further indicate that the introduced\ndynamic CNNs achieve better performance on downstream tasks and scale up well,\nattaining Transformer performance and even outperforming them on AudioSet and\nseveral downstream tasks.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.11713", "title": "Separating Invisible Sounds Toward Universal Audiovisual Scene-Aware\n  Sound Separation", "abstract": "  The audio-visual sound separation field assumes visible sources in videos,\nbut this excludes invisible sounds beyond the camera's view. Current methods\nstruggle with such sounds lacking visible cues. This paper introduces a novel\n\"Audio-Visual Scene-Aware Separation\" (AVSA-Sep) framework. It includes a\nsemantic parser for visible and invisible sounds and a separator for\nscene-informed separation. AVSA-Sep successfully separates both sound types,\nwith joint training and cross-modal alignment enhancing effectiveness.\n", "date": "2023-10-18", "categories": []}
{"id": "2310.09853", "title": "MERTech: Instrument Playing Technique Detection Using Self-Supervised\n  Pretrained Model With Multi-Task Finetuning", "abstract": "  Instrument playing techniques (IPTs) constitute a pivotal component of\nmusical expression. However, the development of automatic IPT detection methods\nsuffers from limited labeled data and inherent class imbalance issues. In this\npaper, we propose to apply a self-supervised learning model pre-trained on\nlarge-scale unlabeled music data and finetune it on IPT detection tasks. This\napproach addresses data scarcity and class imbalance challenges. Recognizing\nthe significance of pitch in capturing the nuances of IPTs and the importance\nof onset in locating IPT events, we investigate multi-task finetuning with\npitch and onset detection as auxiliary tasks. Additionally, we apply a\npost-processing approach for event-level prediction, where an IPT activation\ninitiates an event only if the onset output confirms an onset in that frame.\nOur method outperforms prior approaches in both frame-level and event-level\nmetrics across multiple IPT benchmark datasets. Further experiments demonstrate\nthe efficacy of multi-task finetuning on each IPT class.\n", "date": "2023-10-15", "categories": []}
{"id": "2310.10159", "title": "Joint Music and Language Attention Models for Zero-shot Music Tagging", "abstract": "  Music tagging is a task to predict the tags of music recordings. However,\nprevious music tagging research primarily focuses on close-set music tagging\ntasks which can not be generalized to new tags. In this work, we propose a\nzero-shot music tagging system modeled by a joint music and language attention\n(JMLA) model to address the open-set music tagging problem. The JMLA model\nconsists of an audio encoder modeled by a pretrained masked autoencoder and a\ndecoder modeled by a Falcon7B. We introduce preceiver resampler to convert\narbitrary length audio into fixed length embeddings. We introduce dense\nattention connections between encoder and decoder layers to improve the\ninformation flow between the encoder and decoder layers. We collect a\nlarge-scale music and description dataset from the internet. We propose to use\nChatGPT to convert the raw descriptions into formalized and diverse\ndescriptions to train the JMLA models. Our proposed JMLA system achieves a\nzero-shot audio tagging accuracy of $ 64.82\\% $ on the GTZAN dataset,\noutperforming previous zero-shot systems and achieves comparable results to\nprevious systems on the FMA and the MagnaTagATune datasets.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.08303", "title": "Multimodal Variational Auto-encoder based Audio-Visual Segmentation", "abstract": "  We propose an Explicit Conditional Multimodal Variational Auto-Encoder\n(ECMVAE) for audio-visual segmentation (AVS), aiming to segment sound sources\nin the video sequence. Existing AVS methods focus on implicit feature fusion\nstrategies, where models are trained to fit the discrete samples in the\ndataset. With a limited and less diverse dataset, the resulting performance is\nusually unsatisfactory. In contrast, we address this problem from an effective\nrepresentation learning perspective, aiming to model the contribution of each\nmodality explicitly. Specifically, we find that audio contains critical\ncategory information of the sound producers, and visual data provides candidate\nsound producer(s). Their shared information corresponds to the target sound\nproducer(s) shown in the visual data. In this case, cross-modal shared\nrepresentation learning is especially important for AVS. To achieve this, our\nECMVAE factorizes the representations of each modality with a modality-shared\nrepresentation and a modality-specific representation. An orthogonality\nconstraint is applied between the shared and specific representations to\nmaintain the exclusive attribute of the factorized latent code. Further, a\nmutual information maximization regularizer is introduced to achieve extensive\nexploration of each modality. Quantitative and qualitative evaluations on the\nAVSBench demonstrate the effectiveness of our approach, leading to a new\nstate-of-the-art for AVS, with a 3.84 mIOU performance leap on the challenging\nMS3 subset for multiple sound source segmentation.\n", "date": "2023-10-12", "categories": []}
{"id": "2310.06364", "title": "Noisy-ArcMix: Additive Noisy Angular Margin Loss Combined With Mixup\n  Anomalous Sound Detection", "abstract": "  Unsupervised anomalous sound detection (ASD) aims to identify anomalous\nsounds by learning the features of normal operational sounds and sensing their\ndeviations. Recent approaches have focused on the self-supervised task\nutilizing the classification of normal data, and advanced models have shown\nthat securing representation space for anomalous data is important through\nrepresentation learning yielding compact intra-class and well-separated\nintra-class distributions. However, we show that conventional approaches often\nfail to ensure sufficient intra-class compactness and exhibit angular disparity\nbetween samples and their corresponding centers. In this paper, we propose a\ntraining technique aimed at ensuring intra-class compactness and increasing the\nangle gap between normal and abnormal samples. Furthermore, we present an\narchitecture that extracts features for important temporal regions, enabling\nthe model to learn which time frames should be emphasized or suppressed.\nExperimental results demonstrate that the proposed method achieves the best\nperformance giving 0.90%, 0.83%, and 2.16% improvement in terms of AUC, pAUC,\nand mAUC, respectively, compared to the state-of-the-art method on DCASE 2020\nChallenge Task2 dataset.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.06546", "title": "AutoCycle-VC: Towards Bottleneck-Independent Zero-Shot Cross-Lingual\n  Voice Conversion", "abstract": "  This paper proposes a simple and robust zero-shot voice conversion system\nwith a cycle structure and mel-spectrogram pre-processing. Previous works\nsuffer from information loss and poor synthesis quality due to their reliance\non a carefully designed bottleneck structure. Moreover, models relying solely\non self-reconstruction loss struggled with reproducing different speakers'\nvoices. To address these issues, we suggested a cycle-consistency loss that\nconsiders conversion back and forth between target and source speakers.\nAdditionally, stacked random-shuffled mel-spectrograms and a label smoothing\nmethod are utilized during speaker encoder training to extract a\ntime-independent global speaker representation from speech, which is the key to\na zero-shot conversion. Our model outperforms existing state-of-the-art results\nin both subjective and objective evaluations. Furthermore, it facilitates\ncross-lingual voice conversions and enhances the quality of synthesized speech.\n", "date": "2023-10-10", "categories": []}
{"id": "2211.12195", "title": "Ontology-aware Learning and Evaluation for Audio Tagging", "abstract": "  This study defines a new evaluation metric for audio tagging tasks to\novercome the limitation of the conventional mean average precision (mAP)\nmetric, which treats different kinds of sound as independent classes without\nconsidering their relations. Also, due to the ambiguities in sound labeling,\nthe labels in the training and evaluation set are not guaranteed to be accurate\nand exhaustive, which poses challenges for robust evaluation with mAP. The\nproposed metric, ontology-aware mean average precision (OmAP) addresses the\nweaknesses of mAP by utilizing the AudioSet ontology information during the\nevaluation. Specifically, we reweight the false positive events in the model\nprediction based on the ontology graph distance to the target classes. The OmAP\nmeasure also provides more insights into model performance by evaluations with\ndifferent coarse-grained levels in the ontology graph. We conduct human\nevaluations and demonstrate that OmAP is more consistent with human perception\nthan mAP. To further verify the importance of utilizing the ontology\ninformation, we also propose a novel loss function (OBCE) that reweights binary\ncross entropy (BCE) loss based on the ontology distance. Our experiment shows\nthat OBCE can improve both mAP and OmAP metrics on the AudioSet tagging task.\n", "date": "2022-11-22", "categories": []}
{"id": "2305.17706", "title": "Spot keywords from very noisy and mixed speech", "abstract": "  Most existing keyword spotting research focuses on conditions with slight or\nmoderate noise. In this paper, we try to tackle a more challenging task:\ndetecting keywords buried under strong interfering speech (10 times higher than\nthe keyword in amplitude), and even worse, mixed with other keywords. We\npropose a novel Mix Training (MT) strategy that encourages the model to\ndiscover low-energy keywords from noisy and mixed speech. Experiments were\nconducted with a vanilla CNN and two EfficientNet (B0/B2) architectures. The\nresults evaluated with the Google Speech Command dataset demonstrated that the\nproposed mix training approach is highly effective and outperforms standard\ndata augmentation and mixup training.\n", "date": "2023-05-28", "categories": []}
{"id": "2310.04681", "title": "VoiceExtender: Short-utterance Text-independent Speaker Verification\n  with Guided Diffusion Model", "abstract": "  Speaker verification (SV) performance deteriorates as utterances become\nshorter. To this end, we propose a new architecture called VoiceExtender which\nprovides a promising solution for improving SV performance when handling\nshort-duration speech signals. We use two guided diffusion models, the built-in\nand the external speaker embedding (SE) guided diffusion model, both of which\nutilize a diffusion model-based sample generator that leverages SE guidance to\naugment the speech features based on a short utterance. Extensive experimental\nresults on the VoxCeleb1 dataset show that our method outperforms the baseline,\nwith relative improvements in equal error rate (EER) of 46.1%, 35.7%, 10.4%,\nand 5.7% for the short utterance conditions of 0.5, 1.0, 1.5, and 2.0 seconds,\nrespectively.\n", "date": "2023-10-07", "categories": []}
{"id": "2310.05203", "title": "A Comparative Study of Voice Conversion Models with Large-Scale Speech\n  and Singing Data: The T13 Systems for the Singing Voice Conversion Challenge\n  2023", "abstract": "  This paper presents our systems (denoted as T13) for the singing voice\nconversion challenge (SVCC) 2023. For both in-domain and cross-domain English\nsinging voice conversion (SVC) tasks (Task 1 and Task 2), we adopt a\nrecognition-synthesis approach with self-supervised learning-based\nrepresentation. To achieve data-efficient SVC with a limited amount of target\nsinger/speaker's data (150 to 160 utterances for SVCC 2023), we first train a\ndiffusion-based any-to-any voice conversion model using publicly available\nlarge-scale 750 hours of speech and singing data. Then, we finetune the model\nfor each target singer/speaker of Task 1 and Task 2. Large-scale listening\ntests conducted by SVCC 2023 show that our T13 system achieves competitive\nnaturalness and speaker similarity for the harder cross-domain SVC (Task 2),\nwhich implies the generalization ability of our proposed method. Our objective\nevaluation results show that using large datasets is particularly beneficial\nfor cross-domain SVC.\n", "date": "2023-10-08", "categories": []}
{"id": "2310.05821", "title": "Pre-trained Spatial Priors on Multichannel NMF for Music Source\n  Separation", "abstract": "  This paper presents a novel approach to sound source separation that\nleverages spatial information obtained during the recording setup. Our method\ntrains a spatial mixing filter using solo passages to capture information about\nthe room impulse response and transducer response at each sensor location. This\npre-trained filter is then integrated into a multichannel non-negative matrix\nfactorization (MNMF) scheme to better capture the variances of different sound\nsources. The recording setup used in our experiments is the typical setup for\norchestra recordings, with a main microphone and a close \"cardioid\" or\n\"supercardioid\" microphone for each section of the orchestra. This makes the\nproposed method applicable to many existing recordings. Experiments on\npolyphonic ensembles demonstrate the effectiveness of the proposed framework in\nseparating individual sound sources, improving performance compared to\nconventional MNMF methods.\n", "date": "2023-10-09", "categories": []}
{"id": "2310.04369", "title": "MBTFNet: Multi-Band Temporal-Frequency Neural Network For Singing Voice\n  Enhancement", "abstract": "  A typical neural speech enhancement (SE) approach mainly handles speech and\nnoise mixtures, which is not optimal for singing voice enhancement scenarios.\nMusic source separation (MSS) models treat vocals and various accompaniment\ncomponents equally, which may reduce performance compared to the model that\nonly considers vocal enhancement. In this paper, we propose a novel multi-band\ntemporal-frequency neural network (MBTFNet) for singing voice enhancement,\nwhich particularly removes background music, noise and even backing vocals from\nsinging recordings. MBTFNet combines inter and intra-band modeling for better\nprocessing of full-band signals. Dual-path modeling are introduced to expand\nthe receptive field of the model. We propose an implicit personalized\nenhancement (IPE) stage based on signal-to-noise ratio (SNR) estimation, which\nfurther improves the performance of MBTFNet. Experiments show that our proposed\nmodel significantly outperforms several state-of-the-art SE and MSS models.\n", "date": "2023-10-06", "categories": []}
{"id": "2310.03349", "title": "An Integrated Algorithm for Robust and Imperceptible Audio Adversarial\n  Examples", "abstract": "  Audio adversarial examples are audio files that have been manipulated to fool\nan automatic speech recognition (ASR) system, while still sounding benign to a\nhuman listener. Most methods to generate such samples are based on a two-step\nalgorithm: first, a viable adversarial audio file is produced, then, this is\nfine-tuned with respect to perceptibility and robustness. In this work, we\npresent an integrated algorithm that uses psychoacoustic models and room\nimpulse responses (RIR) in the generation step. The RIRs are dynamically\ncreated by a neural network during the generation process to simulate a\nphysical environment to harden our examples against transformations experienced\nin over-the-air attacks. We compare the different approaches in three\nexperiments: in a simulated environment and in a realistic over-the-air\nscenario to evaluate the robustness, and in a human study to evaluate the\nperceptibility. Our algorithms considering psychoacoustics only or in addition\nto the robustness show an improvement in the signal-to-noise ratio (SNR) as\nwell as in the human perception study, at the cost of an increased word error\nrate (WER).\n", "date": "2023-10-05", "categories": []}
{"id": "2309.05287", "title": "Addressing Feature Imbalance in Sound Source Separation", "abstract": "  Neural networks often suffer from a feature preference problem, where they\ntend to overly rely on specific features to solve a task while disregarding\nother features, even if those neglected features are essential for the task.\nFeature preference problems have primarily been investigated in classification\ntask. However, we observe that feature preference occurs in high-dimensional\nregression task, specifically, source separation. To mitigate feature\npreference in source separation, we propose FEAture BAlancing by Suppressing\nEasy feature (FEABASE). This approach enables efficient data utilization by\nlearning hidden information about the neglected feature. We evaluate our method\nin a multi-channel source separation task, where feature preference between\nspatial feature and timbre feature appears.\n", "date": "2023-09-11", "categories": []}
{"id": "2306.00561", "title": "Masked Autoencoders with Multi-Window Local-Global Attention Are Better\n  Audio Learners", "abstract": "  In this work, we propose a Multi-Window Masked Autoencoder (MW-MAE) fitted\nwith a novel Multi-Window Multi-Head Attention (MW-MHA) module that facilitates\nthe modelling of local-global interactions in every decoder transformer block\nthrough attention heads of several distinct local and global windows. Empirical\nresults on ten downstream audio tasks show that MW-MAEs consistently outperform\nstandard MAEs in overall performance and learn better general-purpose audio\nrepresentations, along with demonstrating considerably better scaling\ncharacteristics. Investigating attention distances and entropies reveals that\nMW-MAE encoders learn heads with broader local and global attention. Analyzing\nattention head feature representations through Projection Weighted Canonical\nCorrelation Analysis (PWCCA) shows that attention heads with the same window\nsizes across the decoder layers of the MW-MAE learn correlated feature\nrepresentations which enables each block to independently capture local and\nglobal information, leading to a decoupled decoder feature hierarchy. Code for\nfeature extraction and downstream experiments along with pre-trained models\nwill be released publically.\n", "date": "2023-06-01", "categories": []}
{"id": "2310.00140", "title": "GASS: Generalizing Audio Source Separation with Large-scale Data", "abstract": "  Universal source separation targets at separating the audio sources of an\narbitrary mix, removing the constraint to operate on a specific domain like\nspeech or music. Yet, the potential of universal source separation is limited\nbecause most existing works focus on mixes with predominantly sound events, and\nsmall training datasets also limit its potential for supervised learning. Here,\nwe study a single general audio source separation (GASS) model trained to\nseparate speech, music, and sound events in a supervised fashion with a\nlarge-scale dataset. We assess GASS models on a diverse set of tasks. Our\nstrong in-distribution results show the feasibility of GASS models, and the\ncompetitive out-of-distribution performance in sound event and speech\nseparation shows its generalization abilities. Yet, it is challenging for GASS\nmodels to generalize for separating out-of-distribution cinematic and music\ncontent. We also fine-tune GASS models on each dataset and consistently\noutperform the ones without pre-training. All fine-tuned models (except the\nmusic separation one) obtain state-of-the-art results in their respective\nbenchmarks.\n", "date": "2023-09-29", "categories": []}
{"id": "2309.17125", "title": "Style Transfer for Non-differentiable Audio Effects", "abstract": "  Digital audio effects are widely used by audio engineers to alter the\nacoustic and temporal qualities of audio data. However, these effects can have\na large number of parameters which can make them difficult to learn for\nbeginners and hamper creativity for professionals. Recently, there have been a\nnumber of efforts to employ progress in deep learning to acquire the low-level\nparameter configurations of audio effects by minimising an objective function\nbetween an input and reference track, commonly referred to as style transfer.\nHowever, current approaches use inflexible black-box techniques or require that\nthe effects under consideration are implemented in an auto-differentiation\nframework. In this work, we propose a deep learning approach to audio\nproduction style matching which can be used with effects implemented in some of\nthe most widely used frameworks, requiring only that the parameters under\nconsideration have a continuous domain. Further, our method includes style\nmatching for various classes of effects, many of which are difficult or\nimpossible to be approximated closely using differentiable functions. We show\nthat our audio embedding approach creates logical encodings of timbral\ninformation, which can be used for a number of downstream tasks. Further, we\nperform a listening test which demonstrates that our approach is able to\nconvincingly style match a multi-band compressor effect.\n", "date": "2023-09-29", "categories": []}
{"id": "2309.15938", "title": "Exploring Self-Supervised Contrastive Learning of Spatial Sound Event\n  Representation", "abstract": "  In this study, we present a simple multi-channel framework for contrastive\nlearning (MC-SimCLR) to encode 'what' and 'where' of spatial audios. MC-SimCLR\nlearns joint spectral and spatial representations from unlabeled spatial\naudios, thereby enhancing both event classification and sound localization in\ndownstream tasks. At its core, we propose a multi-level data augmentation\npipeline that augments different levels of audio features, including waveforms,\nMel spectrograms, and generalized cross-correlation (GCC) features. In\naddition, we introduce simple yet effective channel-wise augmentation methods\nto randomly swap the order of the microphones and mask Mel and GCC channels. By\nusing these augmentations, we find that linear layers on top of the learned\nrepresentation significantly outperform supervised models in terms of both\nevent classification accuracy and localization error. We also perform a\ncomprehensive analysis of the effect of each augmentation method and a\ncomparison of the fine-tuning performance using different amounts of labeled\ndata.\n", "date": "2023-09-27", "categories": []}
{"id": "2309.15977", "title": "Neural Acoustic Context Field: Rendering Realistic Room Impulse Response\n  With Neural Fields", "abstract": "  Room impulse response (RIR), which measures the sound propagation within an\nenvironment, is critical for synthesizing high-fidelity audio for a given\nenvironment. Some prior work has proposed representing RIR as a neural field\nfunction of the sound emitter and receiver positions. However, these methods do\nnot sufficiently consider the acoustic properties of an audio scene, leading to\nunsatisfactory performance. This letter proposes a novel Neural Acoustic\nContext Field approach, called NACF, to parameterize an audio scene by\nleveraging multiple acoustic contexts, such as geometry, material property, and\nspatial information. Driven by the unique properties of RIR, i.e., temporal\nun-smoothness and monotonic energy attenuation, we design a temporal\ncorrelation module and multi-scale energy decay criterion. Experimental results\nshow that NACF outperforms existing field-based methods by a notable margin.\nPlease visit our project page for more qualitative results.\n", "date": "2023-09-27", "categories": []}
{"id": "2211.01299", "title": "Late Audio-Visual Fusion for In-The-Wild Speaker Diarization", "abstract": "  Speaker diarization is well studied for constrained audios but little\nexplored for challenging in-the-wild videos, which have more speakers, shorter\nutterances, and inconsistent on-screen speakers. We address this gap by\nproposing an audio-visual diarization model which combines audio-only and\nvisual-centric sub-systems via late fusion. For audio, we show that an\nattractor-based end-to-end system (EEND-EDA) performs remarkably well when\ntrained with our proposed recipe of a simulated proxy dataset, and propose an\nimproved version, EEND-EDA++, that uses attention in decoding and a speaker\nrecognition loss during training to better handle the larger number of\nspeakers. The visual-centric sub-system leverages facial attributes and\nlip-audio synchrony for identity and speech activity estimation of on-screen\nspeakers. Both sub-systems surpass the state of the art (SOTA) by a large\nmargin, with the fused audio-visual system achieving a new SOTA on the AVA-AVD\nbenchmark.\n", "date": "2022-11-02", "categories": []}
{"id": "2307.14609", "title": "Complete and separate: Conditional separation with missing target source\n  attribute completion", "abstract": "  Recent approaches in source separation leverage semantic information about\ntheir input mixtures and constituent sources that when used in conditional\nseparation models can achieve impressive performance. Most approaches along\nthese lines have focused on simple descriptions, which are not always useful\nfor varying types of input mixtures. In this work, we present an approach in\nwhich a model, given an input mixture and partial semantic information about a\ntarget source, is trained to extract additional semantic data. We then leverage\nthis pre-trained model to improve the separation performance of an uncoupled\nmulti-conditional separation network. Our experiments demonstrate that the\nseparation performance of this multi-conditional model is significantly\nimproved, approaching the performance of an oracle model with complete semantic\ninformation. Furthermore, our approach achieves performance levels that are\ncomparable to those of the best performing specialized single conditional\nmodels, thus providing an easier to use alternative.\n", "date": "2023-07-27", "categories": []}
{"id": "2309.11845", "title": "TMac: Temporal Multi-Modal Graph Learning for Acoustic Event\n  Classification", "abstract": "  Audiovisual data is everywhere in this digital age, which raises higher\nrequirements for the deep learning models developed on them. To well handle the\ninformation of the multi-modal data is the key to a better audiovisual modal.\nWe observe that these audiovisual data naturally have temporal attributes, such\nas the time information for each frame in the video. More concretely, such data\nis inherently multi-modal according to both audio and visual cues, which\nproceed in a strict chronological order. It indicates that temporal information\nis important in multi-modal acoustic event modeling for both intra- and\ninter-modal. However, existing methods deal with each modal feature\nindependently and simply fuse them together, which neglects the mining of\ntemporal relation and thus leads to sub-optimal performance. With this\nmotivation, we propose a Temporal Multi-modal graph learning method for\nAcoustic event Classification, called TMac, by modeling such temporal\ninformation via graph learning techniques. In particular, we construct a\ntemporal graph for each acoustic event, dividing its audio data and video data\ninto multiple segments. Each segment can be considered as a node, and the\ntemporal relationships between nodes can be considered as timestamps on their\nedges. In this case, we can smoothly capture the dynamic information in\nintra-modal and inter-modal. Several experiments are conducted to demonstrate\nTMac outperforms other SOTA models in performance. Our code is available at\nhttps://github.com/MGitHubL/TMac.\n", "date": "2023-09-21", "categories": []}
{"id": "2309.15024", "title": "Synthia's Melody: A Benchmark Framework for Unsupervised Domain\n  Adaptation in Audio", "abstract": "  Despite significant advancements in deep learning for vision and natural\nlanguage, unsupervised domain adaptation in audio remains relatively\nunexplored. We, in part, attribute this to the lack of an appropriate benchmark\ndataset. To address this gap, we present Synthia's melody, a novel audio data\ngeneration framework capable of simulating an infinite variety of 4-second\nmelodies with user-specified confounding structures characterised by musical\nkeys, timbre, and loudness. Unlike existing datasets collected under\nobservational settings, Synthia's melody is free of unobserved biases, ensuring\nthe reproducibility and comparability of experiments. To showcase its utility,\nwe generate two types of distribution shifts-domain shift and sample selection\nbias-and evaluate the performance of acoustic deep learning models under these\nshifts. Our evaluations reveal that Synthia's melody provides a robust testbed\nfor examining the susceptibility of these models to varying levels of\ndistribution shift.\n", "date": "2023-09-26", "categories": []}
{"id": "2303.16342", "title": "Language-Guided Audio-Visual Source Separation via Trimodal Consistency", "abstract": "  We propose a self-supervised approach for learning to perform audio source\nseparation in videos based on natural language queries, using only unlabeled\nvideo and audio pairs as training data. A key challenge in this task is\nlearning to associate the linguistic description of a sound-emitting object to\nits visual features and the corresponding components of the audio waveform, all\nwithout access to annotations during training. To overcome this challenge, we\nadapt off-the-shelf vision-language foundation models to provide pseudo-target\nsupervision via two novel loss functions and encourage a stronger alignment\nbetween the audio, visual and natural language modalities. During inference,\nour approach can separate sounds given text, video and audio input, or given\ntext and audio input alone. We demonstrate the effectiveness of our\nself-supervised approach on three audio-visual separation datasets, including\nMUSIC, SOLOS and AudioSet, where we outperform state-of-the-art strongly\nsupervised approaches despite not using object detectors or text labels during\ntraining.\n", "date": "2023-03-28", "categories": []}
{"id": "2309.13373", "title": "Asca: less audio data is more insightful", "abstract": "  Audio recognition in specialized areas such as birdsong and submarine\nacoustics faces challenges in large-scale pre-training due to the limitations\nin available samples imposed by sampling environments and specificity\nrequirements. While the Transformer model excels in audio recognition, its\ndependence on vast amounts of data becomes restrictive in resource-limited\nsettings. Addressing this, we introduce the Audio Spectrogram Convolution\nAttention (ASCA) based on CoAtNet, integrating a Transformer-convolution hybrid\narchitecture, novel network design, and attention techniques, further augmented\nwith data enhancement and regularization strategies. On the BirdCLEF2023 and\nAudioSet(Balanced), ASCA achieved accuracies of 81.2% and 35.1%, respectively,\nsignificantly outperforming competing methods. The unique structure of our\nmodel enriches output, enabling generalization across various audio detection\ntasks. Our code can be found at https://github.com/LeeCiang/ASCA.\n", "date": "2023-09-23", "categories": []}
{"id": "2309.12581", "title": "Sampling-Frequency-Independent Universal Sound Separation", "abstract": "  This paper proposes a universal sound separation (USS) method capable of\nhandling untrained sampling frequencies (SFs). The USS aims at separating\narbitrary sources of different types and can be the key technique to realize a\nsource separator that can be universally used as a preprocessor for any\ndownstream tasks. To realize a universal source separator, there are two\nessential properties: universalities with respect to source types and recording\nconditions. The former property has been studied in the USS literature, which\nhas greatly increased the number of source types that can be handled by a\nsingle neural network. However, the latter property (e.g., SF) has received\nless attention despite its necessity. Since the SF varies widely depending on\nthe downstream tasks, the universal source separator must handle a wide variety\nof SFs. In this paper, to encompass the two properties, we propose an\nSF-independent (SFI) extension of a computationally efficient USS network,\nSuDoRM-RF. The proposed network uses our previously proposed SFI convolutional\nlayers, which can handle various SFs by generating convolutional kernels in\naccordance with an input SF. Experiments show that signal resampling can\ndegrade the USS performance and the proposed method works more consistently\nthan signal-resampling-based methods for various SFs.\n", "date": "2023-09-22", "categories": []}
{"id": "2309.12242", "title": "Weakly-supervised Automated Audio Captioning via text only training", "abstract": "  In recent years, datasets of paired audio and captions have enabled\nremarkable success in automatically generating descriptions for audio clips,\nnamely Automated Audio Captioning (AAC). However, it is labor-intensive and\ntime-consuming to collect a sufficient number of paired audio and captions.\nMotivated by the recent advances in Contrastive Language-Audio Pretraining\n(CLAP), we propose a weakly-supervised approach to train an AAC model assuming\nonly text data and a pre-trained CLAP model, alleviating the need for paired\ntarget data. Our approach leverages the similarity between audio and text\nembeddings in CLAP. During training, we learn to reconstruct the text from the\nCLAP text embedding, and during inference, we decode using the audio\nembeddings. To mitigate the modality gap between the audio and text embeddings\nwe employ strategies to bridge the gap during training and inference stages. We\nevaluate our proposed method on Clotho and AudioCaps datasets demonstrating its\nability to achieve a relative performance of up to ~$83\\%$ compared to fully\nsupervised approaches trained with paired target data.\n", "date": "2023-09-21", "categories": []}
{"id": "2309.11140", "title": "Investigating Personalization Methods in Text to Music Generation", "abstract": "  In this work, we investigate the personalization of text-to-music diffusion\nmodels in a few-shot setting. Motivated by recent advances in the computer\nvision domain, we are the first to explore the combination of pre-trained\ntext-to-audio diffusers with two established personalization methods. We\nexperiment with the effect of audio-specific data augmentation on the overall\nsystem performance and assess different training strategies. For evaluation, we\nconstruct a novel dataset with prompts and music clips. We consider both\nembedding-based and music-specific metrics for quantitative evaluation, as well\nas a user study for qualitative evaluation. Our analysis shows that similarity\nmetrics are in accordance with user preferences and that current\npersonalization approaches tend to learn rhythmic music constructs more easily\nthan melody. The code, dataset, and example material of this study are open to\nthe research community.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.11218", "title": "Automatic Bat Call Classification using Transformer Networks", "abstract": "  Automatically identifying bat species from their echolocation calls is a\ndifficult but important task for monitoring bats and the ecosystem they live\nin. Major challenges in automatic bat call identification are high call\nvariability, similarities between species, interfering calls and lack of\nannotated data. Many currently available models suffer from relatively poor\nperformance on real-life data due to being trained on single call datasets and,\nmoreover, are often too slow for real-time classification. Here, we propose a\nTransformer architecture for multi-label classification with potential\napplications in real-time classification scenarios. We train our model on\nsynthetically generated multi-species recordings by merging multiple bats calls\ninto a single recording with multiple simultaneous calls. Our approach achieves\na single species accuracy of 88.92% (F1-score of 84.23%) and a multi species\nmacro F1-score of 74.40% on our test set. In comparison to three other tools on\nthe independent and publicly available dataset ChiroVox, our model achieves at\nleast 25.82% better accuracy for single species classification and at least\n6.9% better macro F1-score for multi species classification.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.08773", "title": "Enhance audio generation controllability through representation\n  similarity regularization", "abstract": "  This paper presents an innovative approach to enhance control over audio\ngeneration by emphasizing the alignment between audio and text representations\nduring model training. In the context of language model-based audio generation,\nthe model leverages input from both textual and audio token representations to\npredict subsequent audio tokens. However, the current configuration lacks\nexplicit regularization to ensure the alignment between the chosen text\nrepresentation and the language model's predictions. Our proposal involves the\nincorporation of audio and text representation regularization, particularly\nduring the classifier-free guidance (CFG) phase, where the text condition is\nexcluded from cross attention during language model training. The aim of this\nproposed representation regularization is to minimize discrepancies in audio\nand text similarity compared to other samples within the same training batch.\nExperimental results on both music and audio generation tasks demonstrate that\nour proposed methods lead to improvements in objective metrics for both audio\nand music generation, as well as an enhancement in the human perception for\naudio generation.\n", "date": "2023-09-15", "categories": []}
{"id": "2309.09470", "title": "Face-Driven Zero-Shot Voice Conversion with Memory-based Face-Voice\n  Alignment", "abstract": "  This paper presents a novel task, zero-shot voice conversion based on face\nimages (zero-shot FaceVC), which aims at converting the voice characteristics\nof an utterance from any source speaker to a newly coming target speaker,\nsolely relying on a single face image of the target speaker. To address this\ntask, we propose a face-voice memory-based zero-shot FaceVC method. This method\nleverages a memory-based face-voice alignment module, in which slots act as the\nbridge to align these two modalities, allowing for the capture of voice\ncharacteristics from face images. A mixed supervision strategy is also\nintroduced to mitigate the long-standing issue of the inconsistency between\ntraining and inference phases for voice conversion tasks. To obtain\nspeaker-independent content-related representations, we transfer the knowledge\nfrom a pretrained zero-shot voice conversion model to our zero-shot FaceVC\nmodel. Considering the differences between FaceVC and traditional voice\nconversion tasks, systematic subjective and objective metrics are designed to\nthoroughly evaluate the homogeneity, diversity and consistency of voice\ncharacteristics controlled by face images. Through extensive experiments, we\ndemonstrate the superiority of our proposed method on the zero-shot FaceVC\ntask. Samples are presented on our demo website.\n", "date": "2023-09-18", "categories": []}
{"id": "2309.08141", "title": "Audio Difference Learning for Audio Captioning", "abstract": "  This study introduces a novel training paradigm, audio difference learning,\nfor improving audio captioning. The fundamental concept of the proposed\nlearning method is to create a feature representation space that preserves the\nrelationship between audio, enabling the generation of captions that detail\nintricate audio information. This method employs a reference audio along with\nthe input audio, both of which are transformed into feature representations via\na shared encoder. Captions are then generated from these differential features\nto describe their differences. Furthermore, a unique technique is proposed that\ninvolves mixing the input audio with additional audio, and using the additional\naudio as a reference. This results in the difference between the mixed audio\nand the reference audio reverting back to the original input audio. This allows\nthe original input's caption to be used as the caption for their difference,\neliminating the need for additional annotations for the differences. In the\nexperiments using the Clotho and ESC50 datasets, the proposed method\ndemonstrated an improvement in the SPIDEr score by 7% compared to conventional\nmethods.\n", "date": "2023-09-15", "categories": []}
{"id": "2309.08146", "title": "Syn-Att: Synthetic Speech Attribution via Semi-Supervised Unknown\n  Multi-Class Ensemble of CNNs", "abstract": "  With the huge technological advances introduced by deep learning in audio &\nspeech processing, many novel synthetic speech techniques achieved incredible\nrealistic results. As these methods generate realistic fake human voices, they\ncan be used in malicious acts such as people imitation, fake news, spreading,\nspoofing, media manipulations, etc. Hence, the ability to detect synthetic or\nnatural speech has become an urgent necessity. Moreover, being able to tell\nwhich algorithm has been used to generate a synthetic speech track can be of\npreeminent importance to track down the culprit. In this paper, a novel\nstrategy is proposed to attribute a synthetic speech track to the generator\nthat is used to synthesize it. The proposed detector transforms the audio into\nlog-mel spectrogram, extracts features using CNN, and classifies it between\nfive known and unknown algorithms, utilizing semi-supervision and ensemble to\nimprove its robustness and generalizability significantly. The proposed\ndetector is validated on two evaluation datasets consisting of a total of\n18,000 weakly perturbed (Eval 1) & 10,000 strongly perturbed (Eval 2) synthetic\nspeeches. The proposed method outperforms other top teams in accuracy by 12-13%\non Eval 2 and 1-2% on Eval 1, in the IEEE SP Cup challenge at ICASSP 2022.\n", "date": "2023-09-15", "categories": []}
{"id": "2309.08208", "title": "HM-Conformer: A Conformer-based audio deepfake detection system with\n  hierarchical pooling and multi-level classification token aggregation methods", "abstract": "  Audio deepfake detection (ADD) is the task of detecting spoofing attacks\ngenerated by text-to-speech or voice conversion systems. Spoofing evidence,\nwhich helps to distinguish between spoofed and bona-fide utterances, might\nexist either locally or globally in the input features. To capture these, the\nConformer, which consists of Transformers and CNN, possesses a suitable\nstructure. However, since the Conformer was designed for sequence-to-sequence\ntasks, its direct application to ADD tasks may be sub-optimal. To tackle this\nlimitation, we propose HM-Conformer by adopting two components: (1)\nHierarchical pooling method progressively reducing the sequence length to\neliminate duplicated information (2) Multi-level classification token\naggregation method utilizing classification tokens to gather information from\ndifferent blocks. Owing to these components, HM-Conformer can efficiently\ndetect spoofing evidence by processing various sequence lengths and aggregating\nthem. In experimental results on the ASVspoof 2021 Deepfake dataset,\nHM-Conformer achieved a 15.71% EER, showing competitive performance compared to\nrecent systems.\n", "date": "2023-09-15", "categories": []}
{"id": "2309.07314", "title": "AudioSR: Versatile Audio Super-resolution at Scale", "abstract": "  Audio super-resolution is a fundamental task that predicts high-frequency\ncomponents for low-resolution audio, enhancing audio quality in digital\napplications. Previous methods have limitations such as the limited scope of\naudio types (e.g., music, speech) and specific bandwidth settings they can\nhandle (e.g., 4kHz to 8kHz). In this paper, we introduce a diffusion-based\ngenerative model, AudioSR, that is capable of performing robust audio\nsuper-resolution on versatile audio types, including sound effects, music, and\nspeech. Specifically, AudioSR can upsample any input audio signal within the\nbandwidth range of 2kHz to 16kHz to a high-resolution audio signal at 24kHz\nbandwidth with a sampling rate of 48kHz. Extensive objective evaluation on\nvarious audio super-resolution benchmarks demonstrates the strong result\nachieved by the proposed model. In addition, our subjective evaluation shows\nthat AudioSR can acts as a plug-and-play module to enhance the generation\nquality of a wide range of audio generative models, including AudioLDM,\nFastspeech2, and MusicGen. Our code and demo are available at\nhttps://audioldm.github.io/audiosr.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.06661", "title": "Sound field decomposition based on two-stage neural networks", "abstract": "  A method for sound field decomposition based on neural networks is proposed.\nThe method comprises two stages: a sound field separation stage and a\nsingle-source localization stage. In the first stage, the sound pressure at\nmicrophones synthesized by multiple sources is separated into one excited by\neach sound source. In the second stage, the source location is obtained as a\nregression from the sound pressure at microphones consisting of a single sound\nsource. The estimated location is not affected by discretization because the\nsecond stage is designed as a regression rather than a classification. Datasets\nare generated by simulation using Green's function, and the neural network is\ntrained for each frequency. Numerical experiments reveal that, compared with\nconventional methods, the proposed method can achieve higher\nsource-localization accuracy and higher sound-field-reconstruction accuracy.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.06960", "title": "PhantomSound: Black-Box, Query-Efficient Audio Adversarial Attack via\n  Split-Second Phoneme Injection", "abstract": "  In this paper, we propose PhantomSound, a query-efficient black-box attack\ntoward voice assistants. Existing black-box adversarial attacks on voice\nassistants either apply substitution models or leverage the intermediate model\noutput to estimate the gradients for crafting adversarial audio samples.\nHowever, these attack approaches require a significant amount of queries with a\nlengthy training stage. PhantomSound leverages the decision-based attack to\nproduce effective adversarial audios, and reduces the number of queries by\noptimizing the gradient estimation. In the experiments, we perform our attack\nagainst 4 different speech-to-text APIs under 3 real-world scenarios to\ndemonstrate the real-time attack impact. The results show that PhantomSound is\npractical and robust in attacking 5 popular commercial voice controllable\ndevices over the air, and is able to bypass 3 liveness detection mechanisms\nwith >95% success rate. The benchmark result shows that PhantomSound can\ngenerate adversarial examples and launch the attack in a few minutes. We\nsignificantly enhance the query efficiency and reduce the cost of a successful\nuntargeted and targeted adversarial attack by 93.1% and 65.5% compared with the\nstate-of-the-art black-box attacks, using merely ~300 queries (~5 minutes) and\n~1,500 queries (~25 minutes), respectively.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.05843", "title": "Optimizing Audio Augmentations for Contrastive Learning of\n  Health-Related Acoustic Signals", "abstract": "  Health-related acoustic signals, such as cough and breathing sounds, are\nrelevant for medical diagnosis and continuous health monitoring. Most existing\nmachine learning approaches for health acoustics are trained and evaluated on\nspecific tasks, limiting their generalizability across various healthcare\napplications. In this paper, we leverage a self-supervised learning framework,\nSimCLR with a Slowfast NFNet backbone, for contrastive learning of health\nacoustics. A crucial aspect of optimizing Slowfast NFNet for this application\nlies in identifying effective audio augmentations. We conduct an in-depth\nanalysis of various audio augmentation strategies and demonstrate that an\nappropriate augmentation strategy enhances the performance of the Slowfast\nNFNet audio encoder across a diverse set of health acoustic tasks. Our findings\nreveal that when augmentations are combined, they can produce synergistic\neffects that exceed the benefits seen when each is applied individually.\n", "date": "2023-09-11", "categories": []}
{"id": "2309.04762", "title": "AudRandAug: Random Image Augmentations for Audio Classification", "abstract": "  Data augmentation has proven to be effective in training neural networks.\nRecently, a method called RandAug was proposed, randomly selecting data\naugmentation techniques from a predefined search space. RandAug has\ndemonstrated significant performance improvements for image-related tasks while\nimposing minimal computational overhead. However, no prior research has\nexplored the application of RandAug specifically for audio data augmentation,\nwhich converts audio into an image-like pattern. To address this gap, we\nintroduce AudRandAug, an adaptation of RandAug for audio data. AudRandAug\nselects data augmentation policies from a dedicated audio search space. To\nevaluate the effectiveness of AudRandAug, we conducted experiments using\nvarious models and datasets. Our findings indicate that AudRandAug outperforms\nother existing data augmentation methods regarding accuracy performance.\n", "date": "2023-09-09", "categories": []}
{"id": "2309.03340", "title": "Parameter Efficient Audio Captioning With Faithful Guidance Using\n  Audio-text Shared Latent Representation", "abstract": "  There has been significant research on developing pretrained transformer\narchitectures for multimodal-to-text generation tasks. Albeit performance\nimprovements, such models are frequently overparameterized, hence suffer from\nhallucination and large memory footprint making them challenging to deploy on\nedge devices. In this paper, we address both these issues for the application\nof automated audio captioning. First, we propose a data augmentation technique\nfor generating hallucinated audio captions and show that similarity based on an\naudio-text shared latent space is suitable for detecting hallucination. Then,\nwe propose a parameter efficient inference time faithful decoding algorithm\nthat enables smaller audio captioning models with performance equivalent to\nlarger models trained with more data. During the beam decoding step, the\nsmaller model utilizes an audio-text shared latent representation to\nsemantically align the generated text with corresponding input audio. Faithful\nguidance is introduced into the beam probability by incorporating the cosine\nsimilarity between latent representation projections of greedy rolled out\nintermediate beams and audio clip. We show the efficacy of our algorithm on\nbenchmark datasets and evaluate the proposed scheme against baselines using\nconventional audio captioning and semantic similarity metrics while\nillustrating tradeoffs between performance and complexity.\n", "date": "2023-09-06", "categories": []}
{"id": "2309.03544", "title": "MVD:A Novel Methodology and Dataset for Acoustic Vehicle Type\n  Classification", "abstract": "  Rising urban populations have led to a surge in vehicle use and made traffic\nmonitoring and management indispensable. Acoustic traffic monitoring (ATM)\noffers a cost-effective and efficient alternative to more computationally\nexpensive methods of monitoring traffic such as those involving computer vision\ntechnologies. In this paper, we present MVD and MVDA: two open datasets for the\ndevelopment of acoustic traffic monitoring and vehicle-type classification\nalgorithms, which contain audio recordings of moving vehicles. The dataset\ncontain four classes- Trucks, Cars, Motorbikes, and a No-vehicle class.\nAdditionally, we propose a novel and efficient way to accurately classify these\nacoustic signals using cepstrum and spectrum based local and global audio\nfeatures, and a multi-input neural network. Experimental results show that our\nmethodology improves upon the established baselines of previous works and\nachieves an accuracy of 91.98% and 96.66% on MVD and MVDA Datasets,\nrespectively. Finally, the proposed model was deployed through an Android\napplication to make it accessible for testing and demonstrate its efficacy.\n", "date": "2023-09-07", "categories": []}
{"id": "2309.03884", "title": "Zero-Shot Audio Captioning via Audibility Guidance", "abstract": "  The task of audio captioning is similar in essence to tasks such as image and\nvideo captioning. However, it has received much less attention. We propose\nthree desiderata for captioning audio -- (i) fluency of the generated text,\n(ii) faithfulness of the generated text to the input audio, and the somewhat\nrelated (iii) audibility, which is the quality of being able to be perceived\nbased only on audio. Our method is a zero-shot method, i.e., we do not learn to\nperform captioning. Instead, captioning occurs as an inference process that\ninvolves three networks that correspond to the three desired qualities: (i) A\nLarge Language Model, in our case, for reasons of convenience, GPT-2, (ii) A\nmodel that provides a matching score between an audio file and a text, for\nwhich we use a multimodal matching network called ImageBind, and (iii) A text\nclassifier, trained using a dataset we collected automatically by instructing\nGPT-4 with prompts designed to direct the generation of both audible and\ninaudible sentences. We present our results on the AudioCap dataset,\ndemonstrating that audibility guidance significantly enhances performance\ncompared to the baseline, which lacks this objective.\n", "date": "2023-09-07", "categories": []}
{"id": "2309.02232", "title": "FSD: An Initial Chinese Dataset for Fake Song Detection", "abstract": "  Singing voice synthesis and singing voice conversion have significantly\nadvanced, revolutionizing musical experiences. However, the rise of \"Deepfake\nSongs\" generated by these technologies raises concerns about authenticity.\nUnlike Audio DeepFake Detection (ADD), the field of song deepfake detection\nlacks specialized datasets or methods for song authenticity verification. In\nthis paper, we initially construct a Chinese Fake Song Detection (FSD) dataset\nto investigate the field of song deepfake detection. The fake songs in the FSD\ndataset are generated by five state-of-the-art singing voice synthesis and\nsinging voice conversion methods. Our initial experiments on FSD revealed the\nineffectiveness of existing speech-trained ADD models for the task of song\ndeepFake detection. Thus, we employ the FSD dataset for the training of ADD\nmodels. We subsequently evaluate these models under two scenarios: one with the\noriginal songs and another with separated vocal tracks. Experiment results show\nthat song-trained ADD models exhibit a 38.58% reduction in average equal error\nrate compared to speech-trained ADD models on the FSD test set.\n", "date": "2023-09-05", "categories": []}
{"id": "2306.12714", "title": "Toward Leveraging Pre-Trained Self-Supervised Frontends for Automatic\n  Singing Voice Understanding Tasks: Three Case Studies", "abstract": "  Automatic singing voice understanding tasks, such as singer identification,\nsinging voice transcription, and singing technique classification, benefit from\ndata-driven approaches that utilize deep learning techniques. These approaches\nwork well even under the rich diversity of vocal and noisy samples owing to\ntheir representation ability. However, the limited availability of labeled data\nremains a significant obstacle to achieving satisfactory performance. In recent\nyears, self-supervised learning models (SSL models) have been trained using\nlarge amounts of unlabeled data in the field of speech processing and music\nclassification. By fine-tuning these models for the target tasks, comparable\nperformance to conventional supervised learning can be achieved with limited\ntraining data. Therefore, in this paper, we investigate the effectiveness of\nSSL models for various singing voice recognition tasks. We report the results\nof experiments comparing SSL models for three different tasks (i.e., singer\nidentification, singing voice transcription, and singing technique\nclassification) as initial exploration and aim to discuss these findings.\nExperimental results show that each SSL model achieves comparable performance\nand sometimes outperforms compared to state-of-the-art methods on each task. We\nalso conducted a layer-wise analysis to further understand the behavior of the\nSSL models.\n", "date": "2023-06-22", "categories": []}
{"id": "2309.00878", "title": "Pretraining Representations for Bioacoustic Few-shot Detection using\n  Supervised Contrastive Learning", "abstract": "  Deep learning has been widely used recently for sound event detection and\nclassification. Its success is linked to the availability of sufficiently large\ndatasets, possibly with corresponding annotations when supervised learning is\nconsidered. In bioacoustic applications, most tasks come with few labelled\ntraining data, because annotating long recordings is time consuming and costly.\nTherefore supervised learning is not the best suited approach to solve\nbioacoustic tasks. The bioacoustic community recasted the problem of sound\nevent detection within the framework of few-shot learning, i.e. training a\nsystem with only few labeled examples. The few-shot bioacoustic sound event\ndetection task in the DCASE challenge focuses on detecting events in long audio\nrecordings given only five annotated examples for each class of interest. In\nthis paper, we show that learning a rich feature extractor from scratch can be\nachieved by leveraging data augmentation using a supervised contrastive\nlearning framework. We highlight the ability of this framework to transfer well\nfor five-shot event detection on previously unseen classes in the training\ndata. We obtain an F-score of 63.46\\% on the validation set and 42.7\\% on the\ntest set, ranking second in the DCASE challenge. We provide an ablation study\nfor the critical choices of data augmentation techniques as well as for the\nlearning strategy applied on the training set.\n", "date": "2023-09-02", "categories": []}
{"id": "2308.16892", "title": "ReZero: Region-customizable Sound Extraction", "abstract": "  We introduce region-customizable sound extraction (ReZero), a general and\nflexible framework for the multi-channel region-wise sound extraction (R-SE)\ntask. R-SE task aims at extracting all active target sounds (e.g., human\nspeech) within a specific, user-defined spatial region, which is different from\nconventional and existing tasks where a blind separation or a fixed, predefined\nspatial region are typically assumed. The spatial region can be defined as an\nangular window, a sphere, a cone, or other geometric patterns. Being a solution\nto the R-SE task, the proposed ReZero framework includes (1) definitions of\ndifferent types of spatial regions, (2) methods for region feature extraction\nand aggregation, and (3) a multi-channel extension of the band-split RNN\n(BSRNN) model specified for the R-SE task. We design experiments for different\nmicrophone array geometries, different types of spatial regions, and\ncomprehensive ablation studies on different system configurations. Experimental\nresults on both simulated and real-recorded data demonstrate the effectiveness\nof ReZero. Demos are available at https://innerselfm.github.io/rezero/.\n", "date": "2023-08-31", "categories": []}
{"id": "2305.11582", "title": "What You Hear Is What You See: Audio Quality Metrics From Image Quality\n  Metrics", "abstract": "  In this study, we investigate the feasibility of utilizing state-of-the-art\nimage perceptual metrics for evaluating audio signals by representing them as\nspectrograms. The encouraging outcome of the proposed approach is based on the\nsimilarity between the neural mechanisms in the auditory and visual pathways.\nFurthermore, we customise one of the metrics which has a psychoacoustically\nplausible architecture to account for the peculiarities of sound signals. We\nevaluate the effectiveness of our proposed metric and several baseline metrics\nusing a music dataset, with promising results in terms of the correlation\nbetween the metrics and the perceived quality of audio as rated by human\nevaluators.\n", "date": "2023-05-19", "categories": []}
{"id": "2308.15726", "title": "AGS: An Dataset and Taxonomy for Domestic Scene Sound Event Recognition", "abstract": "  Environmental sound scene and sound event recognition is important for the\nrecognition of suspicious events in indoor and outdoor environments (such as\nnurseries, smart homes, nursing homes, etc.) and is a fundamental task involved\nin many audio surveillance applications. In particular, there is no public\ncommon data set for the research field of sound event recognition for the data\nset of the indoor environmental sound scene. Therefore, this paper proposes a\ndata set (called as AGS) for the home environment sound. This data set\nconsiders various types of overlapping audio in the scene, background noise.\nMoreover, based on the proposed data set, this paper compares and analyzes the\nadvanced methods for sound event recognition, and then illustrates the\nreliability of the data set proposed in this paper, and studies the challenges\nraised by the new data set. Our proposed AGS and the source code of the\ncorresponding baselines at https://github.com/taolunzu11/AGS .\n", "date": "2023-08-30", "categories": []}
{"id": "2308.15061", "title": "AIoT-Based Drum Transcription Robot using Convolutional Neural Networks", "abstract": "  With the development of information technology, robot technology has made\ngreat progress in various fields. These new technologies enable robots to be\nused in industry, agriculture, education and other aspects. In this paper, we\npropose a drum robot that can automatically complete music transcription in\nreal-time, which is based on AIoT and fog computing technology. Specifically,\nthis drum robot system consists of a cloud node for data storage, edge nodes\nfor real-time computing, and data-oriented execution application nodes. In\norder to analyze drumming music and realize drum transcription, we further\npropose a light-weight convolutional neural network model to classify drums,\nwhich can be more effectively deployed in terminal devices for fast edge\ncalculations. The experimental results show that the proposed system can\nachieve more competitive performance and enjoy a variety of smart applications\nand services.\n", "date": "2023-08-29", "categories": []}
{"id": "2308.15090", "title": "Killing two birds with one stone: Can an audio captioning system also be\n  used for audio-text retrieval?", "abstract": "  Automated Audio Captioning (AAC) aims to develop systems capable of\ndescribing an audio recording using a textual sentence. In contrast, Audio-Text\nRetrieval (ATR) systems seek to find the best matching audio recording(s) for a\ngiven textual query (Text-to-Audio) or vice versa (Audio-to-Text). These tasks\nrequire different types of systems: AAC employs a sequence-to-sequence model,\nwhile ATR utilizes a ranking model that compares audio and text representations\nwithin a shared projection subspace. However, this work investigates the\nrelationship between AAC and ATR by exploring the ATR capabilities of an\nunmodified AAC system, without fine-tuning for the new task. Our AAC system\nconsists of an audio encoder (ConvNeXt-Tiny) trained on AudioSet for audio\ntagging, and a transformer decoder responsible for generating sentences. For\nAAC, it achieves a high SPIDEr-FL score of 0.298 on Clotho and 0.472 on\nAudioCaps on average. For ATR, we propose using the standard Cross-Entropy loss\nvalues obtained for any audio/caption pair. Experimental results on the Clotho\nand AudioCaps datasets demonstrate decent recall values using this simple\napproach. For instance, we obtained a Text-to-Audio R@1 value of 0.382 for\nAu-dioCaps, which is above the current state-of-the-art method without external\ndata. Interestingly, we observe that normalizing the loss values was necessary\nfor Audio-to-Text retrieval.\n", "date": "2023-08-29", "categories": []}
{"id": "2308.07221", "title": "AudioFormer: Audio Transformer learns audio feature representations from\n  discrete acoustic codes", "abstract": "  We propose a method named AudioFormer,which learns audio feature\nrepresentations through the acquisition of discrete acoustic codes and\nsubsequently fine-tunes them for audio classification tasks. Initially,we\nintroduce a novel perspective by considering the audio classification task as a\nform of natural language understanding (NLU). Leveraging an existing neural\naudio codec model,we generate discrete acoustic codes and utilize them to train\na masked language model (MLM),thereby obtaining audio feature representations.\nFurthermore,we pioneer the integration of a Multi-Positive sample Contrastive\n(MPC) learning approach. This method enables the learning of joint\nrepresentations among multiple discrete acoustic codes within the same audio\ninput. In our experiments,we treat discrete acoustic codes as textual data and\ntrain a masked language model using a cloze-like methodology,ultimately\nderiving high-quality audio representations. Notably,the MPC learning technique\neffectively captures collaborative representations among distinct positive\nsamples. Our research outcomes demonstrate that AudioFormer attains\nsignificantly improved performance compared to prevailing monomodal audio\nclassification models across multiple datasets,and even outperforms\naudio-visual multimodal classification models on select datasets.\nSpecifically,our approach achieves remarkable results on datasets including\nAudioSet (2M,20K),and FSD50K,with performance scores of 53.9,45.1,and\n65.6,respectively. We have openly shared both the code and models:\nhttps://github.com/LZH-0225/AudioFormer.git.\n", "date": "2023-08-14", "categories": []}
{"id": "2308.12370", "title": "AdVerb: Visually Guided Audio Dereverberation", "abstract": "  We present AdVerb, a novel audio-visual dereverberation framework that uses\nvisual cues in addition to the reverberant sound to estimate clean audio.\nAlthough audio-only dereverberation is a well-studied problem, our approach\nincorporates the complementary visual modality to perform audio\ndereverberation. Given an image of the environment where the reverberated sound\nsignal has been recorded, AdVerb employs a novel geometry-aware cross-modal\ntransformer architecture that captures scene geometry and audio-visual\ncross-modal relationship to generate a complex ideal ratio mask, which, when\napplied to the reverberant audio predicts the clean sound. The effectiveness of\nour method is demonstrated through extensive quantitative and qualitative\nevaluations. Our approach significantly outperforms traditional audio-only and\naudio-visual baselines on three downstream tasks: speech enhancement, speech\nrecognition, and speaker verification, with relative improvements in the range\nof 18% - 82% on the LibriSpeech test-clean set. We also achieve highly\nsatisfactory RT60 error scores on the AVSpeech dataset.\n", "date": "2023-08-23", "categories": []}
{"id": "2308.12408", "title": "An Initial Exploration: Learning to Generate Realistic Audio for Silent\n  Video", "abstract": "  Generating realistic audio effects for movies and other media is a\nchallenging task that is accomplished today primarily through physical\ntechniques known as Foley art. Foley artists create sounds with common objects\n(e.g., boxing gloves, broken glass) in time with video as it is playing to\ngenerate captivating audio tracks. In this work, we aim to develop a\ndeep-learning based framework that does much the same - observes video in it's\nnatural sequence and generates realistic audio to accompany it. Notably, we\nhave reason to believe this is achievable due to advancements in realistic\naudio generation techniques conditioned on other inputs (e.g., Wavenet\nconditioned on text). We explore several different model architectures to\naccomplish this task that process both previously-generated audio and video\ncontext. These include deep-fusion CNN, dilated Wavenet CNN with visual\ncontext, and transformer-based architectures. We find that the\ntransformer-based architecture yields the most promising results, matching\nlow-frequencies to visual patterns effectively, but failing to generate more\nnuanced waveforms.\n", "date": "2023-08-23", "categories": []}
{"id": "2308.12526", "title": "UNISOUND System for VoxCeleb Speaker Recognition Challenge 2023", "abstract": "  This report describes the UNISOUND submission for Track1 and Track2 of\nVoxCeleb Speaker Recognition Challenge 2023 (VoxSRC 2023). We submit the same\nsystem on Track 1 and Track 2, which is trained with only VoxCeleb2-dev.\nLarge-scale ResNet and RepVGG architectures are developed for the challenge. We\npropose a consistency-aware score calibration method, which leverages the\nstability of audio voiceprints in similarity score by a Consistency Measure\nFactor (CMF). CMF brings a huge performance boost in this challenge. Our final\nsystem is a fusion of six models and achieves the first place in Track 1 and\nsecond place in Track 2 of VoxSRC 2023. The minDCF of our submission is 0.0855\nand the EER is 1.5880%.\n", "date": "2023-08-24", "categories": []}
{"id": "2308.12599", "title": "Exploiting Time-Frequency Conformers for Music Audio Enhancement", "abstract": "  With the proliferation of video platforms on the internet, recording musical\nperformances by mobile devices has become commonplace. However, these\nrecordings often suffer from degradation such as noise and reverberation, which\nnegatively impact the listening experience. Consequently, the necessity for\nmusic audio enhancement (referred to as music enhancement from this point\nonward), involving the transformation of degraded audio recordings into\npristine high-quality music, has surged to augment the auditory experience. To\naddress this issue, we propose a music enhancement system based on the\nConformer architecture that has demonstrated outstanding performance in speech\nenhancement tasks. Our approach explores the attention mechanisms of the\nConformer and examines their performance to discover the best approach for the\nmusic enhancement task. Our experimental results show that our proposed model\nachieves state-of-the-art performance on single-stem music enhancement.\nFurthermore, our system can perform general music enhancement with multi-track\nmixtures, which has not been examined in previous work.\n", "date": "2023-08-24", "categories": []}
{"id": "2308.12734", "title": "Real-time Detection of AI-Generated Speech for DeepFake Voice Conversion", "abstract": "  There are growing implications surrounding generative AI in the speech domain\nthat enable voice cloning and real-time voice conversion from one individual to\nanother. This technology poses a significant ethical threat and could lead to\nbreaches of privacy and misrepresentation, thus there is an urgent need for\nreal-time detection of AI-generated speech for DeepFake Voice Conversion. To\naddress the above emerging issues, the DEEP-VOICE dataset is generated in this\nstudy, comprised of real human speech from eight well-known figures and their\nspeech converted to one another using Retrieval-based Voice Conversion.\nPresenting as a binary classification problem of whether the speech is real or\nAI-generated, statistical analysis of temporal audio features through t-testing\nreveals that there are significantly different distributions. Hyperparameter\noptimisation is implemented for machine learning models to identify the source\nof speech. Following the training of 208 individual machine learning models\nover 10-fold cross validation, it is found that the Extreme Gradient Boosting\nmodel can achieve an average classification accuracy of 99.3% and can classify\nspeech in real-time, at around 0.004 milliseconds given one second of speech.\nAll data generated for this study is released publicly for future research on\nAI speech detection.\n", "date": "2023-08-24", "categories": []}
{"id": "2308.11800", "title": "Complex-valued neural networks for voice anti-spoofing", "abstract": "  Current anti-spoofing and audio deepfake detection systems use either\nmagnitude spectrogram-based features (such as CQT or Melspectrograms) or raw\naudio processed through convolution or sinc-layers. Both methods have\ndrawbacks: magnitude spectrograms discard phase information, which affects\naudio naturalness, and raw-feature-based models cannot use traditional\nexplainable AI methods. This paper proposes a new approach that combines the\nbenefits of both methods by using complex-valued neural networks to process the\ncomplex-valued, CQT frequency-domain representation of the input audio. This\nmethod retains phase information and allows for explainable AI methods. Results\nshow that this approach outperforms previous methods on the \"In-the-Wild\"\nanti-spoofing dataset and enables interpretation of the results through\nexplainable AI. Ablation studies confirm that the model has learned to use\nphase information to detect voice spoofing.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.11923", "title": "Audio Difference Captioning Utilizing Similarity-Discrepancy\n  Disentanglement", "abstract": "  We proposed Audio Difference Captioning (ADC) as a new extension task of\naudio captioning for describing the semantic differences between input pairs of\nsimilar but slightly different audio clips. The ADC solves the problem that\nconventional audio captioning sometimes generates similar captions for similar\naudio clips, failing to describe the difference in content. We also propose a\ncross-attention-concentrated transformer encoder to extract differences by\ncomparing a pair of audio clips and a similarity-discrepancy disentanglement to\nemphasize the difference in the latent space. To evaluate the proposed methods,\nwe built an AudioDiffCaps dataset consisting of pairs of similar but slightly\ndifferent audio clips with human-annotated descriptions of their differences.\nThe experiment with the AudioDiffCaps dataset showed that the proposed methods\nsolve the ADC task effectively and improve the attention weights to extract the\ndifference by visualizing them in the transformer encoder.\n", "date": "2023-08-23", "categories": []}
{"id": "2209.04360", "title": "A Semi-Supervised Algorithm for Improving the Consistency of\n  Crowdsourced Datasets: The COVID-19 Case Study on Respiratory Disorder\n  Classification", "abstract": "  Cough audio signal classification is a potentially useful tool in screening\nfor respiratory disorders, such as COVID-19. Since it is dangerous to collect\ndata from patients with such contagious diseases, many research teams have\nturned to crowdsourcing to quickly gather cough sound data, as it was done to\ngenerate the COUGHVID dataset. The COUGHVID dataset enlisted expert physicians\nto diagnose the underlying diseases present in a limited number of uploaded\nrecordings. However, this approach suffers from potential mislabeling of the\ncoughs, as well as notable disagreement between experts. In this work, we use a\nsemi-supervised learning (SSL) approach to improve the labeling consistency of\nthe COUGHVID dataset and the robustness of COVID-19 versus healthy cough sound\nclassification. First, we leverage existing SSL expert knowledge aggregation\ntechniques to overcome the labeling inconsistencies and sparsity in the\ndataset. Next, our SSL approach is used to identify a subsample of re-labeled\nCOUGHVID audio samples that can be used to train or augment future cough\nclassification models. The consistency of the re-labeled data is demonstrated\nin that it exhibits a high degree of class separability, 3x higher than that of\nthe user-labeled data, despite the expert label inconsistency present in the\noriginal dataset. Furthermore, the spectral differences in the user-labeled\naudio segments are amplified in the re-labeled data, resulting in significantly\ndifferent power spectral densities between healthy and COVID-19 coughs, which\ndemonstrates both the increased consistency of the new dataset and its\nexplainability from an acoustic perspective. Finally, we demonstrate how the\nre-labeled dataset can be used to train a cough classifier. This SSL approach\ncan be used to combine the medical knowledge of several experts to improve the\ndatabase consistency for any diagnostic classification task.\n", "date": "2022-09-09", "categories": []}
{"id": "2308.10281", "title": "The DKU-DUKEECE System for the Manipulation Region Location Task of ADD\n  2023", "abstract": "  This paper introduces our system designed for Track 2, which focuses on\nlocating manipulated regions, in the second Audio Deepfake Detection Challenge\n(ADD 2023). Our approach involves the utilization of multiple detection systems\nto identify splicing regions and determine their authenticity. Specifically, we\ntrain and integrate two frame-level systems: one for boundary detection and the\nother for deepfake detection. Additionally, we employ a third VAE model trained\nexclusively on genuine data to determine the authenticity of a given audio\nclip. Through the fusion of these three systems, our top-performing solution\nfor the ADD challenge achieves an impressive 82.23% sentence accuracy and an F1\nscore of 60.66%. This results in a final ADD score of 0.6713, securing the\nfirst rank in Track 2 of ADD 2023.\n", "date": "2023-08-20", "categories": []}
{"id": "2308.07293", "title": "DiffSED: Sound Event Detection with Denoising Diffusion", "abstract": "  Sound Event Detection (SED) aims to predict the temporal boundaries of all\nthe events of interest and their class labels, given an unconstrained audio\nsample. Taking either the splitand-classify (i.e., frame-level) strategy or the\nmore principled event-level modeling approach, all existing methods consider\nthe SED problem from the discriminative learning perspective. In this work, we\nreformulate the SED problem by taking a generative learning perspective.\nSpecifically, we aim to generate sound temporal boundaries from noisy proposals\nin a denoising diffusion process, conditioned on a target audio sample. During\ntraining, our model learns to reverse the noising process by converting noisy\nlatent queries to the groundtruth versions in the elegant Transformer decoder\nframework. Doing so enables the model generate accurate event boundaries from\neven noisy queries during inference. Extensive experiments on the Urban-SED and\nEPIC-Sounds datasets demonstrate that our model significantly outperforms\nexisting alternatives, with 40+% faster convergence in training.\n", "date": "2023-08-14", "categories": []}
{"id": "2308.08181", "title": "ChinaTelecom System Description to VoxCeleb Speaker Recognition\n  Challenge 2023", "abstract": "  This technical report describes ChinaTelecom system for Track 1 (closed) of\nthe VoxCeleb2023 Speaker Recognition Challenge (VoxSRC 2023). Our system\nconsists of several ResNet variants trained only on VoxCeleb2, which were fused\nfor better performance later. Score calibration was also applied for each\nvariant and the fused system. The final submission achieved minDCF of 0.1066\nand EER of 1.980%.\n", "date": "2023-08-16", "categories": []}
{"id": "2210.16192", "title": "Pretraining Respiratory Sound Representations using Metadata and\n  Contrastive Learning", "abstract": "  Methods based on supervised learning using annotations in an end-to-end\nfashion have been the state-of-the-art for classification problems. However,\nthey may be limited in their generalization capability, especially in the low\ndata regime. In this study, we address this issue using supervised contrastive\nlearning combined with available metadata to solve multiple pretext tasks that\nlearn a good representation of data. We apply our approach on respiratory sound\nclassification. This task is suited for this setting as demographic information\nsuch as sex and age are correlated with presence of lung diseases, and learning\na system that implicitly encode this information may better detect anomalies.\nSupervised contrastive learning is a paradigm that learns similar\nrepresentations to samples sharing the same class labels and dissimilar\nrepresentations to samples with different class labels. The feature extractor\nlearned using this paradigm extract useful features from the data, and we show\nthat it outperforms cross-entropy in classifying respiratory anomalies in two\ndifferent datasets. We also show that learning representations using only\nmetadata, without class labels, obtains similar performance as using cross\nentropy with those labels only. In addition, when combining class labels with\nmetadata using multiple supervised contrastive learning, an extension of\nsupervised contrastive learning solving an additional task of grouping patients\nwithin the same sex and age group, more informative features are learned. This\nwork suggests the potential of using multiple metadata sources in supervised\ncontrastive settings, in particular in settings with class imbalance and few\ndata. Our code is released at https://github.com/ilyassmoummad/scl_icbhi2017\n", "date": "2022-10-27", "categories": []}
{"id": "2302.07640", "title": "Detection and classification of vocal productions in large scale audio\n  recordings", "abstract": "  We propose an automatic data processing pipeline to extract vocal productions\nfrom large-scale natural audio recordings and classify these vocal productions.\nThe pipeline is based on a deep neural network and adresses both issues\nsimultaneously. Though a series of computationel steps (windowing, creation of\na noise class, data augmentation, re-sampling, transfer learning, Bayesian\noptimisation), it automatically trains a neural network without requiring a\nlarge sample of labeled data and important computing resources. Our end-to-end\nmethodology can handle noisy recordings made under different recording\nconditions. We test it on two different natural audio data sets, one from a\ngroup of Guinea baboons recorded from a primate research center and one from\nhuman babies recorded at home. The pipeline trains a model on 72 and 77 minutes\nof labeled audio recordings, with an accuracy of 94.58% and 99.76%. It is then\nused to process 443 and 174 hours of natural continuous recordings and it\ncreates two new databases of 38.8 and 35.2 hours, respectively. We discuss the\nstrengths and limitations of this approach that can be applied to any massive\naudio recording.\n", "date": "2023-02-14", "categories": []}
{"id": "2306.09807", "title": "FALL-E: A Foley Sound Synthesis Model and Strategies", "abstract": "  This paper introduces FALL-E, a foley synthesis system and its\ntraining/inference strategies. The FALL-E model employs a cascaded approach\ncomprising low-resolution spectrogram generation, spectrogram super-resolution,\nand a vocoder. We trained every sound-related model from scratch using our\nextensive datasets, and utilized a pre-trained language model. We conditioned\nthe model with dataset-specific texts, enabling it to learn sound quality and\nrecording environment based on text input. Moreover, we leveraged external\nlanguage models to improve text descriptions of our datasets and performed\nprompt engineering for quality, coherence, and diversity. FALL-E was evaluated\nby an objective measure as well as listening tests in the DCASE 2023 challenge\nTask 7. The submission achieved the second place on average, while achieving\nthe best score for diversity, second place for audio quality, and third place\nfor class fitness.\n", "date": "2023-06-16", "categories": []}
{"id": "2308.04978", "title": "Transferable Models for Bioacoustics with Human Language Supervision", "abstract": "  Passive acoustic monitoring offers a scalable, non-invasive method for\ntracking global biodiversity and anthropogenic impacts on species. Although\ndeep learning has become a vital tool for processing this data, current models\nare inflexible, typically cover only a handful of species, and are limited by\ndata scarcity. In this work, we propose BioLingual, a new model for\nbioacoustics based on contrastive language-audio pretraining. We first\naggregate bioacoustic archives into a language-audio dataset, called\nAnimalSpeak, with over a million audio-caption pairs holding information on\nspecies, vocalization context, and animal behavior. After training on this\ndataset to connect language and audio representations, our model can identify\nover a thousand species' calls across taxa, complete bioacoustic tasks\nzero-shot, and retrieve animal vocalization recordings from natural text\nqueries. When fine-tuned, BioLingual sets a new state-of-the-art on nine tasks\nin the Benchmark of Animal Sounds. Given its broad taxa coverage and ability to\nbe flexibly queried in human language, we believe this model opens new\nparadigms in ecological monitoring and research, including free-text search on\nthe world's acoustic monitoring archives. We open-source our models, dataset,\nand code.\n", "date": "2023-08-09", "categories": []}
{"id": "2308.04169", "title": "Dual input neural networks for positional sound source localization", "abstract": "  In many signal processing applications, metadata may be advantageously used\nin conjunction with a high dimensional signal to produce a desired output. In\nthe case of classical Sound Source Localization (SSL) algorithms, information\nfrom a high dimensional, multichannel audio signals received by many\ndistributed microphones is combined with information describing acoustic\nproperties of the scene, such as the microphones' coordinates in space, to\nestimate the position of a sound source. We introduce Dual Input Neural\nNetworks (DI-NNs) as a simple and effective way to model these two data types\nin a neural network. We train and evaluate our proposed DI-NN on scenarios of\nvarying difficulty and realism and compare it against an alternative\narchitecture, a classical Least-Squares (LS) method as well as a classical\nConvolutional Recurrent Neural Network (CRNN). Our results show that the DI-NN\nsignificantly outperforms the baselines, achieving a five times lower\nlocalization error than the LS method and two times lower than the CRNN in a\ntest dataset of real recordings.\n", "date": "2023-08-08", "categories": []}
{"id": "2308.04258", "title": "Advancing Natural-Language Based Audio Retrieval with PaSST and Large\n  Audio-Caption Data Sets", "abstract": "  This work presents a text-to-audio-retrieval system based on pre-trained text\nand spectrogram transformers. Our method projects recordings and textual\ndescriptions into a shared audio-caption space in which related examples from\ndifferent modalities are close. Through a systematic analysis, we examine how\neach component of the system influences retrieval performance. As a result, we\nidentify two key components that play a crucial role in driving performance:\nthe self-attention-based audio encoder for audio embedding and the utilization\nof additional human-generated and synthetic data sets during pre-training. We\nfurther experimented with augmenting ClothoV2 captions with available keywords\nto increase their variety; however, this only led to marginal improvements. Our\nsystem ranked first in the 2023's DCASE Challenge, and it outperforms the\ncurrent state of the art on the ClothoV2 benchmark by 5.6 pp. mAP@10.\n", "date": "2023-08-08", "categories": []}
{"id": "2305.10270", "title": "Boosting Local Spectro-Temporal Features for Speech Analysis", "abstract": "  We introduce the problem of phone classification in the context of speech\nrecognition, and explore several sets of local spectro-temporal features that\ncan be used for phone classification. In particular, we present some\npreliminary results for phone classification using two sets of features that\nare commonly used for object detection: Haar features and SVM-classified\nHistograms of Gradients (HoG).\n", "date": "2023-05-17", "categories": []}
{"id": "2308.03300", "title": "Do You Remember? Overcoming Catastrophic Forgetting for Fake Audio\n  Detection", "abstract": "  Current fake audio detection algorithms have achieved promising performances\non most datasets. However, their performance may be significantly degraded when\ndealing with audio of a different dataset. The orthogonal weight modification\nto overcome catastrophic forgetting does not consider the similarity of genuine\naudio across different datasets. To overcome this limitation, we propose a\ncontinual learning algorithm for fake audio detection to overcome catastrophic\nforgetting, called Regularized Adaptive Weight Modification (RAWM). When\nfine-tuning a detection network, our approach adaptively computes the direction\nof weight modification according to the ratio of genuine utterances and fake\nutterances. The adaptive modification direction ensures the network can\neffectively detect fake audio on the new dataset while preserving its knowledge\nof old model, thus mitigating catastrophic forgetting. In addition, genuine\naudio collected from quite different acoustic conditions may skew their feature\ndistribution, so we introduce a regularization constraint to force the network\nto remember the old distribution in this regard. Our method can easily be\ngeneralized to related fields, like speech emotion recognition. We also\nevaluate our approach across multiple datasets and obtain a significant\nperformance improvement on cross-dataset experiments.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.03332", "title": "Improving Deep Attractor Network by BGRU and GMM for Speech Separation", "abstract": "  Deep Attractor Network (DANet) is the state-of-the-art technique in speech\nseparation field, which uses Bidirectional Long Short-Term Memory (BLSTM), but\nthe complexity of the DANet model is very high. In this paper, a simplified and\npowerful DANet model is proposed using Bidirectional Gated neural network\n(BGRU) instead of BLSTM. The Gaussian Mixture Model (GMM) other than the\nk-means was applied in DANet as a clustering algorithm to reduce the complexity\nand increase the learning speed and accuracy. The metrics used in this paper\nare Signal to Distortion Ratio (SDR), Signal to Interference Ratio (SIR),\nSignal to Artifact Ratio (SAR), and Perceptual Evaluation Speech Quality (PESQ)\nscore. Two speaker mixture datasets from TIMIT corpus were prepared to evaluate\nthe proposed model, and the system achieved 12.3 dB and 2.94 for SDR and PESQ\nscores respectively, which were better than the original DANet model. Other\nimprovements were 20.7% and 17.9% in the number of parameters and time\ntraining, respectively. The model was applied on mixed Arabic speech signals\nand the results were better than that in English.\n", "date": "2023-08-07", "categories": []}
{"id": "2303.11020", "title": "DS-TDNN: Dual-stream Time-delay Neural Network with Global-aware Filter\n  for Speaker Verification", "abstract": "  Conventional time-delay neural networks (TDNNs) struggle to handle long-range\ncontext, their ability to represent speaker information is therefore limited in\nlong utterances. Existing solutions either depend on increasing model\ncomplexity or try to balance between local features and global context to\naddress this issue. To effectively leverage the long-term dependencies of audio\nsignals and constrain model complexity, we introduce a novel module called\nGlobal-aware Filter layer (GF layer) in this work, which employs a set of\nlearnable transform-domain filters between a 1D discrete Fourier transform and\nits inverse transform to capture global context. Additionally, we develop a\ndynamic filtering strategy and a sparse regularization method to enhance the\nperformance of the GF layer and prevent overfitting. Based on the GF layer, we\npresent a dual-stream TDNN architecture called DS-TDNN for automatic speaker\nverification (ASV), which utilizes two unique branches to extract both local\nand global features in parallel and employs an efficient strategy to fuse\ndifferent-scale information. Experiments on the Voxceleb and SITW databases\ndemonstrate that the DS-TDNN achieves a relative improvement of 10\\% together\nwith a relative decline of 20\\% in computational cost over the ECAPA-TDNN in\nspeaker verification task. This improvement will become more evident as the\nutterance's duration grows. Furthermore, the DS-TDNN also beats popular deep\nresidual models and attention-based systems on utterances of arbitrary length.\n", "date": "2023-03-20", "categories": []}
{"id": "2305.01864", "title": "Unsupervised Improvement of Audio-Text Cross-Modal Representations", "abstract": "  Recent advances in using language models to obtain cross-modal audio-text\nrepresentations have overcome the limitations of conventional training\napproaches that use predefined labels. This has allowed the community to make\nprogress in tasks like zero-shot classification, which would otherwise not be\npossible. However, learning such representations requires a large amount of\nhuman-annotated audio-text pairs. In this paper, we study unsupervised\napproaches to improve the learning framework of such representations with\nunpaired text and audio. We explore domain-unspecific and domain-specific\ncuration methods to create audio-text pairs that we use to further improve the\nmodel. We also show that when domain-specific curation is used in conjunction\nwith a soft-labeled contrastive loss, we are able to obtain significant\nimprovement in terms of zero-shot classification performance on downstream\nsound event classification or acoustic scene classification tasks.\n", "date": "2023-05-03", "categories": []}
{"id": "2304.03323", "title": "DSVAE: Interpretable Disentangled Representation for Synthetic Speech\n  Detection", "abstract": "  Tools to generate high quality synthetic speech signal that is perceptually\nindistinguishable from speech recorded from human speakers are easily\navailable. Several approaches have been proposed for detecting synthetic\nspeech. Many of these approaches use deep learning methods as a black box\nwithout providing reasoning for the decisions they make. This limits the\ninterpretability of these approaches. In this paper, we propose Disentangled\nSpectrogram Variational Auto Encoder (DSVAE) which is a two staged trained\nvariational autoencoder that processes spectrograms of speech using\ndisentangled representation learning to generate interpretable representations\nof a speech signal for detecting synthetic speech. DSVAE also creates an\nactivation map to highlight the spectrogram regions that discriminate synthetic\nand bona fide human speech signals. We evaluated the representations obtained\nfrom DSVAE using the ASVspoof2019 dataset. Our experimental results show high\naccuracy (>98%) on detecting synthetic speech from 6 known and 10 out of 11\nunknown speech synthesizers. We also visualize the representation obtained from\nDSVAE for 17 different speech synthesizers and verify that they are indeed\ninterpretable and discriminate bona fide and synthetic speech from each of the\nsynthesizers.\n", "date": "2023-04-06", "categories": []}
{"id": "2307.15913", "title": "Moisesdb: A dataset for source separation beyond 4-stems", "abstract": "  In this paper, we introduce the MoisesDB dataset for musical source\nseparation. It consists of 240 tracks from 45 artists, covering twelve musical\ngenres. For each song, we provide its individual audio sources, organized in a\ntwo-level hierarchical taxonomy of stems. This will facilitate building and\nevaluating fine-grained source separation systems that go beyond the limitation\nof using four stems (drums, bass, other, and vocals) due to lack of data. To\nfacilitate the adoption of this dataset, we publish an easy-to-use Python\nlibrary to download, process and use MoisesDB. Alongside a thorough\ndocumentation and analysis of the dataset contents, this work provides baseline\nresults for open-source separation models for varying separation granularities\n(four, five, and six stems), and discuss their results.\n", "date": "2023-07-29", "categories": []}
{"id": "2307.15101", "title": "Detection of Children Abuse by Voice and Audio Classification by\n  Short-Time Fourier Transform Machine Learning implemented on Nvidia Edge GPU\n  device", "abstract": "  The safety of children in children home has become an increasing social\nconcern, and the purpose of this experiment is to use machine learning applied\nto detect the scenarios of child abuse to increase the safety of children. This\nexperiment uses machine learning to classify and recognize a child's voice and\npredict whether the current sound made by the child is crying, screaming or\nlaughing. If a child is found to be crying or screaming, an alert is\nimmediately sent to the relevant personnel so that they can perceive what the\nchild may be experiencing in a surveillance blind spot and respond in a timely\nmanner. Together with a hybrid use of video image classification, the accuracy\nof child abuse detection can be significantly increased. This greatly reduces\nthe likelihood that a child will receive violent abuse in the nursery and\nallows personnel to stop an imminent or incipient child abuse incident in time.\nThe datasets collected from this experiment is entirely from sounds recorded on\nsite at the children home, including crying, laughing, screaming sound and\nbackground noises. These sound files are transformed into spectrograms using\nShort-Time Fourier Transform, and then these image data are imported into a CNN\nneural network for classification, and the final trained model can achieve an\naccuracy of about 92% for sound detection.\n", "date": "2023-07-27", "categories": []}
{"id": "2209.03143", "title": "AudioLM: a Language Modeling Approach to Audio Generation", "abstract": "  We introduce AudioLM, a framework for high-quality audio generation with\nlong-term consistency. AudioLM maps the input audio to a sequence of discrete\ntokens and casts audio generation as a language modeling task in this\nrepresentation space. We show how existing audio tokenizers provide different\ntrade-offs between reconstruction quality and long-term structure, and we\npropose a hybrid tokenization scheme to achieve both objectives. Namely, we\nleverage the discretized activations of a masked language model pre-trained on\naudio to capture long-term structure and the discrete codes produced by a\nneural audio codec to achieve high-quality synthesis. By training on large\ncorpora of raw audio waveforms, AudioLM learns to generate natural and coherent\ncontinuations given short prompts. When trained on speech, and without any\ntranscript or annotation, AudioLM generates syntactically and semantically\nplausible speech continuations while also maintaining speaker identity and\nprosody for unseen speakers. Furthermore, we demonstrate how our approach\nextends beyond speech by generating coherent piano music continuations, despite\nbeing trained without any symbolic representation of music.\n", "date": "2022-09-07", "categories": []}
{"id": "2307.13788", "title": "Histogram Layer Time Delay Neural Networks for Passive Sonar\n  Classification", "abstract": "  Underwater acoustic target detection in remote marine sensing operations is\nchallenging due to complex sound wave propagation. Despite the availability of\nreliable sonar systems, target recognition remains a difficult problem. Various\nmethods address improved target recognition. However, most struggle to\ndisentangle the high-dimensional, non-linear patterns in the observed target\nrecordings. In this work, a novel method combines a time delay neural network\nand histogram layer to incorporate statistical contexts for improved feature\nlearning and underwater acoustic target classification. The proposed method\noutperforms the baseline model, demonstrating the utility in incorporating\nstatistical contexts for passive sonar target recognition. The code for this\nwork is publicly available.\n", "date": "2023-07-25", "categories": []}
{"id": "2307.12343", "title": "Self-Supervised Learning for Audio-Based Emotion Recognition", "abstract": "  Emotion recognition models using audio input data can enable the development\nof interactive systems with applications in mental healthcare, marketing,\ngaming, and social media analysis. While the field of affective computing using\naudio data is rich, a major barrier to achieve consistently high-performance\nmodels is the paucity of available training labels. Self-supervised learning\n(SSL) is a family of methods which can learn despite a scarcity of supervised\nlabels by predicting properties of the data itself. To understand the utility\nof self-supervised learning for audio-based emotion recognition, we have\napplied self-supervised learning pre-training to the classification of emotions\nfrom the CMU- MOSEI's acoustic modality. Unlike prior papers that have\nexperimented with raw acoustic data, our technique has been applied to encoded\nacoustic data. Our model is first pretrained to uncover the randomly-masked\ntimestamps of the acoustic data. The pre-trained model is then fine-tuned using\na small sample of annotated data. The performance of the final model is then\nevaluated via several evaluation metrics against a baseline deep learning model\nwith an identical backbone architecture. We find that self-supervised learning\nconsistently improves the performance of the model across all metrics. This\nwork shows the utility of self-supervised learning for affective computing,\ndemonstrating that self-supervised learning is most useful when the number of\ntraining examples is small, and that the effect is most pronounced for emotions\nwhich are easier to classify such as happy, sad and anger. This work further\ndemonstrates that self-supervised learning works when applied to embedded\nfeature representations rather than the traditional approach of pre-training on\nthe raw input space.\n", "date": "2023-07-23", "categories": []}
{"id": "2306.09382", "title": "Sound Demixing Challenge 2023 Music Demixing Track Technical Report:\n  TFC-TDF-UNet v3", "abstract": "  In this report, we present our award-winning solutions for the Music Demixing\nTrack of Sound Demixing Challenge 2023. First, we propose TFC-TDF-UNet v3, a\ntime-efficient music source separation model that achieves state-of-the-art\nresults on the MUSDB benchmark. We then give full details regarding our\nsolutions for each Leaderboard, including a loss masking approach for\nnoise-robust training. Code for reproducing model training and final\nsubmissions is available at github.com/kuielab/sdx23.\n", "date": "2023-06-15", "categories": []}
{"id": "2307.11450", "title": "Topic Identification For Spontaneous Speech: Enriching Audio Features\n  With Embedded Linguistic Information", "abstract": "  Traditional topic identification solutions from audio rely on an automatic\nspeech recognition system (ASR) to produce transcripts used as input to a\ntext-based model. These approaches work well in high-resource scenarios, where\nthere are sufficient data to train both components of the pipeline. However, in\nlow-resource situations, the ASR system, even if available, produces\nlow-quality transcripts, leading to a bad text-based classifier. Moreover,\nspontaneous speech containing hesitations can further degrade the performance\nof the ASR model. In this paper, we investigate alternatives to the standard\ntext-only solutions by comparing audio-only and hybrid techniques of jointly\nutilising text and audio features. The models evaluated on spontaneous Finnish\nspeech demonstrate that purely audio-based solutions are a viable option when\nASR components are not available, while the hybrid multi-modal solutions\nachieve the best results.\n", "date": "2023-07-21", "categories": []}
{"id": "2307.10773", "title": "Music Genre Classification with ResNet and Bi-GRU Using Visual\n  Spectrograms", "abstract": "  Music recommendation systems have emerged as a vital component to enhance\nuser experience and satisfaction for the music streaming services, which\ndominates music consumption. The key challenge in improving these recommender\nsystems lies in comprehending the complexity of music data, specifically for\nthe underpinning music genre classification. The limitations of manual genre\nclassification have highlighted the need for a more advanced system, namely the\nAutomatic Music Genre Classification (AMGC) system. While traditional machine\nlearning techniques have shown potential in genre classification, they heavily\nrely on manually engineered features and feature selection, failing to capture\nthe full complexity of music data. On the other hand, deep learning\nclassification architectures like the traditional Convolutional Neural Networks\n(CNN) are effective in capturing the spatial hierarchies but struggle to\ncapture the temporal dynamics inherent in music data. To address these\nchallenges, this study proposes a novel approach using visual spectrograms as\ninput, and propose a hybrid model that combines the strength of the Residual\nneural Network (ResNet) and the Gated Recurrent Unit (GRU). This model is\ndesigned to provide a more comprehensive analysis of music data, offering the\npotential to improve the music recommender systems through achieving a more\ncomprehensive analysis of music data and hence potentially more accurate genre\nclassification.\n", "date": "2023-07-20", "categories": []}
{"id": "2307.09795", "title": "From West to East: Who can understand the music of the others better?", "abstract": "  Recent developments in MIR have led to several benchmark deep learning models\nwhose embeddings can be used for a variety of downstream tasks. At the same\ntime, the vast majority of these models have been trained on Western pop/rock\nmusic and related styles. This leads to research questions on whether these\nmodels can be used to learn representations for different music cultures and\nstyles, or whether we can build similar music audio embedding models trained on\ndata from different cultures or styles. To that end, we leverage transfer\nlearning methods to derive insights about the similarities between the\ndifferent music cultures to which the data belongs to. We use two Western music\ndatasets, two traditional/folk datasets coming from eastern Mediterranean\ncultures, and two datasets belonging to Indian art music. Three deep audio\nembedding models are trained and transferred across domains, including two\nCNN-based and a Transformer-based architecture, to perform auto-tagging for\neach target domain dataset. Experimental results show that competitive\nperformance is achieved in all domains via transfer learning, while the best\nsource dataset varies for each music culture. The implementation and the\ntrained models are both provided in a public repository.\n", "date": "2023-07-19", "categories": []}
{"id": "2307.09286", "title": "FlexiAST: Flexibility is What AST Needs", "abstract": "  The objective of this work is to give patch-size flexibility to Audio\nSpectrogram Transformers (AST). Recent advancements in ASTs have shown superior\nperformance in various audio-based tasks. However, the performance of standard\nASTs degrades drastically when evaluated using different patch sizes from that\nused during training. As a result, AST models are typically re-trained to\naccommodate changes in patch sizes. To overcome this limitation, this paper\nproposes a training procedure to provide flexibility to standard AST models\nwithout architectural changes, allowing them to work with various patch sizes\nat the inference stage - FlexiAST. This proposed training approach simply\nutilizes random patch size selection and resizing of patch and positional\nembedding weights. Our experiments show that FlexiAST gives similar performance\nto standard AST models while maintaining its evaluation ability at various\npatch sizes on different datasets for audio classification tasks.\n", "date": "2023-07-18", "categories": []}
{"id": "2212.08071", "title": "MAViL: Masked Audio-Video Learners", "abstract": "  We present Masked Audio-Video Learners (MAViL) to train audio-visual\nrepresentations. Our approach learns with three complementary forms of\nself-supervision: (1) reconstruction of masked audio and video input data, (2)\nintra- and inter-modal contrastive learning with masking, and (3) self-training\nby reconstructing joint audio-video contextualized features learned from the\nfirst two objectives. Pre-training with MAViL not only enables the model to\nperform well in audio-visual classification and retrieval tasks but also\nimproves representations of each modality in isolation, without using\ninformation from the other modality for fine-tuning or inference. Empirically,\nMAViL sets a new state-of-the-art on AudioSet (53.1 mAP) and VGGSound (67.1%\naccuracy). For the first time, a self-supervised audio-visual model outperforms\nones that use external supervision on these benchmarks.\n", "date": "2022-12-15", "categories": []}
{"id": "2307.07265", "title": "AudioInceptionNeXt: TCL AI LAB Submission to EPIC-SOUND\n  Audio-Based-Interaction-Recognition Challenge 2023", "abstract": "  This report presents the technical details of our submission to the 2023\nEpic-Kitchen EPIC-SOUNDS Audio-Based Interaction Recognition Challenge. The\ntask is to learn the mapping from audio samples to their corresponding action\nlabels. To achieve this goal, we propose a simple yet effective single-stream\nCNN-based architecture called AudioInceptionNeXt that operates on the\ntime-frequency log-mel-spectrogram of the audio samples. Motivated by the\ndesign of the InceptionNeXt, we propose parallel multi-scale depthwise\nseparable convolutional kernels in the AudioInceptionNeXt block, which enable\nthe model to learn the time and frequency information more effectively. The\nlarge-scale separable kernels capture the long duration of activities and the\nglobal frequency semantic information, while the small-scale separable kernels\ncapture the short duration of activities and local details of frequency\ninformation. Our approach achieved 55.43% of top-1 accuracy on the challenge\ntest set, ranked as 1st on the public leaderboard. Codes are available\nanonymously at https://github.com/StevenLauHKHK/AudioInceptionNeXt.git.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.07426", "title": "Real-time Percussive Technique Recognition and Embedding Learning for\n  the Acoustic Guitar", "abstract": "  Real-time music information retrieval (RT-MIR) has much potential to augment\nthe capabilities of traditional acoustic instruments. We develop RT-MIR\ntechniques aimed at augmenting percussive fingerstyle, which blends acoustic\nguitar playing with guitar body percussion. We formulate several design\nobjectives for RT-MIR systems for augmented instrument performance: (i) causal\nconstraint, (ii) perceptually negligible action-to-sound latency, (iii) control\nintimacy support, (iv) synthesis control support. We present and evaluate\nreal-time guitar body percussion recognition and embedding learning techniques\nbased on convolutional neural networks (CNNs) and CNNs jointly trained with\nvariational autoencoders (VAEs). We introduce a taxonomy of guitar body\npercussion based on hand part and location. We follow a cross-dataset\nevaluation approach by collecting three datasets labelled according to the\ntaxonomy. The embedding quality of the models is assessed using KL-Divergence\nacross distributions corresponding to different taxonomic classes. Results\nindicate that the networks are strong classifiers especially in a simplified\n2-class recognition task, and the VAEs yield improved class separation compared\nto CNNs as evidenced by increased KL-Divergence across distributions. We argue\nthat the VAE embedding quality could support control intimacy and rich\ninteraction when the latent space's parameters are used to control an external\nsynthesis engine. Further design challenges around generalisation to different\ndatasets have been identified.\n", "date": "2023-07-13", "categories": []}
{"id": "2307.07436", "title": "Towards spoken dialect identification of Irish", "abstract": "  The Irish language is rich in its diversity of dialects and accents. This\ncompounds the difficulty of creating a speech recognition system for the\nlow-resource language, as such a system must contend with a high degree of\nvariability with limited corpora. A recent study investigating dialect bias in\nIrish ASR found that balanced training corpora gave rise to unequal dialect\nperformance, with performance for the Ulster dialect being consistently worse\nthan for the Connacht or Munster dialects. Motivated by this, the present\nexperiments investigate spoken dialect identification of Irish, with a view to\nincorporating such a system into the speech recognition pipeline. Two acoustic\nclassification models are tested, XLS-R and ECAPA-TDNN, in conjunction with a\ntext-based classifier using a pretrained Irish-language BERT model. The\nECAPA-TDNN, particularly a model pretrained for language identification on the\nVoxLingua107 dataset, performed best overall, with an accuracy of 73%. This was\nfurther improved to 76% by fusing the model's outputs with the text-based\nmodel. The Ulster dialect was most accurately identified, with an accuracy of\n94%, however the model struggled to disambiguate between the Connacht and\nMunster dialects, suggesting a more nuanced approach may be necessary to\nrobustly distinguish between the dialects of Irish.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.06860", "title": "AnuraSet: A dataset for benchmarking Neotropical anuran calls\n  identification in passive acoustic monitoring", "abstract": "  Global change is predicted to induce shifts in anuran acoustic behavior,\nwhich can be studied through passive acoustic monitoring (PAM). Understanding\nchanges in calling behavior requires the identification of anuran species,\nwhich is challenging due to the particular characteristics of neotropical\nsoundscapes. In this paper, we introduce a large-scale multi-species dataset of\nanuran amphibians calls recorded by PAM, that comprises 27 hours of expert\nannotations for 42 different species from two Brazilian biomes. We provide open\naccess to the dataset, including the raw recordings, experimental setup code,\nand a benchmark with a baseline model of the fine-grained categorization\nproblem. Additionally, we highlight the challenges of the dataset to encourage\nmachine learning researchers to solve the problem of anuran call identification\ntowards conservation policy. All our experiments and resources can be found on\nour GitHub repository https://github.com/soundclim/anuraset.\n", "date": "2023-07-11", "categories": []}
{"id": "2307.04686", "title": "VampNet: Music Generation via Masked Acoustic Token Modeling", "abstract": "  We introduce VampNet, a masked acoustic token modeling approach to music\nsynthesis, compression, inpainting, and variation. We use a variable masking\nschedule during training which allows us to sample coherent music from the\nmodel by applying a variety of masking approaches (called prompts) during\ninference. VampNet is non-autoregressive, leveraging a bidirectional\ntransformer architecture that attends to all tokens in a forward pass. With\njust 36 sampling passes, VampNet can generate coherent high-fidelity musical\nwaveforms. We show that by prompting VampNet in various ways, we can apply it\nto tasks like music compression, inpainting, outpainting, continuation, and\nlooping with variation (vamping). Appropriately prompted, VampNet is capable of\nmaintaining style, genre, instrumentation, and other high-level aspects of the\nmusic. This flexible prompting capability makes VampNet a powerful music\nco-creation tool. Code and audio samples are available online.\n", "date": "2023-07-10", "categories": []}
{"id": "2307.05161", "title": "On the Effectiveness of Speech Self-supervised Learning for Music", "abstract": "  Self-supervised learning (SSL) has shown promising results in various speech\nand natural language processing applications. However, its efficacy in music\ninformation retrieval (MIR) still remains largely unexplored. While previous\nSSL models pre-trained on music recordings may have been mostly closed-sourced,\nrecent speech models such as wav2vec2.0 have shown promise in music modelling.\nNevertheless, research exploring the effectiveness of applying speech SSL\nmodels to music recordings has been limited. We explore the music adaption of\nSSL with two distinctive speech-related models, data2vec1.0 and Hubert, and\nrefer to them as music2vec and musicHuBERT, respectively. We train $12$ SSL\nmodels with 95M parameters under various pre-training configurations and\nsystematically evaluate the MIR task performances with 13 different MIR tasks.\nOur findings suggest that training with music data can generally improve\nperformance on MIR tasks, even when models are trained using paradigms designed\nfor speech. However, we identify the limitations of such existing\nspeech-oriented designs, especially in modelling polyphonic information. Based\non the experimental results, empirical suggestions are also given for designing\nfuture musical SSL strategies and paradigms.\n", "date": "2023-07-11", "categories": []}
{"id": "2303.16897", "title": "Physics-Driven Diffusion Models for Impact Sound Synthesis from Videos", "abstract": "  Modeling sounds emitted from physical object interactions is critical for\nimmersive perceptual experiences in real and virtual worlds. Traditional\nmethods of impact sound synthesis use physics simulation to obtain a set of\nphysics parameters that could represent and synthesize the sound. However, they\nrequire fine details of both the object geometries and impact locations, which\nare rarely available in the real world and can not be applied to synthesize\nimpact sounds from common videos. On the other hand, existing video-driven deep\nlearning-based approaches could only capture the weak correspondence between\nvisual content and impact sounds since they lack of physics knowledge. In this\nwork, we propose a physics-driven diffusion model that can synthesize\nhigh-fidelity impact sound for a silent video clip. In addition to the video\ncontent, we propose to use additional physics priors to guide the impact sound\nsynthesis procedure. The physics priors include both physics parameters that\nare directly estimated from noisy real-world impact sound examples without\nsophisticated setup and learned residual parameters that interpret the sound\nenvironment via neural networks. We further implement a novel diffusion model\nwith specific training and inference strategies to combine physics priors and\nvisual information for impact sound synthesis. Experimental results show that\nour model outperforms several existing systems in generating realistic impact\nsounds. More importantly, the physics-based representations are fully\ninterpretable and transparent, thus enabling us to perform sound editing\nflexibly.\n", "date": "2023-03-29", "categories": []}
{"id": "2307.04292", "title": "A Demand-Driven Perspective on Generative Audio AI", "abstract": "  To achieve successful deployment of AI research, it is crucial to understand\nthe demands of the industry. In this paper, we present the results of a survey\nconducted with professional audio engineers, in order to determine research\npriorities and define various research tasks. We also summarize the current\nchallenges in audio quality and controllability based on the survey. Our\nanalysis emphasizes that the availability of datasets is currently the main\nbottleneck for achieving high-quality audio generation. Finally, we suggest\npotential solutions for some revealed issues with empirical evidence.\n", "date": "2023-07-10", "categories": []}
{"id": "2307.04305", "title": "Automatic Piano Transcription with Hierarchical Frequency-Time\n  Transformer", "abstract": "  Taking long-term spectral and temporal dependencies into account is essential\nfor automatic piano transcription. This is especially helpful when determining\nthe precise onset and offset for each note in the polyphonic piano content. In\nthis case, we may rely on the capability of self-attention mechanism in\nTransformers to capture these long-term dependencies in the frequency and time\naxes. In this work, we propose hFT-Transformer, which is an automatic music\ntranscription method that uses a two-level hierarchical frequency-time\nTransformer architecture. The first hierarchy includes a convolutional block in\nthe time axis, a Transformer encoder in the frequency axis, and a Transformer\ndecoder that converts the dimension in the frequency axis. The output is then\nfed into the second hierarchy which consists of another Transformer encoder in\nthe time axis. We evaluated our method with the widely used MAPS and MAESTRO\nv3.0.0 datasets, and it demonstrated state-of-the-art performance on all the\nF1-scores of the metrics among Frame, Note, Note with Offset, and Note with\nOffset and Velocity estimations.\n", "date": "2023-07-10", "categories": []}
{"id": "2307.04604", "title": "EchoVest: Real-Time Sound Classification and Depth Perception Expressed\n  through Transcutaneous Electrical Nerve Stimulation", "abstract": "  Over 1.5 billion people worldwide live with hearing impairment. Despite\nvarious technologies that have been created for individuals with such\ndisabilities, most of these technologies are either extremely expensive or\ninaccessible for everyday use in low-medium income countries. In order to\ncombat this issue, we have developed a new assistive device, EchoVest, for\nblind/deaf people to intuitively become more aware of their environment.\nEchoVest transmits vibrations to the user's body by utilizing transcutaneous\nelectric nerve stimulation (TENS) based on the source of the sounds. EchoVest\nalso provides various features, including sound localization, sound\nclassification, noise reduction, and depth perception. We aimed to outperform\nCNN-based machine-learning models, the most commonly used machine learning\nmodel for classification tasks, in accuracy and computational costs. To do so,\nwe developed and employed a novel audio pipeline that adapts the Audio\nSpectrogram Transformer (AST) model, an attention-based model, for our sound\nclassification purposes, and Fast Fourier Transforms for noise reduction. The\napplication of Otsu's Method helped us find the optimal thresholds for\nbackground noise sound filtering and gave us much greater accuracy. In order to\ncalculate direction and depth accurately, we applied Complex Time Difference of\nArrival algorithms and SOTA localization. Our last improvement was to use blind\nsource separation to make our algorithms applicable to multiple microphone\ninputs. The final algorithm achieved state-of-the-art results on numerous\ncheckpoints, including a 95.7\\% accuracy on the ESC-50 dataset for\nenvironmental sound classification.\n", "date": "2023-07-10", "categories": []}
{"id": "2306.14422", "title": "The Singing Voice Conversion Challenge 2023", "abstract": "  We present the latest iteration of the voice conversion challenge (VCC)\nseries, a bi-annual scientific event aiming to compare and understand different\nvoice conversion (VC) systems based on a common dataset. This year we shifted\nour focus to singing voice conversion (SVC), thus named the challenge the\nSinging Voice Conversion Challenge (SVCC). A new database was constructed for\ntwo tasks, namely in-domain and cross-domain SVC. The challenge was run for two\nmonths, and in total we received 26 submissions, including 2 baselines. Through\na large-scale crowd-sourced listening test, we observed that for both tasks,\nalthough human-level naturalness was achieved by the top system, no team was\nable to obtain a similarity score as high as the target speakers. Also, as\nexpected, cross-domain SVC is harder than in-domain SVC, especially in the\nsimilarity aspect. We also investigated whether existing objective measurements\nwere able to predict perceptual performance, and found that only few of them\ncould reach a significant correlation.\n", "date": "2023-06-26", "categories": []}
{"id": "2307.02975", "title": "Transfer Learning for the Efficient Detection of COVID-19 from\n  Smartphone Audio Data", "abstract": "  Disease detection from smartphone data represents an open research challenge\nin mobile health (m-health) systems. COVID-19 and its respiratory symptoms are\nan important case study in this area and their early detection is a potential\nreal instrument to counteract the pandemic situation. The efficacy of this\nsolution mainly depends on the performances of AI algorithms applied to the\ncollected data and their possible implementation directly on the users' mobile\ndevices. Considering these issues, and the limited amount of available data, in\nthis paper we present the experimental evaluation of 3 different deep learning\nmodels, compared also with hand-crafted features, and of two main approaches of\ntransfer learning in the considered scenario: both feature extraction and\nfine-tuning. Specifically, we considered VGGish, YAMNET, and\nL\\textsuperscript{3}-Net (including 12 different configurations) evaluated\nthrough user-independent experiments on 4 different datasets (13,447 samples in\ntotal). Results clearly show the advantages of L\\textsuperscript{3}-Net in all\nthe experimental settings as it overcomes the other solutions by 12.3\\% in\nterms of Precision-Recall AUC as features extractor, and by 10\\% when the model\nis fine-tuned. Moreover, we note that to fine-tune only the fully-connected\nlayers of the pre-trained models generally leads to worse performances, with an\naverage drop of 6.6\\% with respect to feature extraction. %highlighting the\nneed for further investigations. Finally, we evaluate the memory footprints of\nthe different models for their possible applications on commercial mobile\ndevices.\n", "date": "2023-07-06", "categories": []}
{"id": "2305.00417", "title": "Transformer-based Sequence Labeling for Audio Classification based on\n  MFCCs", "abstract": "  Audio classification is vital in areas such as speech and music recognition.\nFeature extraction from the audio signal, such as Mel-Spectrograms and MFCCs,\nis a critical step in audio classification. These features are transformed into\nspectrograms for classification. Researchers have explored various techniques,\nincluding traditional machine and deep learning methods to classify\nspectrograms, but these can be computationally expensive. To simplify this\nprocess, a more straightforward approach inspired by sequence classification in\nNLP can be used. This paper proposes a Transformer-encoder-based model for\naudio classification using MFCCs. The model was benchmarked against the ESC-50,\nSpeech Commands v0.02 and UrbanSound8k datasets and has shown strong\nperformance, with the highest accuracy of 95.2% obtained upon training the\nmodel on the UrbanSound8k dataset. The model consisted of a mere 127,544 total\nparameters, making it light-weight yet highly efficient at the audio\nclassification task.\n", "date": "2023-04-30", "categories": []}
{"id": "2307.02083", "title": "Leveraging multilingual transfer for unsupervised semantic acoustic word\n  embeddings", "abstract": "  Acoustic word embeddings (AWEs) are fixed-dimensional vector representations\nof speech segments that encode phonetic content so that different realisations\nof the same word have similar embeddings. In this paper we explore semantic AWE\nmodelling. These AWEs should not only capture phonetics but also the meaning of\na word (similar to textual word embeddings). We consider the scenario where we\nonly have untranscribed speech in a target language. We introduce a number of\nstrategies leveraging a pre-trained multilingual AWE model -- a phonetic AWE\nmodel trained on labelled data from multiple languages excluding the target.\nOur best semantic AWE approach involves clustering word segments using the\nmultilingual AWE model, deriving soft pseudo-word labels from the cluster\ncentroids, and then training a Skipgram-like model on the soft vectors. In an\nintrinsic word similarity task measuring semantics, this multilingual transfer\napproach outperforms all previous semantic AWE methods. We also show -- for the\nfirst time -- that AWEs can be used for downstream semantic query-by-example\nsearch.\n", "date": "2023-07-05", "categories": []}
{"id": "2211.02133", "title": "Streaming Audio-Visual Speech Recognition with Alignment Regularization", "abstract": "  In this work, we propose a streaming AV-ASR system based on a hybrid\nconnectionist temporal classification (CTC)/attention neural network\narchitecture. The audio and the visual encoder neural networks are both based\non the conformer architecture, which is made streamable using chunk-wise\nself-attention (CSA) and causal convolution. Streaming recognition with a\ndecoder neural network is realized by using the triggered attention technique,\nwhich performs time-synchronous decoding with joint CTC/attention scoring.\nAdditionally, we propose a novel alignment regularization technique that\npromotes synchronization of the audio and visual encoder, which in turn results\nin better word error rates (WERs) at all SNR levels for streaming and offline\nAV-ASR models. The proposed AV-ASR model achieves WERs of 2.0% and 2.6% on the\nLip Reading Sentences 3 (LRS3) dataset in an offline and online setup,\nrespectively, which both present state-of-the-art results when no external\ntraining data are used.\n", "date": "2022-11-03", "categories": []}
{"id": "2306.17203", "title": "Diff-Foley: Synchronized Video-to-Audio Synthesis with Latent Diffusion\n  Models", "abstract": "  The Video-to-Audio (V2A) model has recently gained attention for its\npractical application in generating audio directly from silent videos,\nparticularly in video/film production. However, previous methods in V2A have\nlimited generation quality in terms of temporal synchronization and\naudio-visual relevance. We present Diff-Foley, a synchronized Video-to-Audio\nsynthesis method with a latent diffusion model (LDM) that generates\nhigh-quality audio with improved synchronization and audio-visual relevance. We\nadopt contrastive audio-visual pretraining (CAVP) to learn more temporally and\nsemantically aligned features, then train an LDM with CAVP-aligned visual\nfeatures on spectrogram latent space. The CAVP-aligned features enable LDM to\ncapture the subtler audio-visual correlation via a cross-attention module. We\nfurther significantly improve sample quality with `double guidance'. Diff-Foley\nachieves state-of-the-art V2A performance on current large scale V2A dataset.\nFurthermore, we demonstrate Diff-Foley practical applicability and\ngeneralization capabilities via downstream finetuning. Project Page: see\nhttps://diff-foley.github.io/\n", "date": "2023-06-29", "categories": []}
{"id": "2305.19567", "title": "DC CoMix TTS: An End-to-End Expressive TTS with Discrete Code\n  Collaborated with Mixer", "abstract": "  Despite the huge successes made in neutral TTS, content-leakage remains a\nchallenge. In this paper, we propose a new input representation and simple\narchitecture to achieve improved prosody modeling. Inspired by the recent\nsuccess in the use of discrete code in TTS, we introduce discrete code to the\ninput of the reference encoder. Specifically, we leverage the vector quantizer\nfrom the audio compression model to exploit the diverse acoustic information it\nhas already been trained on. In addition, we apply the modified MLP-Mixer to\nthe reference encoder, making the architecture lighter. As a result, we train\nthe prosody transfer TTS in an end-to-end manner. We prove the effectiveness of\nour method through both subjective and objective evaluations. We demonstrate\nthat the reference encoder learns better speaker-independent prosody when\ndiscrete code is utilized as input in the experiments. In addition, we obtain\ncomparable results even when fewer parameters are inputted.\n", "date": "2023-05-31", "categories": []}
{"id": "2306.16054", "title": "Improving Primate Sounds Classification using Binary Presorting for Deep\n  Learning", "abstract": "  In the field of wildlife observation and conservation, approaches involving\nmachine learning on audio recordings are becoming increasingly popular.\nUnfortunately, available datasets from this field of research are often not\noptimal learning material; Samples can be weakly labeled, of different lengths\nor come with a poor signal-to-noise ratio. In this work, we introduce a\ngeneralized approach that first relabels subsegments of MEL spectrogram\nrepresentations, to achieve higher performances on the actual multi-class\nclassification tasks. For both the binary pre-sorting and the classification,\nwe make use of convolutional neural networks (CNN) and various\ndata-augmentation techniques. We showcase the results of this approach on the\nchallenging \\textit{ComparE 2021} dataset, with the task of classifying between\ndifferent primate species sounds, and report significantly higher Accuracy and\nUAR scores in contrast to comparatively equipped model baselines.\n", "date": "2023-06-28", "categories": []}
{"id": "2306.15212", "title": "TranssionADD: A multi-frame reinforcement based sequence tagging model\n  for audio deepfake detection", "abstract": "  Thanks to recent advancements in end-to-end speech modeling technology, it\nhas become increasingly feasible to imitate and clone a user`s voice. This\nleads to a significant challenge in differentiating between authentic and\nfabricated audio segments. To address the issue of user voice abuse and misuse,\nthe second Audio Deepfake Detection Challenge (ADD 2023) aims to detect and\nanalyze deepfake speech utterances. Specifically, Track 2, named the\nManipulation Region Location (RL), aims to pinpoint the location of manipulated\nregions in audio, which can be present in both real and generated audio\nsegments. We propose our novel TranssionADD system as a solution to the\nchallenging problem of model robustness and audio segment outliers in the trace\ncompetition. Our system provides three unique contributions: 1) we adapt\nsequence tagging task for audio deepfake detection; 2) we improve model\ngeneralization by various data augmentation techniques; 3) we incorporate\nmulti-frame detection (MFD) module to overcome limited representation provided\nby a single frame and use isolated-frame penalty (IFP) loss to handle outliers\nin segments. Our best submission achieved 2nd place in Track 2, demonstrating\nthe effectiveness and robustness of our proposed system.\n", "date": "2023-06-27", "categories": []}
{"id": "2304.06868", "title": "Tempo vs. Pitch: understanding self-supervised tempo estimation", "abstract": "  Self-supervision methods learn representations by solving pretext tasks that\ndo not require human-generated labels, alleviating the need for time-consuming\nannotations. These methods have been applied in computer vision, natural\nlanguage processing, environmental sound analysis, and recently in music\ninformation retrieval, e.g. for pitch estimation. Particularly in the context\nof music, there are few insights about the fragility of these models regarding\ndifferent distributions of data, and how they could be mitigated. In this\npaper, we explore these questions by dissecting a self-supervised model for\npitch estimation adapted for tempo estimation via rigorous experimentation with\nsynthetic data. Specifically, we study the relationship between the input\nrepresentation and data distribution for self-supervised tempo estimation.\n", "date": "2023-04-14", "categories": []}
{"id": "2306.14191", "title": "PrimaDNN': A Characteristics-aware DNN Customization for Singing\n  Technique Detection", "abstract": "  Professional vocalists modulate their voice timbre or pitch to make their\nvocal performance more expressive. Such fluctuations are called singing\ntechniques. Automatic detection of singing techniques from audio tracks can be\nbeneficial to understand how each singer expresses the performance, yet it can\nalso be difficult due to the wide variety of the singing techniques. A deep\nneural network (DNN) model can handle such variety; however, there might be a\npossibility that considering the characteristics of the data improves the\nperformance of singing technique detection. In this paper, we propose PrimaDNN,\na CRNN model with a characteristics-oriented improvement. The features of the\nmodel are: 1) input feature representation based on auxiliary pitch information\nand multi-resolution mel spectrograms, 2) Convolution module based on the\nSqueeze-and-excitation (SENet) and the Instance normalization. In the results\nof J-POP singing technique detection, PrimaDNN achieved the best results of\n44.9% at the overall macro-F measure, compared to conventional works. We also\nfound that the contribution of each component varies depending on the type of\nsinging technique.\n", "date": "2023-06-25", "categories": []}
{"id": "2306.14647", "title": "Mono-to-stereo through parametric stereo generation", "abstract": "  Generating a stereophonic presentation from a monophonic audio signal is a\nchallenging open task, especially if the goal is to obtain a realistic spatial\nimaging with a specific panning of sound elements. In this work, we propose to\nconvert mono to stereo by means of predicting parametric stereo (PS) parameters\nusing both nearest neighbor and deep network approaches. In combination with\nPS, we also propose to model the task with generative approaches, allowing to\nsynthesize multiple and equally-plausible stereo renditions from the same mono\nsignal. To achieve this, we consider both autoregressive and masked token\nmodelling approaches. We provide evidence that the proposed PS-based models\noutperform a competitive classical decorrelation baseline and that, within a PS\nprediction framework, modern generative models outshine equivalent\nnon-generative counterparts. Overall, our work positions both PS and generative\nmodelling as strong and appealing methodologies for mono-to-stereo upmixing. A\ndiscussion of the limitations of these approaches is also provided.\n", "date": "2023-06-26", "categories": []}
{"id": "2211.04772", "title": "Efficient Large-scale Audio Tagging via Transformer-to-CNN Knowledge\n  Distillation", "abstract": "  Audio Spectrogram Transformer models rule the field of Audio Tagging,\noutrunning previously dominating Convolutional Neural Networks (CNNs). Their\nsuperiority is based on the ability to scale up and exploit large-scale\ndatasets such as AudioSet. However, Transformers are demanding in terms of\nmodel size and computational requirements compared to CNNs. We propose a\ntraining procedure for efficient CNNs based on offline Knowledge Distillation\n(KD) from high-performing yet complex transformers. The proposed training\nschema and the efficient CNN design based on MobileNetV3 results in models\noutperforming previous solutions in terms of parameter and computational\nefficiency and prediction performance. We provide models of different\ncomplexity levels, scaling from low-complexity models up to a new\nstate-of-the-art performance of .483 mAP on AudioSet. Source Code available at:\nhttps://github.com/fschmid56/EfficientAT\n", "date": "2022-11-09", "categories": []}
{"id": "2306.12443", "title": "DEPAC: a Corpus for Depression and Anxiety Detection from Speech", "abstract": "  Mental distress like depression and anxiety contribute to the largest\nproportion of the global burden of diseases. Automated diagnosis systems of\nsuch disorders, empowered by recent innovations in Artificial Intelligence, can\npave the way to reduce the sufferings of the affected individuals. Development\nof such systems requires information-rich and balanced corpora. In this work,\nwe introduce a novel mental distress analysis audio dataset DEPAC, labeled\nbased on established thresholds on depression and anxiety standard screening\ntools. This large dataset comprises multiple speech tasks per individual, as\nwell as relevant demographic information. Alongside, we present a feature set\nconsisting of hand-curated acoustic and linguistic features, which were found\neffective in identifying signs of mental illnesses in human speech. Finally, we\njustify the quality and effectiveness of our proposed audio corpus and feature\nset in predicting depression severity by comparing the performance of baseline\nmachine learning models built on this dataset with baseline models trained on\nother well-known depression corpora.\n", "date": "2023-06-20", "categories": []}
{"id": "2306.12957", "title": "Siamese SIREN: Audio Compression with Implicit Neural Representations", "abstract": "  Implicit Neural Representations (INRs) have emerged as a promising method for\nrepresenting diverse data modalities, including 3D shapes, images, and audio.\nWhile recent research has demonstrated successful applications of INRs in image\nand 3D shape compression, their potential for audio compression remains largely\nunexplored. Motivated by this, we present a preliminary investigation into the\nuse of INRs for audio compression. Our study introduces Siamese SIREN, a novel\napproach based on the popular SIREN architecture. Our experimental results\nindicate that Siamese SIREN achieves superior audio reconstruction fidelity\nwhile utilizing fewer network parameters compared to previous INR\narchitectures.\n", "date": "2023-06-22", "categories": []}
{"id": "2306.11764", "title": "On Frequency-Wise Normalizations for Better Recording Device\n  Generalization in Audio Spectrogram Transformers", "abstract": "  Varying conditions between the data seen at training and at application time\nremain a major challenge for machine learning. We study this problem in the\ncontext of Acoustic Scene Classification (ASC) with mismatching recording\ndevices. Previous works successfully employed frequency-wise normalization of\ninputs and hidden layer activations in convolutional neural networks to reduce\nthe recording device discrepancy. The main objective of this work was to adopt\nfrequency-wise normalization for Audio Spectrogram Transformers (ASTs), which\nhave recently become the dominant model architecture in ASC. To this end, we\nfirst investigate how recording device characteristics are encoded in the\nhidden layer activations of ASTs. We find that recording device information is\ninitially encoded in the frequency dimension; however, after the first\nself-attention block, it is largely transformed into the token dimension. Based\non this observation, we conjecture that suppressing recording device\ncharacteristics in the input spectrogram is the most effective. We propose a\nfrequency-centering operation for spectrograms that improves the ASC\nperformance on unseen recording devices on average by up to 18.2 percentage\npoints.\n", "date": "2023-06-20", "categories": []}
{"id": "2302.04959", "title": "Hypernetworks build Implicit Neural Representations of Sounds", "abstract": "  Implicit Neural Representations (INRs) are nowadays used to represent\nmultimedia signals across various real-life applications, including image\nsuper-resolution, image compression, or 3D rendering. Existing methods that\nleverage INRs are predominantly focused on visual data, as their application to\nother modalities, such as audio, is nontrivial due to the inductive biases\npresent in architectural attributes of image-based INR models. To address this\nlimitation, we introduce HyperSound, the first meta-learning approach to\nproduce INRs for audio samples that leverages hypernetworks to generalize\nbeyond samples observed in training. Our approach reconstructs audio samples\nwith quality comparable to other state-of-the-art models and provides a viable\nalternative to contemporary sound representations used in deep neural networks\nfor audio processing, such as spectrograms.\n", "date": "2023-02-09", "categories": []}
{"id": "2303.04104", "title": "An Inception-Residual-Based Architecture with Multi-Objective Loss for\n  Detecting Respiratory Anomalies", "abstract": "  This paper presents a deep learning system applied for detecting anomalies\nfrom respiratory sound recordings. Initially, our system begins with audio\nfeature extraction using Gammatone and Continuous Wavelet transformation. This\nstep aims to transform the respiratory sound input into a two-dimensional\nspectrogram where both spectral and temporal features are presented. Then, our\nproposed system integrates Inception-residual-based backbone models combined\nwith multi-head attention and multi-objective loss to classify respiratory\nanomalies. Instead of applying a simple concatenation approach by combining\nresults from various spectrograms, we propose a Linear combination, which has\nthe ability to regulate equally the contribution of each individual spectrogram\nthroughout the training process. To evaluate the performance, we conducted\nexperiments over the benchmark dataset of SPRSound (The Open-Source SJTU\nPaediatric Respiratory Sound) proposed by the IEEE BioCAS 2022 challenge. As\nregards the Score computed by an average between the average score and harmonic\nscore, our proposed system gained significant improvements of 9.7%, 15.8%,\n17.8%, and 16.1% in Task 1-1, Task 1-2, Task 2-1, and Task 2-2, respectively,\ncompared to the challenge baseline system. Notably, we achieved the Top-1\nperformance in Task 2-1 and Task 2-2 with the highest Score of 74.5% and 53.9%,\nrespectively.\n", "date": "2023-03-07", "categories": []}
{"id": "2305.15571", "title": "Sound Design Strategies for Latent Audio Space Explorations Using Deep\n  Learning Architectures", "abstract": "  The research in Deep Learning applications in sound and music computing have\ngathered an interest in the recent years; however, there is still a missing\nlink between these new technologies and on how they can be incorporated into\nreal-world artistic practices. In this work, we explore a well-known Deep\nLearning architecture called Variational Autoencoders (VAEs). These\narchitectures have been used in many areas for generating latent spaces where\ndata points are organized so that similar data points locate closer to each\nother. Previously, VAEs have been used for generating latent timbre spaces or\nlatent spaces of symbolic music excepts. Applying VAE to audio features of\ntimbre requires a vocoder to transform the timbre generated by the network to\nan audio signal, which is computationally expensive. In this work, we apply\nVAEs to raw audio data directly while bypassing audio feature extraction. This\napproach allows the practitioners to use any audio recording while giving\nflexibility and control over the aesthetics through dataset curation. The lower\ncomputation time in audio signal generation allows the raw audio approach to be\nincorporated into real-time applications. In this work, we propose three\nstrategies to explore latent spaces of audio and timbre for sound design\napplications. By doing so, our aim is to initiate a conversation on artistic\napproaches and strategies to utilize latent audio spaces in sound and music\npractices.\n", "date": "2023-05-24", "categories": []}
{"id": "2306.09906", "title": "Correlation Clustering of Bird Sounds", "abstract": "  Bird sound classification is the task of relating any sound recording to\nthose species of bird that can be heard in the recording. Here, we study bird\nsound clustering, the task of deciding for any pair of sound recordings whether\nthe same species of bird can be heard in both. We address this problem by first\nlearning, from a training set, probabilities of pairs of recordings being\nrelated in this way, and then inferring a maximally probable partition of a\ntest set by correlation clustering. We address the following questions: How\naccurate is this clustering, compared to a classification of the test set? How\ndo the clusters thus inferred relate to the clusters obtained by\nclassification? How accurate is this clustering when applied to recordings of\nbird species not heard during training? How effective is this clustering in\nseparating, from bird sounds, environmental noise not heard during training?\n", "date": "2023-06-16", "categories": []}
{"id": "2306.10684", "title": "Visually-Guided Sound Source Separation with Audio-Visual Predictive\n  Coding", "abstract": "  The framework of visually-guided sound source separation generally consists\nof three parts: visual feature extraction, multimodal feature fusion, and sound\nsignal processing. An ongoing trend in this field has been to tailor involved\nvisual feature extractor for informative visual guidance and separately devise\nmodule for feature fusion, while utilizing U-Net by default for sound analysis.\nHowever, such divide-and-conquer paradigm is parameter inefficient and,\nmeanwhile, may obtain suboptimal performance as jointly optimizing and\nharmonizing various model components is challengeable. By contrast, this paper\npresents a novel approach, dubbed audio-visual predictive coding (AVPC), to\ntackle this task in a parameter efficient and more effective manner. The\nnetwork of AVPC features a simple ResNet-based video analysis network for\nderiving semantic visual features, and a predictive coding-based sound\nseparation network that can extract audio features, fuse multimodal\ninformation, and predict sound separation masks in the same architecture. By\niteratively minimizing the prediction error between features, AVPC integrates\naudio and visual information recursively, leading to progressively improved\nperformance. In addition, we develop a valid self-supervised learning strategy\nfor AVPC via co-predicting two audio-visual representations of the same sound\nsource. Extensive evaluations demonstrate that AVPC outperforms several\nbaselines in separating musical instrument sounds, while reducing the model\nsize significantly. Code is available at:\nhttps://github.com/zjsong/Audio-Visual-Predictive-Coding.\n", "date": "2023-06-19", "categories": []}
{"id": "2306.09944", "title": "RealImpact: A Dataset of Impact Sound Fields for Real Objects", "abstract": "  Objects make unique sounds under different perturbations, environment\nconditions, and poses relative to the listener. While prior works have modeled\nimpact sounds and sound propagation in simulation, we lack a standard dataset\nof impact sound fields of real objects for audio-visual learning and\ncalibration of the sim-to-real gap. We present RealImpact, a large-scale\ndataset of real object impact sounds recorded under controlled conditions.\nRealImpact contains 150,000 recordings of impact sounds of 50 everyday objects\nwith detailed annotations, including their impact locations, microphone\nlocations, contact force profiles, material labels, and RGBD images. We make\npreliminary attempts to use our dataset as a reference to current simulation\nmethods for estimating object impact sounds that match the real world.\nMoreover, we demonstrate the usefulness of our dataset as a testbed for\nacoustic and audio-visual learning via the evaluation of two benchmark tasks,\nincluding listener location classification and visual acoustic matching.\n", "date": "2023-06-16", "categories": []}
{"id": "2306.09979", "title": "Evaluation of Speech Representations for MOS prediction", "abstract": "  In this paper, we evaluate feature extraction models for predicting speech\nquality. We also propose a model architecture to compare embeddings of\nsupervised learning and self-supervised learning models with embeddings of\nspeaker verification models to predict the metric MOS. Our experiments were\nperformed on the VCC2018 dataset and a Brazilian-Portuguese dataset called\nBRSpeechMOS, which was created for this work. The results show that the Whisper\nmodel is appropriate in all scenarios: with both the VCC2018 and BRSpeech- MOS\ndatasets. Among the supervised and self-supervised learning models using\nBRSpeechMOS, Whisper-Small achieved the best linear correlation of 0.6980, and\nthe speaker verification model, SpeakerNet, had linear correlation of 0.6963.\nUsing VCC2018, the best supervised and self-supervised learning model,\nWhisper-Large, achieved linear correlation of 0.7274, and the best model\nspeaker verification, TitaNet, achieved a linear correlation of 0.6933.\nAlthough the results of the speaker verification models are slightly lower, the\nSpeakerNet model has only 5M parameters, making it suitable for real-time\napplications, and the TitaNet model produces an embedding of size 192, the\nsmallest among all the evaluated models. The experiment results are\nreproducible with publicly available source-code1 .\n", "date": "2023-06-16", "categories": []}
{"id": "2306.04956", "title": "Adaptive Fake Audio Detection with Low-Rank Model Squeezing", "abstract": "  The rapid advancement of spoofing algorithms necessitates the development of\nrobust detection methods capable of accurately identifying emerging fake audio.\nTraditional approaches, such as finetuning on new datasets containing these\nnovel spoofing algorithms, are computationally intensive and pose a risk of\nimpairing the acquired knowledge of known fake audio types. To address these\nchallenges, this paper proposes an innovative approach that mitigates the\nlimitations associated with finetuning. We introduce the concept of training\nlow-rank adaptation matrices tailored specifically to the newly emerging fake\naudio types. During the inference stage, these adaptation matrices are combined\nwith the existing model to generate the final prediction output. Extensive\nexperimentation is conducted to evaluate the efficacy of the proposed method.\nThe results demonstrate that our approach effectively preserves the prediction\naccuracy of the existing model for known fake audio types. Furthermore, our\napproach offers several advantages, including reduced storage memory\nrequirements and lower equal error rates compared to conventional finetuning\nmethods, particularly on specific spoofing algorithms.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.08010", "title": "Domain Information Control at Inference Time for Acoustic Scene\n  Classification", "abstract": "  Domain shift is considered a challenge in machine learning as it causes\nsignificant degradation of model performance. In the Acoustic Scene\nClassification task (ASC), domain shift is mainly caused by different recording\ndevices. Several studies have already targeted domain generalization to improve\nthe performance of ASC models on unseen domains, such as new devices. Recently,\nthe Controllable Gate Adapter ConGater has been proposed in Natural Language\nProcessing to address the biased training data problem. ConGater allows\ncontrolling the debiasing process at inference time. ConGater's main advantage\nis the continuous and selective debiasing of a trained model, during inference.\nIn this work, we adapt ConGater to the audio spectrogram transformer for an\nacoustic scene classification task. We show that ConGater can be used to\nselectively adapt the learned representations to be invariant to device domain\nshifts such as recording devices. Our analysis shows that ConGater can\nprogressively remove device information from the learned representations and\nimprove the model generalization, especially under domain shift conditions\n(e.g. unseen devices). We show that information removal can be extended to both\ndevice and location domain. Finally, we demonstrate ConGater's ability to\nenhance specific device performance without further training.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.08012", "title": "A Novel Scheme to classify Read and Spontaneous Speech", "abstract": "  The COVID-19 pandemic has led to an increased use of remote telephonic\ninterviews, making it important to distinguish between scripted and spontaneous\nspeech in audio recordings. In this paper, we propose a novel scheme for\nidentifying read and spontaneous speech. Our approach uses a pre-trained\nDeepSpeech audio-to-alphabet recognition engine to generate a sequence of\nalphabets from the audio. From these alphabets, we derive features that allow\nus to discriminate between read and spontaneous speech. Our experimental\nresults show that even a small set of self-explanatory features can effectively\nclassify the two types of speech very effectively.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.08798", "title": "MPSA-DenseNet: A novel deep learning model for English accent\n  classification", "abstract": "  This paper presents three innovative deep learning models for English accent\nclassification: Multi-DenseNet, PSA-DenseNet, and MPSE-DenseNet, that combine\nmulti-task learning and the PSA module attention mechanism with DenseNet. We\napplied these models to data collected from six dialects of English across\nnative English speaking regions (Britain, the United States, Scotland) and\nnonnative English speaking regions (China, Germany, India). Our experimental\nresults show a significant improvement in classification accuracy, particularly\nwith MPSA-DenseNet, which outperforms all other models, including DenseNet and\nEPSA models previously used for accent identification. Our findings indicate\nthat MPSA-DenseNet is a highly promising model for accurately identifying\nEnglish accents.\n", "date": "2023-06-15", "categories": []}
{"id": "2306.08998", "title": "Team AcieLee: Technical Report for EPIC-SOUNDS Audio-Based Interaction\n  Recognition Challenge 2023", "abstract": "  In this report, we describe the technical details of our submission to the\nEPIC-SOUNDS Audio-Based Interaction Recognition Challenge 2023, by Team\n\"AcieLee\" (username: Yuqi\\_Li). The task is to classify the audio caused by\ninteractions between objects, or from events of the camera wearer. We conducted\nexhaustive experiments and found learning rate step decay, backbone frozen,\nlabel smoothing and focal loss contribute most to the performance improvement.\nAfter training, we combined multiple models from different stages and\nintegrated them into a single model by assigning fusion weights. This proposed\nmethod allowed us to achieve 3rd place in the CVPR 2023 workshop of EPIC-SOUNDS\nAudio-Based Interaction Recognition Challenge.\n", "date": "2023-06-15", "categories": []}
{"id": "2306.09025", "title": "CoverHunter: Cover Song Identification with Refined Attention and\n  Alignments", "abstract": "  Abstract: Cover song identification (CSI) focuses on finding the same music\nwith different versions in reference anchors given a query track. In this\npaper, we propose a novel system named CoverHunter that overcomes the\nshortcomings of existing detection schemes by exploring richer features with\nrefined attention and alignments. CoverHunter contains three key modules: 1) A\nconvolution-augmented transformer (i.e., Conformer) structure that captures\nboth local and global feature interactions in contrast to previous methods\nmainly relying on convolutional neural networks; 2) An attention-based time\npooling module that further exploits the attention in the time dimension; 3) A\nnovel coarse-to-fine training scheme that first trains a network to roughly\nalign the song chunks and then refines the network by training on the aligned\nchunks. At the same time, we also summarize some important training tricks used\nin our system that help achieve better results. Experiments on several standard\nCSI datasets show that our method significantly improves over state-of-the-art\nmethods with an embedding size of 128 (2.3% on SHS100K-TEST and 17.7% on\nDaTacos).\n", "date": "2023-06-15", "categories": []}
{"id": "2306.09223", "title": "Few-shot bioacoustic event detection at the DCASE 2023 challenge", "abstract": "  Few-shot bioacoustic event detection consists in detecting sound events of\nspecified types, in varying soundscapes, while having access to only a few\nexamples of the class of interest. This task ran as part of the DCASE challenge\nfor the third time this year with an evaluation set expanded to include new\nanimal species, and a new rule: ensemble models were no longer allowed. The\n2023 few shot task received submissions from 6 different teams with F-scores\nreaching as high as 63% on the evaluation set. Here we describe the task,\nfocusing on describing the elements that differed from previous years. We also\ntake a look back at past editions to describe how the task has evolved. Not\nonly have the F-score results steadily improved (40% to 60% to 63%), but the\ntype of systems proposed have also become more complex. Sound event detection\nsystems are no longer simple variations of the baselines provided: multiple\nfew-shot learning methodologies are still strong contenders for the task.\n", "date": "2023-06-15", "categories": []}
{"id": "2306.07744", "title": "Contrastive Learning-Based Audio to Lyrics Alignment for Multiple\n  Languages", "abstract": "  Lyrics alignment gained considerable attention in recent years.\nState-of-the-art systems either re-use established speech recognition toolkits,\nor design end-to-end solutions involving a Connectionist Temporal\nClassification (CTC) loss. However, both approaches suffer from specific\nweaknesses: toolkits are known for their complexity, and CTC systems use a loss\ndesigned for transcription which can limit alignment accuracy. In this paper,\nwe use instead a contrastive learning procedure that derives cross-modal\nembeddings linking the audio and text domains. This way, we obtain a novel\nsystem that is simple to train end-to-end, can make use of weakly annotated\ntraining data, jointly learns a powerful text model, and is tailored to\nalignment. The system is not only the first to yield an average absolute error\nbelow 0.2 seconds on the standard Jamendo dataset but it is also robust to\nother languages, even when trained on English data only. Finally, we release\nword-level alignments for the JamendoLyrics Multi-Lang dataset.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.07820", "title": "Unsupervised speech enhancement with deep dynamical generative speech\n  and noise models", "abstract": "  This work builds on a previous work on unsupervised speech enhancement using\na dynamical variational autoencoder (DVAE) as the clean speech model and\nnon-negative matrix factorization (NMF) as the noise model. We propose to\nreplace the NMF noise model with a deep dynamical generative model (DDGM)\ndepending either on the DVAE latent variables, or on the noisy observations, or\non both. This DDGM can be trained in three configurations: noise-agnostic,\nnoise-dependent and noise adaptation after noise-dependent training.\nExperimental results show that the proposed method achieves competitive\nperformance compared to state-of-the-art unsupervised speech enhancement\nmethods, while the noise-dependent training configuration yields a much more\ntime-efficient inference process.\n", "date": "2023-06-13", "categories": []}
{"id": "2212.14597", "title": "Defense Against Adversarial Attacks on Audio DeepFake Detection", "abstract": "  Audio DeepFakes (DF) are artificially generated utterances created using deep\nlearning, with the primary aim of fooling the listeners in a highly convincing\nmanner. Their quality is sufficient to pose a severe threat in terms of\nsecurity and privacy, including the reliability of news or defamation. Multiple\nneural network-based methods to detect generated speech have been proposed to\nprevent the threats. In this work, we cover the topic of adversarial attacks,\nwhich decrease the performance of detectors by adding superficial (difficult to\nspot by a human) changes to input data. Our contribution contains evaluating\nthe robustness of 3 detection architectures against adversarial attacks in two\nscenarios (white-box and using transferability) and enhancing it later by using\nadversarial training performed by our novel adaptive training. Moreover, one of\nthe investigated architectures is RawNet3, which, to the best of our knowledge,\nwe adapted for the first time to DeepFake detection.\n", "date": "2022-12-30", "categories": []}
{"id": "2306.05617", "title": "Low-rank Adaptation Method for Wav2vec2-based Fake Audio Detection", "abstract": "  Self-supervised speech models are a rapidly developing research topic in fake\naudio detection. Many pre-trained models can serve as feature extractors,\nlearning richer and higher-level speech features. However,when fine-tuning\npre-trained models, there is often a challenge of excessively long training\ntimes and high memory consumption, and complete fine-tuning is also very\nexpensive. To alleviate this problem, we apply low-rank adaptation(LoRA) to the\nwav2vec2 model, freezing the pre-trained model weights and injecting a\ntrainable rank-decomposition matrix into each layer of the transformer\narchitecture, greatly reducing the number of trainable parameters for\ndownstream tasks. Compared with fine-tuning with Adam on the wav2vec2 model\ncontaining 317M training parameters, LoRA achieved similar performance by\nreducing the number of trainable parameters by 198 times.\n", "date": "2023-06-09", "categories": []}
{"id": "2306.05915", "title": "Speaker Embeddings as Individuality Proxy for Voice Stress Detection", "abstract": "  Since the mental states of the speaker modulate speech, stress introduced by\ncognitive or physical loads could be detected in the voice. The existing voice\nstress detection benchmark has shown that the audio embeddings extracted from\nthe Hybrid BYOL-S self-supervised model perform well. However, the benchmark\nonly evaluates performance separately on each dataset, but does not evaluate\nperformance across the different types of stress and different languages.\nMoreover, previous studies found strong individual differences in stress\nsusceptibility. This paper presents the design and development of voice stress\ndetection, trained on more than 100 speakers from 9 language groups and five\ndifferent types of stress. We address individual variabilities in voice stress\nanalysis by adding speaker embeddings to the hybrid BYOL-S features. The\nproposed method significantly improves voice stress detection performance with\nan input audio length of only 3-5 seconds.\n", "date": "2023-06-09", "categories": []}
{"id": "2306.05004", "title": "VIFS: An End-to-End Variational Inference for Foley Sound Synthesis", "abstract": "  The goal of DCASE 2023 Challenge Task 7 is to generate various sound clips\nfor Foley sound synthesis (FSS) by \"category-to-sound\" approach. \"Category\" is\nexpressed by a single index while corresponding \"sound\" covers diverse and\ndifferent sound examples. To generate diverse sounds for a given category, we\nadopt VITS, a text-to-speech (TTS) model with variational inference. In\naddition, we apply various techniques from speech synthesis including PhaseAug\nand Avocodo. Different from TTS models which generate short pronunciation from\nphonemes and speaker identity, the category-to-sound problem requires\ngenerating diverse sounds just from a category index. To compensate for the\ndifference while maintaining consistency within each audio clip, we heavily\nmodified the prior encoder to enhance consistency with posterior latent\nvariables. This introduced additional Gaussian on the prior encoder which\npromotes variance within the category. With these modifications, we propose\nVIFS, variational inference for end-to-end Foley sound synthesis, which\ngenerates diverse high-quality sounds.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.05245", "title": "Matching Latent Encoding for Audio-Text based Keyword Spotting", "abstract": "  Using audio and text embeddings jointly for Keyword Spotting (KWS) has shown\nhigh-quality results, but the key challenge of how to semantically align two\nembeddings for multi-word keywords of different sequence lengths remains\nlargely unsolved. In this paper, we propose an audio-text-based end-to-end\nmodel architecture for flexible keyword spotting (KWS), which builds upon\nlearned audio and text embeddings. Our architecture uses a novel dynamic\nprogramming-based algorithm, Dynamic Sequence Partitioning (DSP), to optimally\npartition the audio sequence into the same length as the word-based text\nsequence using the monotonic alignment of spoken content. Our proposed model\nconsists of an encoder block to get audio and text embeddings, a projector\nblock to project individual embeddings to a common latent space, and an\naudio-text aligner containing a novel DSP algorithm, which aligns the audio and\ntext embeddings to determine if the spoken content is the same as the text.\nExperimental results show that our DSP is more effective than other\npartitioning schemes, and the proposed architecture outperformed the\nstate-of-the-art results on the public dataset in terms of Area Under the ROC\nCurve (AUC) and Equal-Error-Rate (EER) by 14.4 % and 28.9%, respectively.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.04268", "title": "Multi-microphone Automatic Speech Segmentation in Meetings Based on\n  Circular Harmonics Features", "abstract": "  Speaker diarization is the task of answering Who spoke and when? in an audio\nstream. Pipeline systems rely on speech segmentation to extract speakers'\nsegments and achieve robust speaker diarization. This paper proposes a common\nframework to solve three segmentation tasks in the distant speech scenario:\nVoice Activity Detection (VAD), Overlapped Speech Detection (OSD), and Speaker\nChange Detection (SCD). In the literature, a few studies investigate the\nmulti-microphone distant speech scenario. In this work, we propose a new set of\nspatial features based on direction-of-arrival estimations in the circular\nharmonic domain (CH-DOA). These spatial features are extracted from\nmulti-microphone audio data and combined with standard acoustic features.\nExperiments on the AMI meeting corpus show that CH-DOA can improve the\nsegmentation while being robust in the case of deactivated microphones.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.04286", "title": "A Mask Free Neural Network for Monaural Speech Enhancement", "abstract": "  In speech enhancement, the lack of clear structural characteristics in the\ntarget speech phase requires the use of conservative and cumbersome network\nframeworks. It seems difficult to achieve competitive performance using direct\nmethods and simple network architectures. However, we propose the MFNet, a\ndirect and simple network that can not only map speech but also map reverse\nnoise. This network is constructed by stacking global local former blocks\n(GLFBs), which combine the advantages of Mobileblock for global processing and\nMetaformer architecture for local interaction. Our experimental results\ndemonstrate that our network using mapping method outperforms masking methods,\nand direct mapping of reverse noise is the optimal solution in strong noise\nenvironments. In a horizontal comparison on the 2020 Deep Noise Suppression\n(DNS) challenge test set without reverberation, to the best of our knowledge,\nMFNet is the current state-of-the-art (SOTA) mapping model.\n", "date": "2023-06-07", "categories": []}
{"id": "2210.14536", "title": "Position tracking of a varying number of sound sources with sliding\n  permutation invariant training", "abstract": "  Recent data- and learning-based sound source localization (SSL) methods have\nshown strong performance in challenging acoustic scenarios. However, little\nwork has been done on adapting such methods to track consistently multiple\nsources appearing and disappearing, as would occur in reality. In this paper,\nwe present a new training strategy for deep learning SSL models with a\nstraightforward implementation based on the mean squared error of the optimal\nassociation between estimated and reference positions in the preceding time\nframes. It optimizes the desired properties of a tracking system: handling a\ntime-varying number of sources and ordering localization estimates according to\ntheir trajectories, minimizing identity switches (IDSs). Evaluation on\nsimulated data of multiple reverberant moving sources and on two model\narchitectures proves its effectiveness on reducing identity switches without\ncompromising frame-wise localization accuracy.\n", "date": "2022-10-26", "categories": []}
{"id": "2306.01864", "title": "Discovering COVID-19 Coughing and Breathing Patterns from Unlabeled Data\n  Using Contrastive Learning with Varying Pre-Training Domains", "abstract": "  Rapid discovery of new diseases, such as COVID-19 can enable a timely\nepidemic response, preventing the large-scale spread and protecting public\nhealth. However, limited research efforts have been taken on this problem. In\nthis paper, we propose a contrastive learning-based modeling approach for\nCOVID-19 coughing and breathing pattern discovery from non-COVID coughs. To\nvalidate our models, extensive experiments have been conducted using four large\naudio datasets and one image dataset. We further explore the effects of\ndifferent factors, such as domain relevance and augmentation order on the\npre-trained models. Our results show that the proposed model can effectively\ndistinguish COVID-19 coughing and breathing from unlabeled data and labeled\nnon-COVID coughs with an accuracy of up to 0.81 and 0.86, respectively.\nFindings from this work will guide future research to detect an outbreak of a\nnew disease early.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.02902", "title": "N-Shot Benchmarking of Whisper on Diverse Arabic Speech Recognition", "abstract": "  Whisper, the recently developed multilingual weakly supervised model, is\nreported to perform well on multiple speech recognition benchmarks in both\nmonolingual and multilingual settings. However, it is not clear how Whisper\nwould fare under diverse conditions even on languages it was evaluated on such\nas Arabic. In this work, we address this gap by comprehensively evaluating\nWhisper on several varieties of Arabic speech for the ASR task. Our evaluation\ncovers most publicly available Arabic speech data and is performed under n-shot\n(zero-, few-, and full) finetuning. We also investigate the robustness of\nWhisper under completely novel conditions, such as in dialect-accented standard\nArabic and in unseen dialects for which we develop evaluation data. Our\nexperiments show that although Whisper zero-shot outperforms fully finetuned\nXLS-R models on all datasets, its performance deteriorates significantly in the\nzero-shot setting for five unseen dialects (i.e., Algeria, Jordan, Palestine,\nUAE, and Yemen).\n", "date": "2023-06-05", "categories": []}
{"id": "2306.01332", "title": "Differentiable Grey-box Modelling of Phaser Effects using Frame-based\n  Spectral Processing", "abstract": "  Machine learning approaches to modelling analog audio effects have seen\nintensive investigation in recent years, particularly in the context of\nnon-linear time-invariant effects such as guitar amplifiers. For modulation\neffects such as phasers, however, new challenges emerge due to the presence of\nthe low-frequency oscillator which controls the slowly time-varying nature of\nthe effect. Existing approaches have either required foreknowledge of this\ncontrol signal, or have been non-causal in implementation. This work presents a\ndifferentiable digital signal processing approach to modelling phaser effects\nin which the underlying control signal and time-varying spectral response of\nthe effect are jointly learned. The proposed model processes audio in short\nframes to implement a time-varying filter in the frequency domain, with a\ntransfer function based on typical analog phaser circuit topology. We show that\nthe model can be trained to emulate an analog reference device, while retaining\ninterpretable and adjustable parameters. The frame duration is an important\nhyper-parameter of the proposed model, so an investigation was carried out into\nits effect on model accuracy. The optimal frame length depends on both the rate\nand transient decay-time of the target effect, but the frame length can be\naltered at inference time without a significant change in accuracy.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.01428", "title": "Improved DeepFake Detection Using Whisper Features", "abstract": "  With a recent influx of voice generation methods, the threat introduced by\naudio DeepFake (DF) is ever-increasing. Several different detection methods\nhave been presented as a countermeasure. Many methods are based on so-called\nfront-ends, which, by transforming the raw audio, emphasize features crucial\nfor assessing the genuineness of the audio sample. Our contribution contains\ninvestigating the influence of the state-of-the-art Whisper automatic speech\nrecognition model as a DF detection front-end. We compare various combinations\nof Whisper and well-established front-ends by training 3 detection models\n(LCNN, SpecRNet, and MesoNet) on a widely used ASVspoof 2021 DF dataset and\nlater evaluating them on the DF In-The-Wild dataset. We show that using\nWhisper-based features improves the detection for each model and outperforms\nrecent results on the In-The-Wild dataset by reducing Equal Error Rate by 21%.\n", "date": "2023-06-02", "categories": []}
{"id": "2210.06324", "title": "SQuId: Measuring Speech Naturalness in Many Languages", "abstract": "  Much of text-to-speech research relies on human evaluation, which incurs\nheavy costs and slows down the development process. The problem is particularly\nacute in heavily multilingual applications, where recruiting and polling judges\ncan take weeks. We introduce SQuId (Speech Quality Identification), a\nmultilingual naturalness prediction model trained on over a million ratings and\ntested in 65 locales-the largest effort of this type to date. The main insight\nis that training one model on many locales consistently outperforms mono-locale\nbaselines. We present our task, the model, and show that it outperforms a\ncompetitive baseline based on w2v-BERT and VoiceMOS by 50.0%. We then\ndemonstrate the effectiveness of cross-locale transfer during fine-tuning and\nhighlight its effect on zero-shot locales, i.e., locales for which there is no\nfine-tuning data. Through a series of analyses, we highlight the role of\nnon-linguistic effects such as sound artifacts in cross-locale transfer.\nFinally, we present the effect of our design decision, e.g., model size,\npre-training diversity, and language rebalancing with several ablation\nexperiments.\n", "date": "2022-10-12", "categories": []}
{"id": "2305.19953", "title": "Multi-Dataset Co-Training with Sharpness-Aware Optimization for Audio\n  Anti-spoofing", "abstract": "  Audio anti-spoofing for automatic speaker verification aims to safeguard\nusers' identities from spoofing attacks. Although state-of-the-art spoofing\ncountermeasure(CM) models perform well on specific datasets, they lack\ngeneralization when evaluated with different datasets. To address this\nlimitation, previous studies have explored large pre-trained models, which\nrequire significant resources and time. We aim to develop a compact but\nwell-generalizing CM model that can compete with large pre-trained models. Our\napproach involves multi-dataset co-training and sharpness-aware minimization,\nwhich has not been investigated in this domain. Extensive experiments reveal\nthat proposed method yield competitive results across various datasets while\nutilizing 4,000 times less parameters than the large pre-trained models.\n", "date": "2023-05-31", "categories": []}
{"id": "2306.00160", "title": "Audio-Visual Speech Separation in Noisy Environments with a Lightweight\n  Iterative Model", "abstract": "  We propose Audio-Visual Lightweight ITerative model (AVLIT), an effective and\nlightweight neural network that uses Progressive Learning (PL) to perform\naudio-visual speech separation in noisy environments. To this end, we adopt the\nAsynchronous Fully Recurrent Convolutional Neural Network (A-FRCNN), which has\nshown successful results in audio-only speech separation. Our architecture\nconsists of an audio branch and a video branch, with iterative A-FRCNN blocks\nsharing weights for each modality. We evaluated our model in a controlled\nenvironment using the NTCD-TIMIT dataset and in-the-wild using a synthetic\ndataset that combines LRS3 and WHAM!. The experiments demonstrate the\nsuperiority of our model in both settings with respect to various audio-only\nand audio-visual baselines. Furthermore, the reduced footprint of our model\nmakes it suitable for low resource applications.\n", "date": "2023-05-31", "categories": []}
{"id": "2306.00410", "title": "Towards hate speech detection in low-resource languages: Comparing ASR\n  to acoustic word embeddings on Wolof and Swahili", "abstract": "  We consider hate speech detection through keyword spotting on radio\nbroadcasts. One approach is to build an automatic speech recognition (ASR)\nsystem for the target low-resource language. We compare this to using acoustic\nword embedding (AWE) models that map speech segments to a space where matching\nwords have similar vectors. We specifically use a multilingual AWE model\ntrained on labelled data from well-resourced languages to spot keywords in data\nin the unseen target language. In contrast to ASR, the AWE approach only\nrequires a few keyword exemplars. In controlled experiments on Wolof and\nSwahili where training and test data are from the same domain, an ASR model\ntrained on just five minutes of data outperforms the AWE approach. But in an\nin-the-wild test on Swahili radio broadcasts with actual hate speech keywords,\nthe AWE model (using one minute of template data) is more robust, giving\nsimilar performance to an ASR system trained on 30 hours of labelled data.\n", "date": "2023-06-01", "categories": []}
{"id": "2306.00489", "title": "Speech inpainting: Context-based speech synthesis guided by video", "abstract": "  Audio and visual modalities are inherently connected in speech signals: lip\nmovements and facial expressions are correlated with speech sounds. This\nmotivates studies that incorporate the visual modality to enhance an acoustic\nspeech signal or even restore missing audio information. Specifically, this\npaper focuses on the problem of audio-visual speech inpainting, which is the\ntask of synthesizing the speech in a corrupted audio segment in a way that it\nis consistent with the corresponding visual content and the uncorrupted audio\ncontext. We present an audio-visual transformer-based deep learning model that\nleverages visual cues that provide information about the content of the\ncorrupted audio. It outperforms the previous state-of-the-art audio-visual\nmodel and audio-only baselines. We also show how visual features extracted with\nAV-HuBERT, a large audio-visual transformer for speech recognition, are\nsuitable for synthesizing speech.\n", "date": "2023-06-01", "categories": []}
{"id": "2211.02940", "title": "Effective Audio Classification Network Based on Paired Inverse Pyramid\n  Structure and Dense MLP Block", "abstract": "  Recently, massive architectures based on Convolutional Neural Network (CNN)\nand self-attention mechanisms have become necessary for audio classification.\nWhile these techniques are state-of-the-art, these works' effectiveness can\nonly be guaranteed with huge computational costs and parameters, large amounts\nof data augmentation, transfer from large datasets and some other tricks. By\nutilizing the lightweight nature of audio, we propose an efficient network\nstructure called Paired Inverse Pyramid Structure (PIP) and a network called\nPaired Inverse Pyramid Structure MLP Network (PIPMN). The PIPMN reaches 96\\% of\nEnvironmental Sound Classification (ESC) accuracy on the UrbanSound8K dataset\nand 93.2\\% of Music Genre Classification (MGC) on the GTAZN dataset, with only\n1 million parameters. Both of the results are achieved without data\naugmentation or model transfer. Public code is available at:\nhttps://github.com/JNAIC/PIPMN\n", "date": "2022-11-05", "categories": []}
{"id": "2305.19304", "title": "Audio classification using ML methods", "abstract": "  Machine Learning systems have achieved outstanding performance in different\ndomains. In this paper machine learning methods have been applied to\nclassification task to classify music genre. The code shows how to extract\nfeatures from audio files and classify them using supervised learning into 2\ngenres namely classical and metal. Algorithms used are LogisticRegression, SVC\nusing different kernals (linear, sigmoid, rbf and poly), KNeighborsClassifier ,\nRandomForestClassifier, DecisionTreeClassifier and GaussianNB.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.19458", "title": "A Unified Audio-Visual Learning Framework for Localization, Separation,\n  and Recognition", "abstract": "  The ability to accurately recognize, localize and separate sound sources is\nfundamental to any audio-visual perception task. Historically, these abilities\nwere tackled separately, with several methods developed independently for each\ntask. However, given the interconnected nature of source localization,\nseparation, and recognition, independent models are likely to yield suboptimal\nperformance as they fail to capture the interdependence between these tasks. To\naddress this problem, we propose a unified audio-visual learning framework\n(dubbed OneAVM) that integrates audio and visual cues for joint localization,\nseparation, and recognition. OneAVM comprises a shared audio-visual encoder and\ntask-specific decoders trained with three objectives. The first objective\naligns audio and visual representations through a localized audio-visual\ncorrespondence loss. The second tackles visual source separation using a\ntraditional mix-and-separate framework. Finally, the third objective reinforces\nvisual feature separation and localization by mixing images in pixel space and\naligning their representations with those of all corresponding sound sources.\nExtensive experiments on MUSIC, VGG-Instruments, VGG-Music, and VGGSound\ndatasets demonstrate the effectiveness of OneAVM for all three tasks,\naudio-visual source localization, separation, and nearest neighbor recognition,\nand empirically demonstrate a strong positive transfer between them.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.19602", "title": "Learning Music Sequence Representation from Text Supervision", "abstract": "  Music representation learning is notoriously difficult for its complex\nhuman-related concepts contained in the sequence of numerical signals. To\nexcavate better MUsic SEquence Representation from labeled audio, we propose a\nnovel text-supervision pre-training method, namely MUSER. MUSER adopts an\naudio-spectrum-text tri-modal contrastive learning framework, where the text\ninput could be any form of meta-data with the help of text templates while the\nspectrum is derived from an audio sequence. Our experiments reveal that MUSER\ncould be more flexibly adapted to downstream tasks compared with the current\ndata-hungry pre-training method, and it only requires 0.056% of pre-training\ndata to achieve the state-of-the-art performance.\n", "date": "2023-05-31", "categories": []}
{"id": "2305.19769", "title": "Attention-Based Methods For Audio Question Answering", "abstract": "  Audio question answering (AQA) is the task of producing natural language\nanswers when a system is provided with audio and natural language questions. In\nthis paper, we propose neural network architectures based on self-attention and\ncross-attention for the AQA task. The self-attention layers extract powerful\naudio and textual representations. The cross-attention maps audio features that\nare relevant to the textual features to produce answers. All our models are\ntrained on the recently proposed Clotho-AQA dataset for both binary yes/no\nquestions and single-word answer questions. Our results clearly show\nimprovement over the reference method reported in the original paper. On the\nyes/no binary classification task, our proposed model achieves an accuracy of\n68.3% compared to 62.7% in the reference model. For the single-word answers\nmulticlass classifier, our model produces a top-1 and top-5 accuracy of 57.9%\nand 99.8% compared to 54.2% and 93.7% in the reference model respectively. We\nfurther discuss some of the challenges in the Clotho-AQA dataset such as the\npresence of the same answer word in multiple tenses, singular and plural forms,\nand the presence of specific and generic answers to the same question. We\naddress these issues and present a revised version of the dataset.\n", "date": "2023-05-31", "categories": []}
{"id": "2211.07710", "title": "End-to-End Speech to Intent Prediction to improve E-commerce Customer\n  Support Voicebot in Hindi and English", "abstract": "  Automation of on-call customer support relies heavily on accurate and\nefficient speech-to-intent (S2I) systems. Building such systems using\nmulti-component pipelines can pose various challenges because they require\nlarge annotated datasets, have higher latency, and have complex deployment.\nThese pipelines are also prone to compounding errors. To overcome these\nchallenges, we discuss an end-to-end (E2E) S2I model for customer support\nvoicebot task in a bilingual setting. We show how we can solve E2E intent\nclassification by leveraging a pre-trained automatic speech recognition (ASR)\nmodel with slight modification and fine-tuning on small annotated datasets.\nExperimental results show that our best E2E model outperforms a conventional\npipeline by a relative ~27% on the F1 score.\n", "date": "2022-10-26", "categories": []}
{"id": "2305.01319", "title": "Long-Term Rhythmic Video Soundtracker", "abstract": "  We consider the problem of generating musical soundtracks in sync with\nrhythmic visual cues. Most existing works rely on pre-defined music\nrepresentations, leading to the incompetence of generative flexibility and\ncomplexity. Other methods directly generating video-conditioned waveforms\nsuffer from limited scenarios, short lengths, and unstable generation quality.\nTo this end, we present Long-Term Rhythmic Video Soundtracker (LORIS), a novel\nframework to synthesize long-term conditional waveforms. Specifically, our\nframework consists of a latent conditional diffusion probabilistic model to\nperform waveform synthesis. Furthermore, a series of context-aware conditioning\nencoders are proposed to take temporal information into consideration for a\nlong-term generation. Notably, we extend our model's applicability from dances\nto multiple sports scenarios such as floor exercise and figure skating. To\nperform comprehensive evaluations, we establish a benchmark for rhythmic video\nsoundtracks including the pre-processed dataset, improved evaluation metrics,\nand robust generative baselines. Extensive experiments show that our model\ngenerates long-term soundtracks with state-of-the-art musical quality and\nrhythmic correspondence. Codes are available at\n\\url{https://github.com/OpenGVLab/LORIS}.\n", "date": "2023-05-02", "categories": []}
{"id": "2305.18602", "title": "From `Snippet-lects' to Doculects and Dialects: Leveraging Neural\n  Representations of Speech for Placing Audio Signals in a Language Landscape", "abstract": "  XLSR-53 a multilingual model of speech, builds a vector representation from\naudio, which allows for a range of computational treatments. The experiments\nreported here use this neural representation to estimate the degree of\ncloseness between audio files, ultimately aiming to extract relevant linguistic\nproperties. We use max-pooling to aggregate the neural representations from a\n\"snippet-lect\" (the speech in a 5-second audio snippet) to a \"doculect\" (the\nspeech in a given resource), then to dialects and languages. We use data from\ncorpora of 11 dialects belonging to 5 less-studied languages. Similarity\nmeasurements between the 11 corpora bring out greatest closeness between those\nthat are known to be dialects of the same language. The findings suggest that\n(i) dialect/language can emerge among the various parameters characterizing\naudio files and (ii) estimates of overall phonetic/phonological closeness can\nbe obtained for a little-resourced or fully unknown language. The findings help\nshed light on the type of information captured by neural representations of\nspeech and how it can be extracted from these representations\n", "date": "2023-05-29", "categories": []}
{"id": "2305.18665", "title": "E-PANNs: Sound Recognition Using Efficient Pre-trained Audio Neural\n  Networks", "abstract": "  Sounds carry an abundance of information about activities and events in our\neveryday environment, such as traffic noise, road works, music, or people\ntalking. Recent machine learning methods, such as convolutional neural networks\n(CNNs), have been shown to be able to automatically recognize sound activities,\na task known as audio tagging. One such method, pre-trained audio neural\nnetworks (PANNs), provides a neural network which has been pre-trained on over\n500 sound classes from the publicly available AudioSet dataset, and can be used\nas a baseline or starting point for other tasks. However, the existing PANNs\nmodel has a high computational complexity and large storage requirement. This\ncould limit the potential for deploying PANNs on resource-constrained devices,\nsuch as on-the-edge sound sensors, and could lead to high energy consumption if\nmany such devices were deployed. In this paper, we reduce the computational\ncomplexity and memory requirement of the PANNs model by taking a pruning\napproach to eliminate redundant parameters from the PANNs model. The resulting\nEfficient PANNs (E-PANNs) model, which requires 36\\% less computations and 70\\%\nless memory, also slightly improves the sound recognition (audio tagging)\nperformance. The code for the E-PANNs model has been released under an open\nsource license.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.19255", "title": "A Stutter Seldom Comes Alone -- Cross-Corpus Stuttering Detection as a\n  Multi-label Problem", "abstract": "  Most stuttering detection and classification research has viewed stuttering\nas a multi-class classification problem or a binary detection task for each\ndysfluency type; however, this does not match the nature of stuttering, in\nwhich one dysfluency seldom comes alone but rather co-occurs with others. This\npaper explores multi-language and cross-corpus end-to-end stuttering detection\nas a multi-label problem using a modified wav2vec 2.0 system with an\nattention-based classification head and multi-task learning. We evaluate the\nmethod using combinations of three datasets containing English and German\nstuttered speech, one containing speech modified by fluency shaping. The\nexperimental results and an error analysis show that multi-label stuttering\ndetection systems trained on cross-corpus and multi-language data achieve\ncompetitive results but performance on samples with multiple labels stays below\nover-all detection results.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.17993", "title": "Multi-Scale Attention for Audio Question Answering", "abstract": "  Audio question answering (AQA), acting as a widely used proxy task to explore\nscene understanding, has got more attention. The AQA is challenging for it\nrequires comprehensive temporal reasoning from different scales' events of an\naudio scene. However, existing methods mostly extend the structures of visual\nquestion answering task to audio ones in a simple pattern but may not perform\nwell when perceiving a fine-grained audio scene. To this end, we present a\nMulti-scale Window Attention Fusion Model (MWAFM) consisting of an asynchronous\nhybrid attention module and a multi-scale window attention module. The former\nis designed to aggregate unimodal and cross-modal temporal contexts, while the\nlatter captures sound events of varying lengths and their temporal dependencies\nfor a more comprehensive understanding. Extensive experiments are conducted to\ndemonstrate that the proposed MWAFM can effectively explore temporal\ninformation to facilitate AQA in the fine-grained scene.Code:\nhttps://github.com/GeWu-Lab/MWAFM\n", "date": "2023-05-29", "categories": []}
{"id": "2305.18283", "title": "CommonAccent: Exploring Large Acoustic Pretrained Models for Accent\n  Classification Based on Common Voice", "abstract": "  Despite the recent advancements in Automatic Speech Recognition (ASR), the\nrecognition of accented speech still remains a dominant problem. In order to\ncreate more inclusive ASR systems, research has shown that the integration of\naccent information, as part of a larger ASR framework, can lead to the\nmitigation of accented speech errors. We address multilingual accent\nclassification through the ECAPA-TDNN and Wav2Vec 2.0/XLSR architectures which\nhave been proven to perform well on a variety of speech-related downstream\ntasks. We introduce a simple-to-follow recipe aligned to the SpeechBrain\ntoolkit for accent classification based on Common Voice 7.0 (English) and\nCommon Voice 11.0 (Italian, German, and Spanish). Furthermore, we establish new\nstate-of-the-art for English accent classification with as high as 95%\naccuracy. We also study the internal categorization of the Wav2Vev 2.0\nembeddings through t-SNE, noting that there is a level of clustering based on\nphonological similarity. (Our recipe is open-source in the SpeechBrain toolkit,\nsee: https://github.com/speechbrain/speechbrain/tree/develop/recipes)\n", "date": "2023-05-29", "categories": []}
{"id": "2206.02432", "title": "Online Neural Diarization of Unlimited Numbers of Speakers Using Global\n  and Local Attractors", "abstract": "  A method to perform offline and online speaker diarization for an unlimited\nnumber of speakers is described in this paper. End-to-end neural diarization\n(EEND) has achieved overlap-aware speaker diarization by formulating it as a\nmulti-label classification problem. It has also been extended for a flexible\nnumber of speakers by introducing speaker-wise attractors. However, the output\nnumber of speakers of attractor-based EEND is empirically capped; it cannot\ndeal with cases where the number of speakers appearing during inference is\nhigher than that during training because its speaker counting is trained in a\nfully supervised manner. Our method, EEND-GLA, solves this problem by\nintroducing unsupervised clustering into attractor-based EEND. In the method,\nthe input audio is first divided into short blocks, then attractor-based\ndiarization is performed for each block, and finally, the results of each block\nare clustered on the basis of the similarity between locally-calculated\nattractors. While the number of output speakers is limited within each block,\nthe total number of speakers estimated for the entire input can be higher than\nthe limitation. To use EEND-GLA in an online manner, our method also extends\nthe speaker-tracing buffer, which was originally proposed to enable online\ninference of conventional EEND. We introduce a block-wise buffer update to make\nthe speaker-tracing buffer compatible with EEND-GLA. Finally, to improve online\ndiarization, our method improves the buffer update method and revisits the\nvariable chunk-size training of EEND. The experimental results demonstrate that\nEEND-GLA can perform speaker diarization of an unseen number of speakers in\nboth offline and online inferences.\n", "date": "2022-06-06", "categories": []}
{"id": "2305.16353", "title": "Betray Oneself: A Novel Audio DeepFake Detection Model via\n  Mono-to-Stereo Conversion", "abstract": "  Audio Deepfake Detection (ADD) aims to detect the fake audio generated by\ntext-to-speech (TTS), voice conversion (VC) and replay, etc., which is an\nemerging topic. Traditionally we take the mono signal as input and focus on\nrobust feature extraction and effective classifier design. However, the\ndual-channel stereo information in the audio signal also includes important\ncues for deepfake, which has not been studied in the prior work. In this paper,\nwe propose a novel ADD model, termed as M2S-ADD, that attempts to discover\naudio authenticity cues during the mono-to-stereo conversion process. We first\nprojects the mono to a stereo signal using a pretrained stereo synthesizer,\nthen employs a dual-branch neural architecture to process the left and right\nchannel signals, respectively. In this way, we effectively reveal the artifacts\nin the fake audio, thus improve the ADD performance. The experiments on the\nASVspoof2019 database show that M2S-ADD outperforms all baselines that input\nmono. We release the source code at \\url{https://github.com/AI-S2-Lab/M2S-ADD}.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16862", "title": "Neural modeling of magnetic tape recorders", "abstract": "  The sound of magnetic recording media, such as open-reel and cassette tape\nrecorders, is still sought after by today's sound practitioners due to the\nimperfections embedded in the physics of the magnetic recording process. This\npaper proposes a method for digitally emulating this character using neural\nnetworks. The signal chain of the proposed system consists of three main\ncomponents: the hysteretic nonlinearity and filtering jointly produced by the\nmagnetic recording process as well as the record and playback amplifiers, the\nfluctuating delay originating from the tape transport, and the combined\nadditive noise component from various electromagnetic origins. In our approach,\nthe hysteretic nonlinear block is modeled using a recurrent neural network,\nwhile the delay trajectories and the noise component are generated using\nseparate diffusion models, which employ U-net deep convolutional neural\nnetworks. According to the conducted objective evaluation, the proposed\narchitecture faithfully captures the character of the magnetic tape recorder.\nThe results of this study can be used to construct virtual replicas of vintage\nsound recording devices with applications in music production and audio\nantiquing tasks.\n", "date": "2023-05-26", "categories": []}
{"id": "2206.01542", "title": "Detecting the Severity of Major Depressive Disorder from Speech: A Novel\n  HARD-Training Methodology", "abstract": "  Major Depressive Disorder (MDD) is a common worldwide mental health issue\nwith high associated socioeconomic costs. The prediction and automatic\ndetection of MDD can, therefore, make a huge impact on society. Speech, as a\nnon-invasive, easy to collect signal, is a promising marker to aid the\ndiagnosis and assessment of MDD. In this regard, speech samples were collected\nas part of the Remote Assessment of Disease and Relapse in Major Depressive\nDisorder (RADAR-MDD) research programme. RADAR-MDD was an observational cohort\nstudy in which speech and other digital biomarkers were collected from a cohort\nof individuals with a history of MDD in Spain, United Kingdom and the\nNetherlands. In this paper, the RADAR-MDD speech corpus was taken as an\nexperimental framework to test the efficacy of a Sequence-to-Sequence model\nwith a local attention mechanism in a two-class depression severity\nclassification paradigm. Additionally, a novel training method, HARD-Training,\nis proposed. It is a methodology based on the selection of more ambiguous\nsamples for the model training, and inspired by the curriculum learning\nparadigm. HARD-Training was found to consistently improve - with an average\nincrement of 8.6% - the performance of our classifiers for both of two speech\nelicitation tasks used and each collection site of the RADAR-MDD speech corpus.\nWith this novel methodology, our Sequence-to-Sequence model was able to\neffectively detect MDD severity regardless of language. Finally, recognising\nthe need for greater awareness of potential algorithmic bias, we conduct an\nadditional analysis of our results separately for each gender.\n", "date": "2022-06-02", "categories": []}
{"id": "2212.01393", "title": "Continual Learning for On-Device Speech Recognition using Disentangled\n  Conformers", "abstract": "  Automatic speech recognition research focuses on training and evaluating on\nstatic datasets. Yet, as speech models are increasingly deployed on personal\ndevices, such models encounter user-specific distributional shifts. To simulate\nthis real-world scenario, we introduce LibriContinual, a continual learning\nbenchmark for speaker-specific domain adaptation derived from LibriVox\naudiobooks, with data corresponding to 118 individual speakers and 6 train\nsplits per speaker of different sizes. Additionally, current speech recognition\nmodels and continual learning algorithms are not optimized to be\ncompute-efficient. We adapt a general-purpose training algorithm NetAug for ASR\nand create a novel Conformer variant called the DisConformer (Disentangled\nConformer). This algorithm produces ASR models consisting of a frozen 'core'\nnetwork for general-purpose use and several tunable 'augment' networks for\nspeaker-specific tuning. Using such models, we propose a novel\ncompute-efficient continual learning algorithm called DisentangledCL. Our\nexperiments show that the DisConformer models significantly outperform\nbaselines on general ASR i.e. LibriSpeech (15.58% rel. WER on test-other). On\nspeaker-specific LibriContinual they significantly outperform\ntrainable-parameter-matched baselines (by 20.65% rel. WER on test) and even\nmatch fully finetuned baselines in some settings.\n", "date": "2022-12-02", "categories": []}
{"id": "2305.15719", "title": "Efficient Neural Music Generation", "abstract": "  Recent progress in music generation has been remarkably advanced by the\nstate-of-the-art MusicLM, which comprises a hierarchy of three LMs,\nrespectively, for semantic, coarse acoustic, and fine acoustic modelings. Yet,\nsampling with the MusicLM requires processing through these LMs one by one to\nobtain the fine-grained acoustic tokens, making it computationally expensive\nand prohibitive for a real-time generation. Efficient music generation with a\nquality on par with MusicLM remains a significant challenge. In this paper, we\npresent MeLoDy (M for music; L for LM; D for diffusion), an LM-guided diffusion\nmodel that generates music audios of state-of-the-art quality meanwhile\nreducing 95.7% or 99.6% forward passes in MusicLM, respectively, for sampling\n10s or 30s music. MeLoDy inherits the highest-level LM from MusicLM for\nsemantic modeling, and applies a novel dual-path diffusion (DPD) model and an\naudio VAE-GAN to efficiently decode the conditioning semantic tokens into\nwaveform. DPD is proposed to simultaneously model the coarse and fine acoustics\nby incorporating the semantic information into segments of latents effectively\nvia cross-attention at each denoising step. Our experimental results suggest\nthe superiority of MeLoDy, not only in its practical advantages on sampling\nspeed and infinitely continuable generation, but also in its state-of-the-art\nmusicality, audio quality, and text correlation.\n  Our samples are available at https://Efficient-MeLoDy.github.io/.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.15816", "title": "DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled\n  Representation and Prior Mixup for Verified Robust Voice Conversion", "abstract": "  Diffusion-based generative models have exhibited powerful generative\nperformance in recent years. However, as many attributes exist in the data\ndistribution and owing to several limitations of sharing the model parameters\nacross all levels of the generation process, it remains challenging to control\nspecific styles for each attribute. To address the above problem, this paper\npresents decoupled denoising diffusion models (DDDMs) with disentangled\nrepresentations, which can control the style for each attribute in generative\nmodels. We apply DDDMs to voice conversion (VC) tasks to address the challenges\nof disentangling and controlling each speech attribute (e.g., linguistic\ninformation, intonation, and timbre). First, we use a self-supervised\nrepresentation to disentangle the speech representation. Subsequently, the\nDDDMs are applied to resynthesize the speech from the disentangled\nrepresentations for denoising with respect to each attribute. Moreover, we also\npropose the prior mixup for robust voice style transfer, which uses the\nconverted representation of the mixed style as a prior distribution for the\ndiffusion models. The experimental results reveal that our method outperforms\npublicly available VC models. Furthermore, we show that our method provides\nrobust generative performance regardless of the model size. Audio samples are\navailable https://hayeong0.github.io/DDDM-VC-demo/.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.04476", "title": "AlignSTS: Speech-to-Singing Conversion via Cross-Modal Alignment", "abstract": "  The speech-to-singing (STS) voice conversion task aims to generate singing\nsamples corresponding to speech recordings while facing a major challenge: the\nalignment between the target (singing) pitch contour and the source (speech)\ncontent is difficult to learn in a text-free situation. This paper proposes\nAlignSTS, an STS model based on explicit cross-modal alignment, which views\nspeech variance such as pitch and content as different modalities. Inspired by\nthe mechanism of how humans will sing the lyrics to the melody, AlignSTS: 1)\nadopts a novel rhythm adaptor to predict the target rhythm representation to\nbridge the modality gap between content and pitch, where the rhythm\nrepresentation is computed in a simple yet effective way and is quantized into\na discrete space; and 2) uses the predicted rhythm representation to re-align\nthe content based on cross-attention and conducts a cross-modal fusion for\nre-synthesize. Extensive experiments show that AlignSTS achieves superior\nperformance in terms of both objective and subjective metrics. Audio samples\nare available at https://alignsts.github.io.\n", "date": "2023-05-08", "categories": []}
{"id": "2305.13331", "title": "A New Benchmark of Aphasia Speech Recognition and Detection Based on\n  E-Branchformer and Multi-task Learning", "abstract": "  Aphasia is a language disorder that affects the speaking ability of millions\nof patients. This paper presents a new benchmark for Aphasia speech recognition\nand detection tasks using state-of-the-art speech recognition techniques with\nthe AphsiaBank dataset. Specifically, we introduce two multi-task learning\nmethods based on the CTC/Attention architecture to perform both tasks\nsimultaneously. Our system achieves state-of-the-art speaker-level detection\naccuracy (97.3%), and a relative WER reduction of 11% for moderate Aphasia\npatients. In addition, we demonstrate the generalizability of our approach by\napplying it to another disordered speech database, the DementiaBank Pitt\ncorpus. We will make our all-in-one recipes and pre-trained model publicly\navailable to facilitate reproducibility. Our standardized data preprocessing\npipeline and open-source recipes enable researchers to compare results\ndirectly, promoting progress in disordered speech processing.\n", "date": "2023-05-19", "categories": []}
{"id": "2305.12712", "title": "LEAN: Light and Efficient Audio Classification Network", "abstract": "  Over the past few years, audio classification task on large-scale dataset\nsuch as AudioSet has been an important research area. Several deeper\nConvolution-based Neural networks have shown compelling performance notably\nVggish, YAMNet, and Pretrained Audio Neural Network (PANN). These models are\navailable as pretrained architecture for transfer learning as well as specific\naudio task adoption. In this paper, we propose a lightweight on-device deep\nlearning-based model for audio classification, LEAN. LEAN consists of a raw\nwaveform-based temporal feature extractor called as Wave Encoder and\nlogmel-based Pretrained YAMNet. We show that using a combination of trainable\nwave encoder, Pretrained YAMNet along with cross attention-based temporal\nrealignment, results in competitive performance on downstream audio\nclassification tasks with lesser memory footprints and hence making it suitable\nfor resource constraints devices such as mobile, edge devices, etc . Our\nproposed system achieves on-device mean average precision(mAP) of .445 with a\nmemory footprint of a mere 4.5MB on the FSD50K dataset which is an improvement\nof 22% over baseline on-device mAP on same dataset.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.12903", "title": "DiffAVA: Personalized Text-to-Audio Generation with Visual Alignment", "abstract": "  Text-to-audio (TTA) generation is a recent popular problem that aims to\nsynthesize general audio given text descriptions. Previous methods utilized\nlatent diffusion models to learn audio embedding in a latent space with text\nembedding as the condition. However, they ignored the synchronization between\naudio and visual content in the video, and tended to generate audio mismatching\nfrom video frames. In this work, we propose a novel and personalized\ntext-to-sound generation approach with visual alignment based on latent\ndiffusion models, namely DiffAVA, that can simply fine-tune lightweight\nvisual-text alignment modules with frozen modality-specific encoders to update\nvisual-aligned text embeddings as the condition. Specifically, our DiffAVA\nleverages a multi-head attention transformer to aggregate temporal information\nfrom video features, and a dual multi-modal residual network to fuse temporal\nvisual representations with text embeddings. Then, a contrastive learning\nobjective is applied to match visual-aligned text embeddings with audio\nfeatures. Experimental results on the AudioCaps dataset demonstrate that the\nproposed DiffAVA can achieve competitive performance on visual-aligned\ntext-to-audio generation.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.13262", "title": "Modulation Extraction for LFO-driven Audio Effects", "abstract": "  Low frequency oscillator (LFO) driven audio effects such as phaser, flanger,\nand chorus, modify an input signal using time-varying filters and delays,\nresulting in characteristic sweeping or widening effects. It has been shown\nthat these effects can be modeled using neural networks when conditioned with\nthe ground truth LFO signal. However, in most cases, the LFO signal is not\naccessible and measurement from the audio signal is nontrivial, hindering the\nmodeling process. To address this, we propose a framework capable of extracting\narbitrary LFO signals from processed audio across multiple digital audio\neffects, parameter settings, and instrument configurations. Since our system\nimposes no restrictions on the LFO signal shape, we demonstrate its ability to\nextract quasiperiodic, combined, and distorted modulation signals that are\nrelevant to effect modeling. Furthermore, we show how coupling the extraction\nmodel with a simple processing network enables training of end-to-end black-box\nmodels of unseen analog or digital LFO-driven audio effects using only dry and\nwet audio pairs, overcoming the need to access the audio effect or internal LFO\nsignal. We make our code available and provide the trained audio effect models\nin a real-time VST plugin.\n", "date": "2023-05-22", "categories": []}
{"id": "2211.01515", "title": "MAST: Multiscale Audio Spectrogram Transformers", "abstract": "  We present Multiscale Audio Spectrogram Transformer (MAST) for audio\nclassification, which brings the concept of multiscale feature hierarchies to\nthe Audio Spectrogram Transformer (AST). Given an input audio spectrogram, we\nfirst patchify and project it into an initial temporal resolution and embedding\ndimension, post which the multiple stages in MAST progressively expand the\nembedding dimension while reducing the temporal resolution of the input. We use\na pyramid structure that allows early layers of MAST operating at a high\ntemporal resolution but low embedding space to model simple low-level acoustic\ninformation and deeper temporally coarse layers to model high-level acoustic\ninformation with high-dimensional embeddings. We also extend our approach to\npresent a new Self-Supervised Learning (SSL) method called SS-MAST, which\ncalculates a symmetric contrastive loss between latent representations from a\nstudent and a teacher encoder, leveraging patch-drop, a novel audio\naugmentation approach that we introduce. In practice, MAST significantly\noutperforms AST by an average accuracy of 3.4% across 8 speech and non-speech\ntasks from the LAPE Benchmark, achieving state-of-the-art results on keyword\nspotting in Speech Commands. Additionally, our proposed SS-MAST achieves an\nabsolute average improvement of 2.6% over the previously proposed SSAST.\n", "date": "2022-11-02", "categories": []}
{"id": "2211.01519", "title": "SLICER: Learning universal audio representations using low-resource\n  self-supervised pre-training", "abstract": "  We present a new Self-Supervised Learning (SSL) approach to pre-train\nencoders on unlabeled audio data that reduces the need for large amounts of\nlabeled data for audio and speech classification. Our primary aim is to learn\naudio representations that can generalize across a large variety of speech and\nnon-speech tasks in a low-resource un-labeled audio pre-training setting.\nInspired by the recent success of clustering and contrasting learning paradigms\nfor SSL-based speech representation learning, we propose SLICER (Symmetrical\nLearning of Instance and Cluster-level Efficient Representations), which brings\ntogether the best of both clustering and contrasting learning paradigms. We use\na symmetric loss between latent representations from student and teacher\nencoders and simultaneously solve instance and cluster-level contrastive\nlearning tasks. We obtain cluster representations online by just projecting the\ninput spectrogram into an output subspace with dimensions equal to the number\nof clusters. In addition, we propose a novel mel-spectrogram augmentation\nprocedure, k-mix, based on mixup, which does not require labels and aids\nunsupervised representation learning for audio. Overall, SLICER achieves\nstate-of-the-art results on the LAPE Benchmark \\cite{9868132}, significantly\noutperforming DeLoRes-M and other prior approaches, which are pre-trained on\n$10\\times$ larger of unsupervised data. We will make all our codes available on\nGitHub.\n", "date": "2022-11-02", "categories": []}
{"id": "2303.05668", "title": "UNFUSED: UNsupervised Finetuning Using SElf supervised Distillation", "abstract": "  In this paper, we introduce UnFuSeD, a novel approach to leverage\nself-supervised learning and reduce the need for large amounts of labeled data\nfor audio classification. Unlike prior works, which directly fine-tune a\nself-supervised pre-trained encoder on a target dataset, we use the encoder to\ngenerate pseudo-labels for unsupervised fine-tuning before the actual\nfine-tuning step. We first train an encoder using a novel self-supervised\nlearning algorithm (SSL) on an unlabeled audio dataset. Then, we use that\nencoder to generate pseudo-labels on our target task dataset via clustering the\nextracted representations. These pseudo-labels are then used to guide\nself-distillation on a randomly initialized model, which we call unsupervised\nfine-tuning. Finally, the resultant encoder is then fine-tuned on our target\ntask dataset. Through UnFuSeD, we propose the first system that moves away from\ngeneric SSL paradigms in literature, which pre-train and fine-tune the same\nencoder, and present a novel self-distillation-based system to leverage SSL\npre-training for low-resource audio classification. In practice, UnFuSeD\nachieves state-of-the-art results on the LAPE Benchmark, significantly\noutperforming all our baselines. Additionally, UnFuSeD allows us to achieve\nthis at a 40% reduction in the number of parameters over the previous\nstate-of-the-art system. We make all our codes publicly available.\n", "date": "2023-03-10", "categories": []}
{"id": "2305.10659", "title": "Use of Speech Impairment Severity for Dysarthric Speech Recognition", "abstract": "  A key challenge in dysarthric speech recognition is the speaker-level\ndiversity attributed to both speaker-identity associated factors such as\ngender, and speech impairment severity. Most prior researches on addressing\nthis issue focused on using speaker-identity only. To this end, this paper\nproposes a novel set of techniques to use both severity and speaker-identity in\ndysarthric speech recognition: a) multitask training incorporating severity\nprediction error; b) speaker-severity aware auxiliary feature adaptation; and\nc) structured LHUC transforms separately conditioned on speaker-identity and\nseverity. Experiments conducted on UASpeech suggest incorporating additional\nspeech impairment severity into state-of-the-art hybrid DNN, E2E Conformer and\npre-trained Wav2vec 2.0 ASR systems produced statistically significant WER\nreductions up to 4.78% (14.03% relative). Using the best system the lowest\npublished WER of 17.82% (51.25% on very low intelligibility) was obtained on\nUASpeech.\n", "date": "2023-05-18", "categories": []}
{"id": "2212.03023", "title": "FretNet: Continuous-Valued Pitch Contour Streaming for Polyphonic Guitar\n  Tablature Transcription", "abstract": "  In recent years, the task of Automatic Music Transcription (AMT), whereby\nvarious attributes of music notes are estimated from audio, has received\nincreasing attention. At the same time, the related task of Multi-Pitch\nEstimation (MPE) remains a challenging but necessary component of almost all\nAMT approaches, even if only implicitly. In the context of AMT, pitch\ninformation is typically quantized to the nominal pitches of the Western music\nscale. Even in more general contexts, MPE systems typically produce pitch\npredictions with some degree of quantization. In certain applications of AMT,\nsuch as Guitar Tablature Transcription (GTT), it is more meaningful to estimate\ncontinuous-valued pitch contours. Guitar tablature has the capacity to\nrepresent various playing techniques, some of which involve pitch modulation.\nContemporary approaches to AMT do not adequately address pitch modulation, and\noffer only less quantization at the expense of more model complexity. In this\npaper, we present a GTT formulation that estimates continuous-valued pitch\ncontours, grouping them according to their string and fret of origin. We\ndemonstrate that for this task, the proposed method significantly improves the\nresolution of MPE and simultaneously yields tablature estimation results\ncompetitive with baseline models.\n", "date": "2022-12-06", "categories": []}
{"id": "2305.09690", "title": "A Whisper transformer for audio captioning trained with synthetic\n  captions and transfer learning", "abstract": "  The field of audio captioning has seen significant advancements in recent\nyears, driven by the availability of large-scale audio datasets and\nadvancements in deep learning techniques. In this technical report, we present\nour approach to audio captioning, focusing on the use of a pretrained\nspeech-to-text Whisper model and pretraining on synthetic captions. We discuss\nour training procedures and present our experiments' results, which include\nmodel size variations, dataset mixtures, and other hyperparameters. Our\nfindings demonstrate the impact of different training strategies on the\nperformance of the audio captioning model. Our code and trained models are\npublicly available on GitHub and Hugging Face Hub.\n", "date": "2023-05-15", "categories": []}
{"id": "2305.09463", "title": "Low-complexity deep learning frameworks for acoustic scene\n  classification using teacher-student scheme and multiple spectrograms", "abstract": "  In this technical report, a low-complexity deep learning system for acoustic\nscene classification (ASC) is presented. The proposed system comprises two main\nphases: (Phase I) Training a teacher network; and (Phase II) training a student\nnetwork using distilled knowledge from the teacher. In the first phase, the\nteacher, which presents a large footprint model, is trained. After training the\nteacher, the embeddings, which are the feature map of the second last layer of\nthe teacher, are extracted. In the second phase, the student network, which\npresents a low complexity model, is trained with the embeddings extracted from\nthe teacher. Our experiments conducted on DCASE 2023 Task 1 Development dataset\nhave fulfilled the requirement of low-complexity and achieved the best\nclassification accuracy of 57.4%, improving DCASE baseline by 14.5%.\n", "date": "2023-05-16", "categories": []}
{"id": "2305.09636", "title": "SoundStorm: Efficient Parallel Audio Generation", "abstract": "  We present SoundStorm, a model for efficient, non-autoregressive audio\ngeneration. SoundStorm receives as input the semantic tokens of AudioLM, and\nrelies on bidirectional attention and confidence-based parallel decoding to\ngenerate the tokens of a neural audio codec. Compared to the autoregressive\ngeneration approach of AudioLM, our model produces audio of the same quality\nand with higher consistency in voice and acoustic conditions, while being two\norders of magnitude faster. SoundStorm generates 30 seconds of audio in 0.5\nseconds on a TPU-v4. We demonstrate the ability of our model to scale audio\ngeneration to longer sequences by synthesizing high-quality, natural dialogue\nsegments, given a transcript annotated with speaker turns and a short prompt\nwith the speakers' voices.\n", "date": "2023-05-16", "categories": []}
{"id": "2207.04156", "title": "Automated Audio Captioning and Language-Based Audio Retrieval", "abstract": "  This project involved participation in the DCASE 2022 Competition (Task 6)\nwhich had two subtasks: (1) Automated Audio Captioning and (2) Language-Based\nAudio Retrieval. The first subtask involved the generation of a textual\ndescription for audio samples, while the goal of the second was to find audio\nsamples within a fixed dataset that match a given description. For both\nsubtasks, the Clotho dataset was used. The models were evaluated on BLEU1,\nBLEU2, BLEU3, ROUGEL, METEOR, CIDEr, SPICE, and SPIDEr scores for audio\ncaptioning and R1, R5, R10 and mARP10 scores for audio retrieval. We have\nconducted a handful of experiments that modify the baseline models for these\ntasks. Our final architecture for Automated Audio Captioning is close to the\nbaseline performance, while our model for Language-Based Audio Retrieval has\nsurpassed its counterpart.\n", "date": "2022-07-08", "categories": []}
{"id": "2305.07132", "title": "Tackling Interpretability in Audio Classification Networks with\n  Non-negative Matrix Factorization", "abstract": "  This paper tackles two major problem settings for interpretability of audio\nprocessing networks, post-hoc and by-design interpretation. For post-hoc\ninterpretation, we aim to interpret decisions of a network in terms of\nhigh-level audio objects that are also listenable for the end-user. This is\nextended to present an inherently interpretable model with high performance. To\nthis end, we propose a novel interpreter design that incorporates non-negative\nmatrix factorization (NMF). In particular, an interpreter is trained to\ngenerate a regularized intermediate embedding from hidden layers of a target\nnetwork, learnt as time-activations of a pre-learnt NMF dictionary. Our\nmethodology allows us to generate intuitive audio-based interpretations that\nexplicitly enhance parts of the input signal most relevant for a network's\ndecision. We demonstrate our method's applicability on a variety of\nclassification tasks, including multi-label data for real-world audio and\nmusic.\n", "date": "2023-05-11", "categories": []}
{"id": "2211.00497", "title": "Modelling black-box audio effects with time-varying feature modulation", "abstract": "  Deep learning approaches for black-box modelling of audio effects have shown\npromise, however, the majority of existing work focuses on nonlinear effects\nwith behaviour on relatively short time-scales, such as guitar amplifiers and\ndistortion. While recurrent and convolutional architectures can theoretically\nbe extended to capture behaviour at longer time scales, we show that simply\nscaling the width, depth, or dilation factor of existing architectures does not\nresult in satisfactory performance when modelling audio effects such as fuzz\nand dynamic range compression. To address this, we propose the integration of\ntime-varying feature-wise linear modulation into existing temporal\nconvolutional backbones, an approach that enables learnable adaptation of the\nintermediate activations. We demonstrate that our approach more accurately\ncaptures long-range dependencies for a range of fuzz and compressor\nimplementations across both time and frequency domain metrics. We provide sound\nexamples, source code, and pretrained models to faciliate reproducibility.\n", "date": "2022-11-01", "categories": []}
{"id": "2305.05780", "title": "Enhancing Gappy Speech Audio Signals with Generative Adversarial\n  Networks", "abstract": "  Gaps, dropouts and short clips of corrupted audio are a common problem and\nparticularly annoying when they occur in speech. This paper uses machine\nlearning to regenerate gaps of up to 320ms in an audio speech signal. Audio\nregeneration is translated into image regeneration by transforming audio into a\nMel-spectrogram and using image in-painting to regenerate the gaps. The full\nMel-spectrogram is then transferred back to audio using the Parallel-WaveGAN\nvocoder and integrated into the audio stream. Using a sample of 1300 spoken\naudio clips of between 1 and 10 seconds taken from the publicly-available\nLJSpeech dataset our results show regeneration of audio gaps in close to real\ntime using GANs with a GPU equipped system. As expected, the smaller the gap in\nthe audio, the better the quality of the filled gaps. On a gap of 240ms the\naverage mean opinion score (MOS) for the best performing models was 3.737, on a\nscale of 1 (worst) to 5 (best) which is sufficient for a human to perceive as\nclose to uninterrupted human speech.\n", "date": "2023-05-09", "categories": []}
{"id": "2305.05591", "title": "AudioSlots: A slot-centric generative model for audio separation", "abstract": "  In a range of recent works, object-centric architectures have been shown to\nbe suitable for unsupervised scene decomposition in the vision domain. Inspired\nby these methods we present AudioSlots, a slot-centric generative model for\nblind source separation in the audio domain. AudioSlots is built using\npermutation-equivariant encoder and decoder networks. The encoder network based\non the Transformer architecture learns to map a mixed audio spectrogram to an\nunordered set of independent source embeddings. The spatial broadcast decoder\nnetwork learns to generate the source spectrograms from the source embeddings.\nWe train the model in an end-to-end manner using a permutation invariant loss\nfunction. Our results on Libri2Mix speech separation constitute a proof of\nconcept that this approach shows promise. We discuss the results and\nlimitations of our approach in detail, and further outline potential ways to\novercome the limitations and directions for future work.\n", "date": "2023-05-09", "categories": []}
{"id": "2208.07282", "title": "Differentiable WORLD Synthesizer-based Neural Vocoder With Application\n  To End-To-End Audio Style Transfer", "abstract": "  In this paper, we propose a differentiable WORLD synthesizer and demonstrate\nits use in end-to-end audio style transfer tasks such as (singing) voice\nconversion and the DDSP timbre transfer task. Accordingly, our baseline\ndifferentiable synthesizer has no model parameters, yet it yields adequate\nsynthesis quality. We can extend the baseline synthesizer by appending\nlightweight black-box postnets which apply further processing to the baseline\noutput in order to improve fidelity. An alternative differentiable approach\nconsiders extraction of the source excitation spectrum directly, which can\nimprove naturalness albeit for a narrower class of style transfer applications.\nThe acoustic feature parameterization used by our approaches has the added\nbenefit that it naturally disentangles pitch and timbral information so that\nthey can be modeled separately. Moreover, as there exists a robust means of\nestimating these acoustic features from monophonic audio sources, it allows for\nparameter loss terms to be added to an end-to-end objective function, which can\nhelp convergence and/or further stabilize (adversarial) training.\n", "date": "2022-08-15", "categories": []}
{"id": "2210.00943", "title": "Simple Pooling Front-ends For Efficient Audio Classification", "abstract": "  Recently, there has been increasing interest in building efficient audio\nneural networks for on-device scenarios. Most existing approaches are designed\nto reduce the size of audio neural networks using methods such as model\npruning. In this work, we show that instead of reducing model size using\ncomplex methods, eliminating the temporal redundancy in the input audio\nfeatures (e.g., mel-spectrogram) could be an effective approach for efficient\naudio classification. To do so, we proposed a family of simple pooling\nfront-ends (SimPFs) which use simple non-parametric pooling operations to\nreduce the redundant information within the mel-spectrogram. We perform\nextensive experiments on four audio classification tasks to evaluate the\nperformance of SimPFs. Experimental results show that SimPFs can achieve a\nreduction in more than half of the number of floating point operations (FLOPs)\nfor off-the-shelf audio neural networks, with negligible degradation or even\nsome improvements in audio classification performance.\n", "date": "2022-10-03", "categories": []}
{"id": "2211.07302", "title": "MedleyVox: An Evaluation Dataset for Multiple Singing Voices Separation", "abstract": "  Separation of multiple singing voices into each voice is a rarely studied\narea in music source separation research. The absence of a benchmark dataset\nhas hindered its progress. In this paper, we present an evaluation dataset and\nprovide baseline studies for multiple singing voices separation. First, we\nintroduce MedleyVox, an evaluation dataset for multiple singing voices\nseparation. We specify the problem definition in this dataset by categorizing\nit into i) unison, ii) duet, iii) main vs. rest, and iv) N-singing separation.\nSecond, to overcome the absence of existing multi-singing datasets for a\ntraining purpose, we present a strategy for construction of multiple singing\nmixtures using various single-singing datasets. Third, we propose the improved\nsuper-resolution network (iSRNet), which greatly enhances initial estimates of\nseparation networks. Jointly trained with the Conv-TasNet and the multi-singing\nmixture construction strategy, the proposed iSRNet achieved comparable\nperformance to ideal time-frequency masks on duet and unison subsets of\nMedleyVox. Audio samples, the dataset, and codes are available on our website\n(https://github.com/jeonchangbin49/MedleyVox).\n", "date": "2022-11-14", "categories": []}
{"id": "2305.02382", "title": "Learning to Detect Novel and Fine-Grained Acoustic Sequences Using\n  Pretrained Audio Representations", "abstract": "  This work investigates pretrained audio representations for few shot Sound\nEvent Detection. We specifically address the task of few shot detection of\nnovel acoustic sequences, or sound events with semantically meaningful temporal\nstructure, without assuming access to non-target audio. We develop procedures\nfor pretraining suitable representations, and methods which transfer them to\nour few shot learning scenario. Our experiments evaluate the general purpose\nutility of our pretrained representations on AudioSet, and the utility of\nproposed few shot methods via tasks constructed from real-world acoustic\nsequences. Our pretrained embeddings are suitable to the proposed task, and\nenable multiple aspects of our few shot framework.\n", "date": "2023-05-03", "categories": []}
{"id": "2211.03643", "title": "Egocentric Audio-Visual Noise Suppression", "abstract": "  This paper studies audio-visual noise suppression for egocentric videos --\nwhere the speaker is not captured in the video. Instead, potential noise\nsources are visible on screen with the camera emulating the off-screen\nspeaker's view of the outside world. This setting is different from prior work\nin audio-visual speech enhancement that relies on lip and facial visuals. In\nthis paper, we first demonstrate that egocentric visual information is helpful\nfor noise suppression. We compare object recognition and action\nclassification-based visual feature extractors and investigate methods to align\naudio and visual representations. Then, we examine different fusion strategies\nfor the aligned features, and locations within the noise suppression model to\nincorporate visual information. Experiments demonstrate that visual features\nare most helpful when used to generate additive correction masks. Finally, in\norder to ensure that the visual features are discriminative with respect to\ndifferent noise types, we introduce a multi-task learning framework that\njointly optimizes audio-visual noise suppression and video-based acoustic event\ndetection. This proposed multi-task framework outperforms the audio-only\nbaseline on all metrics, including a 0.16 PESQ improvement. Extensive ablations\nreveal the improved performance of the proposed model with multiple active\ndistractors, overall noise types, and across different SNRs.\n", "date": "2022-11-07", "categories": []}
{"id": "2211.00878", "title": "Neural Fourier Shift for Binaural Speech Rendering", "abstract": "  We present a neural network for rendering binaural speech from given monaural\naudio, position, and orientation of the source. Most of the previous works have\nfocused on synthesizing binaural speeches by conditioning the positions and\norientations in the feature space of convolutional neural networks. These\nsynthesis approaches are powerful in estimating the target binaural speeches\neven for in-the-wild data but are difficult to generalize for rendering the\naudio from out-of-distribution domains. To alleviate this, we propose Neural\nFourier Shift (NFS), a novel network architecture that enables binaural speech\nrendering in the Fourier space. Specifically, utilizing a geometric time delay\nbased on the distance between the source and the receiver, NFS is trained to\npredict the delays and scales of various early reflections. NFS is efficient in\nboth memory and computational cost, is interpretable, and operates\nindependently of the source domain by its design. Experimental results show\nthat NFS performs comparable to the previous studies on the benchmark dataset,\neven with its 25 times lighter memory and 6 times fewer calculations.\n", "date": "2022-11-02", "categories": []}
{"id": "2304.14882", "title": "The ACM Multimedia 2023 Computational Paralinguistics Challenge: Emotion\n  Share & Requests", "abstract": "  The ACM Multimedia 2023 Computational Paralinguistics Challenge addresses two\ndifferent problems for the first time in a research competition under\nwell-defined conditions: In the Emotion Share Sub-Challenge, a regression on\nspeech has to be made; and in the Requests Sub-Challenges, requests and\ncomplaints need to be detected. We describe the Sub-Challenges, baseline\nfeature extraction, and classifiers based on the usual ComPaRE features, the\nauDeep toolkit, and deep feature extraction from pre-trained CNNs using the\nDeepSpectRum toolkit; in addition, wav2vec2 models are used.\n", "date": "2023-04-28", "categories": []}
{"id": "2305.00426", "title": "Transfer of knowledge among instruments in automatic music transcription", "abstract": "  Automatic music transcription (AMT) is one of the most challenging tasks in\nthe music information retrieval domain. It is the process of converting an\naudio recording of music into a symbolic representation containing information\nabout the notes, chords, and rhythm. Current research in this domain focuses on\ndeveloping new models based on transformer architecture or using methods to\nperform semi-supervised training, which gives outstanding results, but the\ncomputational cost of training such models is enormous.\n  This work shows how to employ easily generated synthesized audio data\nproduced by software synthesizers to train a universal model. It is a good base\nfor further transfer learning to quickly adapt transcription model for other\ninstruments. Achieved results prove that using synthesized data for training\nmay be a good base for pretraining general-purpose models, where the task of\ntranscription is not focused on one instrument.\n", "date": "2023-04-30", "categories": []}
{"id": "2210.15387", "title": "Automatic Severity Classification of Dysarthric speech by using\n  Self-supervised Model with Multi-task Learning", "abstract": "  Automatic assessment of dysarthric speech is essential for sustained\ntreatments and rehabilitation. However, obtaining atypical speech is\nchallenging, often leading to data scarcity issues. To tackle the problem, we\npropose a novel automatic severity assessment method for dysarthric speech,\nusing the self-supervised model in conjunction with multi-task learning.\nWav2vec 2.0 XLS-R is jointly trained for two different tasks: severity\nclassification and auxiliary automatic speech recognition (ASR). For the\nbaseline experiments, we employ hand-crafted acoustic features and machine\nlearning classifiers such as SVM, MLP, and XGBoost. Explored on the Korean\ndysarthric speech QoLT database, our model outperforms the traditional baseline\nmethods, with a relative percentage increase of 1.25% for F1-score. In\naddition, the proposed model surpasses the model trained without ASR head,\nachieving 10.61% relative percentage improvements. Furthermore, we present how\nmulti-task learning affects the severity classification performance by\nanalyzing the latent representations and regularization effect.\n", "date": "2022-10-27", "categories": []}
{"id": "2304.14019", "title": "XAI-based Comparison of Input Representations for Audio Event\n  Classification", "abstract": "  Deep neural networks are a promising tool for Audio Event Classification. In\ncontrast to other data like natural images, there are many sensible and\nnon-obvious representations for audio data, which could serve as input to these\nmodels. Due to their black-box nature, the effect of different input\nrepresentations has so far mostly been investigated by measuring classification\nperformance. In this work, we leverage eXplainable AI (XAI), to understand the\nunderlying classification strategies of models trained on different input\nrepresentations. Specifically, we compare two model architectures with regard\nto relevant input features used for Audio Event Detection: one directly\nprocesses the signal as the raw waveform, and the other takes in its\ntime-frequency spectrogram representation. We show how relevance heatmaps\nobtained via \"Siren\"{Layer-wise Relevance Propagation} uncover\nrepresentation-dependent decision strategies. With these insights, we can make\na well-informed decision about the best input representation in terms of\nrobustness and representativity and confirm that the model's classification\nstrategies align with human requirements.\n", "date": "2023-04-27", "categories": []}
{"id": "2304.11472", "title": "A Comparative Study of Pre-trained Speech and Audio Embeddings for\n  Speech Emotion Recognition", "abstract": "  Pre-trained models (PTMs) have shown great promise in the speech and audio\ndomain. Embeddings leveraged from these models serve as inputs for learning\nalgorithms with applications in various downstream tasks. One such crucial task\nis Speech Emotion Recognition (SER) which has a wide range of applications,\nincluding dynamic analysis of customer calls, mental health assessment, and\npersonalized language learning. PTM embeddings have helped advance SER,\nhowever, a comprehensive comparison of these PTM embeddings that consider\nmultiple facets such as embedding model architecture, data used for\npre-training, and the pre-training procedure being followed is missing. A\nthorough comparison of PTM embeddings will aid in the faster and more efficient\ndevelopment of models and enable their deployment in real-world scenarios. In\nthis work, we exploit this research gap and perform a comparative analysis of\nembeddings extracted from eight speech and audio PTMs (wav2vec 2.0, data2vec,\nwavLM, UniSpeech-SAT, wav2clip, YAMNet, x-vector, ECAPA). We perform an\nextensive empirical analysis with four speech emotion datasets (CREMA-D, TESS,\nSAVEE, Emo-DB) by training three algorithms (XGBoost, Random Forest, FCN) on\nthe derived embeddings. The results of our study indicate that the best\nperformance is achieved by algorithms trained on embeddings derived from PTMs\ntrained for speaker recognition followed by wav2clip and UniSpeech-SAT. This\ncan relay that the top performance by embeddings from speaker recognition PTMs\nis most likely due to the model taking up information about numerous speech\nfeatures such as tone, accent, pitch, and so on during its speaker recognition\ntraining. Insights from this work will assist future studies in their selection\nof embeddings for applications related to SER.\n", "date": "2023-04-22", "categories": []}
{"id": "2304.11708", "title": "Sound-based drone fault classification using multitask learning", "abstract": "  The drone has been used for various purposes, including military\napplications, aerial photography, and pesticide spraying. However, the drone is\nvulnerable to external disturbances, and malfunction in propellers and motors\ncan easily occur. To improve the safety of drone operations, one should detect\nthe mechanical faults of drones in real-time. This paper proposes a sound-based\ndeep neural network (DNN) fault classifier and drone sound dataset. The dataset\nwas constructed by collecting the operating sounds of drones from microphones\nmounted on three different drones in an anechoic chamber. The dataset includes\nvarious operating conditions of drones, such as flight directions (front, back,\nright, left, clockwise, counterclockwise) and faults on propellers and motors.\nThe drone sounds were then mixed with noises recorded in five different spots\non the university campus, with a signal-to-noise ratio (SNR) varying from 10 dB\nto 15 dB. Using the acquired dataset, we train a DNN classifier, 1DCNN-ResNet,\nthat classifies the types of mechanical faults and their locations from\nshort-time input waveforms. We employ multitask learning (MTL) and incorporate\nthe direction classification task as an auxiliary task to make the classifier\nlearn more general audio features. The test over unseen data reveals that the\nproposed multitask model can successfully classify faults in drones and\noutperforms single-task models even with less training data.\n", "date": "2023-04-23", "categories": []}
{"id": "2304.11907", "title": "Advancing underwater acoustic target recognition via adaptive data\n  pruning and smoothness-inducing regularization", "abstract": "  Underwater acoustic recognition for ship-radiated signals has high practical\napplication value due to the ability to recognize non-line-of-sight targets.\nHowever, due to the difficulty of data acquisition, the collected signals are\nscarce in quantity and mainly composed of mechanical periodic noise. According\nto the experiments, we observe that the repeatability of periodic signals leads\nto a double-descent phenomenon, which indicates a significant local bias toward\nrepeated samples. To address this issue, we propose a strategy based on\ncross-entropy to prune excessively similar segments in training data.\nFurthermore, to compensate for the reduction of training data, we generate\nnoisy samples and apply smoothness-inducing regularization based on KL\ndivergence to mitigate overfitting. Experiments show that our proposed data\npruning and regularization strategy can bring stable benefits and our framework\nsignificantly outperforms the state-of-the-art in low-resource scenarios.\n", "date": "2023-04-24", "categories": []}
{"id": "2211.02250", "title": "Real-Time Target Sound Extraction", "abstract": "  We present the first neural network model to achieve real-time and streaming\ntarget sound extraction. To accomplish this, we propose Waveformer, an\nencoder-decoder architecture with a stack of dilated causal convolution layers\nas the encoder, and a transformer decoder layer as the decoder. This hybrid\narchitecture uses dilated causal convolutions for processing large receptive\nfields in a computationally efficient manner while also leveraging the\ngeneralization performance of transformer-based architectures. Our evaluations\nshow as much as 2.2-3.3 dB improvement in SI-SNRi compared to the prior models\nfor this task while having a 1.2-4x smaller model size and a 1.5-2x lower\nruntime. We provide code, dataset, and audio samples:\nhttps://waveformer.cs.washington.edu/.\n", "date": "2022-11-04", "categories": []}
{"id": "2304.08811", "title": "Towards the Transferable Audio Adversarial Attack via Ensemble Methods", "abstract": "  In recent years, deep learning (DL) models have achieved significant progress\nin many domains, such as autonomous driving, facial recognition, and speech\nrecognition. However, the vulnerability of deep learning models to adversarial\nattacks has raised serious concerns in the community because of their\ninsufficient robustness and generalization. Also, transferable attacks have\nbecome a prominent method for black-box attacks. In this work, we explore the\npotential factors that impact adversarial examples (AEs) transferability in\nDL-based speech recognition. We also discuss the vulnerability of different DL\nsystems and the irregular nature of decision boundaries. Our results show a\nremarkable difference in the transferability of AEs between speech and images,\nwith the data relevance being low in images but opposite in speech recognition.\nMotivated by dropout-based ensemble approaches, we propose random gradient\nensembles and dynamic gradient-weighted ensembles, and we evaluate the impact\nof ensembles on the transferability of AEs. The results show that the AEs\ncreated by both approaches are valid for transfer to the black box API.\n", "date": "2023-04-18", "categories": []}
{"id": "2304.07449", "title": "Self-supervised Auxiliary Loss for Metric Learning in Music\n  Similarity-based Retrieval and Auto-tagging", "abstract": "  In the realm of music information retrieval, similarity-based retrieval and\nauto-tagging serve as essential components. Given the limitations and\nnon-scalability of human supervision signals, it becomes crucial for models to\nlearn from alternative sources to enhance their performance. Self-supervised\nlearning, which exclusively relies on learning signals derived from music audio\ndata, has demonstrated its efficacy in the context of auto-tagging. In this\nstudy, we propose a model that builds on the self-supervised learning approach\nto address the similarity-based retrieval challenge by introducing our method\nof metric learning with a self-supervised auxiliary loss. Furthermore,\ndiverging from conventional self-supervised learning methodologies, we\ndiscovered the advantages of concurrently training the model with both\nself-supervision and supervision signals, without freezing pre-trained models.\nWe also found that refraining from employing augmentation during the\nfine-tuning phase yields better results. Our experimental results confirm that\nthe proposed methodology enhances retrieval and tagging performance metrics in\ntwo distinct scenarios: one where human-annotated tags are consistently\navailable for all music tracks, and another where such tags are accessible only\nfor a subset of tracks.\n", "date": "2023-04-15", "categories": []}
{"id": "2304.08490", "title": "Conditional Generation of Audio from Video via Foley Analogies", "abstract": "  The sound effects that designers add to videos are designed to convey a\nparticular artistic effect and, thus, may be quite different from a scene's\ntrue sound. Inspired by the challenges of creating a soundtrack for a video\nthat differs from its true sound, but that nonetheless matches the actions\noccurring on screen, we propose the problem of conditional Foley. We present\nthe following contributions to address this problem. First, we propose a\npretext task for training our model to predict sound for an input video clip\nusing a conditional audio-visual clip sampled from another time within the same\nsource video. Second, we propose a model for generating a soundtrack for a\nsilent input video, given a user-supplied example that specifies what the video\nshould \"sound like\". We show through human studies and automated evaluation\nmetrics that our model successfully generates sound from video, while varying\nits output according to the content of a supplied example. Project site:\nhttps://xypb.github.io/CondFoleyGen/\n", "date": "2023-04-17", "categories": []}
{"id": "2210.07839", "title": "Contrastive Audio-Visual Masked Autoencoder", "abstract": "  In this paper, we first extend the recent Masked Auto-Encoder (MAE) model\nfrom a single modality to audio-visual multi-modalities. Subsequently, we\npropose the Contrastive Audio-Visual Masked Auto-Encoder (CAV-MAE) by combining\ncontrastive learning and masked data modeling, two major self-supervised\nlearning frameworks, to learn a joint and coordinated audio-visual\nrepresentation. Our experiments show that the contrastive audio-visual\ncorrespondence learning objective not only enables the model to perform\naudio-visual retrieval tasks, but also helps the model learn a better joint\nrepresentation. As a result, our fully self-supervised pretrained CAV-MAE\nachieves a new SOTA accuracy of 65.9% on VGGSound, and is comparable with the\nprevious best supervised pretrained model on AudioSet in the audio-visual event\nclassification task. Code and pretrained models are at\nhttps://github.com/yuangongnd/cav-mae.\n", "date": "2022-10-02", "categories": []}
{"id": "2211.02247", "title": "Music Mixing Style Transfer: A Contrastive Learning Approach to\n  Disentangle Audio Effects", "abstract": "  We propose an end-to-end music mixing style transfer system that converts the\nmixing style of an input multitrack to that of a reference song. This is\nachieved with an encoder pre-trained with a contrastive objective to extract\nonly audio effects related information from a reference music recording. All\nour models are trained in a self-supervised manner from an already-processed\nwet multitrack dataset with an effective data preprocessing method that\nalleviates the data scarcity of obtaining unprocessed dry data. We analyze the\nproposed encoder for the disentanglement capability of audio effects and also\nvalidate its performance for mixing style transfer through both objective and\nsubjective evaluations. From the results, we show the proposed system not only\nconverts the mixing style of multitrack audio close to a reference but is also\nrobust with mixture-wise style transfer upon using a music source separation\nmodel.\n", "date": "2022-11-04", "categories": []}
{"id": "2304.03588", "title": "Anomalous Sound Detection using Audio Representation with Machine ID\n  based Contrastive Learning Pretraining", "abstract": "  Existing contrastive learning methods for anomalous sound detection refine\nthe audio representation of each audio sample by using the contrast between the\nsamples' augmentations (e.g., with time or frequency masking). However, they\nmight be biased by the augmented data, due to the lack of physical properties\nof machine sound, thereby limiting the detection performance. This paper uses\ncontrastive learning to refine audio representations for each machine ID,\nrather than for each audio sample. The proposed two-stage method uses\ncontrastive learning to pretrain the audio representation model by\nincorporating machine ID and a self-supervised ID classifier to fine-tune the\nlearnt model, while enhancing the relation between audio features from the same\nID. Experiments show that our method outperforms the state-of-the-art methods\nusing contrastive learning or self-supervised classification in overall anomaly\ndetection performance and stability on DCASE 2020 Challenge Task2 dataset.\n", "date": "2023-04-07", "categories": []}
{"id": "2304.03416", "title": "To Wake-up or Not to Wake-up: Reducing Keyword False Alarm by Successive\n  Refinement", "abstract": "  Keyword spotting systems continuously process audio streams to detect\nkeywords. One of the most challenging tasks in designing such systems is to\nreduce False Alarm (FA) which happens when the system falsely registers a\nkeyword despite the keyword not being uttered. In this paper, we propose a\nsimple yet elegant solution to this problem that follows from the law of total\nprobability. We show that existing deep keyword spotting mechanisms can be\nimproved by Successive Refinement, where the system first classifies whether\nthe input audio is speech or not, followed by whether the input is keyword-like\nor not, and finally classifies which keyword was uttered. We show across\nmultiple models with size ranging from 13K parameters to 2.41M parameters, the\nsuccessive refinement technique reduces FA by up to a factor of 8 on in-domain\nheld-out FA data, and up to a factor of 7 on out-of-domain (OOD) FA data.\nFurther, our proposed approach is \"plug-and-play\" and can be applied to any\ndeep keyword spotting model.\n", "date": "2023-04-06", "categories": []}
{"id": "2301.00508", "title": "EmoGator: A New Open Source Vocal Burst Dataset with Baseline Machine\n  Learning Classification Methodologies", "abstract": "  Vocal Bursts -- short, non-speech vocalizations that convey emotions, such as\nlaughter, cries, sighs, moans, and groans -- are an often-overlooked aspect of\nspeech emotion recognition, but an important aspect of human vocal\ncommunication. One barrier to study of these interesting vocalizations is a\nlack of large datasets. I am pleased to introduce the EmoGator dataset, which\nconsists of 32,130 samples from 357 speakers, 16.9654 hours of audio; each\nsample classified into one of 30 distinct emotion categories by the speaker.\nSeveral different approaches to construct classifiers to identify emotion\ncategories will be discussed, and directions for future research will be\nsuggested. Data set is available for download from\nhttps://github.com/fredbuhl/EmoGator.\n", "date": "2023-01-02", "categories": []}
{"id": "2304.02714", "title": "Learning Stage-wise GANs for Whistle Extraction in Time-Frequency\n  Spectrograms", "abstract": "  Whistle contour extraction aims to derive animal whistles from time-frequency\nspectrograms as polylines. For toothed whales, whistle extraction results can\nserve as the basis for analyzing animal abundance, species identity, and social\nactivities. During the last few decades, as long-term recording systems have\nbecome affordable, automated whistle extraction algorithms were proposed to\nprocess large volumes of recording data. Recently, a deep learning-based method\ndemonstrated superior performance in extracting whistles under varying noise\nconditions. However, training such networks requires a large amount of\nlabor-intensive annotation, which is not available for many species. To\novercome this limitation, we present a framework of stage-wise generative\nadversarial networks (GANs), which compile new whistle data suitable for deep\nmodel training via three stages: generation of background noise in the\nspectrogram, generation of whistle contours, and generation of whistle signals.\nBy separating the generation of different components in the samples, our\nframework composes visually promising whistle data and labels even when few\nexpert annotated data are available. Regardless of the amount of\nhuman-annotated data, the proposed data augmentation framework leads to a\nconsistent improvement in performance of the whistle extraction model, with a\nmaximum increase of 1.69 in the whistle extraction mean F1-score. Our\nstage-wise GAN also surpasses one single GAN in improving whistle extraction\nmodels with augmented data. The data and code will be available at\nhttps://github.com/Paul-LiPu/CompositeGAN\\_WhistleAugment.\n", "date": "2023-04-05", "categories": []}
{"id": "2304.02916", "title": "Efficient Audio Captioning Transformer with Patchout and Text Guidance", "abstract": "  Automated audio captioning is multi-modal translation task that aim to\ngenerate textual descriptions for a given audio clip. In this paper we propose\na full Transformer architecture that utilizes Patchout as proposed in [1],\nsignificantly reducing the computational complexity and avoiding overfitting.\nThe caption generation is partly conditioned on textual AudioSet tags extracted\nby a pre-trained classification model which is fine-tuned to maximize the\nsemantic similarity between AudioSet labels and ground truth captions. To\nmitigate the data scarcity problem of Automated Audio Captioning we introduce\ntransfer learning from an upstream audio-related task and an enlarged in-domain\ndataset. Moreover, we propose a method to apply Mixup augmentation for AAC.\nAblation studies are carried out to investigate how Patchout and text guidance\ncontribute to the final performance. The results show that the proposed\ntechniques improve the performance of our system and while reducing the\ncomputational complexity. Our proposed method received the Judges Award at the\nTask6A of DCASE Challenge 2022.\n", "date": "2023-04-06", "categories": []}
{"id": "2304.00830", "title": "AUDIT: Audio Editing by Following Instructions with Latent Diffusion\n  Models", "abstract": "  Audio editing is applicable for various purposes, such as adding background\nsound effects, replacing a musical instrument, and repairing damaged audio.\nRecently, some diffusion-based methods achieved zero-shot audio editing by\nusing a diffusion and denoising process conditioned on the text description of\nthe output audio. However, these methods still have some problems: 1) they have\nnot been trained on editing tasks and cannot ensure good editing effects; 2)\nthey can erroneously modify audio segments that do not require editing; 3) they\nneed a complete description of the output audio, which is not always available\nor necessary in practical scenarios. In this work, we propose AUDIT, an\ninstruction-guided audio editing model based on latent diffusion models.\nSpecifically, AUDIT has three main design features: 1) we construct triplet\ntraining data (instruction, input audio, output audio) for different audio\nediting tasks and train a diffusion model using instruction and input (to be\nedited) audio as conditions and generating output (edited) audio; 2) it can\nautomatically learn to only modify segments that need to be edited by comparing\nthe difference between the input and output audio; 3) it only needs edit\ninstructions instead of full target audio descriptions as text input. AUDIT\nachieves state-of-the-art results in both objective and subjective metrics for\nseveral audio editing tasks (e.g., adding, dropping, replacement, inpainting,\nsuper-resolution). Demo samples are available at https://audit-demo.github.io/.\n", "date": "2023-04-03", "categories": []}
{"id": "2304.02160", "title": "Pac-HuBERT: Self-Supervised Music Source Separation via Primitive\n  Auditory Clustering and Hidden-Unit BERT", "abstract": "  In spite of the progress in music source separation research, the small\namount of publicly-available clean source data remains a constant limiting\nfactor for performance. Thus, recent advances in self-supervised learning\npresent a largely-unexplored opportunity for improving separation models by\nleveraging unlabelled music data. In this paper, we propose a self-supervised\nlearning framework for music source separation inspired by the HuBERT speech\nrepresentation model. We first investigate the potential impact of the original\nHuBERT model by inserting an adapted version of it into the well-known Demucs\nV2 time-domain separation model architecture. We then propose a\ntime-frequency-domain self-supervised model, Pac-HuBERT (for primitive auditory\nclustering HuBERT), that we later use in combination with a Res-U-Net decoder\nfor source separation. Pac-HuBERT uses primitive auditory features of music as\nunsupervised clustering labels to initialize the self-supervised pretraining\nprocess using the Free Music Archive (FMA) dataset. The resulting framework\nachieves better source-to-distortion ratio (SDR) performance on the MusDB18\ntest set than the original Demucs V2 and Res-U-Net models. We further\ndemonstrate that it can boost performance with small amounts of supervised\ndata. Ultimately, our proposed framework is an effective solution to the\nchallenge of limited clean source data for music source separation.\n", "date": "2023-04-04", "categories": []}
{"id": "2303.17949", "title": "Unsupervised Anomaly Detection and Localization of Machine Audio: A\n  GAN-based Approach", "abstract": "  Automatic detection of machine anomaly remains challenging for machine\nlearning. We believe the capability of generative adversarial network (GAN)\nsuits the need of machine audio anomaly detection, yet rarely has this been\ninvestigated by previous work. In this paper, we propose AEGAN-AD, a totally\nunsupervised approach in which the generator (also an autoencoder) is trained\nto reconstruct input spectrograms. It is pointed out that the denoising nature\nof reconstruction deprecates its capacity. Thus, the discriminator is\nredesigned to aid the generator during both training stage and detection stage.\nThe performance of AEGAN-AD on the dataset of DCASE 2022 Challenge TASK 2\ndemonstrates the state-of-the-art result on five machine types. A novel anomaly\nlocalization method is also investigated. Source code available at:\nwww.github.com/jianganbai/AEGAN-AD\n", "date": "2023-03-31", "categories": []}
{"id": "2303.17056", "title": "Audio-Visual Grouping Network for Sound Localization from Mixtures", "abstract": "  Sound source localization is a typical and challenging task that predicts the\nlocation of sound sources in a video. Previous single-source methods mainly\nused the audio-visual association as clues to localize sounding objects in each\nimage. Due to the mixed property of multiple sound sources in the original\nspace, there exist rare multi-source approaches to localizing multiple sources\nsimultaneously, except for one recent work using a contrastive random walk in\nthe graph with images and separated sound as nodes. Despite their promising\nperformance, they can only handle a fixed number of sources, and they cannot\nlearn compact class-aware representations for individual sources. To alleviate\nthis shortcoming, in this paper, we propose a novel audio-visual grouping\nnetwork, namely AVGN, that can directly learn category-wise semantic features\nfor each source from the input audio mixture and image to localize multiple\nsources simultaneously. Specifically, our AVGN leverages learnable audio-visual\nclass tokens to aggregate class-aware source features. Then, the aggregated\nsemantic features for each source can be used as guidance to localize the\ncorresponding visual regions. Compared to existing multi-source methods, our\nnew framework can localize a flexible number of sources and disentangle\ncategory-aware audio-visual representations for individual sound sources. We\nconduct extensive experiments on MUSIC, VGGSound-Instruments, and VGG-Sound\nSources benchmarks. The results demonstrate that the proposed AVGN can achieve\nstate-of-the-art sounding object localization performance on both single-source\nand multi-source scenarios. Code is available at\n\\url{https://github.com/stoneMo/AVGN}.\n", "date": "2023-03-29", "categories": []}
{"id": "2303.11302", "title": "Learning Audio-Visual Source Localization via False Negative Aware\n  Contrastive Learning", "abstract": "  Self-supervised audio-visual source localization aims to locate sound-source\nobjects in video frames without extra annotations. Recent methods often\napproach this goal with the help of contrastive learning, which assumes only\nthe audio and visual contents from the same video are positive samples for each\nother. However, this assumption would suffer from false negative samples in\nreal-world training. For example, for an audio sample, treating the frames from\nthe same audio class as negative samples may mislead the model and therefore\nharm the learned representations e.g., the audio of a siren wailing may\nreasonably correspond to the ambulances in multiple images). Based on this\nobservation, we propose a new learning strategy named False Negative Aware\nContrastive (FNAC) to mitigate the problem of misleading the training with such\nfalse negative samples. Specifically, we utilize the intra-modal similarities\nto identify potentially similar samples and construct corresponding adjacency\nmatrices to guide contrastive learning. Further, we propose to strengthen the\nrole of true negative samples by explicitly leveraging the visual features of\nsound sources to facilitate the differentiation of authentic sounding source\nregions. FNAC achieves state-of-the-art performances on Flickr-SoundNet,\nVGG-Sound, and AVSBench, which demonstrates the effectiveness of our method in\nmitigating the false negative issue. The code is available at\n\\url{https://github.com/OpenNLPLab/FNAC_AVL}.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.15042", "title": "Partially Adaptive Multichannel Joint Reduction of Ego-noise and\n  Environmental Noise", "abstract": "  Human-robot interaction relies on a noise-robust audio processing module\ncapable of estimating target speech from audio recordings impacted by\nenvironmental noise, as well as self-induced noise, so-called ego-noise. While\nexternal ambient noise sources vary from environment to environment, ego-noise\nis mainly caused by the internal motors and joints of a robot. Ego-noise and\nenvironmental noise reduction are often decoupled, i.e., ego-noise reduction is\nperformed without considering environmental noise. Recently, a variational\nautoencoder (VAE)-based speech model has been combined with a fully adaptive\nnon-negative matrix factorization (NMF) noise model to recover clean speech\nunder different environmental noise disturbances. However, its enhancement\nperformance is limited in adverse acoustic scenarios involving, e.g. ego-noise.\nIn this paper, we propose a multichannel partially adaptive scheme to jointly\nmodel ego-noise and environmental noise utilizing the VAE-NMF framework, where\nwe take advantage of spatially and spectrally structured characteristics of\nego-noise by pre-training the ego-noise model, while retaining the ability to\nadapt to unknown environmental noise. Experimental results show that our\nproposed approach outperforms the methods based on a completely fixed scheme\nand a fully adaptive scheme when ego-noise and environmental noise are present\nsimultaneously.\n", "date": "2023-03-27", "categories": []}
{"id": "2303.13272", "title": "Frame-Level Multi-Label Playing Technique Detection Using Multi-Scale\n  Network and Self-Attention Mechanism", "abstract": "  Instrument playing technique (IPT) is a key element of musical presentation.\nHowever, most of the existing works for IPT detection only concern monophonic\nmusic signals, yet little has been done to detect IPTs in polyphonic\ninstrumental solo pieces with overlapping IPTs or mixed IPTs. In this paper, we\nformulate it as a frame-level multi-label classification problem and apply it\nto Guzheng, a Chinese plucked string instrument. We create a new dataset,\nGuzheng\\_Tech99, containing Guzheng recordings and onset, offset, pitch, IPT\nannotations of each note. Because different IPTs vary a lot in their lengths,\nwe propose a new method to solve this problem using multi-scale network and\nself-attention. The multi-scale network extracts features from different\nscales, and the self-attention mechanism applied to the feature maps at the\ncoarsest scale further enhances the long-range feature extraction. Our approach\noutperforms existing works by a large margin, indicating its effectiveness in\nIPT detection.\n", "date": "2023-03-23", "categories": []}
{"id": "2303.12187", "title": "Practice of the conformer enhanced AUDIO-VISUAL HUBERT on Mandarin and\n  English", "abstract": "  Considering the bimodal nature of human speech perception, lips, and teeth\nmovement has a pivotal role in automatic speech recognition. Benefiting from\nthe correlated and noise-invariant visual information, audio-visual recognition\nsystems enhance robustness in multiple scenarios. In previous work,\naudio-visual HuBERT appears to be the finest practice incorporating modality\nknowledge. This paper outlines a mixed methodology, named conformer enhanced\nAV-HuBERT, boosting the AV-HuBERT system's performance a step further. Compared\nwith baseline AV-HuBERT, our method in the one-phase evaluation of clean and\nnoisy conditions achieves 7% and 16% relative WER reduction on the English AVSR\nbenchmark dataset LRS3. Furthermore, we establish a novel 1000h Mandarin AVSR\ndataset CSTS. On top of the baseline AV-HuBERT, we exceed the WeNet ASR system\nby 14% and 18% relatively on MISP and CMLR by pre-training with this dataset.\nThe conformer-enhanced AV-HuBERT we proposed brings 7% on MISP and 6% CER\nreduction on CMLR, compared with the baseline AV-HuBERT system.\n", "date": "2023-02-28", "categories": []}
{"id": "2303.10757", "title": "Multiscale Audio Spectrogram Transformer for Efficient Audio\n  Classification", "abstract": "  Audio event has a hierarchical architecture in both time and frequency and\ncan be grouped together to construct more abstract semantic audio classes. In\nthis work, we develop a multiscale audio spectrogram Transformer (MAST) that\nemploys hierarchical representation learning for efficient audio\nclassification. Specifically, MAST employs one-dimensional (and\ntwo-dimensional) pooling operators along the time (and frequency domains) in\ndifferent stages, and progressively reduces the number of tokens and increases\nthe feature dimensions. MAST significantly outperforms AST~\\cite{gong2021ast}\nby 22.2\\%, 4.4\\% and 4.7\\% on Kinetics-Sounds, Epic-Kitchens-100 and VGGSound\nin terms of the top-1 accuracy without external training data. On the\ndownloaded AudioSet dataset, which has over 20\\% missing audios, MAST also\nachieves slightly better accuracy than AST. In addition, MAST is 5x more\nefficient in terms of multiply-accumulates (MACs) with 42\\% reduction in the\nnumber of parameters compared to AST. Through clustering metrics and\nvisualizations, we demonstrate that the proposed MAST can learn semantically\nmore separable feature representations from audio signals.\n", "date": "2023-03-19", "categories": []}
{"id": "2303.08362", "title": "Transfer Learning Based Diagnosis and Analysis of Lung Sound Aberrations", "abstract": "  With the development of computer -systems that can collect and analyze\nenormous volumes of data, the medical profession is establishing several\nnon-invasive tools. This work attempts to develop a non-invasive technique for\nidentifying respiratory sounds acquired by a stethoscope and voice recording\nsoftware via machine learning techniques. This study suggests a trained and\nproven CNN-based approach for categorizing respiratory sounds. A visual\nrepresentation of each audio sample is constructed, allowing resource\nidentification for classification using methods like those used to effectively\ndescribe visuals. We used a technique called Mel Frequency Cepstral\nCoefficients (MFCCs). Here, features are retrieved and categorized via VGG16\n(transfer learning) and prediction is accomplished using 5-fold\ncross-validation. Employing various data splitting techniques, Respiratory\nSound Database obtained cutting-edge results, including accuracy of 95%,\nprecision of 88%, recall score of 86%, and F1 score of 81%. The ICBHI dataset\nis used to train and test the model.\n", "date": "2023-03-15", "categories": []}
{"id": "2303.07449", "title": "Blind Acoustic Room Parameter Estimation Using Phase Features", "abstract": "  Modeling room acoustics in a field setting involves some degree of blind\nparameter estimation from noisy and reverberant audio. Modern approaches\nleverage convolutional neural networks (CNNs) in tandem with time-frequency\nrepresentation. Using short-time Fourier transforms to develop these\nspectrogram-like features has shown promising results, but this method\nimplicitly discards a significant amount of audio information in the phase\ndomain. Inspired by recent works in speech enhancement, we propose utilizing\nnovel phase-related features to extend recent approaches to blindly estimate\nthe so-called \"reverberation fingerprint\" parameters, namely, volume and RT60.\nThe addition of these features is shown to outperform existing methods that\nrely solely on magnitude-based spectral features across a wide range of\nacoustics spaces. We evaluate the effectiveness of the deployment of these\nnovel features in both single-parameter and multi-parameter estimation\nstrategies, using a novel dataset that consists of publicly available room\nimpulse responses (RIRs), synthesized RIRs, and in-house measurements of real\nacoustic spaces.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.07643", "title": "Feature-Rich Audio Model Inversion for Data-Free Knowledge Distillation\n  Towards General Sound Classification", "abstract": "  Data-Free Knowledge Distillation (DFKD) has recently attracted growing\nattention in the academic community, especially with major breakthroughs in\ncomputer vision. Despite promising results, the technique has not been well\napplied to audio and signal processing. Due to the variable duration of audio\nsignals, it has its own unique way of modeling. In this work, we propose\nfeature-rich audio model inversion (FRAMI), a data-free knowledge distillation\nframework for general sound classification tasks. It first generates\nhigh-quality and feature-rich Mel-spectrograms through a feature-invariant\ncontrastive loss. Then, the hidden states before and after the statistics\npooling layer are reused when knowledge distillation is performed on these\nfeature-rich samples. Experimental results on the Urbansound8k, ESC-50, and\naudioMNIST datasets demonstrate that FRAMI can generate feature-rich samples.\nMeanwhile, the accuracy of the student model is further improved by reusing the\nhidden state and significantly outperforms the baseline method.\n", "date": "2023-03-14", "categories": []}
{"id": "2210.11328", "title": "Play It Back: Iterative Attention for Audio Recognition", "abstract": "  A key function of auditory cognition is the association of characteristic\nsounds with their corresponding semantics over time. Humans attempting to\ndiscriminate between fine-grained audio categories, often replay the same\ndiscriminative sounds to increase their prediction confidence. We propose an\nend-to-end attention-based architecture that through selective repetition\nattends over the most discriminative sounds across the audio sequence. Our\nmodel initially uses the full audio sequence and iteratively refines the\ntemporal segments replayed based on slot attention. At each playback, the\nselected segments are replayed using a smaller hop length which represents\nhigher resolution features within these segments. We show that our method can\nconsistently achieve state-of-the-art performance across three\naudio-classification benchmarks: AudioSet, VGG-Sound, and EPIC-KITCHENS-100.\n", "date": "2022-10-20", "categories": []}
{"id": "2211.05103", "title": "Accidental Learners: Spoken Language Identification in Multilingual\n  Self-Supervised Models", "abstract": "  In this paper, we extend previous self-supervised approaches for language\nidentification by experimenting with Conformer based architecture in a\nmultilingual pre-training paradigm. We find that pre-trained speech models\noptimally encode language discriminatory information in lower layers. Further,\nwe demonstrate that the embeddings obtained from these layers are significantly\nrobust to classify unseen languages and different acoustic environments without\nadditional training. After fine-tuning a pre-trained Conformer model on the\nVoxLingua107 dataset, we achieve results similar to current state-of-the-art\nsystems for language identification. More, our model accomplishes this with 5x\nless parameters. We open-source the model through the NVIDIA NeMo toolkit.\n", "date": "2022-11-09", "categories": []}
{"id": "2211.10999", "title": "LA-VocE: Low-SNR Audio-visual Speech Enhancement using Neural Vocoders", "abstract": "  Audio-visual speech enhancement aims to extract clean speech from a noisy\nenvironment by leveraging not only the audio itself but also the target\nspeaker's lip movements. This approach has been shown to yield improvements\nover audio-only speech enhancement, particularly for the removal of interfering\nspeech. Despite recent advances in speech synthesis, most audio-visual\napproaches continue to use spectral mapping/masking to reproduce the clean\naudio, often resulting in visual backbones added to existing speech enhancement\narchitectures. In this work, we propose LA-VocE, a new two-stage approach that\npredicts mel-spectrograms from noisy audio-visual speech via a\ntransformer-based architecture, and then converts them into waveform audio\nusing a neural vocoder (HiFi-GAN). We train and evaluate our framework on\nthousands of speakers and 11+ different languages, and study our model's\nability to adapt to different levels of background noise and speech\ninterference. Our experiments show that LA-VocE outperforms existing methods\naccording to multiple metrics, particularly under very noisy scenarios.\n", "date": "2022-11-20", "categories": []}
{"id": "2301.02886", "title": "Perceptual-Neural-Physical Sound Matching", "abstract": "  Sound matching algorithms seek to approximate a target waveform by parametric\naudio synthesis. Deep neural networks have achieved promising results in\nmatching sustained harmonic tones. However, the task is more challenging when\ntargets are nonstationary and inharmonic, e.g., percussion. We attribute this\nproblem to the inadequacy of loss function. On one hand, mean square error in\nthe parametric domain, known as \"P-loss\", is simple and fast but fails to\naccommodate the differing perceptual significance of each parameter. On the\nother hand, mean square error in the spectrotemporal domain, known as \"spectral\nloss\", is perceptually motivated and serves in differentiable digital signal\nprocessing (DDSP). Yet, spectral loss is a poor predictor of pitch intervals\nand its gradient may be computationally expensive; hence a slow convergence.\nAgainst this conundrum, we present Perceptual-Neural-Physical loss (PNP). PNP\nis the optimal quadratic approximation of spectral loss while being as fast as\nP-loss during training. We instantiate PNP with physical modeling synthesis as\ndecoder and joint time-frequency scattering transform (JTFS) as spectral\nrepresentation. We demonstrate its potential on matching synthetic drum sounds\nin comparison with other loss functions.\n", "date": "2023-01-07", "categories": []}
{"id": "2303.06806", "title": "Neural Diarization with Non-autoregressive Intermediate Attractors", "abstract": "  End-to-end neural diarization (EEND) with encoder-decoder-based attractors\n(EDA) is a promising method to handle the whole speaker diarization problem\nsimultaneously with a single neural network. While the EEND model can produce\nall frame-level speaker labels simultaneously, it disregards output label\ndependency. In this work, we propose a novel EEND model that introduces the\nlabel dependency between frames. The proposed method generates\nnon-autoregressive intermediate attractors to produce speaker labels at the\nlower layers and conditions the subsequent layers with these labels. While the\nproposed model works in a non-autoregressive manner, the speaker labels are\nrefined by referring to the whole sequence of intermediate labels. The\nexperiments with the two-speaker CALLHOME dataset show that the intermediate\nlabels with the proposed non-autoregressive intermediate attractors boost the\ndiarization performance. The proposed method with the deeper network benefits\nmore from the intermediate labels, resulting in better performance and training\nthroughput than EEND-EDA.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.07143", "title": "Multi-Microphone Speaker Separation by Spatial Regions", "abstract": "  We consider the task of region-based source separation of reverberant\nmulti-microphone recordings. We assume pre-defined spatial regions with a\nsingle active source per region. The objective is to estimate the signals from\nthe individual spatial regions as captured by a reference microphone while\nretaining a correspondence between signals and spatial regions. We propose a\ndata-driven approach using a modified version of a state-of-the-art network,\nwhere different layers model spatial and spectro-temporal information. The\nnetwork is trained to enforce a fixed mapping of regions to network outputs.\nUsing speech from LibriMix, we construct a data set specifically designed to\ncontain the region information. Additionally, we train the network with\npermutation invariant training. We show that both training methods result in a\nfixed mapping of regions to network outputs, achieve comparable performance,\nand that the networks exploit spatial information. The proposed network\noutperforms a baseline network by 1.5 dB in scale-invariant\nsignal-to-distortion ratio.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.05678", "title": "Improving Weakly Supervised Sound Event Detection with Causal\n  Intervention", "abstract": "  Existing weakly supervised sound event detection (WSSED) work has not\nexplored both types of co-occurrences simultaneously, i.e., some sound events\noften co-occur, and their occurrences are usually accompanied by specific\nbackground sounds, so they would be inevitably entangled, causing\nmisclassification and biased localization results with only clip-level\nsupervision. To tackle this issue, we first establish a structural causal model\n(SCM) to reveal that the context is the main cause of co-occurrence confounders\nthat mislead the model to learn spurious correlations between frames and\nclip-level labels. Based on the causal analysis, we propose a causal\nintervention (CI) method for WSSED to remove the negative impact of\nco-occurrence confounders by iteratively accumulating every possible context of\neach class and then re-projecting the contexts to the frame-level features for\nmaking the event boundary clearer. Experiments show that our method effectively\nimproves the performance on multiple datasets and can generalize to various\nbaseline models.\n", "date": "2023-03-10", "categories": []}
{"id": "2303.05023", "title": "X-SepFormer: End-to-end Speaker Extraction Network with Explicit\n  Optimization on Speaker Confusion", "abstract": "  Target speech extraction (TSE) systems are designed to extract target speech\nfrom a multi-talker mixture. The popular training objective for most prior TSE\nnetworks is to enhance reconstruction performance of extracted speech waveform.\nHowever, it has been reported that a TSE system delivers high reconstruction\nperformance may still suffer low-quality experience problems in practice. One\nsuch experience problem is wrong speaker extraction (called speaker confusion,\nSC), which leads to strong negative experience and hampers effective\nconversations. To mitigate the imperative SC issue, we reformulate the training\nobjective and propose two novel loss schemes that explore the metric of\nreconstruction improvement performance defined at small chunk-level and\nleverage the metric associated distribution information. Both loss schemes aim\nto encourage a TSE network to pay attention to those SC chunks based on the\nsaid distribution information. On this basis, we present X-SepFormer, an\nend-to-end TSE model with proposed loss schemes and a backbone of SepFormer.\nExperimental results on the benchmark WSJ0-2mix dataset validate the\neffectiveness of our proposals, showing consistent improvements on SC errors\n(by 14.8% relative). Moreover, with SI-SDRi of 19.4 dB and PESQ of 3.81, our\nbest system significantly outperforms the current SOTA systems and offers the\ntop TSE results reported till date on the WSJ0-2mix.\n", "date": "2023-03-09", "categories": []}
{"id": "2211.05385", "title": "GANStrument: Adversarial Instrument Sound Synthesis with Pitch-invariant\n  Instance Conditioning", "abstract": "  We propose GANStrument, a generative adversarial model for instrument sound\nsynthesis. Given a one-shot sound as input, it is able to generate pitched\ninstrument sounds that reflect the timbre of the input within an interactive\ntime. By exploiting instance conditioning, GANStrument achieves better fidelity\nand diversity of synthesized sounds and generalization ability to various\ninputs. In addition, we introduce an adversarial training scheme for a\npitch-invariant feature extractor that significantly improves the pitch\naccuracy and timbre consistency. Experimental results show that GANStrument\noutperforms strong baselines that do not use instance conditioning in terms of\ngeneration quality and input editability. Qualitative examples are available\nonline.\n", "date": "2022-11-10", "categories": []}
{"id": "2303.03591", "title": "Approach to Learning Generalized Audio Representation Through Batch\n  Embedding Covariance Regularization and Constant-Q Transforms", "abstract": "  General-purpose embedding is highly desirable for few-shot even zero-shot\nlearning in many application scenarios, including audio tasks. In order to\nunderstand representations better, we conducted a thorough error analysis and\nvisualization of HEAR 2021 submission results. Inspired by the analysis, this\nwork experiments with different front-end audio preprocessing methods,\nincluding Constant-Q Transform (CQT) and Short-time Fourier transform (STFT),\nand proposes a Batch Embedding Covariance Regularization (BECR) term to uncover\na more holistic simulation of the frequency information received by the human\nauditory system. We tested the models on the suite of HEAR 2021 tasks, which\nencompass a broad category of tasks. Preliminary results show (1) the proposed\nBECR can incur a more dispersed embedding on the test set, (2) BECR improves\nthe PaSST model without extra computation complexity, and (3) STFT\npreprocessing outperforms CQT in all tasks we tested.\nGithub:https://github.com/ankitshah009/general_audio_embedding_hear_2021\n", "date": "2023-03-07", "categories": []}
{"id": "2303.03666", "title": "Face: Fast, Accurate and Context-Aware Audio Annotation and\n  Classification", "abstract": "  This paper presents a context-aware framework for feature selection and\nclassification procedures to realize a fast and accurate audio event annotation\nand classification. The context-aware design starts with exploring feature\nextraction techniques to find an appropriate combination to select a set\nresulting in remarkable classification accuracy with minimal computational\neffort. The exploration for feature selection also embraces an investigation of\naudio Tempo representation, an advantageous feature extraction method missed by\nprevious works in the environmental audio classification research scope. The\nproposed annotation method considers outlier, inlier, and hard-to-predict data\nsamples to realize context-aware Active Learning, leading to the average\naccuracy of 90% when only 15% of data possess initial annotation. Our proposed\nalgorithm for sound classification obtained average prediction accuracy of\n98.05% on the UrbanSound8K dataset. The notebooks containing our source codes\nand implementation results are available at https://github.com/gitmehrdad/FACE.\n", "date": "2023-03-07", "categories": []}
{"id": "2303.03732", "title": "A Multi-Stage Triple-Path Method for Speech Separation in Noisy and\n  Reverberant Environments", "abstract": "  In noisy and reverberant environments, the performance of deep learning-based\nspeech separation methods drops dramatically because previous methods are not\ndesigned and optimized for such situations. To address this issue, we propose a\nmulti-stage end-to-end learning method that decouples the difficult speech\nseparation problem in noisy and reverberant environments into three\nsub-problems: speech denoising, separation, and de-reverberation. The\nprobability and speed of searching for the optimal solution of the speech\nseparation model are improved by reducing the solution space. Moreover, since\nthe channel information of the audio sequence in the time domain is crucial for\nspeech separation, we propose a triple-path structure capable of modeling the\nchannel dimension of audio sequences. Experimental results show that the\nproposed multi-stage triple-path method can improve the performance of speech\nseparation models at the cost of little model parameter increment.\n", "date": "2023-03-07", "categories": []}
{"id": "2207.04213", "title": "Dual-Path Cross-Modal Attention for better Audio-Visual Speech\n  Extraction", "abstract": "  Audio-visual target speech extraction, which aims to extract a certain\nspeaker's speech from the noisy mixture by looking at lip movements, has made\nsignificant progress combining time-domain speech separation models and visual\nfeature extractors (CNN). One problem of fusing audio and video information is\nthat they have different time resolutions. Most current research upsamples the\nvisual features along the time dimension so that audio and video features are\nable to align in time. However, we believe that lip movement should mostly\ncontain long-term, or phone-level information. Based on this assumption, we\npropose a new way to fuse audio-visual features. We observe that for DPRNN\n\\cite{dprnn}, the interchunk dimension's time resolution could be very close to\nthe time resolution of video frames. Like \\cite{sepformer}, the LSTM in DPRNN\nis replaced by intra-chunk and inter-chunk self-attention, but in the proposed\nalgorithm, inter-chunk attention incorporates the visual features as an\nadditional feature stream. This prevents the upsampling of visual cues,\nresulting in more efficient audio-visual fusion. The result shows we achieve\nsuperior results compared with other time-domain based audio-visual fusion\nmodels.\n", "date": "2022-07-09", "categories": []}
{"id": "2210.12108", "title": "Adversarial Permutation Invariant Training for Universal Sound\n  Separation", "abstract": "  Universal sound separation consists of separating mixes with arbitrary sounds\nof different types, and permutation invariant training (PIT) is used to train\nsource agnostic models that do so. In this work, we complement PIT with\nadversarial losses but find it challenging with the standard formulation used\nin speech source separation. We overcome this challenge with a novel\nI-replacement context-based adversarial loss, and by training with multiple\ndiscriminators. Our experiments show that by simply improving the loss (keeping\nthe same model and dataset) we obtain a non-negligible improvement of 1.4 dB\nSI-SNRi in the reverberant FUSS dataset. We also find adversarial PIT to be\neffective at reducing spectral holes, ubiquitous in mask-based separation\nmodels, which highlights the potential relevance of adversarial losses for\nsource separation.\n", "date": "2022-10-21", "categories": []}
{"id": "2212.07065", "title": "CLIPSep: Learning Text-queried Sound Separation with Noisy Unlabeled\n  Videos", "abstract": "  Recent years have seen progress beyond domain-specific sound separation for\nspeech or music towards universal sound separation for arbitrary sounds. Prior\nwork on universal sound separation has investigated separating a target sound\nout of an audio mixture given a text query. Such text-queried sound separation\nsystems provide a natural and scalable interface for specifying arbitrary\ntarget sounds. However, supervised text-queried sound separation systems\nrequire costly labeled audio-text pairs for training. Moreover, the audio\nprovided in existing datasets is often recorded in a controlled environment,\ncausing a considerable generalization gap to noisy audio in the wild. In this\nwork, we aim to approach text-queried universal sound separation by using only\nunlabeled data. We propose to leverage the visual modality as a bridge to learn\nthe desired audio-textual correspondence. The proposed CLIPSep model first\nencodes the input query into a query vector using the contrastive\nlanguage-image pretraining (CLIP) model, and the query vector is then used to\ncondition an audio separation model to separate out the target sound. While the\nmodel is trained on image-audio pairs extracted from unlabeled videos, at test\ntime we can instead query the model with text inputs in a zero-shot setting,\nthanks to the joint language-image embedding learned by the CLIP model.\nFurther, videos in the wild often contain off-screen sounds and background\nnoise that may hinder the model from learning the desired audio-textual\ncorrespondence. To address this problem, we further propose an approach called\nnoise invariant training for training a query-based sound separation model on\nnoisy data. Experimental results show that the proposed models successfully\nlearn text-queried universal sound separation using only noisy unlabeled\nvideos, even achieving competitive performance against a supervised model in\nsome settings.\n", "date": "2022-12-14", "categories": []}
{"id": "2303.02396", "title": "A General Framework for Learning Procedural Audio Models of\n  Environmental Sounds", "abstract": "  This paper introduces the Procedural (audio) Variational autoEncoder (ProVE)\nframework as a general approach to learning Procedural Audio PA models of\nenvironmental sounds with an improvement to the realism of the synthesis while\nmaintaining provision of control over the generated sound through adjustable\nparameters. The framework comprises two stages: (i) Audio Class Representation,\nin which a latent representation space is defined by training an audio\nautoencoder, and (ii) Control Mapping, in which a joint function of\nstatic/temporal control variables derived from the audio and a random sample of\nuniform noise is learned to replace the audio encoder. We demonstrate the use\nof ProVE through the example of footstep sound effects on various surfaces. Our\nresults show that ProVE models outperform both classical PA models and an\nadversarial-based approach in terms of sound fidelity, as measured by Fr\\'echet\nAudio Distance (FAD), Maximum Mean Discrepancy (MMD), and subjective\nevaluations, making them feasible tools for sound design workflows.\n", "date": "2023-03-04", "categories": []}
{"id": "2303.02599", "title": "Hybrid Y-Net Architecture for Singing Voice Separation", "abstract": "  This research paper presents a novel deep learning-based neural network\narchitecture, named Y-Net, for achieving music source separation. The proposed\narchitecture performs end-to-end hybrid source separation by extracting\nfeatures from both spectrogram and waveform domains. Inspired by the U-Net\narchitecture, Y-Net predicts a spectrogram mask to separate vocal sources from\na mixture signal. Our results demonstrate the effectiveness of the proposed\narchitecture for music source separation with fewer parameters. Overall, our\nwork presents a promising approach for improving the accuracy and efficiency of\nmusic source separation.\n", "date": "2023-03-05", "categories": []}
{"id": "2303.01507", "title": "Defending against Adversarial Audio via Diffusion Model", "abstract": "  Deep learning models have been widely used in commercial acoustic systems in\nrecent years. However, adversarial audio examples can cause abnormal behaviors\nfor those acoustic systems, while being hard for humans to perceive. Various\nmethods, such as transformation-based defenses and adversarial training, have\nbeen proposed to protect acoustic systems from adversarial attacks, but they\nare less effective against adaptive attacks. Furthermore, directly applying the\nmethods from the image domain can lead to suboptimal results because of the\nunique properties of audio data. In this paper, we propose an adversarial\npurification-based defense pipeline, AudioPure, for acoustic systems via\noff-the-shelf diffusion models. Taking advantage of the strong generation\nability of diffusion models, AudioPure first adds a small amount of noise to\nthe adversarial audio and then runs the reverse sampling step to purify the\nnoisy audio and recover clean audio. AudioPure is a plug-and-play method that\ncan be directly applied to any pretrained classifier without any fine-tuning or\nre-training. We conduct extensive experiments on speech command recognition\ntask to evaluate the robustness of AudioPure. Our method is effective against\ndiverse adversarial attacks (e.g. $\\mathcal{L}_2$ or\n$\\mathcal{L}_\\infty$-norm). It outperforms the existing methods under both\nstrong adaptive white-box and black-box attacks bounded by $\\mathcal{L}_2$ or\n$\\mathcal{L}_\\infty$-norm (up to +20\\% in robust accuracy). Besides, we also\nevaluate the certified robustness for perturbations bounded by\n$\\mathcal{L}_2$-norm via randomized smoothing. Our pipeline achieves a higher\ncertified accuracy than baselines.\n", "date": "2023-03-02", "categories": []}
{"id": "2303.01884", "title": "AutoMatch: A Large-scale Audio Beat Matching Benchmark for Boosting Deep\n  Learning Assistant Video Editing", "abstract": "  The explosion of short videos has dramatically reshaped the manners people\nsocialize, yielding a new trend for daily sharing and access to the latest\ninformation. These rich video resources, on the one hand, benefited from the\npopularization of portable devices with cameras, but on the other, they can not\nbe independent of the valuable editing work contributed by numerous video\ncreators. In this paper, we investigate a novel and practical problem, namely\naudio beat matching (ABM), which aims to recommend the proper transition time\nstamps based on the background music. This technique helps to ease the\nlabor-intensive work during video editing, saving energy for creators so that\nthey can focus more on the creativity of video content. We formally define the\nABM problem and its evaluation protocol. Meanwhile, a large-scale audio\ndataset, i.e., the AutoMatch with over 87k finely annotated background music,\nis presented to facilitate this newly opened research direction. To further lay\nsolid foundations for the following study, we also propose a novel model termed\nBeatX to tackle this challenging task. Alongside, we creatively present the\nconcept of label scope, which eliminates the data imbalance issues and assigns\nadaptive weights for the ground truth during the training procedure in one\nstop. Though plentiful short video platforms have flourished for a long time,\nthe relevant research concerning this scenario is not sufficient, and to the\nbest of our knowledge, AutoMatch is the first large-scale dataset to tackle the\naudio beat matching problem. We hope the released dataset and our competitive\nbaseline can encourage more attention to this line of research. The dataset and\ncodes will be made publicly available.\n", "date": "2023-03-03", "categories": []}
{"id": "2208.01214", "title": "Audio Deepfake Detection Based on a Combination of F0 Information and\n  Real Plus Imaginary Spectrogram Features", "abstract": "  Recently, pioneer research works have proposed a large number of acoustic\nfeatures (log power spectrogram, linear frequency cepstral coefficients,\nconstant Q cepstral coefficients, etc.) for audio deepfake detection, obtaining\ngood performance, and showing that different subbands have different\ncontributions to audio deepfake detection. However, this lacks an explanation\nof the specific information in the subband, and these features also lose\ninformation such as phase. Inspired by the mechanism of synthetic speech, the\nfundamental frequency (F0) information is used to improve the quality of\nsynthetic speech, while the F0 of synthetic speech is still too average, which\ndiffers significantly from that of real speech. It is expected that F0 can be\nused as important information to discriminate between bonafide and fake speech,\nwhile this information cannot be used directly due to the irregular\ndistribution of F0. Insteadly, the frequency band containing most of F0 is\nselected as the input feature. Meanwhile, to make full use of the phase and\nfull-band information, we also propose to use real and imaginary spectrogram\nfeatures as complementary input features and model the disjoint subbands\nseparately. Finally, the results of F0, real and imaginary spectrogram features\nare fused. Experimental results on the ASVspoof 2019 LA dataset show that our\nproposed system is very effective for the audio deepfake detection task,\nachieving an equivalent error rate (EER) of 0.43%, which surpasses almost all\nsystems.\n", "date": "2022-08-02", "categories": []}
{"id": "2211.13956", "title": "Learning General Audio Representations with Large-Scale Training of\n  Patchout Audio Transformers", "abstract": "  The success of supervised deep learning methods is largely due to their\nability to learn relevant features from raw data. Deep Neural Networks (DNNs)\ntrained on large-scale datasets are capable of capturing a diverse set of\nfeatures, and learning a representation that can generalize onto unseen tasks\nand datasets that are from the same domain. Hence, these models can be used as\npowerful feature extractors, in combination with shallower models as\nclassifiers, for smaller tasks and datasets where the amount of training data\nis insufficient for learning an end-to-end model from scratch. During the past\nyears, Convolutional Neural Networks (CNNs) have largely been the method of\nchoice for audio processing. However, recently attention-based transformer\nmodels have demonstrated great potential in supervised settings, outperforming\nCNNs. In this work, we investigate the use of audio transformers trained on\nlarge-scale datasets to learn general-purpose representations. We study how the\ndifferent setups in these audio transformers affect the quality of their\nembeddings. We experiment with the models' time resolution, extracted embedding\nlevel, and receptive fields in order to see how they affect performance on a\nvariety of tasks and datasets, following the HEAR 2021 NeurIPS challenge\nevaluation setup. Our results show that representations extracted by audio\ntransformers outperform CNN representations. Furthermore, we will show that\ntransformers trained on Audioset can be extremely effective representation\nextractors for a wide range of downstream tasks.\n", "date": "2022-11-25", "categories": []}
{"id": "2302.08639", "title": "Improving Transformer-based Networks With Locality For Automatic Speaker\n  Verification", "abstract": "  Recently, Transformer-based architectures have been explored for speaker\nembedding extraction. Although the Transformer employs the self-attention\nmechanism to efficiently model the global interaction between token embeddings,\nit is inadequate for capturing short-range local context, which is essential\nfor the accurate extraction of speaker information. In this study, we enhance\nthe Transformer with the enhanced locality modeling in two directions. First,\nwe propose the Locality-Enhanced Conformer (LE-Confomer) by introducing\ndepth-wise convolution and channel-wise attention into the Conformer blocks.\nSecond, we present the Speaker Swin Transformer (SST) by adapting the Swin\nTransformer, originally proposed for vision tasks, into speaker embedding\nnetwork. We evaluate the proposed approaches on the VoxCeleb datasets and a\nlarge-scale Microsoft internal multilingual (MS-internal) dataset. The proposed\nmodels achieve 0.75% EER on VoxCeleb 1 test set, outperforming the previously\nproposed Transformer-based models and CNN-based models, such as ResNet34 and\nECAPA-TDNN. When trained on the MS-internal dataset, the proposed models\nachieve promising results with 14.6% relative reduction in EER over the\nRes2Net50 model.\n", "date": "2023-02-17", "categories": []}
{"id": "2302.14204", "title": "HalluAudio: Hallucinating Frequency as Concepts for Few-Shot Audio\n  Classification", "abstract": "  Few-shot audio classification is an emerging topic that attracts more and\nmore attention from the research community. Most existing work ignores the\nspecificity of the form of the audio spectrogram and focuses largely on the\nembedding space borrowed from image tasks, while in this work, we aim to take\nadvantage of this special audio format and propose a new method by\nhallucinating high-frequency and low-frequency parts as structured concepts.\nExtensive experiments on ESC-50 and our curated balanced Kaggle18 dataset show\nthe proposed method outperforms the baseline by a notable margin. The way that\nour method hallucinates high-frequency and low-frequency parts also enables its\ninterpretability and opens up new potentials for the few-shot audio\nclassification.\n", "date": "2023-02-27", "categories": []}
{"id": "2210.15283", "title": "On Out-of-Distribution Detection for Audio with Deep Nearest Neighbors", "abstract": "  Out-of-distribution (OOD) detection is concerned with identifying data points\nthat do not belong to the same distribution as the model's training data. For\nthe safe deployment of predictive models in a real-world environment, it is\ncritical to avoid making confident predictions on OOD inputs as it can lead to\npotentially dangerous consequences. However, OOD detection largely remains an\nunder-explored area in the audio (and speech) domain. This is despite the fact\nthat audio is a central modality for many tasks, such as speaker diarization,\nautomatic speech recognition, and sound event detection. To address this, we\npropose to leverage feature-space of the model with deep k-nearest neighbors to\ndetect OOD samples. We show that this simple and flexible method effectively\ndetects OOD inputs across a broad category of audio (and speech) datasets.\nSpecifically, it improves the false positive rate (FPR@TPR95) by 17% and the\nAUROC score by 7% than other prior techniques.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.15887", "title": "Nonparallel High-Quality Audio Super Resolution with Domain Adaptation\n  and Resampling CycleGANs", "abstract": "  Neural audio super-resolution models are typically trained on low- and\nhigh-resolution audio signal pairs. Although these methods achieve highly\naccurate super-resolution if the acoustic characteristics of the input data are\nsimilar to those of the training data, challenges remain: the models suffer\nfrom quality degradation for out-of-domain data, and paired data are required\nfor training. To address these problems, we propose Dual-CycleGAN, a\nhigh-quality audio super-resolution method that can utilize unpaired data based\non two connected cycle consistent generative adversarial networks (CycleGAN).\nOur method decomposes the super-resolution method into domain adaptation and\nresampling processes to handle acoustic mismatch in the unpaired low- and\nhigh-resolution signals. The two processes are then jointly optimized within\nthe CycleGAN framework. Experimental results verify that the proposed method\nsignificantly outperforms conventional methods when paired data are not\navailable. Code and audio samples are available from\nhttps://chomeyama.github.io/DualCycleGAN-Demo/.\n", "date": "2022-10-28", "categories": []}
{"id": "2211.12232", "title": "AERO: Audio Super Resolution in the Spectral Domain", "abstract": "  We present AERO, a audio super-resolution model that processes speech and\nmusic signals in the spectral domain. AERO is based on an encoder-decoder\narchitecture with U-Net like skip connections. We optimize the model using both\ntime and frequency domain loss functions. Specifically, we consider a set of\nreconstruction losses together with perceptual ones in the form of adversarial\nand feature discriminator loss functions. To better handle phase information\nthe proposed method operates over the complex-valued spectrogram using two\nseparate channels. Unlike prior work which mainly considers low and high\nfrequency concatenation for audio super-resolution, the proposed method\ndirectly predicts the full frequency range. We demonstrate high performance\nacross a wide range of sample rates considering both speech and music. AERO\noutperforms the evaluated baselines considering Log-Spectral Distance, ViSQOL,\nand the subjective MUSHRA test. Audio samples and code are available at\nhttps://pages.cs.huji.ac.il/adiyoss-lab/aero\n", "date": "2022-11-22", "categories": []}
{"id": "2212.08571", "title": "Statistical Design and Analysis for Robust Machine Learning: A Case\n  Study from COVID-19", "abstract": "  Since early in the coronavirus disease 2019 (COVID-19) pandemic, there has\nbeen interest in using artificial intelligence methods to predict COVID-19\ninfection status based on vocal audio signals, for example cough recordings.\nHowever, existing studies have limitations in terms of data collection and of\nthe assessment of the performances of the proposed predictive models. This\npaper rigorously assesses state-of-the-art machine learning techniques used to\npredict COVID-19 infection status based on vocal audio signals, using a dataset\ncollected by the UK Health Security Agency. This dataset includes acoustic\nrecordings and extensive study participant meta-data. We provide guidelines on\ntesting the performance of methods to classify COVID-19 infection status based\non acoustic features and we discuss how these can be extended more generally to\nthe development and assessment of predictive methods based on public health\ndatasets.\n", "date": "2022-12-15", "categories": []}
{"id": "2302.13401", "title": "From Audio to Symbolic Encoding", "abstract": "  Automatic music transcription (AMT) aims to convert raw audio to symbolic\nmusic representation. As a fundamental problem of music information retrieval\n(MIR), AMT is considered a difficult task even for trained human experts due to\noverlap of multiple harmonics in the acoustic signal. On the other hand, speech\nrecognition, as one of the most popular tasks in natural language processing,\naims to translate human spoken language to texts. Based on the similar nature\nof AMT and speech recognition (as they both deal with tasks of translating\naudio signal to symbolic encoding), this paper investigated whether a generic\nneural network architecture could possibly work on both tasks. In this paper,\nwe introduced our new neural network architecture built on top of the current\nstate-of-the-art Onsets and Frames, and compared the performances of its\nmultiple variations on AMT task. We also tested our architecture with the task\nof speech recognition. For AMT, our models were able to produce better results\ncompared to the model trained using the state-of-art architecture; however,\nalthough similar architecture was able to be trained on the speech recognition\ntask, it did not generate very ideal result compared to other task-specific\nmodels.\n", "date": "2023-02-26", "categories": []}
{"id": "2302.13542", "title": "Continuous descriptor-based control for deep audio synthesis", "abstract": "  Despite significant advances in deep models for music generation, the use of\nthese techniques remains restricted to expert users. Before being democratized\namong musicians, generative models must first provide expressive control over\nthe generation, as this conditions the integration of deep generative models in\ncreative workflows. In this paper, we tackle this issue by introducing a deep\ngenerative audio model providing expressive and continuous descriptor-based\ncontrol, while remaining lightweight enough to be embedded in a hardware\nsynthesizer. We enforce the controllability of real-time generation by\nexplicitly removing salient musical features in the latent space using an\nadversarial confusion criterion. User-specified features are then reintroduced\nas additional conditioning information, allowing for continuous control of the\ngeneration, akin to a synthesizer knob. We assess the performance of our method\non a wide variety of sounds including instrumental, percussive and speech\nrecordings while providing both timbre and attributes transfer, allowing new\nways of generating sounds.\n", "date": "2023-02-27", "categories": []}
{"id": "2211.10565", "title": "Filterbank Learning for Noise-Robust Small-Footprint Keyword Spotting", "abstract": "  In the context of keyword spotting (KWS), the replacement of handcrafted\nspeech features by learnable features has not yielded superior KWS performance.\nIn this study, we demonstrate that filterbank learning outperforms handcrafted\nspeech features for KWS whenever the number of filterbank channels is severely\ndecreased. Reducing the number of channels might yield certain KWS performance\ndrop, but also a substantial energy consumption reduction, which is key when\ndeploying common always-on KWS on low-resource devices. Experimental results on\na noisy version of the Google Speech Commands Dataset show that filterbank\nlearning adapts to noise characteristics to provide a higher degree of\nrobustness to noise, especially when dropout is integrated. Thus, switching\nfrom typically used 40-channel log-Mel features to 8-channel learned features\nleads to a relative KWS accuracy loss of only 3.5% while simultaneously\nachieving a 6.3x energy consumption reduction.\n", "date": "2022-11-19", "categories": []}
{"id": "2302.12434", "title": "Catch You and I Can: Revealing Source Voiceprint Against Voice\n  Conversion", "abstract": "  Voice conversion (VC) techniques can be abused by malicious parties to\ntransform their audios to sound like a target speaker, making it hard for a\nhuman being or a speaker verification/identification system to trace the source\nspeaker. In this paper, we make the first attempt to restore the source\nvoiceprint from audios synthesized by voice conversion methods with high\ncredit. However, unveiling the features of the source speaker from a converted\naudio is challenging since the voice conversion operation intends to\ndisentangle the original features and infuse the features of the target\nspeaker. To fulfill our goal, we develop Revelio, a representation learning\nmodel, which learns to effectively extract the voiceprint of the source speaker\nfrom converted audio samples. We equip Revelio with a carefully-designed\ndifferential rectification algorithm to eliminate the influence of the target\nspeaker by removing the representation component that is parallel to the\nvoiceprint of the target speaker. We have conducted extensive experiments to\nevaluate the capability of Revelio in restoring voiceprint from audios\nconverted by VQVC, VQVC+, AGAIN, and BNE. The experiments verify that Revelio\nis able to rebuild voiceprints that can be traced to the source speaker by\nspeaker verification and identification systems. Revelio also exhibits robust\nperformance under inter-gender conversion, unseen languages, and telephony\nnetworks.\n", "date": "2023-02-24", "categories": []}
{"id": "2210.05917", "title": "Enemy Spotted: in-game gun sound dataset for gunshot classification and\n  localization", "abstract": "  Recently, deep learning-based methods have drawn huge attention due to their\nsimple yet high performance without domain knowledge in sound classification\nand localization tasks. However, a lack of gun sounds in existing datasets has\nbeen a major obstacle to implementing a support system to spot criminals from\ntheir gunshots by leveraging deep learning models. Since the occurrence of\ngunshot is rare and unpredictable, it is impractical to collect gun sounds in\nthe real world. As an alternative, gun sounds can be obtained from an FPS game\nthat is designed to mimic real-world warfare. The recent FPS game offers a\nrealistic environment where we can safely collect gunshot data while simulating\neven dangerous situations. By exploiting the advantage of the game environment,\nwe construct a gunshot dataset, namely BGG, for the firearm classification and\ngunshot localization tasks. The BGG dataset consists of 37 different types of\nfirearms, distances, and directions between the sound source and a receiver. We\ncarefully verify that the in-game gunshot data has sufficient information to\nidentify the location and type of gunshots by training several sound\nclassification and localization baselines on the BGG dataset. Afterward, we\ndemonstrate that the accuracy of real-world firearm classification and\nlocalization tasks can be enhanced by utilizing the BGG dataset.\n", "date": "2022-10-12", "categories": []}
{"id": "2302.08632", "title": "jazznet: A Dataset of Fundamental Piano Patterns for Music Audio Machine\n  Learning Research", "abstract": "  This paper introduces the jazznet Dataset, a dataset of fundamental jazz\npiano music patterns for developing machine learning (ML) algorithms in music\ninformation retrieval (MIR). The dataset contains 162520 labeled piano\npatterns, including chords, arpeggios, scales, and chord progressions with\ntheir inversions, resulting in more than 26k hours of audio and a total size of\n95GB. The paper explains the dataset's composition, creation, and generation,\nand presents an open-source Pattern Generator using a method called\nDistance-Based Pattern Structures (DBPS), which allows researchers to easily\ngenerate new piano patterns simply by defining the distances between pitches\nwithin the musical patterns. We demonstrate that the dataset can help\nresearchers benchmark new models for challenging MIR tasks, using a\nconvolutional recurrent neural network (CRNN) and a deep convolutional neural\nnetwork. The dataset and code are available via:\nhttps://github.com/tosiron/jazznet.\n", "date": "2023-02-17", "categories": []}
{"id": "2206.04658", "title": "BigVGAN: A Universal Neural Vocoder with Large-Scale Training", "abstract": "  Despite recent progress in generative adversarial network (GAN)-based\nvocoders, where the model generates raw waveform conditioned on acoustic\nfeatures, it is challenging to synthesize high-fidelity audio for numerous\nspeakers across various recording environments. In this work, we present\nBigVGAN, a universal vocoder that generalizes well for various\nout-of-distribution scenarios without fine-tuning. We introduce periodic\nactivation function and anti-aliased representation into the GAN generator,\nwhich brings the desired inductive bias for audio synthesis and significantly\nimproves audio quality. In addition, we train our GAN vocoder at the largest\nscale up to 112M parameters, which is unprecedented in the literature. We\nidentify and address the failure modes in large-scale GAN training for audio,\nwhile maintaining high-fidelity output without over-regularization. Our\nBigVGAN, trained only on clean speech (LibriTTS), achieves the state-of-the-art\nperformance for various zero-shot (out-of-distribution) conditions, including\nunseen speakers, languages, recording environments, singing voices, music, and\ninstrumental audio. We release our code and model at:\nhttps://github.com/NVIDIA/BigVGAN\n", "date": "2022-06-09", "categories": []}
{"id": "2208.00061", "title": "UAVM: Towards Unifying Audio and Visual Models", "abstract": "  Conventional audio-visual models have independent audio and video branches.\nIn this work, we unify the audio and visual branches by designing a Unified\nAudio-Visual Model (UAVM). The UAVM achieves a new state-of-the-art\naudio-visual event classification accuracy of 65.8% on VGGSound. More\ninterestingly, we also find a few intriguing properties of UAVM that the\nmodality-independent counterparts do not have.\n", "date": "2022-07-29", "categories": []}
{"id": "2302.08130", "title": "Personalized Audio Quality Preference Prediction", "abstract": "  This paper proposes to use both audio input and subject information to\npredict the personalized preference of two audio segments with the same content\nin different qualities. A siamese network is used to compare the inputs and\npredict the preference. Several different structures for each side of the\nsiamese network are investigated, and an LDNet with PANNs' CNN6 as the encoder\nand a multi-layer perceptron block as the decoder outperforms a baseline model\nusing only audio input the most, where the overall accuracy grows from 77.56%\nto 78.04%. Experimental results also show that using all the subject\ninformation, including age, gender, and the specifications of headphones or\nearphones, is more effective than using only a part of them.\n", "date": "2023-02-16", "categories": []}
{"id": "2302.03498", "title": "MAC: A unified framework boosting low resource automatic speech\n  recognition", "abstract": "  We propose a unified framework for low resource automatic speech recognition\ntasks named meta audio concatenation (MAC). It is easy to implement and can be\ncarried out in extremely low resource environments. Mathematically, we give a\nclear description of MAC framework from the perspective of bayesian sampling.\nIn this framework, we leverage a novel concatenative synthesis text-to-speech\nsystem to boost the low resource ASR task. By the concatenative synthesis\ntext-to-speech system, we can integrate language pronunciation rules and adjust\nthe TTS process. Furthermore, we propose a broad notion of meta audio set to\nmeet the modeling needs of different languages and different scenes when using\nthe system. Extensive experiments have demonstrated the great effectiveness of\nMAC on low resource ASR tasks. For CTC greedy search, CTC prefix, attention,\nand attention rescoring decode mode in Cantonese ASR task, Taiwanese ASR task,\nand Japanese ASR task the MAC method can reduce the CER by more than 15\\%.\nFurthermore, in the ASR task, MAC beats wav2vec2 (with fine-tuning) on common\nvoice datasets of Cantonese and gets really competitive results on common voice\ndatasets of Taiwanese and Japanese. Among them, it is worth mentioning that we\nachieve a \\textbf{10.9\\%} character error rate (CER) on the common voice\nCantonese ASR task, bringing about \\textbf{30\\%} relative improvement compared\nto the wav2vec2 (with fine-tuning).\n", "date": "2023-02-05", "categories": []}
{"id": "2302.07315", "title": "A dataset for Audio-Visual Sound Event Detection in Movies", "abstract": "  Audio event detection is a widely studied audio processing task, with\napplications ranging from self-driving cars to healthcare. In-the-wild datasets\nsuch as Audioset have propelled research in this field. However, many efforts\ntypically involve manual annotation and verification, which is expensive to\nperform at scale. Movies depict various real-life and fictional scenarios which\nmakes them a rich resource for mining a wide-range of audio events. In this\nwork, we present a dataset of audio events called Subtitle-Aligned Movie Sounds\n(SAM-S). We use publicly-available closed-caption transcripts to automatically\nmine over 110K audio events from 430 movies. We identify three dimensions to\ncategorize audio events: sound, source, quality, and present the steps involved\nto produce a final taxonomy of 245 sounds. We discuss the choices involved in\ngenerating the taxonomy, and also highlight the human-centered nature of sounds\nin our dataset. We establish a baseline performance for audio-only sound\nclassification of 34.76% mean average precision and show that incorporating\nvisual information can further improve the performance by about 5%. Data and\ncode are made available for research at\nhttps://github.com/usc-sail/mica-subtitle-aligned-movie-sounds\n", "date": "2023-02-14", "categories": []}
{"id": "2302.05940", "title": "SemanticAC: Semantics-Assisted Framework for Audio Classification", "abstract": "  In this paper, we propose SemanticAC, a semantics-assisted framework for\nAudio Classification to better leverage the semantic information. Unlike\nconventional audio classification methods that treat class labels as discrete\nvectors, we employ a language model to extract abundant semantics from labels\nand optimize the semantic consistency between audio signals and their labels.\nWe verify that simple textual information from labels and advanced pretraining\nmodels enable more abundant semantic supervision for better performance.\nSpecifically, we design a text encoder to capture the semantic information from\nthe text extension of labels. Then we map the audio signals to align with the\nsemantics of corresponding class labels via an audio encoder and a similarity\ncalculation module so as to enforce the semantic consistency. Extensive\nexperiments on two audio datasets, ESC-50 and US8K demonstrate that our\nproposed method consistently outperforms the compared audio classification\nmethods.\n", "date": "2023-02-12", "categories": []}
{"id": "2302.02845", "title": "Audio Representation Learning by Distilling Video as Privileged\n  Information", "abstract": "  Deep audio representation learning using multi-modal audio-visual data often\nleads to a better performance compared to uni-modal approaches. However, in\nreal-world scenarios both modalities are not always available at the time of\ninference, leading to performance degradation by models trained for multi-modal\ninference. In this work, we propose a novel approach for deep audio\nrepresentation learning using audio-visual data when the video modality is\nabsent at inference. For this purpose, we adopt teacher-student knowledge\ndistillation under the framework of learning using privileged information\n(LUPI). While the previous methods proposed for LUPI use soft-labels generated\nby the teacher, in our proposed method we use embeddings learned by the teacher\nto train the student network. We integrate our method in two different\nsettings: sequential data where the features are divided into multiple segments\nthroughout time, and non-sequential data where the entire features are treated\nas one whole segment. In the non-sequential setting both the teacher and\nstudent networks are comprised of an encoder component and a task header. We\nuse the embeddings produced by the encoder component of the teacher to train\nthe encoder of the student, while the task header of the student is trained\nusing ground-truth labels. In the sequential setting, the networks have an\nadditional aggregation component that is placed between the encoder and task\nheader. We use two sets of embeddings produced by the encoder and aggregation\ncomponent of the teacher to train the student. Similar to the non-sequential\nsetting, the task header of the student network is trained using ground-truth\nlabels. We test our framework on two different audio-visual tasks, namely\nspeaker recognition and speech emotion recognition and show considerable\nimprovements over sole audio-based recognition as well as prior works that use\nLUPI.\n", "date": "2023-02-06", "categories": []}
{"id": "2302.00286", "title": "Jointist: Simultaneous Improvement of Multi-instrument Transcription and\n  Music Source Separation via Joint Training", "abstract": "  In this paper, we introduce Jointist, an instrument-aware multi-instrument\nframework that is capable of transcribing, recognizing, and separating multiple\nmusical instruments from an audio clip. Jointist consists of an instrument\nrecognition module that conditions the other two modules: a transcription\nmodule that outputs instrument-specific piano rolls, and a source separation\nmodule that utilizes instrument information and transcription results. The\njoint training of the transcription and source separation modules serves to\nimprove the performance of both tasks. The instrument module is optional and\ncan be directly controlled by human users. This makes Jointist a flexible\nuser-controllable framework. Our challenging problem formulation makes the\nmodel highly useful in the real world given that modern popular music typically\nconsists of multiple instruments. Its novelty, however, necessitates a new\nperspective on how to evaluate such a model. In our experiments, we assess the\nproposed model from various aspects, providing a new evaluation perspective for\nmulti-instrument transcription. Our subjective listening study shows that\nJointist achieves state-of-the-art performance on popular music, outperforming\nexisting multi-instrument transcription models such as MT3. We conducted\nexperiments on several downstream tasks and found that the proposed method\nimproved transcription by more than 1 percentage points (ppt.), source\nseparation by 5 SDR, downbeat detection by 1.8 ppt., chord recognition by 1.4\nppt., and key estimation by 1.4 ppt., when utilizing transcription results\nobtained from Jointist.\n  Demo available at \\url{https://jointist.github.io/Demo}.\n", "date": "2023-02-01", "categories": []}
{"id": "2206.02671", "title": "Canonical Cortical Graph Neural Networks and its Application for Speech\n  Enhancement in Audio-Visual Hearing Aids", "abstract": "  Despite the recent success of machine learning algorithms, most models face\ndrawbacks when considering more complex tasks requiring interaction between\ndifferent sources, such as multimodal input data and logical time sequences. On\nthe other hand, the biological brain is highly sharpened in this sense,\nempowered to automatically manage and integrate such streams of information. In\nthis context, this work draws inspiration from recent discoveries in brain\ncortical circuits to propose a more biologically plausible self-supervised\nmachine learning approach. This combines multimodal information using\nintra-layer modulations together with Canonical Correlation Analysis, and a\nmemory mechanism to keep track of temporal data, the overall approach termed\nCanonical Cortical Graph Neural networks. This is shown to outperform recent\nstate-of-the-art models in terms of clean audio reconstruction and energy\nefficiency for a benchmark audio-visual speech dataset. The enhanced\nperformance is demonstrated through a reduced and smother neuron firing rate\ndistribution. suggesting that the proposed model is amenable for speech\nenhancement in future audio-visual hearing aid devices.\n", "date": "2022-06-06", "categories": []}
{"id": "2209.12900", "title": "The Efficacy of Self-Supervised Speech Models for Audio Representations", "abstract": "  Self-supervised learning (SSL) speech models, which can serve as powerful\nupstream models to extract meaningful speech representations, have achieved\nunprecedented success in speech representation learning. However, their\neffectiveness on non-speech datasets is relatively less explored. In this work,\nwe propose an ensemble framework, with a combination of ensemble techniques, to\nfuse SSL speech models' embeddings. Extensive experiments on speech and\nnon-speech audio datasets are conducted to investigate the representation\nabilities of our ensemble method and its single constituent model. Ablation\nstudies are carried out to evaluate the performances of different ensemble\ntechniques, such as feature averaging and concatenation. All experiments are\nconducted during NeurIPS 2021 HEAR Challenge as a standard evaluation pipeline\nprovided by competition officials. Results demonstrate SSL speech models'\nstrong abilities on various non-speech tasks, while we also note that they fail\nto deal with fine-grained music tasks, such as pitch classification and note\nonset detection. In addition, feature ensemble is shown to have great potential\non producing more holistic representations, as our proposed framework generally\nsurpasses state-of-the-art SSL speech/audio models and has superior performance\non various datasets compared with other teams in HEAR Challenge. Our code is\navailable at https://github.com/tony10101105/HEAR-2021-NeurIPS-Challenge --\nNTU-GURA.\n", "date": "2022-09-26", "categories": []}
{"id": "2301.12662", "title": "SingSong: Generating musical accompaniments from singing", "abstract": "  We present SingSong, a system that generates instrumental music to accompany\ninput vocals, potentially offering musicians and non-musicians alike an\nintuitive new way to create music featuring their own voice. To accomplish\nthis, we build on recent developments in musical source separation and audio\ngeneration. Specifically, we apply a state-of-the-art source separation\nalgorithm to a large corpus of music audio to produce aligned pairs of vocals\nand instrumental sources. Then, we adapt AudioLM (Borsos et al., 2022) -- a\nstate-of-the-art approach for unconditional audio generation -- to be suitable\nfor conditional \"audio-to-audio\" generation tasks, and train it on the\nsource-separated (vocal, instrumental) pairs. In a pairwise comparison with the\nsame vocal inputs, listeners expressed a significant preference for\ninstrumentals generated by SingSong compared to those from a strong retrieval\nbaseline.\n  Sound examples at https://g.co/magenta/singsong\n", "date": "2023-01-30", "categories": []}
{"id": "2301.10183", "title": "Mesostructures: Beyond Spectrogram Loss in Differentiable Time-Frequency\n  Analysis", "abstract": "  Computer musicians refer to mesostructures as the intermediate levels of\narticulation between the microstructure of waveshapes and the macrostructure of\nmusical forms. Examples of mesostructures include melody, arpeggios,\nsyncopation, polyphonic grouping, and textural contrast. Despite their central\nrole in musical expression, they have received limited attention in deep\nlearning. Currently, autoencoders and neural audio synthesizers are only\ntrained and evaluated at the scale of microstructure: i.e., local amplitude\nvariations up to 100 milliseconds or so. In this paper, we formulate and\naddress the problem of mesostructural audio modeling via a composition of a\ndifferentiable arpeggiator and time-frequency scattering. We empirically\ndemonstrate that time--frequency scattering serves as a differentiable model of\nsimilarity between synthesis parameters that govern mesostructure. By exposing\nthe sensitivity of short-time spectral distances to time alignment, we motivate\nthe need for a time-invariant and multiscale differentiable time--frequency\nmodel of similarity at the level of both local spectra and spectrotemporal\nmodulations.\n", "date": "2023-01-24", "categories": []}
{"id": "2212.14618", "title": "Blind Restoration of Real-World Audio by 1D Operational GANs", "abstract": "  Objective: Despite numerous studies proposed for audio restoration in the\nliterature, most of them focus on an isolated restoration problem such as\ndenoising or dereverberation, ignoring other artifacts. Moreover, assuming a\nnoisy or reverberant environment with limited number of fixed\nsignal-to-distortion ratio (SDR) levels is a common practice. However,\nreal-world audio is often corrupted by a blend of artifacts such as\nreverberation, sensor noise, and background audio mixture with varying types,\nseverities, and duration. In this study, we propose a novel approach for blind\nrestoration of real-world audio signals by Operational Generative Adversarial\nNetworks (Op-GANs) with temporal and spectral objective metrics to enhance the\nquality of restored audio signal regardless of the type and severity of each\nartifact corrupting it. Methods: 1D Operational-GANs are used with generative\nneuron model optimized for blind restoration of any corrupted audio signal.\nResults: The proposed approach has been evaluated extensively over the\nbenchmark TIMIT-RAR (speech) and GTZAN-RAR (non-speech) datasets corrupted with\na random blend of artifacts each with a random severity to mimic real-world\naudio signals. Average SDR improvements of over 7.2 dB and 4.9 dB are achieved,\nrespectively, which are substantial when compared with the baseline methods.\nSignificance: This is a pioneer study in blind audio restoration with the\nunique capability of direct (time-domain) restoration of real-world audio\nwhilst achieving an unprecedented level of performance for a wide SDR range and\nartifact types. Conclusion: 1D Op-GANs can achieve robust and computationally\neffective real-world audio restoration with significantly improved performance.\nThe source codes and the generated real-world audio datasets are shared\npublicly with the research community in a dedicated GitHub repository1.\n", "date": "2022-12-30", "categories": []}
{"id": "2211.11060", "title": "Simultaneously Learning Robust Audio Embeddings and balanced Hash codes\n  for Query-by-Example", "abstract": "  Audio fingerprinting systems must efficiently and robustly identify query\nsnippets in an extensive database. To this end, state-of-the-art systems use\ndeep learning to generate compact audio fingerprints. These systems deploy\nindexing methods, which quantize fingerprints to hash codes in an unsupervised\nmanner to expedite the search. However, these methods generate imbalanced hash\ncodes, leading to their suboptimal performance. Therefore, we propose a\nself-supervised learning framework to compute fingerprints and balanced hash\ncodes in an end-to-end manner to achieve both fast and accurate retrieval\nperformance. We model hash codes as a balanced clustering process, which we\nregard as an instance of the optimal transport problem. Experimental results\nindicate that the proposed approach improves retrieval efficiency while\npreserving high accuracy, particularly at high distortion levels, compared to\nthe competing methods. Moreover, our system is efficient and scalable in\ncomputational load and memory storage.\n", "date": "2022-11-20", "categories": []}
{"id": "2301.07665", "title": "An investigation of the reconstruction capacity of stacked convolutional\n  autoencoders for log-mel-spectrograms", "abstract": "  In audio processing applications, the generation of expressive sounds based\non high-level representations demonstrates a high demand. These representations\ncan be used to manipulate the timbre and influence the synthesis of creative\ninstrumental notes. Modern algorithms, such as neural networks, have inspired\nthe development of expressive synthesizers based on musical instrument timbre\ncompression. Unsupervised deep learning methods can achieve audio compression\nby training the network to learn a mapping from waveforms or spectrograms to\nlow-dimensional representations. This study investigates the use of stacked\nconvolutional autoencoders for the compression of time-frequency audio\nrepresentations for a variety of instruments for a single pitch. Further\nexploration of hyper-parameters and regularization techniques is demonstrated\nto enhance the performance of the initial design. In an unsupervised manner,\nthe network is able to reconstruct a monophonic and harmonic sound based on\nlatent representations. In addition, we introduce an evaluation metric to\nmeasure the similarity between the original and reconstructed samples.\nEvaluating a deep generative model for the synthesis of sound is a challenging\ntask. Our approach is based on the accuracy of the generated frequencies as it\npresents a significant metric for the perception of harmonic sounds. This work\nis expected to accelerate future experiments on audio compression using neural\nautoencoders.\n", "date": "2023-01-18", "categories": []}
{"id": "2206.04780", "title": "Speak Like a Dog: Human to Non-human creature Voice Conversion", "abstract": "  This paper proposes a new voice conversion (VC) task from human speech to\ndog-like speech while preserving linguistic information as an example of human\nto non-human creature voice conversion (H2NH-VC) tasks. Although most VC\nstudies deal with human to human VC, H2NH-VC aims to convert human speech into\nnon-human creature-like speech. Non-parallel VC allows us to develop H2NH-VC,\nbecause we cannot collect a parallel dataset that non-human creatures speak\nhuman language. In this study, we propose to use dogs as an example of a\nnon-human creature target domain and define the \"speak like a dog\" task. To\nclarify the possibilities and characteristics of the \"speak like a dog\" task,\nwe conducted a comparative experiment using existing representative\nnon-parallel VC methods in acoustic features (Mel-cepstral coefficients and\nMel-spectrograms), network architectures (five different kernel-size settings),\nand training criteria (variational autoencoder (VAE)- based and generative\nadversarial network-based). Finally, the converted voices were evaluated using\nmean opinion scores: dog-likeness, sound quality and intelligibility, and\ncharacter error rate (CER). The experiment showed that the employment of the\nMel-spectrogram improved the dog-likeness of the converted speech, while it is\nchallenging to preserve linguistic information. Challenges and limitations of\nthe current VC methods for H2NH-VC are highlighted.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.13310", "title": "Insights Into Deep Non-linear Filters for Improved Multi-channel Speech\n  Enhancement", "abstract": "  The key advantage of using multiple microphones for speech enhancement is\nthat spatial filtering can be used to complement the tempo-spectral processing.\nIn a traditional setting, linear spatial filtering (beamforming) and\nsingle-channel post-filtering are commonly performed separately. In contrast,\nthere is a trend towards employing deep neural networks (DNNs) to learn a joint\nspatial and tempo-spectral non-linear filter, which means that the restriction\nof a linear processing model and that of a separate processing of spatial and\ntempo-spectral information can potentially be overcome. However, the internal\nmechanisms that lead to good performance of such data-driven filters for\nmulti-channel speech enhancement are not well understood. Therefore, in this\nwork, we analyse the properties of a non-linear spatial filter realized by a\nDNN as well as its interdependency with temporal and spectral processing by\ncarefully controlling the information sources (spatial, spectral, and temporal)\navailable to the network. We confirm the superiority of a non-linear spatial\nprocessing model, which outperforms an oracle linear spatial filter in a\nchallenging speaker extraction scenario for a low number of microphones by 0.24\nPOLQA score. Our analyses reveal that in particular spectral information should\nbe processed jointly with spatial information as this increases the spatial\nselectivity of the filter. Our systematic evaluation then leads to a simple\nnetwork architecture, that outperforms state-of-the-art network architectures\non a speaker extraction task by 0.22 POLQA score and by 0.32 POLQA score on the\nCHiME3 data.\n", "date": "2022-06-27", "categories": []}
{"id": "2301.06277", "title": "Improving Target Speaker Extraction with Sparse LDA-transformed Speaker\n  Embeddings", "abstract": "  As a practical alternative of speech separation, target speaker extraction\n(TSE) aims to extract the speech from the desired speaker using additional\nspeaker cue extracted from the speaker. Its main challenge lies in how to\nproperly extract and leverage the speaker cue to benefit the extracted speech\nquality. The cue extraction method adopted in majority existing TSE studies is\nto directly utilize discriminative speaker embedding, which is extracted from\nthe pre-trained models for speaker verification. Although the high speaker\ndiscriminability is a most desirable property for speaker verification task, we\nargue that it may be too sophisticated for TSE. In this study, we propose that\na simplified speaker cue with clear class separability might be preferred for\nTSE. To verify our proposal, we introduce several forms of speaker cues,\nincluding naive speaker embedding (such as, x-vector and xi-vector) and new\nspeaker embeddings produced from sparse LDA-transform. Corresponding TSE models\nare built by integrating these speaker cues with SepFormer (one SOTA speech\nseparation model). Performances of these TSE models are examined on the\nbenchmark WSJ0-2mix dataset. Experimental results validate the effectiveness\nand generalizability of our proposal, showing up to 9.9% relative improvement\nin SI-SDRi. Moreover, with SI-SDRi of 19.4 dB and PESQ of 3.78, our best TSE\nsystem significantly outperforms the current SOTA systems and offers the top\nTSE results reported till date on the WSJ0-2mix.\n", "date": "2023-01-16", "categories": []}
{"id": "2207.06405", "title": "Masked Autoencoders that Listen", "abstract": "  This paper studies a simple extension of image-based Masked Autoencoders\n(MAE) to self-supervised representation learning from audio spectrograms.\nFollowing the Transformer encoder-decoder design in MAE, our Audio-MAE first\nencodes audio spectrogram patches with a high masking ratio, feeding only the\nnon-masked tokens through encoder layers. The decoder then re-orders and\ndecodes the encoded context padded with mask tokens, in order to reconstruct\nthe input spectrogram. We find it beneficial to incorporate local window\nattention in the decoder, as audio spectrograms are highly correlated in local\ntime and frequency bands. We then fine-tune the encoder with a lower masking\nratio on target datasets. Empirically, Audio-MAE sets new state-of-the-art\nperformance on six audio and speech classification tasks, outperforming other\nrecent models that use external supervised pre-training. The code and models\nwill be at https://github.com/facebookresearch/AudioMAE.\n", "date": "2022-07-13", "categories": []}
{"id": "2301.02736", "title": "Using External Off-Policy Speech-To-Text Mappings in Contextual\n  End-To-End Automated Speech Recognition", "abstract": "  Despite improvements to the generalization performance of automated speech\nrecognition (ASR) models, specializing ASR models for downstream tasks remains\na challenging task, primarily due to reduced data availability (necessitating\nincreased data collection), and rapidly shifting data distributions (requiring\nmore frequent model fine-tuning). In this work, we investigate the potential of\nleveraging external knowledge, particularly through off-policy key-value stores\ngenerated with text-to-speech methods, to allow for flexible post-training\nadaptation to new data distributions. In our approach, audio embeddings\ncaptured from text-to-speech, along with semantic text embeddings, are used to\nbias ASR via an approximate k-nearest-neighbor (KNN) based attentive fusion\nstep. Our experiments on LibiriSpeech and in-house voice assistant/search\ndatasets show that the proposed approach can reduce domain adaptation time by\nup to 1K GPU-hours while providing up to 3% WER improvement compared to a\nfine-tuning baseline, suggesting a promising approach for adapting production\nASR systems in challenging zero and few-shot scenarios.\n", "date": "2023-01-06", "categories": []}
{"id": "2301.00145", "title": "Attentional Graph Convolutional Network for Structure-aware Audio-Visual\n  Scene Classification", "abstract": "  Audio-Visual scene understanding is a challenging problem due to the\nunstructured spatial-temporal relations that exist in the audio signals and\nspatial layouts of different objects and various texture patterns in the visual\nimages. Recently, many studies have focused on abstracting features from\nconvolutional neural networks while the learning of explicit semantically\nrelevant frames of sound signals and visual images has been overlooked. To this\nend, we present an end-to-end framework, namely attentional graph convolutional\nnetwork (AGCN), for structure-aware audio-visual scene representation. First,\nthe spectrogram of sound and input image is processed by a backbone network for\nfeature extraction. Then, to build multi-scale hierarchical information of\ninput features, we utilize an attention fusion mechanism to aggregate features\nfrom multiple layers of the backbone network. Notably, to well represent the\nsalient regions and contextual information of audio-visual inputs, the salient\nacoustic graph (SAG) and contextual acoustic graph (CAG), salient visual graph\n(SVG), and contextual visual graph (CVG) are constructed for the audio-visual\nscene representation. Finally, the constructed graphs pass through a graph\nconvolutional network for structure-aware audio-visual scene recognition.\nExtensive experimental results on the audio, visual and audio-visual scene\nrecognition datasets show that promising results have been achieved by the AGCN\nmethods. Visualizing graphs on the spectrograms and images have been presented\nto show the effectiveness of proposed CAG/SAG and CVG/SVG that could focus on\nthe salient and semantic relevant regions.\n", "date": "2022-12-31", "categories": []}
{"id": "2212.13369", "title": "Feature Selection Approaches for Optimising Music Emotion Recognition\n  Methods", "abstract": "  The high feature dimensionality is a challenge in music emotion recognition.\nThere is no common consensus on a relation between audio features and emotion.\nThe MER system uses all available features to recognize emotion; however, this\nis not an optimal solution since it contains irrelevant data acting as noise.\nIn this paper, we introduce a feature selection approach to eliminate redundant\nfeatures for MER. We created a Selected Feature Set (SFS) based on the feature\nselection algorithm (FSA) and benchmarked it by training with two models,\nSupport Vector Regression (SVR) and Random Forest (RF) and comparing them\nagainst with using the Complete Feature Set (CFS). The result indicates that\nthe performance of MER has improved for both Random Forest (RF) and Support\nVector Regression (SVR) models by using SFS. We found using FSA can improve\nperformance in all scenarios, and it has potential benefits for model\nefficiency and stability for MER task.\n", "date": "2022-12-27", "categories": []}
{"id": "2209.14345", "title": "Audio Barlow Twins: Self-Supervised Audio Representation Learning", "abstract": "  The Barlow Twins self-supervised learning objective requires neither negative\nsamples or asymmetric learning updates, achieving results on a par with the\ncurrent state-of-the-art within Computer Vision. As such, we present Audio\nBarlow Twins, a novel self-supervised audio representation learning approach,\nadapting Barlow Twins to the audio domain. We pre-train on the large-scale\naudio dataset AudioSet, and evaluate the quality of the learnt representations\non 18 tasks from the HEAR 2021 Challenge, achieving results which outperform,\nor otherwise are on a par with, the current state-of-the-art for instance\ndiscrimination self-supervised learning approaches to audio representation\nlearning. Code at https://github.com/jonahanton/SSL_audio.\n", "date": "2022-09-28", "categories": []}
{"id": "2212.11277", "title": "Audio Denoising for Robust Audio Fingerprinting", "abstract": "  Music discovery services let users identify songs from short mobile\nrecordings. These solutions are often based on Audio Fingerprinting, and rely\nmore specifically on the extraction of spectral peaks in order to be robust to\na number of distortions. Few works have been done to study the robustness of\nthese algorithms to background noise captured in real environments. In\nparticular, AFP systems still struggle when the signal to noise ratio is low,\ni.e when the background noise is strong. In this project, we tackle this\nproblematic with Deep Learning. We test a new hybrid strategy which consists of\ninserting a denoising DL model in front of a peak-based AFP algorithm. We\nsimulate noisy music recordings using a realistic data augmentation pipeline,\nand train a DL model to denoise them. The denoising model limits the impact of\nbackground noise on the AFP system's extracted peaks, improving its robustness\nto noise. We further propose a novel loss function to adapt the DL model to the\nconsidered AFP system, increasing its precision in terms of retrieved spectral\npeaks. To the best of our knowledge, this hybrid strategy has not been tested\nbefore.\n", "date": "2022-12-21", "categories": []}
{"id": "2212.11377", "title": "ReVISE: Self-Supervised Speech Resynthesis with Visual Input for\n  Universal and Generalized Speech Enhancement", "abstract": "  Prior works on improving speech quality with visual input typically study\neach type of auditory distortion separately (e.g., separation, inpainting,\nvideo-to-speech) and present tailored algorithms. This paper proposes to unify\nthese subjects and study Generalized Speech Enhancement, where the goal is not\nto reconstruct the exact reference clean signal, but to focus on improving\ncertain aspects of speech. In particular, this paper concerns intelligibility,\nquality, and video synchronization. We cast the problem as audio-visual speech\nresynthesis, which is composed of two steps: pseudo audio-visual speech\nrecognition (P-AVSR) and pseudo text-to-speech synthesis (P-TTS). P-AVSR and\nP-TTS are connected by discrete units derived from a self-supervised speech\nmodel. Moreover, we utilize self-supervised audio-visual speech model to\ninitialize P-AVSR. The proposed model is coined ReVISE. ReVISE is the first\nhigh-quality model for in-the-wild video-to-speech synthesis and achieves\nsuperior performance on all LRS3 audio-visual enhancement tasks with a single\nmodel. To demonstrates its applicability in the real world, ReVISE is also\nevaluated on EasyCom, an audio-visual benchmark collected under challenging\nacoustic conditions with only 1.6 hours of training data. Similarly, ReVISE\ngreatly suppresses noise and improves quality. Project page:\nhttps://wnhsu.github.io/ReVISE.\n", "date": "2022-12-21", "categories": []}
{"id": "2212.10093", "title": "Visual Transformers for Primates Classification and Covid Detection", "abstract": "  We apply the vision transformer, a deep machine learning model build around\nthe attention mechanism, on mel-spectrogram representations of raw audio\nrecordings. When adding mel-based data augmentation techniques and\nsample-weighting, we achieve comparable performance on both (PRS and CCS\nchallenge) tasks of ComParE21, outperforming most single model baselines. We\nfurther introduce overlapping vertical patching and evaluate the influence of\nparameter configurations. Index Terms: audio classification, attention,\nmel-spectrogram, unbalanced data-sets, computational paralinguistics\n", "date": "2022-12-20", "categories": []}
{"id": "2212.09058", "title": "BEATs: Audio Pre-Training with Acoustic Tokenizers", "abstract": "  The massive growth of self-supervised learning (SSL) has been witnessed in\nlanguage, vision, speech, and audio domains over the past few years. While\ndiscrete label prediction is widely adopted for other modalities, the\nstate-of-the-art audio SSL models still employ reconstruction loss for\npre-training. Compared with reconstruction loss, semantic-rich discrete label\nprediction encourages the SSL model to abstract the high-level audio semantics\nand discard the redundant details as in human perception. However, a\nsemantic-rich acoustic tokenizer for general audio pre-training is usually not\nstraightforward to obtain, due to the continuous property of audio and\nunavailable phoneme sequences like speech. To tackle this challenge, we propose\nBEATs, an iterative audio pre-training framework to learn Bidirectional Encoder\nrepresentation from Audio Transformers, where an acoustic tokenizer and an\naudio SSL model are optimized by iterations. In the first iteration, we use\nrandom projection as the acoustic tokenizer to train an audio SSL model in a\nmask and label prediction manner. Then, we train an acoustic tokenizer for the\nnext iteration by distilling the semantic knowledge from the pre-trained or\nfine-tuned audio SSL model. The iteration is repeated with the hope of mutual\npromotion of the acoustic tokenizer and audio SSL model. The experimental\nresults demonstrate our acoustic tokenizers can generate discrete labels with\nrich audio semantics and our audio SSL models achieve state-of-the-art results\nacross various audio classification benchmarks, even outperforming previous\nmodels that use more training data and model parameters significantly.\nSpecifically, we set a new state-of-the-art mAP 50.6% on AudioSet-2M for\naudio-only models without using any external data, and 98.1% accuracy on\nESC-50. The code and pre-trained models are available at https://aka.ms/beats.\n", "date": "2022-12-18", "categories": []}
{"id": "2212.04684", "title": "Machine Learning-based Classification of Birds through Birdsong", "abstract": "  Audio sound recognition and classification is used for many tasks and\napplications including human voice recognition, music recognition and audio\ntagging. In this paper we apply Mel Frequency Cepstral Coefficients (MFCC) in\ncombination with a range of machine learning models to identify (Australian)\nbirds from publicly available audio files of their birdsong. We present\napproaches used for data processing and augmentation and compare the results of\nvarious state of the art machine learning models. We achieve an overall\naccuracy of 91% for the top-5 birds from the 30 selected as the case study.\nApplying the models to more challenging and diverse audio files comprising 152\nbird species, we achieve an accuracy of 58%\n", "date": "2022-12-09", "categories": []}
{"id": "2212.03814", "title": "iQuery: Instruments as Queries for Audio-Visual Sound Separation", "abstract": "  Current audio-visual separation methods share a standard architecture design\nwhere an audio encoder-decoder network is fused with visual encoding features\nat the encoder bottleneck. This design confounds the learning of multi-modal\nfeature encoding with robust sound decoding for audio separation. To generalize\nto a new instrument: one must finetune the entire visual and audio network for\nall musical instruments. We re-formulate visual-sound separation task and\npropose Instrument as Query (iQuery) with a flexible query expansion mechanism.\nOur approach ensures cross-modal consistency and cross-instrument\ndisentanglement. We utilize \"visually named\" queries to initiate the learning\nof audio queries and use cross-modal attention to remove potential sound source\ninterference at the estimated waveforms. To generalize to a new instrument or\nevent class, drawing inspiration from the text-prompt design, we insert an\nadditional query as an audio prompt while freezing the attention mechanism.\nExperimental results on three benchmarks demonstrate that our iQuery improves\naudio-visual sound source separation performance.\n", "date": "2022-12-07", "categories": []}
{"id": "2212.04283", "title": "On The Relevance Of The Differences Between HRTF Measurement Setups For\n  Machine Learning", "abstract": "  As spatial audio is enjoying a surge in popularity, data-driven machine\nlearning techniques that have been proven successful in other domains are\nincreasingly used to process head-related transfer function measurements.\nHowever, these techniques require much data, whereas the existing datasets are\nranging from tens to the low hundreds of datapoints. It therefore becomes\nattractive to combine multiple of these datasets, although they are measured\nunder different conditions. In this paper, we first establish the common ground\nbetween a number of datasets, then we investigate potential pitfalls of mixing\ndatasets. We perform a simple experiment to test the relevance of the remaining\ndifferences between datasets when applying machine learning techniques.\nFinally, we pinpoint the most relevant differences.\n", "date": "2022-12-08", "categories": []}
{"id": "2212.02610", "title": "Audio Latent Space Cartography", "abstract": "  We explore the generation of visualisations of audio latent spaces using an\naudio-to-image generation pipeline. We believe this can help with the\ninterpretability of audio latent spaces. We demonstrate a variety of results on\nthe NSynth dataset. A web demo is available.\n", "date": "2022-12-05", "categories": []}
{"id": "2212.02508", "title": "MAP-Music2Vec: A Simple and Effective Baseline for Self-Supervised Music\n  Audio Representation Learning", "abstract": "  The deep learning community has witnessed an exponentially growing interest\nin self-supervised learning (SSL). However, it still remains unexplored how to\nbuild a framework for learning useful representations of raw music waveforms in\na self-supervised manner. In this work, we design Music2Vec, a framework\nexploring different SSL algorithmic components and tricks for music audio\nrecordings. Our model achieves comparable results to the state-of-the-art\n(SOTA) music SSL model Jukebox, despite being significantly smaller with less\nthan 2% of parameters of the latter. The model will be released on\nHuggingface(Please refer to: https://huggingface.co/m-a-p/music2vec-v1)\n", "date": "2022-12-05", "categories": []}
{"id": "2212.01884", "title": "Melody transcription via generative pre-training", "abstract": "  Despite the central role that melody plays in music perception, it remains an\nopen challenge in music information retrieval to reliably detect the notes of\nthe melody present in an arbitrary music recording. A key challenge in melody\ntranscription is building methods which can handle broad audio containing any\nnumber of instrument ensembles and musical styles - existing strategies work\nwell for some melody instruments or styles but not all. To confront this\nchallenge, we leverage representations from Jukebox (Dhariwal et al. 2020), a\ngenerative model of broad music audio, thereby improving performance on melody\ntranscription by $20$% relative to conventional spectrogram features. Another\nobstacle in melody transcription is a lack of training data - we derive a new\ndataset containing $50$ hours of melody transcriptions from crowdsourced\nannotations of broad music. The combination of generative pre-training and a\nnew dataset for this task results in $77$% stronger performance on melody\ntranscription relative to the strongest available baseline. By pairing our new\nmelody transcription approach with solutions for beat detection, key\nestimation, and chord recognition, we build Sheet Sage, a system capable of\ntranscribing human-readable lead sheets directly from music audio.\n  Audio examples can be found at https://chrisdonahue.com/sheetsage and code at\nhttps://github.com/chrisdonahue/sheetsage .\n", "date": "2022-12-04", "categories": []}
{"id": "2210.11068", "title": "Frequency of Interest-based Noise Attenuation Method to Improve Anomaly\n  Detection Performance", "abstract": "  Accurately extracting driving events is the way to maximize computational\nefficiency and anomaly detection performance in the tire frictional nose-based\nanomaly detection task. This study proposes a concise and highly useful method\nfor improving the precision of the event extraction that is hindered by extra\nnoise such as wind noise, which is difficult to characterize clearly due to its\nrandomness. The core of the proposed method is based on the identification of\nthe road friction sound corresponding to the frequency of interest and removing\nthe opposite characteristics with several frequency filters. Our method enables\nprecision maximization of driving event extraction while improving anomaly\ndetection performance by an average of 8.506%. Therefore, we conclude our\nmethod is a practical solution suitable for road surface anomaly detection\npurposes in outdoor edge computing environments.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.03581", "title": "Synthetic Voice Detection and Audio Splicing Detection using\n  SE-Res2Net-Conformer Architecture", "abstract": "  Synthetic voice and splicing audio clips have been generated to spoof\nInternet users and artificial intelligence (AI) technologies such as voice\nauthentication. Existing research work treats spoofing countermeasures as a\nbinary classification problem: bonafide vs. spoof. This paper extends the\nexisting Res2Net by involving the recent Conformer block to further exploit the\nlocal patterns on acoustic features. Experimental results on ASVspoof 2019\ndatabase show that the proposed SE-Res2Net-Conformer architecture is able to\nimprove the spoofing countermeasures performance for the logical access\nscenario.\n  In addition, this paper also proposes to re-formulate the existing audio\nsplicing detection problem. Instead of identifying the complete splicing\nsegments, it is more useful to detect the boundaries of the spliced segments.\nMoreover, a deep learning approach can be used to solve the problem, which is\ndifferent from the previous signal processing techniques.\n", "date": "2022-10-07", "categories": []}
{"id": "2207.07036", "title": "u-HuBERT: Unified Mixed-Modal Speech Pretraining And Zero-Shot Transfer\n  to Unlabeled Modality", "abstract": "  While audio-visual speech models can yield superior performance and\nrobustness compared to audio-only models, their development and adoption are\nhindered by the lack of labeled and unlabeled audio-visual data and the cost to\ndeploy one model per modality. In this paper, we present u-HuBERT, a\nself-supervised pre-training framework that can leverage both multimodal and\nunimodal speech with a unified masked cluster prediction objective. By\nutilizing modality dropout during pre-training, we demonstrate that a single\nfine-tuned model can achieve performance on par or better than the\nstate-of-the-art modality-specific models. Moreover, our model fine-tuned only\non audio can perform well with audio-visual and visual speech input, achieving\nzero-shot modality generalization for multiple speech processing tasks. In\nparticular, our single model yields 1.2%/1.4%/27.2% speech recognition word\nerror rate on LRS3 with audio-visual/audio/visual input. Codes and models are\navailable at https://github.com/facebookresearch/av_hubert\n", "date": "2022-07-14", "categories": []}
{"id": "2211.15524", "title": "Differentiable Dictionary Search: Integrating Linear Mixing with Deep\n  Non-Linear Modelling for Audio Source Separation", "abstract": "  This paper describes several improvements to a new method for signal\ndecomposition that we recently formulated under the name of Differentiable\nDictionary Search (DDS). The fundamental idea of DDS is to exploit a class of\npowerful deep invertible density estimators called normalizing flows, to model\nthe dictionary in a linear decomposition method such as NMF, effectively\ncreating a bijection between the space of dictionary elements and the\nassociated probability space, allowing a differentiable search through the\ndictionary space, guided by the estimated densities. As the initial formulation\nwas a proof of concept with some practical limitations, we will present several\nsteps towards making it scalable, hoping to improve both the computational\ncomplexity of the method and its signal decomposition capabilities. As a\ntestbed for experimental evaluation, we choose the task of frame-level piano\ntranscription, where the signal is to be decomposed into sources whose activity\nis attributed to individual piano notes. To highlight the impact of improved\nnon-linear modelling of sources, we compare variants of our method to a linear\novercomplete NMF baseline. Experimental results will show that even in the\nabsence of additional constraints, our models produce increasingly sparse and\nprecise decompositions, according to two pertinent evaluation measures.\n", "date": "2022-11-28", "categories": []}
{"id": "2206.04006", "title": "Few-Shot Audio-Visual Learning of Environment Acoustics", "abstract": "  Room impulse response (RIR) functions capture how the surrounding physical\nenvironment transforms the sounds heard by a listener, with implications for\nvarious applications in AR, VR, and robotics. Whereas traditional methods to\nestimate RIRs assume dense geometry and/or sound measurements throughout the\nenvironment, we explore how to infer RIRs based on a sparse set of images and\nechoes observed in the space. Towards that goal, we introduce a\ntransformer-based method that uses self-attention to build a rich acoustic\ncontext, then predicts RIRs of arbitrary query source-receiver locations\nthrough cross-attention. Additionally, we design a novel training objective\nthat improves the match in the acoustic signature between the RIR predictions\nand the targets. In experiments using a state-of-the-art audio-visual simulator\nfor 3D environments, we demonstrate that our method successfully generates\narbitrary RIRs, outperforming state-of-the-art methods and -- in a major\ndeparture from traditional methods -- generalizing to novel environments in a\nfew-shot manner. Project: http://vision.cs.utexas.edu/projects/fs_rir.\n", "date": "2022-06-08", "categories": []}
{"id": "2211.12623", "title": "SkipConvGAN: Monaural Speech Dereverberation using Generative\n  Adversarial Networks via Complex Time-Frequency Masking", "abstract": "  With the advancements in deep learning approaches, the performance of speech\nenhancing systems in the presence of background noise have shown significant\nimprovements. However, improving the system's robustness against reverberation\nis still a work in progress, as reverberation tends to cause loss of formant\nstructure due to smearing effects in time and frequency. A wide range of deep\nlearning-based systems either enhance the magnitude response and reuse the\ndistorted phase or enhance complex spectrogram using a complex time-frequency\nmask. Though these approaches have demonstrated satisfactory performance, they\ndo not directly address the lost formant structure caused by reverberation. We\nbelieve that retrieving the formant structure can help improve the efficiency\nof existing systems. In this study, we propose SkipConvGAN - an extension of\nour prior work SkipConvNet. The proposed system's generator network tries to\nestimate an efficient complex time-frequency mask, while the discriminator\nnetwork aids in driving the generator to restore the lost formant structure. We\nevaluate the performance of our proposed system on simulated and real\nrecordings of reverberant speech from the single-channel task of the REVERB\nchallenge corpus. The proposed system shows a consistent improvement across\nmultiple room configurations over other deep learning-based generative\nadversarial frameworks.\n", "date": "2022-11-22", "categories": []}
{"id": "2206.05876", "title": "Description and Discussion on DCASE 2022 Challenge Task 2: Unsupervised\n  Anomalous Sound Detection for Machine Condition Monitoring Applying Domain\n  Generalization Techniques", "abstract": "  We present the task description and discussion on the results of the DCASE\n2022 Challenge Task 2: ``Unsupervised anomalous sound detection (ASD) for\nmachine condition monitoring applying domain generalization techniques''.\nDomain shifts are a critical problem for the application of ASD systems.\nBecause domain shifts can change the acoustic characteristics of data, a model\ntrained in a source domain performs poorly for a target domain. In DCASE 2021\nChallenge Task 2, we organized an ASD task for handling domain shifts. In this\ntask, it was assumed that the occurrences of domain shifts are known. However,\nin practice, the domain of each sample may not be given, and the domain shifts\ncan occur implicitly. In 2022 Task 2, we focus on domain generalization\ntechniques that detects anomalies regardless of the domain shifts.\nSpecifically, the domain of each sample is not given in the test data and only\none threshold is allowed for all domains. Analysis of 81 submissions from 31\nteams revealed two remarkable types of domain generalization techniques: 1)\ndomain-mixing-based approach that obtains generalized representations and 2)\ndomain-classification-based approach that explicitly or implicitly classifies\ndifferent domains to improve detection performance for each domain.\n", "date": "2022-06-13", "categories": []}
{"id": "2211.10666", "title": "VarietySound: Timbre-Controllable Video to Sound Generation via\n  Unsupervised Information Disentanglement", "abstract": "  Video to sound generation aims to generate realistic and natural sound given\na video input. However, previous video-to-sound generation methods can only\ngenerate a random or average timbre without any controls or specializations of\nthe generated sound timbre, leading to the problem that people cannot obtain\nthe desired timbre under these methods sometimes. In this paper, we pose the\ntask of generating sound with a specific timbre given a video input and a\nreference audio sample. To solve this task, we disentangle each target sound\naudio into three components: temporal information, acoustic information, and\nbackground information. We first use three encoders to encode these components\nrespectively: 1) a temporal encoder to encode temporal information, which is\nfed with video frames since the input video shares the same temporal\ninformation as the original audio; 2) an acoustic encoder to encode timbre\ninformation, which takes the original audio as input and discards its temporal\ninformation by a temporal-corrupting operation; and 3) a background encoder to\nencode the residual or background sound, which uses the background part of the\noriginal audio as input. To make the generated result achieve better quality\nand temporal alignment, we also adopt a mel discriminator and a temporal\ndiscriminator for the adversarial training. Our experimental results on the VAS\ndataset demonstrate that our method can generate high-quality audio samples\nwith good synchronization with events in video and high timbre similarity with\nthe reference audio.\n", "date": "2022-11-19", "categories": []}
{"id": "2211.11225", "title": "TimbreCLIP: Connecting Timbre to Text and Images", "abstract": "  We present work in progress on TimbreCLIP, an audio-text cross modal\nembedding trained on single instrument notes. We evaluate the models with a\ncross-modal retrieval task on synth patches. Finally, we demonstrate the\napplication of TimbreCLIP on two tasks: text-driven audio equalization and\ntimbre to image generation.\n", "date": "2022-11-21", "categories": []}
{"id": "2211.09898", "title": "Audio Anti-spoofing Using a Simple Attention Module and Joint\n  Optimization Based on Additive Angular Margin Loss and Meta-learning", "abstract": "  Automatic speaker verification systems are vulnerable to a variety of access\nthreats, prompting research into the formulation of effective spoofing\ndetection systems to act as a gate to filter out such spoofing attacks. This\nstudy introduces a simple attention module to infer 3-dim attention weights for\nthe feature map in a convolutional layer, which then optimizes an energy\nfunction to determine each neuron's importance. With the advancement of both\nvoice conversion and speech synthesis technologies, unseen spoofing attacks are\nconstantly emerging to limit spoofing detection system performance. Here, we\npropose a joint optimization approach based on the weighted additive angular\nmargin loss for binary classification, with a meta-learning training framework\nto develop an efficient system that is robust to a wide range of spoofing\nattacks for model generalization enhancement. As a result, when compared to\ncurrent state-of-the-art systems, our proposed approach delivers a competitive\nresult with a pooled EER of 0.99% and min t-DCF of 0.0289.\n", "date": "2022-11-17", "categories": []}
{"id": "2211.09913", "title": "Multi-source Domain Adaptation for Text-independent Forensic Speaker\n  Recognition", "abstract": "  Adapting speaker recognition systems to new environments is a widely-used\ntechnique to improve a well-performing model learned from large-scale data\ntowards a task-specific small-scale data scenarios. However, previous studies\nfocus on single domain adaptation, which neglects a more practical scenario\nwhere training data are collected from multiple acoustic domains needed in\nforensic scenarios. Audio analysis for forensic speaker recognition offers\nunique challenges in model training with multi-domain training data due to\nlocation/scenario uncertainty and diversity mismatch between reference and\nnaturalistic field recordings. It is also difficult to directly employ\nsmall-scale domain-specific data to train complex neural network architectures\ndue to domain mismatch and performance loss. Fine-tuning is a commonly-used\nmethod for adaptation in order to retrain the model with weights initialized\nfrom a well-trained model. Alternatively, in this study, three novel adaptation\nmethods based on domain adversarial training, discrepancy minimization, and\nmoment-matching approaches are proposed to further promote adaptation\nperformance across multiple acoustic domains. A comprehensive set of\nexperiments are conducted to demonstrate that: 1) diverse acoustic environments\ndo impact speaker recognition performance, which could advance research in\naudio forensics, 2) domain adversarial training learns the discriminative\nfeatures which are also invariant to shifts between domains, 3)\ndiscrepancy-minimizing adaptation achieves effective performance simultaneously\nacross multiple acoustic domains, and 4) moment-matching adaptation along with\ndynamic distribution alignment also significantly promotes speaker recognition\nperformance on each domain, especially for the LENA-field domain with noise\ncompared to all other systems.\n", "date": "2022-11-17", "categories": []}
{"id": "2207.07307", "title": "MIMO-DoAnet: Multi-channel Input and Multiple Outputs DoA Network with\n  Unknown Number of Sound Sources", "abstract": "  Recent neural network based Direction of Arrival (DoA) estimation algorithms\nhave performed well on unknown number of sound sources scenarios. These\nalgorithms are usually achieved by mapping the multi-channel audio input to the\nsingle output (i.e. overall spatial pseudo-spectrum (SPS) of all sources), that\nis called MISO. However, such MISO algorithms strongly depend on empirical\nthreshold setting and the angle assumption that the angles between the sound\nsources are greater than a fixed angle. To address these limitations, we\npropose a novel multi-channel input and multiple outputs DoA network called\nMIMO-DoAnet. Unlike the general MISO algorithms, MIMO-DoAnet predicts the SPS\ncoding of each sound source with the help of the informative spatial covariance\nmatrix. By doing so, the threshold task of detecting the number of sound\nsources becomes an easier task of detecting whether there is a sound source in\neach output, and the serious interaction between sound sources disappears\nduring inference stage. Experimental results show that MIMO-DoAnet achieves\nrelative 18.6% and absolute 13.3%, relative 34.4% and absolute 20.2% F1 score\nimprovement compared with the MISO baseline system in 3, 4 sources scenes. The\nresults also demonstrate MIMO-DoAnet alleviates the threshold setting problem\nand solves the angle assumption problem effectively.\n", "date": "2022-07-15", "categories": []}
{"id": "2211.08715", "title": "Conditional variational autoencoder to improve neural audio synthesis\n  for polyphonic music sound", "abstract": "  Deep generative models for audio synthesis have recently been significantly\nimproved. However, the task of modeling raw-waveforms remains a difficult\nproblem, especially for audio waveforms and music signals. Recently, the\nrealtime audio variational autoencoder (RAVE) method was developed for\nhigh-quality audio waveform synthesis. The RAVE method is based on the\nvariational autoencoder and utilizes the two-stage training strategy.\nUnfortunately, the RAVE model is limited in reproducing wide-pitch polyphonic\nmusic sound. Therefore, to enhance the reconstruction performance, we adopt the\npitch activation data as an auxiliary information to the RAVE model. To handle\nthe auxiliary information, we propose an enhanced RAVE model with a conditional\nvariational autoencoder structure and an additional fully-connected layer. To\nevaluate the proposed structure, we conducted a listening experiment based on\nmultiple stimulus tests with hidden references and an anchor (MUSHRA) with the\nMAESTRO. The obtained results indicate that the proposed model exhibits a more\nsignificant performance and stability improvement than the conventional RAVE\nmodel.\n", "date": "2022-11-16", "categories": []}
{"id": "2211.08872", "title": "McNet: Fuse Multiple Cues for Multichannel Speech Enhancement", "abstract": "  In multichannel speech enhancement, both spectral and spatial information are\nvital for discriminating between speech and noise. How to fully exploit these\ntwo types of information and their temporal dynamics remains an interesting\nresearch problem. As a solution to this problem, this paper proposes a\nmulti-cue fusion network named McNet, which cascades four modules to\nrespectively exploit the full-band spatial, narrow-band spatial, sub-band\nspectral, and full-band spectral information. Experiments show that each module\nin the proposed network has its unique contribution and, as a whole, notably\noutperforms other state-of-the-art methods.\n", "date": "2022-11-16", "categories": []}
{"id": "2211.07951", "title": "Show Me the Instruments: Musical Instrument Retrieval from Mixture Audio", "abstract": "  As digital music production has become mainstream, the selection of\nappropriate virtual instruments plays a crucial role in determining the quality\nof music. To search the musical instrument samples or virtual instruments that\nmake one's desired sound, music producers use their ears to listen and compare\neach instrument sample in their collection, which is time-consuming and\ninefficient. In this paper, we call this task as Musical Instrument Retrieval\nand propose a method for retrieving desired musical instruments using reference\nmusic mixture as a query. The proposed model consists of the Single-Instrument\nEncoder and the Multi-Instrument Encoder, both based on convolutional neural\nnetworks. The Single-Instrument Encoder is trained to classify the instruments\nused in single-track audio, and we take its penultimate layer's activation as\nthe instrument embedding. The Multi-Instrument Encoder is trained to estimate\nmultiple instrument embeddings using the instrument embeddings computed by the\nSingle-Instrument Encoder as a set of target embeddings. For more generalized\ntraining and realistic evaluation, we also propose a new dataset called Nlakh.\nExperimental results showed that the Single-Instrument Encoder was able to\nlearn the mapping from the audio signal of unseen instruments to the instrument\nembedding space and the Multi-Instrument Encoder was able to extract multiple\nembeddings from the mixture of music and retrieve the desired instruments\nsuccessfully. The code used for the experiment and audio samples are available\nat: https://github.com/minju0821/musical_instrument_retrieval\n", "date": "2022-11-15", "categories": []}
{"id": "2211.08141", "title": "SSM-Net: feature learning for Music Structure Analysis using a\n  Self-Similarity-Matrix based loss", "abstract": "  In this paper, we propose a new paradigm to learn audio features for Music\nStructure Analysis (MSA). We train a deep encoder to learn features such that\nthe Self-Similarity-Matrix (SSM) resulting from those approximates a\nground-truth SSM. This is done by minimizing a loss between both SSMs. Since\nthis loss is differentiable w.r.t. its input features we can train the encoder\nin a straightforward way. We successfully demonstrate the use of this training\nparadigm using the Area Under the Curve ROC (AUC) on the RWC-Pop dataset.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.08379", "title": "Music Instrument Classification Reprogrammed", "abstract": "  The performance of approaches to Music Instrument Classification, a popular\ntask in Music Information Retrieval, is often impacted and limited by the lack\nof availability of annotated data for training. We propose to address this\nissue with \"reprogramming,\" a technique that utilizes pre-trained deep and\ncomplex neural networks originally targeting a different task by modifying and\nmapping both the input and output of the pre-trained model. We demonstrate that\nreprogramming can effectively leverage the power of the representation learned\nfor a different task and that the resulting reprogrammed system can perform on\npar or even outperform state-of-the-art systems at a fraction of training\nparameters. Our results, therefore, indicate that reprogramming is a promising\ntechnique potentially applicable to other tasks impeded by data scarcity.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.06646", "title": "Efficient Speech Quality Assessment using Self-supervised Framewise\n  Embeddings", "abstract": "  Automatic speech quality assessment is essential for audio researchers,\ndevelopers, speech and language pathologists, and system quality engineers. The\ncurrent state-of-the-art systems are based on framewise speech features\n(hand-engineered or learnable) combined with time dependency modeling. This\npaper proposes an efficient system with results comparable to the best\nperforming model in the ConferencingSpeech 2022 challenge. Our proposed system\nis characterized by a smaller number of parameters (40-60x), fewer FLOPS\n(100x), lower memory consumption (10-15x), and lower latency (30x). Speech\nquality practitioners can therefore iterate much faster, deploy the system on\nresource-limited hardware, and, overall, the proposed system contributes to\nsustainable machine learning. The paper also concludes that framewise\nembeddings outperform utterance-level embeddings and that multi-task training\nwith acoustic conditions modeling does not degrade speech quality prediction\nwhile providing better interpretation.\n", "date": "2022-11-12", "categories": []}
{"id": "2211.05927", "title": "Optimal Condition Training for Target Source Separation", "abstract": "  Recent research has shown remarkable performance in leveraging multiple\nextraneous conditional and non-mutually exclusive semantic concepts for sound\nsource separation, allowing the flexibility to extract a given target source\nbased on multiple different queries. In this work, we propose a new optimal\ncondition training (OCT) method for single-channel target source separation,\nbased on greedy parameter updates using the highest performing condition among\nequivalent conditions associated with a given target source. Our experiments\nshow that the complementary information carried by the diverse semantic\nconcepts significantly helps to disentangle and isolate sources of interest\nmuch more efficiently compared to single-conditioned models. Moreover, we\npropose a variation of OCT with condition refinement, in which an initial\nconditional vector is adapted to the given mixture and transformed to a more\namenable representation for target source extraction. We showcase the\neffectiveness of OCT on diverse source separation experiments where it improves\nupon permutation invariant models with oracle assignment and obtains\nstate-of-the-art performance in the more challenging task of text-based source\nseparation, outperforming even dedicated text-only conditioned models.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.04060", "title": "High-resolution embedding extractor for speaker diarisation", "abstract": "  Speaker embedding extractors significantly influence the performance of\nclustering-based speaker diarisation systems. Conventionally, only one\nembedding is extracted from each speech segment. However, because of the\nsliding window approach, a segment easily includes two or more speakers owing\nto speaker change points. This study proposes a novel embedding extractor\narchitecture, referred to as a high-resolution embedding extractor (HEE), which\nextracts multiple high-resolution embeddings from each speech segment. Hee\nconsists of a feature-map extractor and an enhancer, where the enhancer with\nthe self-attention mechanism is the key to success. The enhancer of HEE\nreplaces the aggregation process; instead of a global pooling layer, the\nenhancer combines relative information to each frame via attention leveraging\nthe global context. Extracted dense frame-level embeddings can each represent a\nspeaker. Thus, multiple speakers can be represented by different frame-level\nfeatures in each segment. We also propose an artificially generating mixture\ndata training framework to train the proposed HEE. Through experiments on five\nevaluation sets, including four public datasets, the proposed HEE demonstrates\nat least 10% improvement on each evaluation set, except for one dataset, which\nwe analyse that rapid speaker changes less exist.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.04124", "title": "Unsupervised vocal dereverberation with diffusion-based generative\n  models", "abstract": "  Removing reverb from reverberant music is a necessary technique to clean up\naudio for downstream music manipulations. Reverberation of music contains two\ncategories, natural reverb, and artificial reverb. Artificial reverb has a\nwider diversity than natural reverb due to its various parameter setups and\nreverberation types. However, recent supervised dereverberation methods may\nfail because they rely on sufficiently diverse and numerous pairs of\nreverberant observations and retrieved data for training in order to be\ngeneralizable to unseen observations during inference. To resolve these\nproblems, we propose an unsupervised method that can remove a general kind of\nartificial reverb for music without requiring pairs of data for training. The\nproposed method is based on diffusion models, where it initializes the unknown\nreverberation operator with a conventional signal processing technique and\nsimultaneously refines the estimate with the help of diffusion models. We show\nthrough objective and perceptual evaluations that our method outperforms the\ncurrent leading vocal dereverberation benchmarks.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.02718", "title": "SAMO: Speaker Attractor Multi-Center One-Class Learning for Voice\n  Anti-Spoofing", "abstract": "  Voice anti-spoofing systems are crucial auxiliaries for automatic speaker\nverification (ASV) systems. A major challenge is caused by unseen attacks\nempowered by advanced speech synthesis technologies. Our previous research on\none-class learning has improved the generalization ability to unseen attacks by\ncompacting the bona fide speech in the embedding space. However, such\ncompactness lacks consideration of the diversity of speakers. In this work, we\npropose speaker attractor multi-center one-class learning (SAMO), which\nclusters bona fide speech around a number of speaker attractors and pushes away\nspoofing attacks from all the attractors in a high-dimensional embedding space.\nFor training, we propose an algorithm for the co-optimization of bona fide\nspeech clustering and bona fide/spoof classification. For inference, we propose\nstrategies to enable anti-spoofing for speakers without enrollment. Our\nproposed system outperforms existing state-of-the-art single systems with a\nrelative improvement of 38% on equal error rate (EER) on the ASVspoof2019 LA\nevaluation set.\n", "date": "2022-11-04", "categories": []}
{"id": "2211.03202", "title": "\"Seeing Sound\": Audio Classification with the Wigner-Wille Distribution\n  and Convolutional Neural Networks", "abstract": "  With big data becoming increasingly available, IoT hardware becoming widely\nadopted, and AI capabilities becoming more powerful, organizations are\ncontinuously investing in sensing. Data coming from sensor networks are\ncurrently combined with sensor fusion and AI algorithms to drive innovation in\nfields such as self-driving cars. Data from these sensors can be utilized in\nnumerous use cases, including alerts in safety systems of urban settings, for\nevents such as gun shots and explosions. Moreover, diverse types of sensors,\nsuch as sound sensors, can be utilized in low-light conditions or at locations\nwhere a camera is not available. This paper investigates the potential of the\nutilization of sound-sensor data in an urban context. Technically, we propose a\nnovel approach of classifying sound data using the Wigner-Ville distribution\nand Convolutional Neural Networks. In this paper, we report on the performance\nof the approach on open-source datasets. The concept and work presented is\nbased on my doctoral thesis, which was performed as part of the Engineering\nDoctorate program in Data Science at the University of Eindhoven, in\ncollaboration with the Dutch National Police. Additional work on real-world\ndatasets was performed during the thesis, which are not presented here due to\nconfidentiality.\n", "date": "2022-11-06", "categories": []}
{"id": "2210.17327", "title": "Diffusion-based Generative Speech Source Separation", "abstract": "  We propose DiffSep, a new single channel source separation method based on\nscore-matching of a stochastic differential equation (SDE). We craft a tailored\ncontinuous time diffusion-mixing process starting from the separated sources\nand converging to a Gaussian distribution centered on their mixture. This\nformulation lets us apply the machinery of score-based generative modelling.\nFirst, we train a neural network to approximate the score function of the\nmarginal probabilities or the diffusion-mixing process. Then, we use it to\nsolve the reverse time SDE that progressively separates the sources starting\nfrom their mixture. We propose a modified training strategy to handle model\nmismatch and source permutation ambiguity. Experiments on the WSJ0 2mix dataset\ndemonstrate the potential of the method. Furthermore, the method is also\nsuitable for speech enhancement and shows performance competitive with prior\nwork on the VoiceBank-DEMAND dataset.\n", "date": "2022-10-31", "categories": []}
{"id": "2211.00988", "title": "Audio-visual speech enhancement with a deep Kalman filter generative\n  model", "abstract": "  Deep latent variable generative models based on variational autoencoder (VAE)\nhave shown promising performance for audiovisual speech enhancement (AVSE). The\nunderlying idea is to learn a VAEbased audiovisual prior distribution for clean\nspeech data, and then combine it with a statistical noise model to recover a\nspeech signal from a noisy audio recording and video (lip images) of the target\nspeaker. Existing generative models developed for AVSE do not take into account\nthe sequential nature of speech data, which prevents them from fully\nincorporating the power of visual data. In this paper, we present an\naudiovisual deep Kalman filter (AV-DKF) generative model which assumes a\nfirst-order Markov chain model for the latent variables and effectively fuses\naudiovisual data. Moreover, we develop an efficient inference methodology to\nestimate speech signals at test time. We conduct a set of experiments to\ncompare different variants of generative models for speech enhancement. The\nresults demonstrate the superiority of the AV-DKF model compared with both its\naudio-only version and the non-sequential audio-only and audiovisual VAE-based\nmodels.\n", "date": "2022-11-02", "categories": []}
{"id": "2210.15750", "title": "One-Shot Acoustic Matching Of Audio Signals -- Learning to Hear Music In\n  Any Room/ Concert Hall", "abstract": "  The acoustic space in which a sound is created and heard plays an essential\nrole in how that sound is perceived by affording a unique sense of\n\\textit{presence}. Every sound we hear results from successive convolution\noperations intrinsic to the sound source and external factors such as\nmicrophone characteristics and room impulse responses. Typically, researchers\nuse an excitation such as a pistol shot or balloon pop as an impulse signal\nwith which an auralization can be created. The room \"impulse\" responses\nconvolved with the signal of interest can transform the input sound into the\nsound played in the acoustic space of interest. Here we propose a novel\narchitecture that can transform a sound of interest into any other acoustic\nspace(room or hall) of interest by using arbitrary audio recorded as a proxy\nfor a balloon pop. The architecture is grounded in simple signal processing\nideas to learn residual signals from a learned acoustic signature and the input\nsignal. Our framework allows a neural network to adjust gains of every point in\nthe time-frequency representation, giving sound qualitative and quantitative\nresults.\n", "date": "2022-10-27", "categories": []}
{"id": "2208.11460", "title": "Improving Natural-Language-based Audio Retrieval with Transfer Learning\n  and Audio & Text Augmentations", "abstract": "  The absence of large labeled datasets remains a significant challenge in many\napplication areas of deep learning. Researchers and practitioners typically\nresort to transfer learning and data augmentation to alleviate this issue. We\nstudy these strategies in the context of audio retrieval with natural language\nqueries (Task 6b of the DCASE 2022 Challenge). Our proposed system uses\npre-trained embedding models to project recordings and textual descriptions\ninto a shared audio-caption space in which related examples from different\nmodalities are close. We employ various data augmentation techniques on audio\nand text inputs and systematically tune their corresponding hyperparameters\nwith sequential model-based optimization. Our results show that the used\naugmentations strategies reduce overfitting and improve retrieval performance.\n", "date": "2022-08-24", "categories": []}
{"id": "2210.06007", "title": "JukeDrummer: Conditional Beat-aware Audio-domain Drum Accompaniment\n  Generation via Transformer VQ-VAE", "abstract": "  This paper proposes a model that generates a drum track in the audio domain\nto play along to a user-provided drum-free recording. Specifically, using\npaired data of drumless tracks and the corresponding human-made drum tracks, we\ntrain a Transformer model to improvise the drum part of an unseen drumless\nrecording. We combine two approaches to encode the input audio. First, we train\na vector-quantized variational autoencoder (VQ-VAE) to represent the input\naudio with discrete codes, which can then be readily used in a Transformer.\nSecond, using an audio-domain beat tracking model, we compute beat-related\nfeatures of the input audio and use them as embeddings in the Transformer.\nInstead of generating the drum track directly as waveforms, we use a separate\nVQ-VAE to encode the mel-spectrogram of a drum track into another set of\ndiscrete codes, and train the Transformer to predict the sequence of\ndrum-related discrete codes. The output codes are then converted to a\nmel-spectrogram with a decoder, and then to the waveform with a vocoder. We\nreport both objective and subjective evaluations of variants of the proposed\nmodel, demonstrating that the model with beat information generates drum\naccompaniment that is rhythmically and stylistically consistent with the input\naudio.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.16472", "title": "Learning Audio-Visual Dynamics Using Scene Graphs for Audio Source\n  Separation", "abstract": "  There exists an unequivocal distinction between the sound produced by a\nstatic source and that produced by a moving one, especially when the source\nmoves towards or away from the microphone. In this paper, we propose to use\nthis connection between audio and visual dynamics for solving two challenging\ntasks simultaneously, namely: (i) separating audio sources from a mixture using\nvisual cues, and (ii) predicting the 3D visual motion of a sounding source\nusing its separated audio. Towards this end, we present Audio Separator and\nMotion Predictor (ASMP) -- a deep learning framework that leverages the 3D\nstructure of the scene and the motion of sound sources for better audio source\nseparation. At the heart of ASMP is a 2.5D scene graph capturing various\nobjects in the video and their pseudo-3D spatial proximities. This graph is\nconstructed by registering together 2.5D monocular depth predictions from the\n2D video frames and associating the 2.5D scene regions with the outputs of an\nobject detector applied on those frames. The ASMP task is then mathematically\nmodeled as the joint problem of: (i) recursively segmenting the 2.5D scene\ngraph into several sub-graphs, each associated with a constituent sound in the\ninput audio mixture (which is then separated) and (ii) predicting the 3D\nmotions of the corresponding sound sources from the separated audio. To\nempirically evaluate ASMP, we present experiments on two challenging\naudio-visual datasets, viz. Audio Separation in the Wild (ASIW) and Audio\nVisual Event (AVE). Our results demonstrate that ASMP achieves a clear\nimprovement in source separation quality, outperforming prior works on both\ndatasets, while also estimating the direction of motion of the sound sources\nbetter than other methods.\n", "date": "2022-10-29", "categories": []}
{"id": "2207.03572", "title": "The ACII 2022 Affective Vocal Bursts Workshop & Competition:\n  Understanding a critically understudied modality of emotional expression", "abstract": "  The ACII Affective Vocal Bursts Workshop & Competition is focused on\nunderstanding multiple affective dimensions of vocal bursts: laughs, gasps,\ncries, screams, and many other non-linguistic vocalizations central to the\nexpression of emotion and to human communication more generally. This year's\ncompetition comprises four tracks using a large-scale and in-the-wild dataset\nof 59,299 vocalizations from 1,702 speakers. The first, the A-VB-High task,\nrequires competition participants to perform a multi-label regression on a\nnovel model for emotion, utilizing ten classes of richly annotated emotional\nexpression intensities, including; Awe, Fear, and Surprise. The second, the\nA-VB-Two task, utilizes the more conventional 2-dimensional model for emotion,\narousal, and valence. The third, the A-VB-Culture task, requires participants\nto explore the cultural aspects of the dataset, training native-country\ndependent models. Finally, for the fourth task, A-VB-Type, participants should\nrecognize the type of vocal burst (e.g., laughter, cry, grunt) as an 8-class\nclassification. This paper describes the four tracks and baseline systems,\nwhich use state-of-the-art machine learning methods. The baseline performance\nfor each track is obtained by utilizing an end-to-end deep learning model and\nis as follows: for A-VB-High, a mean (over the 10-dimensions) Concordance\nCorrelation Coefficient (CCC) of 0.5687 CCC is obtained; for A-VB-Two, a mean\n(over the 2-dimensions) CCC of 0.5084 is obtained; for A-VB-Culture, a mean CCC\nfrom the four cultures of 0.4401 is obtained; and for A-VB-Type, the baseline\nUnweighted Average Recall (UAR) from the 8-classes is 0.4172 UAR.\n", "date": "2022-07-07", "categories": []}
{"id": "2210.15988", "title": "Spectrograms Are Sequences of Patches", "abstract": "  Self-supervised pre-training models have been used successfully in several\nmachine learning domains. However, only a tiny amount of work is related to\nmusic. In our work, we treat a spectrogram of music as a series of patches and\ndesign a self-supervised model that captures the features of these sequential\npatches: Patchifier, which makes good use of self-supervised learning methods\nfrom both NLP and CV domains. We do not use labeled data for the pre-training\nprocess, only a subset of the MTAT dataset containing 16k music clips. After\npre-training, we apply the model to several downstream tasks. Our model\nachieves a considerably acceptable result compared to other audio\nrepresentation models. Meanwhile, our work demonstrates that it makes sense to\nconsider audio as a series of patch segments.\n", "date": "2022-10-28", "categories": []}
{"id": "2210.15370", "title": "CasNet: Investigating Channel Robustness for Speech Separation", "abstract": "  Recording channel mismatch between training and testing conditions has been\nshown to be a serious problem for speech separation. This situation greatly\nreduces the separation performance, and cannot meet the requirement of daily\nuse. In this study, inheriting the use of our previously constructed TAT-2mix\ncorpus, we address the channel mismatch problem by proposing a channel-aware\naudio separation network (CasNet), a deep learning framework for end-to-end\ntime-domain speech separation. CasNet is implemented on top of TasNet. Channel\nembedding (characterizing channel information in a mixture of multiple\nutterances) generated by Channel Encoder is introduced into the separation\nmodule by the FiLM technique. Through two training strategies, we explore two\nroles that channel embedding may play: 1) a real-life noise disturbance, making\nthe model more robust, or 2) a guide, instructing the separation model to\nretain the desired channel information. Experimental results on TAT-2mix show\nthat CasNet trained with both training strategies outperforms the TasNet\nbaseline, which does not use channel embeddings.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.12635", "title": "Quantitative Evidence on Overlooked Aspects of Enrollment Speaker\n  Embeddings for Target Speaker Separation", "abstract": "  Single channel target speaker separation (TSS) aims at extracting a speaker's\nvoice from a mixture of multiple talkers given an enrollment utterance of that\nspeaker. A typical deep learning TSS framework consists of an upstream model\nthat obtains enrollment speaker embeddings and a downstream model that performs\nthe separation conditioned on the embeddings. In this paper, we look into\nseveral important but overlooked aspects of the enrollment embeddings,\nincluding the suitability of the widely used speaker identification embeddings,\nthe introduction of the log-mel filterbank and self-supervised embeddings, and\nthe embeddings' cross-dataset generalization capability. Our results show that\nthe speaker identification embeddings could lose relevant information due to a\nsub-optimal metric, training objective, or common pre-processing. In contrast,\nboth the filterbank and the self-supervised embeddings preserve the integrity\nof the speaker information, but the former consistently outperforms the latter\nin a cross-dataset evaluation. The competitive separation and generalization\nperformance of the previously overlooked filterbank embedding is consistent\nacross our study, which calls for future research on better upstream features.\n", "date": "2022-10-23", "categories": []}
{"id": "2210.14661", "title": "Full-band General Audio Synthesis with Score-based Diffusion", "abstract": "  Recent works have shown the capability of deep generative models to tackle\ngeneral audio synthesis from a single label, producing a variety of impulsive,\ntonal, and environmental sounds. Such models operate on band-limited signals\nand, as a result of an autoregressive approach, they are typically conformed by\npre-trained latent encoders and/or several cascaded modules. In this work, we\npropose a diffusion-based generative model for general audio synthesis, named\nDAG, which deals with full-band signals end-to-end in the waveform domain.\nResults show the superiority of DAG over existing label-conditioned generators\nin terms of both quality and diversity. More specifically, when compared to the\nstate of the art, the band-limited and full-band versions of DAG achieve\nrelative improvements that go up to 40 and 65%, respectively. We believe DAG is\nflexible enough to accommodate different conditioning schemas while providing\ngood quality synthesis.\n", "date": "2022-10-26", "categories": []}
{"id": "2206.13909", "title": "QTI Submission to DCASE 2021: residual normalization for\n  device-imbalanced acoustic scene classification with efficient design", "abstract": "  This technical report describes the details of our TASK1A submission of the\nDCASE2021 challenge. The goal of the task is to design an audio scene\nclassification system for device-imbalanced datasets under the constraints of\nmodel complexity. This report introduces four methods to achieve the goal.\nFirst, we propose Residual Normalization, a novel feature normalization method\nthat uses instance normalization with a shortcut path to discard unnecessary\ndevice-specific information without losing useful information for\nclassification. Second, we design an efficient architecture, BC-ResNet-Mod, a\nmodified version of the baseline architecture with a limited receptive field.\nThird, we exploit spectrogram-to-spectrogram translation from one to multiple\ndevices to augment training data. Finally, we utilize three model compression\nschemes: pruning, quantization, and knowledge distillation to reduce model\ncomplexity. The proposed system achieves an average test accuracy of 76.3% in\nTAU Urban Acoustic Scenes 2020 Mobile, development dataset with 315k\nparameters, and average test accuracy of 75.3% after compression to 61.0KB of\nnon-zero parameters. We extend this work to [1].\n", "date": "2022-06-28", "categories": []}
{"id": "2207.10441", "title": "Deep Audio Waveform Prior", "abstract": "  Convolutional neural networks contain strong priors for generating natural\nlooking images [1]. These priors enable image denoising, super resolution, and\ninpainting in an unsupervised manner. Previous attempts to demonstrate similar\nideas in audio, namely deep audio priors, (i) use hand picked architectures\nsuch as harmonic convolutions, (ii) only work with spectrogram input, and (iii)\nhave been used mostly for eliminating Gaussian noise [2]. In this work we show\nthat existing SOTA architectures for audio source separation contain deep\npriors even when working with the raw waveform. Deep priors can be discovered\nby training a neural network to generate a single corrupted signal when given\nwhite noise as input. A network with relevant deep priors is likely to generate\na cleaner version of the signal before converging on the corrupted signal. We\ndemonstrate this restoration effect with several corruptions: background noise,\nreverberations, and a gap in the signal (audio inpainting).\n", "date": "2022-07-21", "categories": []}
{"id": "2210.14085", "title": "Audio MFCC-gram Transformers for respiratory insufficiency detection in\n  COVID-19", "abstract": "  This work explores speech as a biomarker and investigates the detection of\nrespiratory insufficiency (RI) by analyzing speech samples. Previous work\n\\cite{spira2021} constructed a dataset of respiratory insufficiency COVID-19\npatient utterances and analyzed it by means of a convolutional neural network\nachieving an accuracy of $87.04\\%$, validating the hypothesis that one can\ndetect RI through speech. Here, we study how Transformer neural network\narchitectures can improve the performance on RI detection. This approach\nenables construction of an acoustic model. By choosing the correct pretraining\ntechnique, we generate a self-supervised acoustic model, leading to improved\nperformance ($96.53\\%$) of Transformers for RI detection.\n", "date": "2022-10-25", "categories": []}
{"id": "2210.13438", "title": "High Fidelity Neural Audio Compression", "abstract": "  We introduce a state-of-the-art real-time, high-fidelity, audio codec\nleveraging neural networks. It consists in a streaming encoder-decoder\narchitecture with quantized latent space trained in an end-to-end fashion. We\nsimplify and speed-up the training by using a single multiscale spectrogram\nadversary that efficiently reduces artifacts and produce high-quality samples.\nWe introduce a novel loss balancer mechanism to stabilize training: the weight\nof a loss now defines the fraction of the overall gradient it should represent,\nthus decoupling the choice of this hyper-parameter from the typical scale of\nthe loss. Finally, we study how lightweight Transformer models can be used to\nfurther compress the obtained representation by up to 40%, while staying faster\nthan real time. We provide a detailed description of the key design choices of\nthe proposed model including: training objective, architectural changes and a\nstudy of various perceptual loss functions. We present an extensive subjective\nevaluation (MUSHRA tests) together with an ablation study for a range of\nbandwidths and audio domains, including speech, noisy-reverberant speech, and\nmusic. Our approach is superior to the baselines methods across all evaluated\nsettings, considering both 24 kHz monophonic and 48 kHz stereophonic audio.\nCode and models are available at github.com/facebookresearch/encodec.\n", "date": "2022-10-24", "categories": []}
{"id": "2210.11722", "title": "Adaptive re-calibration of channel-wise features for Adversarial Audio\n  Classification", "abstract": "  DeepFake Audio, unlike DeepFake images and videos, has been relatively less\nexplored from detection perspective, and the solutions which exist for the\nsynthetic speech classification either use complex networks or dont generalize\nto different varieties of synthetic speech obtained using different generative\nand optimization-based methods. Through this work, we propose a channel-wise\nrecalibration of features using attention feature fusion for synthetic speech\ndetection and compare its performance against different detection methods\nincluding End2End models and Resnet-based models on synthetic speech generated\nusing Text to Speech and Vocoder systems like WaveNet, WaveRNN, Tactotron, and\nWaveGlow. We also experiment with Squeeze Excitation (SE) blocks in our Resnet\nmodels and found that the combination was able to get better performance. In\naddition to the analysis, we also demonstrate that the combination of Linear\nfrequency cepstral coefficients (LFCC) and Mel Frequency cepstral coefficients\n(MFCC) using the attentional feature fusion technique creates better input\nfeatures representations which can help even simpler models generalize well on\nsynthetic speech classification tasks. Our models (Resnet based using feature\nfusion) trained on Fake or Real (FoR) dataset and were able to achieve 95% test\naccuracy with the FoR data, and an average of 90% accuracy with samples we\ngenerated using different generative models after adapting this framework.\n", "date": "2022-10-21", "categories": []}
{"id": "2210.10196", "title": "BirdSoundsDenoising: Deep Visual Audio Denoising for Bird Sounds", "abstract": "  Audio denoising has been explored for decades using both traditional and deep\nlearning-based methods. However, these methods are still limited to either\nmanually added artificial noise or lower denoised audio quality. To overcome\nthese challenges, we collect a large-scale natural noise bird sound dataset. We\nare the first to transfer the audio denoising problem into an image\nsegmentation problem and propose a deep visual audio denoising (DVAD) model.\nWith a total of 14,120 audio images, we develop an audio ImageMask tool and\npropose to use a few-shot generalization strategy to label these images.\nExtensive experimental results demonstrate that the proposed model achieves\nstate-of-the-art performance. We also show that our method can be easily\ngeneralized to speech denoising, audio separation, audio enhancement, and noise\nestimation.\n", "date": "2022-10-18", "categories": []}
{"id": "2210.07508", "title": "Hierarchical Diffusion Models for Singing Voice Neural Vocoder", "abstract": "  Recent progress in deep generative models has improved the quality of neural\nvocoders in speech domain. However, generating a high-quality singing voice\nremains challenging due to a wider variety of musical expressions in pitch,\nloudness, and pronunciations. In this work, we propose a hierarchical diffusion\nmodel for singing voice neural vocoders. The proposed method consists of\nmultiple diffusion models operating in different sampling rates; the model at\nthe lowest sampling rate focuses on generating accurate low-frequency\ncomponents such as pitch, and other models progressively generate the waveform\nat higher sampling rates on the basis of the data at the lower sampling rate\nand acoustic features. Experimental results show that the proposed method\nproduces high-quality singing voices for multiple singers, outperforming\nstate-of-the-art neural vocoders with a similar range of computational costs.\n", "date": "2022-10-14", "categories": []}
{"id": "2210.09529", "title": "A Hybrid System of Sound Event Detection Transformer and Frame-wise\n  Model for DCASE 2022 Task 4", "abstract": "  In this paper, we describe in detail our system for DCASE 2022 Task4. The\nsystem combines two considerably different models: an end-to-end Sound Event\nDetection Transformer (SEDT) and a frame-wise model, Metric Learning and Focal\nLoss CNN (MLFL-CNN). The former is an event-wise model which learns event-level\nrepresentations and predicts sound event categories and boundaries directly,\nwhile the latter is based on the widely adopted frame-classification scheme,\nunder which each frame is classified into event categories and event boundaries\nare obtained by post-processing such as thresholding and smoothing. For SEDT,\nself-supervised pre-training using unlabeled data is applied, and\nsemi-supervised learning is adopted by using an online teacher, which is\nupdated from the student model using the Exponential Moving Average (EMA)\nstrategy and generates reliable pseudo labels for weakly-labeled and unlabeled\ndata. For the frame-wise model, the ICT-TOSHIBA system of DCASE 2021 Task 4 is\nused. Experimental results show that the hybrid system considerably outperforms\neither individual model and achieves psds1 of 0.420 and psds2 of 0.783 on the\nvalidation set without external data. The code is available at\nhttps://github.com/965694547/Hybrid-system-of-frame-wise-model-and-SEDT.\n", "date": "2022-10-18", "categories": []}
{"id": "2210.05037", "title": "Automated Audio Captioning via Fusion of Low- and High- Dimensional\n  Features", "abstract": "  Automated audio captioning (AAC) aims to describe the content of an audio\nclip using simple sentences. Existing AAC methods are developed based on an\nencoder-decoder architecture that success is attributed to the use of a\npre-trained CNN10 called PANNs as the encoder to learn rich audio\nrepresentations. AAC is a highly challenging task due to its high-dimensional\ntalent space involves audio of various scenarios. Existing methods only use the\nhigh-dimensional representation of the PANNs as the input of the decoder.\nHowever, the low-dimension representation may retain as much audio information\nas the high-dimensional representation may be neglected. In addition, although\nthe high-dimensional approach may predict the audio captions by learning from\nexisting audio captions, which lacks robustness and efficiency. To deal with\nthese challenges, a fusion model which integrates low- and high-dimensional\nfeatures AAC framework is proposed. In this paper, a new encoder-decoder\nframework is proposed called the Low- and High-Dimensional Feature Fusion\n(LHDFF) model for AAC. Moreover, in LHDFF, a new PANNs encoder is proposed\ncalled Residual PANNs (RPANNs) by fusing the low-dimensional feature from the\nintermediate convolution layer output and the high-dimensional feature from the\nfinal layer output of PANNs. To fully explore the information of the low- and\nhigh-dimensional fusion feature and high-dimensional feature respectively, we\nproposed dual transformer decoder structures to generate the captions in\nparallel. Especially, a probabilistic fusion approach is proposed that can\nensure the overall performance of the system is improved by concentrating on\nthe respective advantages of the two transformer decoders. Experimental results\nshow that LHDFF achieves the best performance on the Clotho and AudioCaps\ndatasets compared with other existing models\n", "date": "2022-10-10", "categories": []}
{"id": "2210.08610", "title": "Robust, General, and Low Complexity Acoustic Scene Classification\n  Systems and An Effective Visualization for Presenting a Sound Scene Context", "abstract": "  In this paper, we present a comprehensive analysis of Acoustic Scene\nClassification (ASC), the task of identifying the scene of an audio recording\nfrom its acoustic signature. In particular, we firstly propose an\ninception-based and low footprint ASC model, referred to as the ASC baseline.\nThe proposed ASC baseline is then compared with benchmark and high-complexity\nnetwork architectures of MobileNetV1, MobileNetV2, VGG16, VGG19, ResNet50V2,\nResNet152V2, DenseNet121, DenseNet201, and Xception. Next, we improve the ASC\nbaseline by proposing a novel deep neural network architecture which leverages\nresidual-inception architectures and multiple kernels. Given the novel\nresidual-inception (NRI) model, we further evaluate the trade off between the\nmodel complexity and the model accuracy performance. Finally, we evaluate\nwhether sound events occurring in a sound scene recording can help to improve\nASC accuracy, then indicate how a sound scene context is well presented by\ncombining both sound scene and sound event information. We conduct extensive\nexperiments on various ASC datasets, including Crowded Scenes, IEEE AASP\nChallenge on Detection and Classification of Acoustic Scenes and Events (DCASE)\n2018 Task 1A and 1B, 2019 Task 1A and 1B, 2020 Task 1A, 2021 Task 1A, 2022 Task\n1. The experimental results on several different ASC challenges highlight two\nmain achievements; the first is to propose robust, general, and low complexity\nASC systems which are suitable for real-life applications on a wide range of\nedge devices and mobiles; the second is to propose an effective visualization\nmethod for comprehensively presenting a sound scene context.\n", "date": "2022-10-16", "categories": []}
{"id": "2210.07642", "title": "Training speech emotion classifier without categorical annotations", "abstract": "  There are two paradigms of emotion representation, categorical labeling and\ndimensional description in continuous space. Therefore, the emotion recognition\ntask can be treated as a classification or regression. The main aim of this\nstudy is to investigate the relation between these two representations and\npropose a classification pipeline that uses only dimensional annotation. The\nproposed approach contains a regressor model which is trained to predict a\nvector of continuous values in dimensional representation for given speech\naudio. The output of this model can be interpreted as an emotional category\nusing a mapping algorithm. We investigated the performances of a combination of\nthree feature extractors, three neural network architectures, and three mapping\nalgorithms on two different corpora. Our study shows the advantages and\nlimitations of the classification via regression approach.\n", "date": "2022-10-14", "categories": []}
{"id": "2206.13979", "title": "Attack Agnostic Dataset: Towards Generalization and Stabilization of\n  Audio DeepFake Detection", "abstract": "  Audio DeepFakes allow the creation of high-quality, convincing utterances and\ntherefore pose a threat due to its potential applications such as impersonation\nor fake news. Methods for detecting these manipulations should be characterized\nby good generalization and stability leading to robustness against attacks\nconducted with techniques that are not explicitly included in the training. In\nthis work, we introduce Attack Agnostic Dataset - a combination of two audio\nDeepFakes and one anti-spoofing datasets that, thanks to the disjoint use of\nattacks, can lead to better generalization of detection methods. We present a\nthorough analysis of current DeepFake detection methods and consider different\naudio features (front-ends). In addition, we propose a model based on LCNN with\nLFCC and mel-spectrogram front-end, which not only is characterized by a good\ngeneralization and stability results but also shows improvement over LFCC-based\nmode - we decrease standard deviation on all folds and EER in two folds by up\nto 5%.\n", "date": "2022-06-27", "categories": []}
{"id": "2207.11690", "title": "HouseX: A Fine-grained House Music Dataset and its Potential in the\n  Music Industry", "abstract": "  Machine sound classification has been one of the fundamental tasks of music\ntechnology. A major branch of sound classification is the classification of\nmusic genres. However, though covering most genres of music, existing music\ngenre datasets often do not contain fine-grained labels that indicate the\ndetailed sub-genres of music. In consideration of the consistency of genres of\nsongs in a mixtape or in a DJ (live) set, we have collected and annotated a\ndataset of house music that provide 4 sub-genre labels, namely future house,\nbass house, progressive house and melodic house. Experiments show that our\nannotations well exhibit the characteristics of different categories. Also, we\nhave built baseline models that classify the sub-genre based on the\nmel-spectrograms of a track, achieving strongly competitive results. Besides,\nwe have put forward a few application scenarios of our dataset and baseline\nmodel, with a simulated sci-fi tunnel as a short demo built and rendered in a\n3D modeling software, with the colors of the lights automated by the output of\nour model.\n", "date": "2022-07-24", "categories": []}
{"id": "2210.06105", "title": "SpecRNet: Towards Faster and More Accessible Audio DeepFake Detection", "abstract": "  Audio DeepFakes are utterances generated with the use of deep neural\nnetworks. They are highly misleading and pose a threat due to use in fake news,\nimpersonation, or extortion. In this work, we focus on increasing accessibility\nto the audio DeepFake detection methods by providing SpecRNet, a neural network\narchitecture characterized by a quick inference time and low computational\nrequirements. Our benchmark shows that SpecRNet, requiring up to about 40% less\ntime to process an audio sample, provides performance comparable to LCNN\narchitecture - one of the best audio DeepFake detection models. Such a method\ncan not only be used by online multimedia services to verify a large bulk of\ncontent uploaded daily but also, thanks to its low requirements, by average\ncitizens to evaluate materials on their devices. In addition, we provide\nbenchmarks in three unique settings that confirm the correctness of our model.\nThey reflect scenarios of low-resource datasets, detection on short utterances\nand limited attacks benchmark in which we take a closer look at the influence\nof particular attacks on given architectures.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.06368", "title": "Individualized Conditioning and Negative Distances for Speaker\n  Separation", "abstract": "  Speaker separation aims to extract multiple voices from a mixed signal. In\nthis paper, we propose two speaker-aware designs to improve the existing\nspeaker separation solutions. The first model is a speaker conditioning network\nthat integrates speech samples to generate individualized speaker conditions,\nwhich then provide informed guidance for a separation module to produce\nwell-separated outputs.\n  The second design aims to reduce non-target voices in the separated speech.\nTo this end, we propose negative distances to penalize the appearance of any\nnon-target voice in the channel outputs, and positive distances to drive the\nseparated voices closer to the clean targets. We explore two different setups,\nweighted-sum and triplet-like, to integrate these two distances to form a\ncombined auxiliary loss for the separation networks. Experiments conducted on\nLibriMix demonstrate the effectiveness of our proposed models.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.05254", "title": "Deep Spectro-temporal Artifacts for Detecting Synthesized Speech", "abstract": "  The Audio Deep Synthesis Detection (ADD) Challenge has been held to detect\ngenerated human-like speech. With our submitted system, this paper provides an\noverall assessment of track 1 (Low-quality Fake Audio Detection) and track 2\n(Partially Fake Audio Detection). In this paper, spectro-temporal artifacts\nwere detected using raw temporal signals, spectral features, as well as deep\nembedding features. To address track 1, low-quality data augmentation, domain\nadaptation via finetuning, and various complementary feature information fusion\nwere aggregated in our system. Furthermore, we analyzed the clustering\ncharacteristics of subsystems with different features by visualization method\nand explained the effectiveness of our proposed greedy fusion strategy. As for\ntrack 2, frame transition and smoothing were detected using self-supervised\nlearning structure to capture the manipulation of PF attacks in the time\ndomain. We ranked 4th and 5th in track 1 and track 2, respectively.\n", "date": "2022-10-11", "categories": []}
{"id": "2209.12826", "title": "Multi-encoder attention-based architectures for sound recognition with\n  partial visual assistance", "abstract": "  Large-scale sound recognition data sets typically consist of acoustic\nrecordings obtained from multimedia libraries. As a consequence, modalities\nother than audio can often be exploited to improve the outputs of models\ndesigned for associated tasks. Frequently, however, not all contents are\navailable for all samples of such a collection: For example, the original\nmaterial may have been removed from the source platform at some point, and\ntherefore, non-auditory features can no longer be acquired.\n  We demonstrate that a multi-encoder framework can be employed to deal with\nthis issue by applying this method to attention-based deep learning systems,\nwhich are currently part of the state of the art in the domain of sound\nrecognition. More specifically, we show that the proposed model extension can\nsuccessfully be utilized to incorporate partially available visual information\ninto the operational procedures of such networks, which normally only use\nauditory features during training and inference. Experimentally, we verify that\nthe considered approach leads to improved predictions in a number of evaluation\nscenarios pertaining to audio tagging and sound event detection. Additionally,\nwe scrutinize some properties and limitations of the presented technique.\n", "date": "2022-09-26", "categories": []}
{"id": "2210.03799", "title": "Supervised and Unsupervised Learning of Audio Representations for Music\n  Understanding", "abstract": "  In this work, we provide a broad comparative analysis of strategies for\npre-training audio understanding models for several tasks in the music domain,\nincluding labelling of genre, era, origin, mood, instrumentation, key, pitch,\nvocal characteristics, tempo and sonority. Specifically, we explore how the\ndomain of pre-training datasets (music or generic audio) and the pre-training\nmethodology (supervised or unsupervised) affects the adequacy of the resulting\naudio embeddings for downstream tasks.\n  We show that models trained via supervised learning on large-scale\nexpert-annotated music datasets achieve state-of-the-art performance in a wide\nrange of music labelling tasks, each with novel content and vocabularies. This\ncan be done in an efficient manner with models containing less than 100 million\nparameters that require no fine-tuning or reparameterization for downstream\ntasks, making this approach practical for industry-scale audio catalogs.\n  Within the class of unsupervised learning strategies, we show that the domain\nof the training dataset can significantly impact the performance of\nrepresentations learned by the model. We find that restricting the domain of\nthe pre-training dataset to music allows for training with smaller batch sizes\nwhile achieving state-of-the-art in unsupervised learning -- and in some cases,\nsupervised learning -- for music understanding.\n  We also corroborate that, while achieving state-of-the-art performance on\nmany tasks, supervised learning can cause models to specialize to the\nsupervised information provided, somewhat compromising a model's generality.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.02642", "title": "Feasibility on Detecting Door Slamming towards Monitoring Early Signs of\n  Domestic Violence", "abstract": "  By using low-cost microcontrollers and TinyML, we investigate the feasibility\nof detecting potential early warning signs of domestic violence and other\nanti-social behaviors within the home. We created a machine learning model to\ndetermine if a door was closed aggressively by analyzing audio data and feeding\nthis into a convolutional neural network to classify the sample. Under test\nconditions, with no background noise, accuracy of 88.89\\% was achieved,\ndeclining to 87.50\\% when assorted background noises were mixed in at a\nrelative volume of 0.5 times that of the sample. The model is then deployed on\nan Arduino Nano BLE 33 Sense attached to the door, and only begins sampling\nonce an acceleration greater than a predefined threshold acceleration is\ndetected. The predictions made by the model can then be sent via BLE to another\ndevice, such as a smartphone of Raspberry Pi.\n", "date": "2022-10-06", "categories": []}
{"id": "2210.02746", "title": "The Sound of Silence: Efficiency of First Digit Features in Synthetic\n  Audio Detection", "abstract": "  The recent integration of generative neural strategies and audio processing\ntechniques have fostered the widespread of synthetic speech synthesis or\ntransformation algorithms. This capability proves to be harmful in many legal\nand informative processes (news, biometric authentication, audio evidence in\ncourts, etc.). Thus, the development of efficient detection algorithms is both\ncrucial and challenging due to the heterogeneity of forgery techniques.\n  This work investigates the discriminative role of silenced parts in synthetic\nspeech detection and shows how first digit statistics extracted from MFCC\ncoefficients can efficiently enable a robust detection. The proposed procedure\nis computationally-lightweight and effective on many different algorithms since\nit does not rely on large neural detection architecture and obtains an accuracy\nabove 90\\% in most of the classes of the ASVSpoof dataset.\n", "date": "2022-10-06", "categories": []}
{"id": "2209.15174", "title": "Music Source Separation with Band-split RNN", "abstract": "  The performance of music source separation (MSS) models has been greatly\nimproved in recent years thanks to the development of novel neural network\narchitectures and training pipelines. However, recent model designs for MSS\nwere mainly motivated by other audio processing tasks or other research fields,\nwhile the intrinsic characteristics and patterns of the music signals were not\nfully discovered. In this paper, we propose band-split RNN (BSRNN), a\nfrequency-domain model that explictly splits the spectrogram of the mixture\ninto subbands and perform interleaved band-level and sequence-level modeling.\nThe choices of the bandwidths of the subbands can be determined by a priori\nknowledge or expert knowledge on the characteristics of the target source in\norder to optimize the performance on a certain type of target musical\ninstrument. To better make use of unlabeled data, we also describe a\nsemi-supervised model finetuning pipeline that can further improve the\nperformance of the model. Experiment results show that BSRNN trained only on\nMUSDB18-HQ dataset significantly outperforms several top-ranking models in\nMusic Demixing (MDX) Challenge 2021, and the semi-supervised finetuning stage\nfurther improves the performance on all four instrument tracks.\n", "date": "2022-09-30", "categories": []}
{"id": "2209.14335", "title": "Text Independent Speaker Identification System for Access Control", "abstract": "  Even human intelligence system fails to offer 100% accuracy in identifying\nspeeches from a specific individual. Machine intelligence is trying to mimic\nhumans in speaker identification problems through various approaches to speech\nfeature extraction and speech modeling techniques. This paper presents a\ntext-independent speaker identification system that employs Mel Frequency\nCepstral Coefficients (MFCC) for feature extraction and k-Nearest Neighbor\n(kNN) for classification. The maximum cross-validation accuracy obtained was\n60%. This will be improved upon in subsequent research.\n", "date": "2022-09-26", "categories": []}
{"id": "2209.13914", "title": "An Efficient Multitask Learning Architecture for Affective Vocal Burst\n  Analysis", "abstract": "  Affective speech analysis is an ongoing topic of research. A relatively new\nproblem in this field is the analysis of vocal bursts, which are nonverbal\nvocalisations such as laughs or sighs. Current state-of-the-art approaches to\naddress affective vocal burst analysis are mostly based on wav2vec2 or HuBERT\nfeatures. In this paper, we investigate the use of the wav2vec successor\ndata2vec in combination with a multitask learning pipeline to tackle different\nanalysis problems at once. To assess the performance of our efficient multitask\nlearning architecture, we participate in the 2022 ACII Affective Vocal Burst\nChallenge, showing that our approach substantially outperforms the baseline\nestablished there in three different subtasks.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.14098", "title": "Deepfake audio detection by speaker verification", "abstract": "  Thanks to recent advances in deep learning, sophisticated generation tools\nexist, nowadays, that produce extremely realistic synthetic speech. However,\nmalicious uses of such tools are possible and likely, posing a serious threat\nto our society. Hence, synthetic voice detection has become a pressing research\ntopic, and a large variety of detection methods have been recently proposed.\nUnfortunately, they hardly generalize to synthetic audios generated by tools\nnever seen in the training phase, which makes them unfit to face real-world\nscenarios. In this work, we aim at overcoming this issue by proposing a new\ndetection approach that leverages only the biometric characteristics of the\nspeaker, with no reference to specific manipulations. Since the detector is\ntrained only on real data, generalization is automatically ensured. The\nproposed approach can be implemented based on off-the-shelf speaker\nverification tools. We test several such solutions on three popular test sets,\nobtaining good performance, high generalization ability, and high robustness to\naudio impairment.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.14275", "title": "Audio Retrieval with WavText5K and CLAP Training", "abstract": "  Audio-Text retrieval takes a natural language query to retrieve relevant\naudio files in a database. Conversely, Text-Audio retrieval takes an audio file\nas a query to retrieve relevant natural language descriptions. Most of the\nliterature train retrieval systems with one audio captioning dataset, but\nevaluating the benefit of training with multiple datasets is underexplored.\nMoreover, retrieval systems have to learn the alignment between elaborated\nsentences describing audio content of variable length ranging from a few\nseconds to several minutes. In this work, we propose a new collection of web\naudio-text pairs and a new framework for retrieval. First, we provide a new\ncollection of about five thousand web audio-text pairs that we refer to as\nWavText5K. When used to train our retrieval system, WavText5K improved\nperformance more than other audio captioning datasets. Second, our framework\nlearns to connect language and audio content by using a text encoder, two audio\nencoders, and a contrastive learning objective. Combining both audio encoders\nhelps to process variable length audio. The two contributions beat state of the\nart performance for AudioCaps and Clotho on Text-Audio retrieval by a relative\n2% and 16%, and Audio-Text retrieval by 6% and 23%.\n", "date": "2022-09-28", "categories": []}
{"id": "2206.08545", "title": "NU-Wave 2: A General Neural Audio Upsampling Model for Various Sampling\n  Rates", "abstract": "  Conventionally, audio super-resolution models fixed the initial and the\ntarget sampling rates, which necessitate the model to be trained for each pair\nof sampling rates. We introduce NU-Wave 2, a diffusion model for neural audio\nupsampling that enables the generation of 48 kHz audio signals from inputs of\nvarious sampling rates with a single model. Based on the architecture of\nNU-Wave, NU-Wave 2 uses short-time Fourier convolution (STFC) to generate\nharmonics to resolve the main failure modes of NU-Wave, and incorporates\nbandwidth spectral feature transform (BSFT) to condition the bandwidths of\ninputs in the frequency domain. We experimentally demonstrate that NU-Wave 2\nproduces high-resolution audio regardless of the sampling rate of input while\nrequiring fewer parameters than other models. The official code and the audio\nsamples are available at https://mindslab-ai.github.io/nuwave2.\n", "date": "2022-06-17", "categories": []}
{"id": "2209.07384", "title": "Self-Supervised Attention Networks and Uncertainty Loss Weighting for\n  Multi-Task Emotion Recognition on Vocal Bursts", "abstract": "  Vocal bursts play an important role in communicating affect, making them\nvaluable for improving speech emotion recognition. Here, we present our\napproach for classifying vocal bursts and predicting their emotional\nsignificance in the ACII Affective Vocal Burst Workshop & Challenge 2022\n(A-VB). We use a large self-supervised audio model as shared feature extractor\nand compare multiple architectures built on classifier chains and attention\nnetworks, combined with uncertainty loss weighting strategies. Our approach\nsurpasses the challenge baseline by a wide margin on all four tasks.\n", "date": "2022-09-15", "categories": []}
{"id": "2209.10088", "title": "Boosting Star-GANs for Voice Conversion with Contrastive Discriminator", "abstract": "  Nonparallel multi-domain voice conversion methods such as the StarGAN-VCs\nhave been widely applied in many scenarios. However, the training of these\nmodels usually poses a challenge due to their complicated adversarial network\narchitectures. To address this, in this work we leverage the state-of-the-art\ncontrastive learning techniques and incorporate an efficient Siamese network\nstructure into the StarGAN discriminator. Our method is called\nSimSiam-StarGAN-VC and it boosts the training stability and effectively\nprevents the discriminator overfitting issue in the training process. We\nconduct experiments on the Voice Conversion Challenge (VCC 2018) dataset, plus\na user study to validate the performance of our framework. Our experimental\nresults show that SimSiam-StarGAN-VC significantly outperforms existing\nStarGAN-VC methods in terms of both the objective and subjective metrics.\n", "date": "2022-09-21", "categories": []}
{"id": "2209.12045", "title": "Song Emotion Recognition: a Performance Comparison Between Audio\n  Features and Artificial Neural Networks", "abstract": "  When songs are composed or performed, there is often an intent by the\nsinger/songwriter of expressing feelings or emotions through it. For humans,\nmatching the emotiveness in a musical composition or performance with the\nsubjective perception of an audience can be quite challenging. Fortunately, the\nmachine learning approach for this problem is simpler. Usually, it takes a\ndata-set, from which audio features are extracted to present this information\nto a data-driven model, that will, in turn, train to predict what is the\nprobability that a given song matches a target emotion. In this paper, we\nstudied the most common features and models used in recent publications to\ntackle this problem, revealing which ones are best suited for recognizing\nemotion in a cappella songs.\n", "date": "2022-09-24", "categories": []}
{"id": "2209.11527", "title": "An artificial neural network-based system for detecting machine failures\n  using tiny sound data: A case study", "abstract": "  In an effort to advocate the research for a deep learning-based machine\nfailure detection system, we present a case study of our proposed system based\non a tiny sound dataset. Our case study investigates a variational autoencoder\n(VAE) for augmenting a small drill sound dataset from Valmet AB. A Valmet\ndataset contains 134 sounds that have been divided into two categories:\n\"Anomaly\" and \"Normal\" recorded from a drilling machine in Valmet AB, a company\nin Sundsvall, Sweden that supplies equipment and processes for the production\nof biofuels. Using deep learning models to detect failure drills on such a\nsmall sound dataset is typically unsuccessful. We employed a VAE to increase\nthe number of sounds in the tiny dataset by synthesizing new sounds from\noriginal sounds. The augmented dataset was created by combining these\nsynthesized sounds with the original sounds. We used a high-pass filter with a\npassband frequency of 1000 Hz and a low-pass filter with a passband frequency\nof 22\\kern 0.16667em000 Hz to pre-process sounds in the augmented dataset\nbefore transforming them to Mel spectrograms. The pre-trained 2D-CNN Alexnet\nwas then trained using these Mel spectrograms. When compared to using the\noriginal tiny sound dataset to train pre-trained Alexnet, using the augmented\nsound dataset enhanced the CNN model's classification results by 6.62\\%(94.12\\%\nwhen trained on the augmented dataset versus 87.5\\% when trained on the\noriginal dataset).\n", "date": "2022-09-23", "categories": []}
{"id": "2209.11625", "title": "The SpeakIn Speaker Verification System for Far-Field Speaker\n  Verification Challenge 2022", "abstract": "  This paper describes speaker verification (SV) systems submitted by the\nSpeakIn team to the Task 1 and Task 2 of the Far-Field Speaker Verification\nChallenge 2022 (FFSVC2022). SV tasks of the challenge focus on the problem of\nfully supervised far-field speaker verification (Task 1) and semi-supervised\nfar-field speaker verification (Task 2). In Task 1, we used the VoxCeleb and\nFFSVC2020 datasets as train datasets. And for Task 2, we only used the VoxCeleb\ndataset as train set. The ResNet-based and RepVGG-based architectures were\ndeveloped for this challenge. Global statistic pooling structure and MQMHA\npooling structure were used to aggregate the frame-level features across time\nto obtain utterance-level representation. We adopted AM-Softmax and AAM-Softmax\nto classify the resulting embeddings. We innovatively propose a staged transfer\nlearning method. In the pre-training stage we reserve the speaker weights, and\nthere are no positive samples to train them in this stage. Then we fine-tune\nthese weights with both positive and negative samples in the second stage.\nCompared with the traditional transfer learning strategy, this strategy can\nbetter improve the model performance. The Sub-Mean and AS-Norm backend methods\nwere used to solve the problem of domain mismatch. In the fusion stage, three\nmodels were fused in Task1 and two models were fused in Task2. On the FFSVC2022\nleaderboard, the EER of our submission is 3.0049% and the corresponding minDCF\nis 0.2938 in Task1. In Task2, EER and minDCF are 6.2060% and 0.5232\nrespectively. Our approach leads to excellent performance and ranks 1st in both\nchallenge tasks.\n", "date": "2022-09-23", "categories": []}
{"id": "2209.11003", "title": "Predicting pairwise preferences between TTS audio stimuli using parallel\n  ratings data and anti-symmetric twin neural networks", "abstract": "  Automatically predicting the outcome of subjective listening tests is a\nchallenging task. Ratings may vary from person to person even if preferences\nare consistent across listeners. While previous work has focused on predicting\nlisteners' ratings (mean opinion scores) of individual stimuli, we focus on the\nsimpler task of predicting subjective preference given two speech stimuli for\nthe same text. We propose a model based on anti-symmetric twin neural networks,\ntrained on pairs of waveforms and their corresponding preference scores. We\nexplore both attention and recurrent neural nets to account for the fact that\nstimuli in a pair are not time aligned. To obtain a large training set we\nconvert listeners' ratings from MUSHRA tests to values that reflect how often\none stimulus in the pair was rated higher than the other. Specifically, we\nevaluate performance on data obtained from twelve MUSHRA evaluations conducted\nover five years, containing different TTS systems, built from data of different\nspeakers. Our results compare favourably to a state-of-the-art model trained to\npredict MOS scores.\n", "date": "2022-09-22", "categories": []}
{"id": "2209.11061", "title": "Cross-domain Voice Activity Detection with Self-Supervised\n  Representations", "abstract": "  Voice Activity Detection (VAD) aims at detecting speech segments on an audio\nsignal, which is a necessary first step for many today's speech based\napplications. Current state-of-the-art methods focus on training a neural\nnetwork exploiting features directly contained in the acoustics, such as Mel\nFilter Banks (MFBs). Such methods therefore require an extra normalisation step\nto adapt to a new domain where the acoustics is impacted, which can be simply\ndue to a change of speaker, microphone, or environment. In addition, this\nnormalisation step is usually a rather rudimentary method that has certain\nlimitations, such as being highly susceptible to the amount of data available\nfor the new domain. Here, we exploited the crowd-sourced Common Voice (CV)\ncorpus to show that representations based on Self-Supervised Learning (SSL) can\nadapt well to different domains, because they are computed with contextualised\nrepresentations of speech across multiple domains. SSL representations also\nachieve better results than systems based on hand-crafted representations\n(MFBs), and off-the-shelf VADs, with significant improvement in cross-domain\nsettings.\n", "date": "2022-09-22", "categories": []}
{"id": "2209.09010", "title": "The Royalflush System for VoxCeleb Speaker Recognition Challenge 2022", "abstract": "  In this technical report, we describe the Royalflush submissions for the\nVoxCeleb Speaker Recognition Challenge 2022 (VoxSRC-22). Our submissions\ncontain track 1, which is for supervised speaker verification and track 3,\nwhich is for semi-supervised speaker verification. For track 1, we develop a\npowerful U-Net-based speaker embedding extractor with a symmetric architecture.\nThe proposed system achieves 2.06% in EER and 0.1293 in MinDCF on the\nvalidation set. Compared with the state-of-the-art ECAPA-TDNN, it obtains a\nrelative improvement of 20.7% in EER and 22.70% in MinDCF. For track 3, we\nemploy the joint training of source domain supervision and target domain\nself-supervision to get a speaker embedding extractor. The subsequent\nclustering process can obtain target domain pseudo-speaker labels. We adapt the\nspeaker embedding extractor using all source and target domain data in a\nsupervised manner, where it can fully leverage both domain information.\nMoreover, clustering and supervised domain adaptation can be repeated until the\nperformance converges on the validation set. Our final submission is a fusion\nof 10 models and achieves 7.75% EER and 0.3517 MinDCF on the validation set.\n", "date": "2022-09-19", "categories": []}
{"id": "2209.09634", "title": "A Closer Look at Weakly-Supervised Audio-Visual Source Localization", "abstract": "  Audio-visual source localization is a challenging task that aims to predict\nthe location of visual sound sources in a video. Since collecting ground-truth\nannotations of sounding objects can be costly, a plethora of weakly-supervised\nlocalization methods that can learn from datasets with no bounding-box\nannotations have been proposed in recent years, by leveraging the natural\nco-occurrence of audio and visual signals. Despite significant interest,\npopular evaluation protocols have two major flaws. First, they allow for the\nuse of a fully annotated dataset to perform early stopping, thus significantly\nincreasing the annotation effort required for training. Second, current\nevaluation metrics assume the presence of sound sources at all times. This is\nof course an unrealistic assumption, and thus better metrics are necessary to\ncapture the model's performance on (negative) samples with no visible sound\nsources. To accomplish this, we extend the test set of popular benchmarks,\nFlickr SoundNet and VGG-Sound Sources, in order to include negative samples,\nand measure performance using metrics that balance localization accuracy and\nrecall. Using the new protocol, we conducted an extensive evaluation of prior\nmethods, and found that most prior works are not capable of identifying\nnegatives and suffer from significant overfitting problems (rely heavily on\nearly stopping for best results). We also propose a new approach for visual\nsound source localization that addresses both these problems. In particular, we\nfound that, through extreme visual dropout and the use of momentum encoders,\nthe proposed approach combats overfitting effectively, and establishes a new\nstate-of-the-art performance on both Flickr SoundNet and VGG-Sound Source. Code\nand pre-trained models are available at https://github.com/stoneMo/SLAVC.\n", "date": "2022-08-30", "categories": []}
{"id": "2209.08774", "title": "Playing Technique Detection by Fusing Note Onset Information in Guzheng\n  Performance", "abstract": "  The Guzheng is a kind of traditional Chinese instruments with diverse playing\ntechniques. Instrument playing techniques (IPT) play an important role in\nmusical performance. However, most of the existing works for IPT detection show\nlow efficiency for variable-length audio and provide no assurance in the\ngeneralization as they rely on a single sound bank for training and testing. In\nthis study, we propose an end-to-end Guzheng playing technique detection system\nusing Fully Convolutional Networks that can be applied to variable-length\naudio. Because each Guzheng playing technique is applied to a note, a dedicated\nonset detector is trained to divide an audio into several notes and its\npredictions are fused with frame-wise IPT predictions. During fusion, we add\nthe IPT predictions frame by frame inside each note and get the IPT with the\nhighest probability within each note as the final output of that note. We\ncreate a new dataset named GZ_IsoTech from multiple sound banks and real-world\nrecordings for Guzheng performance analysis. Our approach achieves 87.97% in\nframe-level accuracy and 80.76% in note-level F1-score, outperforming existing\nworks by a large margin, which indicates the effectiveness of our proposed\nmethod in IPT detection.\n", "date": "2022-09-19", "categories": []}
{"id": "2209.06581", "title": "Applying wav2vec2 for Speech Recognition on Bengali Common Voices\n  Dataset", "abstract": "  Speech is inherently continuous, where discrete words, phonemes and other\nunits are not clearly segmented, and so speech recognition has been an active\nresearch problem for decades. In this work we have fine-tuned wav2vec 2.0 to\nrecognize and transcribe Bengali speech -- training it on the Bengali Common\nVoice Speech Dataset. After training for 71 epochs, on a training set\nconsisting of 36919 mp3 files, we achieved a training loss of 0.3172 and WER of\n0.2524 on a validation set of size 7,747. Using a 5-gram language model, the\nLevenshtein Distance was 2.6446 on a test set of size 7,747. Then the training\nset and validation set were combined, shuffled and split into 85-15 ratio.\nTraining for 7 more epochs on this combined dataset yielded an improved\nLevenshtein Distance of 2.60753 on the test set. Our model was the best\nperforming one, achieving a Levenshtein Distance of 6.234 on a hidden dataset,\nwhich was 1.1049 units lower than other competing submissions.\n", "date": "2022-09-11", "categories": []}
{"id": "2209.05900", "title": "Binaural Signal Representations for Joint Sound Event Detection and\n  Acoustic Scene Classification", "abstract": "  Sound event detection (SED) and Acoustic scene classification (ASC) are two\nwidely researched audio tasks that constitute an important part of research on\nacoustic scene analysis. Considering shared information between sound events\nand acoustic scenes, performing both tasks jointly is a natural part of a\ncomplex machine listening system. In this paper, we investigate the usefulness\nof several spatial audio features in training a joint deep neural network (DNN)\nmodel performing SED and ASC. Experiments are performed for two different\ndatasets containing binaural recordings and synchronous sound event and\nacoustic scene labels to analyse the differences between performing SED and ASC\nseparately or jointly. The presented results show that the use of specific\nbinaural features, mainly the Generalized Cross Correlation with Phase\nTransform (GCC-phat) and sines and cosines of phase differences, result in a\nbetter performing model in both separate and joint tasks as compared with\nbaseline methods based on logmel energies only.\n", "date": "2022-09-13", "categories": []}
{"id": "2209.04071", "title": "Audio Analytics-based Human Trafficking Detection Framework for\n  Autonomous Vehicles", "abstract": "  Human trafficking is a universal problem, persistent despite numerous efforts\nto combat it globally. Individuals of any age, race, ethnicity, sex, gender\nidentity, sexual orientation, nationality, immigration status, cultural\nbackground, religion, socioeconomic class, and education can be a victim of\nhuman trafficking. With the advancements in technology and the introduction of\nautonomous vehicles (AVs), human traffickers will adopt new ways to transport\nvictims, which could accelerate the growth of organized human trafficking\nnetworks, which can make the detection of trafficking in persons more\nchallenging for law enforcement agencies. The objective of this study is to\ndevelop an innovative audio analytics-based human trafficking detection\nframework for autonomous vehicles. The primary contributions of this study are\nto: (i) define four non-trivial, feasible, and realistic human trafficking\nscenarios for AVs; (ii) create a new and comprehensive audio dataset related to\nhuman trafficking with five classes i.e., crying, screaming, car door banging,\ncar noise, and conversation; and (iii) develop a deep 1-D Convolution Neural\nNetwork (CNN) architecture for audio data classification related to human\ntrafficking. We have also conducted a case study using the new audio dataset\nand evaluated the audio classification performance of the deep 1-D CNN. Our\nanalyses reveal that the deep 1-D CNN can distinguish sound coming from a human\ntrafficking victim from a non-human trafficking sound with an accuracy of 95%,\nwhich proves the efficacy of our framework.\n", "date": "2022-09-09", "categories": []}
{"id": "2209.04109", "title": "MATT: A Multiple-instance Attention Mechanism for Long-tail Music Genre\n  Classification", "abstract": "  Imbalanced music genre classification is a crucial task in the Music\nInformation Retrieval (MIR) field for identifying the long-tail, data-poor\ngenre based on the related music audio segments, which is very prevalent in\nreal-world scenarios. Most of the existing models are designed for\nclass-balanced music datasets, resulting in poor performance in accuracy and\ngeneralization when identifying the music genres at the tail of the\ndistribution. Inspired by the success of introducing Multi-instance Learning\n(MIL) in various classification tasks, we propose a novel mechanism named\nMulti-instance Attention (MATT) to boost the performance for identifying tail\nclasses. Specifically, we first construct the bag-level datasets by generating\nthe album-artist pair bags. Second, we leverage neural networks to encode the\nmusic audio segments. Finally, under the guidance of a multi-instance attention\nmechanism, the neural network-based models could select the most informative\ngenre to match the given music segment. Comprehensive experimental results on a\nlarge-scale music genre benchmark dataset with long-tail distribution\ndemonstrate MATT significantly outperforms other state-of-the-art baselines.\n", "date": "2022-09-09", "categories": []}
{"id": "2209.03711", "title": "What Did I Just Hear? Detecting Pornographic Sounds in Adult Videos\n  Using Neural Networks", "abstract": "  Audio-based pornographic detection enables efficient adult content filtering\nwithout sacrificing performance by exploiting distinct spectral\ncharacteristics. To improve it, we explore pornographic sound modeling based on\ndifferent neural architectures and acoustic features. We find that CNN trained\non log mel spectrogram achieves the best performance on Pornography-800\ndataset. Our experiment results also show that log mel spectrogram allows\nbetter representations for the models to recognize pornographic sounds.\nFinally, to classify whole audio waveforms rather than segments, we employ\nvoting segment-to-audio technique that yields the best audio-level detection\nresults.\n", "date": "2022-09-08", "categories": []}
{"id": "2209.01374", "title": "Identify The Beehive Sound Using Deep Learning", "abstract": "  Flowers play an essential role in removing the duller from the environment.\nThe life cycle of the flowering plants involves pollination, fertilization,\nflowering, seed-formation, dispersion, and germination. Honeybees pollinate\napproximately 75% of all flowering plants. Environmental pollution, climate\nchange, natural landscape demolition, and so on, threaten the natural habitats,\nthus continuously reducing the number of honeybees. As a result, several\nresearchers are attempting to resolve this issue. Applying acoustic\nclassification to recordings of beehive sounds may be a way of detecting\nchanges within them. In this research, we use deep learning techniques, namely\nSequential Neural Network, Convolutional Neural Network, and Recurrent Neural\nNetwork, on the recorded sounds to classify bee sounds from the nonbeehive\nnoises. In addition, we perform a comparative study among some popular non-deep\nlearning techniques, namely Support Vector Machine, Decision Tree, Random\nForest, and Na\\\"ive Bayes, with the deep learning techniques. The techniques\nare also verified on the combined recorded sounds (25-75% noises).\n", "date": "2022-09-03", "categories": []}
{"id": "2209.01478", "title": "Equivariant Self-Supervision for Musical Tempo Estimation", "abstract": "  Self-supervised methods have emerged as a promising avenue for representation\nlearning in the recent years since they alleviate the need for labeled\ndatasets, which are scarce and expensive to acquire. Contrastive methods are a\npopular choice for self-supervision in the audio domain, and typically provide\na learning signal by forcing the model to be invariant to some transformations\nof the input. These methods, however, require measures such as negative\nsampling or some form of regularisation to be taken to prevent the model from\ncollapsing on trivial solutions. In this work, instead of invariance, we\npropose to use equivariance as a self-supervision signal to learn audio tempo\nrepresentations from unlabelled data. We derive a simple loss function that\nprevents the network from collapsing on a trivial solution during training,\nwithout requiring any form of regularisation or negative sampling. Our\nexperiments show that it is possible to learn meaningful representations for\ntempo estimation by solely relying on equivariant self-supervision, achieving\nperformance comparable with supervised methods on several benchmarks. As an\nadded benefit, our method only requires moderate compute resources and\ntherefore remains accessible to a wide research community.\n", "date": "2022-09-03", "categories": []}
{"id": "2209.00672", "title": "Exploring traditional machine learning for identification of\n  pathological auscultations", "abstract": "  Today, data collection has improved in various areas, and the medical domain\nis no exception. Auscultation, as an important diagnostic technique for\nphysicians, due to the progress and availability of digital stethoscopes, lends\nitself well to applications of machine learning. Due to the large number of\nauscultations performed, the availability of data opens up an opportunity for\nmore effective analysis of sounds where prognostic accuracy even among experts\nremains low. In this study, digital 6-channel auscultations of 45 patients were\nused in various machine learning scenarios, with the aim of distinguishing\nbetween normal and anomalous pulmonary sounds. Audio features (such as\nfundamental frequencies F0-4, loudness, HNR, DFA, as well as descriptive\nstatistics of log energy, RMS and MFCC) were extracted using the Python library\nSurfboard. Windowing and feature aggregation and concatenation strategies were\nused to prepare data for tree-based ensemble models in unsupervised (fair-cut\nforest) and supervised (random forest) machine learning settings. The\nevaluation was carried out using 9-fold stratified cross-validation repeated 30\ntimes. Decision fusion by averaging outputs for a subject was tested and found\nto be useful. Supervised models showed a consistent advantage over unsupervised\nones, achieving mean AUC ROC of 0.691 (accuracy 71.11%, Kappa 0.416, F1-score\n0.771) in side-based detection and mean AUC ROC of 0.721 (accuracy 68.89%,\nKappa 0.371, F1-score 0.650) in patient-based detection.\n", "date": "2022-09-01", "categories": []}
{"id": "2209.00934", "title": "TB or not TB? Acoustic cough analysis for tuberculosis classification", "abstract": "  In this work, we explore recurrent neural network architectures for\ntuberculosis (TB) cough classification. In contrast to previous unsuccessful\nattempts to implement deep architectures in this domain, we show that a basic\nbidirectional long short-term memory network (BiLSTM) can achieve improved\nperformance. In addition, we show that by performing greedy feature selection\nin conjunction with a newly-proposed attention-based architecture that learns\npatient invariant features, substantially better generalisation can be achieved\ncompared to a baseline and other considered architectures. Furthermore, this\nattention mechanism allows an inspection of the temporal regions of the audio\nsignal considered to be important for classification to be performed. Finally,\nwe develop a neural style transfer technique to infer idealised inputs which\ncan subsequently be analysed. We find distinct differences between the\nidealised power spectra of TB and non-TB coughs, which provide clues about the\norigin of the features in the audio signal.\n", "date": "2022-09-02", "categories": []}
{"id": "2209.00130", "title": "Evaluating generative audio systems and their metrics", "abstract": "  Recent years have seen considerable advances in audio synthesis with deep\ngenerative models. However, the state-of-the-art is very difficult to quantify;\ndifferent studies often use different evaluation methodologies and different\nmetrics when reporting results, making a direct comparison to other systems\ndifficult if not impossible. Furthermore, the perceptual relevance and meaning\nof the reported metrics in most cases unknown, prohibiting any conclusive\ninsights with respect to practical usability and audio quality. This paper\npresents a study that investigates state-of-the-art approaches side-by-side\nwith (i) a set of previously proposed objective metrics for audio\nreconstruction, and with (ii) a listening study. The results indicate that\ncurrently used objective metrics are insufficient to describe the perceptual\nquality of current systems.\n", "date": "2022-08-31", "categories": []}
{"id": "2208.11428", "title": "Automatic music mixing with deep learning and out-of-domain data", "abstract": "  Music mixing traditionally involves recording instruments in the form of\nclean, individual tracks and blending them into a final mixture using audio\neffects and expert knowledge (e.g., a mixing engineer). The automation of music\nproduction tasks has become an emerging field in recent years, where rule-based\nmethods and machine learning approaches have been explored. Nevertheless, the\nlack of dry or clean instrument recordings limits the performance of such\nmodels, which is still far from professional human-made mixes. We explore\nwhether we can use out-of-domain data such as wet or processed multitrack music\nrecordings and repurpose it to train supervised deep learning models that can\nbridge the current gap in automatic mixing quality. To achieve this we propose\na novel data preprocessing method that allows the models to perform automatic\nmusic mixing. We also redesigned a listening test method for evaluating music\nmixing systems. We validate our results through such subjective tests using\nhighly experienced mixing engineers as participants.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.13285", "title": "Computing with Hypervectors for Efficient Speaker Identification", "abstract": "  We introduce a method to identify speakers by computing with high-dimensional\nrandom vectors. Its strengths are simplicity and speed. With only 1.02k active\nparameters and a 128-minute pass through the training data we achieve Top-1 and\nTop-5 scores of 31% and 52% on the VoxCeleb1 dataset of 1,251 speakers. This is\nin contrast to CNN models requiring several million parameters and orders of\nmagnitude higher computational complexity for only a 2$\\times$ gain in\ndiscriminative power as measured in mutual information. An additional 92\nseconds of training with Generalized Learning Vector Quantization (GLVQ) raises\nthe scores to 48% and 67%. A trained classifier classifies 1 second of speech\nin 5.7 ms. All processing was done on standard CPU-based machines.\n", "date": "2022-08-28", "categories": []}
{"id": "2208.12387", "title": "Music Separation Enhancement with Generative Modeling", "abstract": "  Despite phenomenal progress in recent years, state-of-the-art music\nseparation systems produce source estimates with significant perceptual\nshortcomings, such as adding extraneous noise or removing harmonics. We propose\na post-processing model (the Make it Sound Good (MSG) post-processor) to\nenhance the output of music source separation systems. We apply our\npost-processing model to state-of-the-art waveform-based and spectrogram-based\nmusic source separators, including a separator unseen by MSG during training.\nOur analysis of the errors produced by source separators shows that waveform\nmodels tend to introduce more high-frequency noise, while spectrogram models\ntend to lose transients and high frequency content. We introduce objective\nmeasures to quantify both kinds of errors and show MSG improves the source\nreconstruction of both kinds of errors. Crowdsourced subjective evaluations\ndemonstrate that human listeners prefer source estimates of bass and drums that\nhave been post-processed by MSG.\n", "date": "2022-08-26", "categories": []}
{"id": "2208.12782", "title": "Mel Spectrogram Inversion with Stable Pitch", "abstract": "  Vocoders are models capable of transforming a low-dimensional spectral\nrepresentation of an audio signal, typically the mel spectrogram, to a\nwaveform. Modern speech generation pipelines use a vocoder as their final\ncomponent. Recent vocoder models developed for speech achieve a high degree of\nrealism, such that it is natural to wonder how they would perform on music\nsignals. Compared to speech, the heterogeneity and structure of the musical\nsound texture offers new challenges. In this work we focus on one specific\nartifact that some vocoder models designed for speech tend to exhibit when\napplied to music: the perceived instability of pitch when synthesizing\nsustained notes. We argue that the characteristic sound of this artifact is due\nto the lack of horizontal phase coherence, which is often the result of using a\ntime-domain target space with a model that is invariant to time-shifts, such as\na convolutional neural network. We propose a new vocoder model that is\nspecifically designed for music. Key to improving the pitch stability is the\nchoice of a shift-invariant target space that consists of the magnitude\nspectrum and the phase gradient. We discuss the reasons that inspired us to\nre-formulate the vocoder task, outline a working example, and evaluate it on\nmusical signals. Our method results in 60% and 10% improved reconstruction of\nsustained notes and chords with respect to existing models, using a novel\nharmonic error metric.\n", "date": "2022-08-26", "categories": []}
{"id": "2208.12086", "title": "A Study on Broadcast Networks for Music Genre Classification", "abstract": "  Due to the increased demand for music streaming/recommender services and the\nrecent developments of music information retrieval frameworks, Music Genre\nClassification (MGC) has attracted the community's attention. However,\nconvolutional-based approaches are known to lack the ability to efficiently\nencode and localize temporal features. In this paper, we study the\nbroadcast-based neural networks aiming to improve the localization and\ngeneralizability under a small set of parameters (about 180k) and investigate\ntwelve variants of broadcast networks discussing the effect of block\nconfiguration, pooling method, activation function, normalization mechanism,\nlabel smoothing, channel interdependency, LSTM block inclusion, and variants of\ninception schemes. Our computational experiments using relevant datasets such\nas GTZAN, Extended Ballroom, HOMBURG, and Free Music Archive (FMA) show\nstate-of-the-art classification accuracies in Music Genre Classification. Our\napproach offers insights and the potential to enable compact and generalizable\nbroadcast networks for music and audio classification.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.11402", "title": "Improved Zero-Shot Audio Tagging & Classification with Patchout\n  Spectrogram Transformers", "abstract": "  Standard machine learning models for tagging and classifying acoustic signals\ncannot handle classes that were not seen during training. Zero-Shot (ZS)\nlearning overcomes this restriction by predicting classes based on adaptable\nclass descriptions. This study sets out to investigate the effectiveness of\nself-attention-based audio embedding architectures for ZS learning. To this\nend, we compare the very recent patchout spectrogram transformer with two\nclassic convolutional architectures. We evaluate these three architectures on\nthree tasks and on three different benchmark datasets: general-purpose tagging\non AudioSet, environmental sound classification on ESC-50, and instrument\ntagging on OpenMIC. Our results show that the self-attention-based embedding\nmethods outperform both compared convolutional architectures in all of these\nsettings. By designing training and test data accordingly, we observe that\nprediction performance suffers significantly when the `semantic distance'\nbetween training and new test classes is large, an effect that will deserve\nmore detailed investigations.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.10659", "title": "Fall Detection from Audios with Audio Transformers", "abstract": "  Fall detection for the elderly is a well-researched problem with several\nproposed solutions, including wearable and non-wearable techniques. While the\nexisting techniques have excellent detection rates, their adoption by the\ntarget population is lacking due to the need for wearing devices and user\nprivacy concerns. Our paper provides a novel, non-wearable, non-intrusive, and\nscalable solution for fall detection, deployed on an autonomous mobile robot\nequipped with a microphone. The proposed method uses ambient sound input\nrecorded in people's homes. We specifically target the bathroom environment as\nit is highly prone to falls and where existing techniques cannot be deployed\nwithout jeopardizing user privacy. The present work develops a solution based\non a Transformer architecture that takes noisy sound input from bathrooms and\nclassifies it into fall/no-fall class with an accuracy of 0.8673. Further, the\nproposed approach is extendable to other indoor environments, besides bathrooms\nand is suitable for deploying in elderly homes, hospitals, and rehabilitation\nfacilities without requiring the user to wear any device or be constantly\n\"watched\" by the sensors.\n", "date": "2022-08-23", "categories": []}
{"id": "2208.09618", "title": "Fully Automated End-to-End Fake Audio Detection", "abstract": "  The existing fake audio detection systems often rely on expert experience to\ndesign the acoustic features or manually design the hyperparameters of the\nnetwork structure. However, artificial adjustment of the parameters can have a\nrelatively obvious influence on the results. It is almost impossible to\nmanually set the best set of parameters. Therefore this paper proposes a fully\nautomated end-toend fake audio detection method. We first use wav2vec\npre-trained model to obtain a high-level representation of the speech.\nFurthermore, for the network structure, we use a modified version of the\ndifferentiable architecture search (DARTS) named light-DARTS. It learns deep\nspeech representations while automatically learning and optimizing complex\nneural structures consisting of convolutional operations and residual blocks.\nThe experimental results on the ASVspoof 2019 LA dataset show that our proposed\nsystem achieves an equal error rate (EER) of 1.08%, which outperforms the\nstate-of-the-art single system.\n", "date": "2022-08-20", "categories": []}
{"id": "2208.09646", "title": "An Initial Investigation for Detecting Vocoder Fingerprints of Fake\n  Audio", "abstract": "  Many effective attempts have been made for fake audio detection. However,\nthey can only provide detection results but no countermeasures to curb this\nharm. For many related practical applications, what model or algorithm\ngenerated the fake audio also is needed. Therefore, We propose a new problem\nfor detecting vocoder fingerprints of fake audio. Experiments are conducted on\nthe datasets synthesized by eight state-of-the-art vocoders. We have\npreliminarily explored the features and model architectures. The t-SNE\nvisualization shows that different vocoders generate distinct vocoder\nfingerprints.\n", "date": "2022-08-20", "categories": []}
{"id": "2207.13843", "title": "Deep Learning-Based Acoustic Mosquito Detection in Noisy Conditions\n  Using Trainable Kernels and Augmentations", "abstract": "  In this paper, we demonstrate a unique recipe to enhance the effectiveness of\naudio machine learning approaches by fusing pre-processing techniques into a\ndeep learning model. Our solution accelerates training and inference\nperformance by optimizing hyper-parameters through training instead of costly\nrandom searches to build a reliable mosquito detector from audio signals. The\nexperiments and the results presented here are part of the MOS C submission of\nthe ACM 2022 challenge. Our results outperform the published baseline by 212%\non the unpublished test set. We believe that this is one of the best real-world\nexamples of building a robust bio-acoustic system that provides reliable\nmosquito detection in noisy conditions.\n", "date": "2022-07-28", "categories": []}
{"id": "2208.09096", "title": "Representation Learning for the Automatic Indexing of Sound Effects\n  Libraries", "abstract": "  Labeling and maintaining a commercial sound effects library is a\ntime-consuming task exacerbated by databases that continually grow in size and\nundergo taxonomy updates. Moreover, sound search and taxonomy creation are\ncomplicated by non-uniform metadata, an unrelenting problem even with the\nintroduction of a new industry standard, the Universal Category System. To\naddress these problems and overcome dataset-dependent limitations that inhibit\nthe successful training of deep learning models, we pursue representation\nlearning to train generalized embeddings that can be used for a wide variety of\nsound effects libraries and are a taxonomy-agnostic representation of sound. We\nshow that a task-specific but dataset-independent representation can\nsuccessfully address data issues such as class imbalance, inconsistent class\nlabels, and insufficient dataset size, outperforming established\nrepresentations such as OpenL3. Detailed experimental results show the impact\nof metric learning approaches and different cross-dataset training methods on\nrepresentational effectiveness.\n", "date": "2022-08-18", "categories": []}
{"id": "2208.09201", "title": "Improving Post-Processing of Audio Event Detectors Using Reinforcement\n  Learning", "abstract": "  We apply post-processing to the class probability distribution outputs of\naudio event classification models and employ reinforcement learning to jointly\ndiscover the optimal parameters for various stages of a post-processing stack,\nsuch as the classification thresholds and the kernel sizes of median filtering\nalgorithms used to smooth out model predictions. To achieve this we define a\nreinforcement learning environment where: 1) a state is the class probability\ndistribution provided by the model for a given audio sample, 2) an action is\nthe choice of a candidate optimal value for each parameter of the\npost-processing stack, 3) the reward is based on the classification accuracy\nmetric we aim to optimize, which is the audio event-based macro F1-score in our\ncase. We apply our post-processing to the class probability distribution\noutputs of two audio event classification models submitted to the DCASE Task4\n2020 challenge. We find that by using reinforcement learning to discover the\noptimal per-class parameters for the post-processing stack that is applied to\nthe outputs of audio event classification models, we can improve the audio\nevent-based macro F1-score (the main metric used in the DCASE challenge to\ncompare audio event classification accuracy) by 4-5% compared to using the same\npost-processing stack with manually tuned parameters.\n", "date": "2022-08-19", "categories": []}
{"id": "2208.07994", "title": "Enhancing Audio Perception of Music By AI Picked Room Acoustics", "abstract": "  Every sound that we hear is the result of successive convolutional operations\n(e.g. room acoustics, microphone characteristics, resonant properties of the\ninstrument itself, not to mention characteristics and limitations of the sound\nreproduction system). In this work we seek to determine the best room in which\nto perform a particular piece using AI. Additionally, we use room acoustics as\na way to enhance the perceptual qualities of a given sound. Historically, rooms\n(particularly Churches and concert halls) were designed to host and serve\nspecific musical functions. In some cases the architectural acoustical\nqualities enhanced the music performed there. We try to mimic this, as a first\nstep, by designating room impulse responses that would correlate to producing\nenhanced sound quality for particular music. A convolutional architecture is\nfirst trained to take in an audio sample and mimic the ratings of experts with\nabout 78 % accuracy for various instrument families and notes for perceptual\nqualities. This gives us a scoring function for any audio sample which can rate\nthe perceptual pleasantness of a note automatically. Now, via a library of\nabout 60,000 synthetic impulse responses mimicking all kinds of room,\nmaterials, etc, we use a simple convolution operation, to transform the sound\nas if it was played in a particular room. The perceptual evaluator is used to\nrank the musical sounds, and yield the \"best room or the concert hall\" to play\na sound. As a byproduct it can also use room acoustics to turn a poor quality\nsound into a \"good\" sound.\n", "date": "2022-08-16", "categories": []}
{"id": "2208.08354", "title": "Extract fundamental frequency based on CNN combined with PYIN", "abstract": "  This paper refers to the extraction of multiple fundamental frequencies\n(multiple F0) based on PYIN, an algorithm for extracting the fundamental\nfrequency (F0) of monophonic music, and a trained convolutional neural networks\n(CNN) model, where a pitch salience function of the input signal is produced to\nestimate the multiple F0. The implementation of these two algorithms and their\ncorresponding advantages and disadvantages are discussed in this article.\nAnalysing the different performance of these two methods, PYIN is applied to\nsupplement the F0 extracted from the trained CNN model to combine the\nadvantages of these two algorithms. For evaluation, four pieces played by two\nviolins are used, and the performance of the models are evaluated accoring to\nthe flatness of the F0 curve extracted. The result shows the combined model\noutperforms the original algorithms when extracting F0 from monophonic music\nand polyphonic music.\n", "date": "2022-08-17", "categories": []}
{"id": "2207.06958", "title": "Proceedings of the ICML 2022 Expressive Vocalizations Workshop and\n  Competition: Recognizing, Generating, and Personalizing Vocal Bursts", "abstract": "  This is the Proceedings of the ICML Expressive Vocalization (ExVo)\nCompetition. The ExVo competition focuses on understanding and generating vocal\nbursts: laughs, gasps, cries, and other non-verbal vocalizations that are\ncentral to emotional expression and communication. ExVo 2022, included three\ncompetition tracks using a large-scale dataset of 59,201 vocalizations from\n1,702 speakers. The first, ExVo-MultiTask, requires participants to train a\nmulti-task model to recognize expressed emotions and demographic traits from\nvocal bursts. The second, ExVo-Generate, requires participants to train a\ngenerative model that produces vocal bursts conveying ten different emotions.\nThe third, ExVo-FewShot, requires participants to leverage few-shot learning\nincorporating speaker identity to train a model for the recognition of 10\nemotions conveyed by vocal bursts.\n", "date": "2022-07-14", "categories": []}
{"id": "2206.12772", "title": "Exploiting Transformation Invariance and Equivariance for\n  Self-supervised Sound Localisation", "abstract": "  We present a simple yet effective self-supervised framework for audio-visual\nrepresentation learning, to localize the sound source in videos. To understand\nwhat enables to learn useful representations, we systematically investigate the\neffects of data augmentations, and reveal that (1) composition of data\naugmentations plays a critical role, i.e. explicitly encouraging the\naudio-visual representations to be invariant to various transformations~({\\em\ntransformation invariance}); (2) enforcing geometric consistency substantially\nimproves the quality of learned representations, i.e. the detected sound source\nshould follow the same transformation applied on input video frames~({\\em\ntransformation equivariance}). Extensive experiments demonstrate that our model\nsignificantly outperforms previous methods on two sound localization\nbenchmarks, namely, Flickr-SoundNet and VGG-Sound. Additionally, we also\nevaluate audio retrieval and cross-modal retrieval tasks. In both cases, our\nself-supervised models demonstrate superior retrieval performances, even\ncompetitive with the supervised approach in audio retrieval. This reveals the\nproposed framework learns strong multi-modal representations that are\nbeneficial to sound localisation and generalization to further applications.\n\\textit{All codes will be available}.\n", "date": "2022-06-26", "categories": []}
{"id": "2208.06127", "title": "An investigation on selecting audio pre-trained models for audio\n  captioning", "abstract": "  Audio captioning is a task that generates description of audio based on\ncontent. Pre-trained models are widely used in audio captioning due to high\ncomplexity. Unless a comprehensive system is re-trained, it is hard to\ndetermine how well pre-trained models contribute to audio captioning system. To\nprevent the time consuming and energy consuming process of retraining, it is\nnecessary to propose a preditor of performance for the pre-trained model in\naudio captioning. In this paper, a series of pre-trained models are\ninvestigated for the correlation between extracted audio features and the\nperformance of audio captioning. A couple of predictor is proposed based on the\nexperiment results.The result demonstrates that the kurtosis and skewness of\naudio features extracted may act as an indicator of the performance of audio\ncaptioning systems for pre-trained audio due to the high correlation between\nkurtosis and skewness of audio features and the performance of audio captioning\nsystems.\n", "date": "2022-08-12", "categories": []}
{"id": "2208.06169", "title": "DDX7: Differentiable FM Synthesis of Musical Instrument Sounds", "abstract": "  FM Synthesis is a well-known algorithm used to generate complex timbre from a\ncompact set of design primitives. Typically featuring a MIDI interface, it is\nusually impractical to control it from an audio source. On the other hand,\nDifferentiable Digital Signal Processing (DDSP) has enabled nuanced audio\nrendering by Deep Neural Networks (DNNs) that learn to control differentiable\nsynthesis layers from arbitrary sound inputs. The training process involves a\ncorpus of audio for supervision, and spectral reconstruction loss functions.\nSuch functions, while being great to match spectral amplitudes, present a lack\nof pitch direction which can hinder the joint optimization of the parameters of\nFM synthesizers. In this paper, we take steps towards enabling continuous\ncontrol of a well-established FM synthesis architecture from an audio input.\nFirstly, we discuss a set of design constraints that ease spectral optimization\nof a differentiable FM synthesizer via a standard reconstruction loss. Next, we\npresent Differentiable DX7 (DDX7), a lightweight architecture for neural FM\nresynthesis of musical instrument sounds in terms of a compact set of\nparameters. We train the model on instrument samples extracted from the URMP\ndataset, and quantitatively demonstrate its comparable audio quality against\nselected benchmarks.\n", "date": "2022-08-12", "categories": []}
{"id": "2208.03393", "title": "Chronological Self-Training for Real-Time Speaker Diarization", "abstract": "  Diarization partitions an audio stream into segments based on the voices of\nthe speakers. Real-time diarization systems that include an enrollment step\nshould limit enrollment training samples to reduce user interaction time.\nAlthough training on a small number of samples yields poor performance, we show\nthat the accuracy can be improved dramatically using a chronological\nself-training approach. We studied the tradeoff between training time and\nclassification performance and found that 1 second is sufficient to reach over\n95% accuracy. We evaluated on 700 audio conversation files of about 10 minutes\neach from 6 different languages and demonstrated average diarization error\nrates as low as 10%.\n", "date": "2022-08-05", "categories": []}
{"id": "2208.02765", "title": "Keyword Spotting System and Evaluation of Pruning and Quantization\n  Methods on Low-power Edge Microcontrollers", "abstract": "  Keyword spotting (KWS) is beneficial for voice-based user interactions with\nlow-power devices at the edge. The edge devices are usually always-on, so edge\ncomputing brings bandwidth savings and privacy protection. The devices\ntypically have limited memory spaces, computational performances, power and\ncosts, for example, Cortex-M based microcontrollers. The challenge is to meet\nthe high computation and low-latency requirements of deep learning on these\ndevices. This paper firstly shows our small-footprint KWS system running on\nSTM32F7 microcontroller with Cortex-M7 core @216MHz and 512KB static RAM. Our\nselected convolutional neural network (CNN) architecture has simplified number\nof operations for KWS to meet the constraint of edge devices. Our baseline\nsystem generates classification results for each 37ms including real-time audio\nfeature extraction part. This paper further evaluates the actual performance\nfor different pruning and quantization methods on microcontroller, including\ndifferent granularity of sparsity, skipping zero weights, weight-prioritized\nloop order, and SIMD instruction. The result shows that for microcontrollers,\nthere are considerable challenges for accelerate unstructured pruned models,\nand the structured pruning is more friendly than unstructured pruning. The\nresult also verified that the performance improvement for quantization and SIMD\ninstruction.\n", "date": "2022-08-04", "categories": []}
{"id": "2207.13192", "title": "Perception-Aware Attack: Creating Adversarial Music via\n  Reverse-Engineering Human Perception", "abstract": "  Recently, adversarial machine learning attacks have posed serious security\nthreats against practical audio signal classification systems, including speech\nrecognition, speaker recognition, and music copyright detection. Previous\nstudies have mainly focused on ensuring the effectiveness of attacking an audio\nsignal classifier via creating a small noise-like perturbation on the original\nsignal. It is still unclear if an attacker is able to create audio signal\nperturbations that can be well perceived by human beings in addition to its\nattack effectiveness. This is particularly important for music signals as they\nare carefully crafted with human-enjoyable audio characteristics.\n  In this work, we formulate the adversarial attack against music signals as a\nnew perception-aware attack framework, which integrates human study into\nadversarial attack design. Specifically, we conduct a human study to quantify\nthe human perception with respect to a change of a music signal. We invite\nhuman participants to rate their perceived deviation based on pairs of original\nand perturbed music signals, and reverse-engineer the human perception process\nby regression analysis to predict the human-perceived deviation given a\nperturbed signal. The perception-aware attack is then formulated as an\noptimization problem that finds an optimal perturbation signal to minimize the\nprediction of perceived deviation from the regressed human perception model. We\nuse the perception-aware framework to design a realistic adversarial music\nattack against YouTube's copyright detector. Experiments show that the\nperception-aware attack produces adversarial music with significantly better\nperceptual quality than prior work.\n", "date": "2022-07-26", "categories": []}
{"id": "2207.11529", "title": "Low-complexity CNNs for Acoustic Scene Classification", "abstract": "  This paper presents a low-complexity framework for acoustic scene\nclassification (ASC). Most of the frameworks designed for ASC use convolutional\nneural networks (CNNs) due to their learning ability and improved performance\ncompared to hand-engineered features. However, CNNs are resource hungry due to\ntheir large size and high computational complexity. Therefore, CNNs are\ndifficult to deploy on resource constrained devices. This paper addresses the\nproblem of reducing the computational complexity and memory requirement in\nCNNs. We propose a low-complexity CNN architecture, and apply pruning and\nquantization to further reduce the parameters and memory. We then propose an\nensemble framework that combines various low-complexity CNNs to improve the\noverall performance. An experimental evaluation of the proposed framework is\nperformed on the publicly available DCASE 2022 Task 1 that focuses on ASC. The\nproposed ensemble framework has approximately 60K parameters, requires 19M\nmultiply-accumulate operations and improves the performance by approximately\n2-4 percentage points compared to the DCASE 2022 Task 1 baseline network.\n", "date": "2022-07-23", "categories": []}
{"id": "2207.11964", "title": "ConceptBeam: Concept Driven Target Speech Extraction", "abstract": "  We propose a novel framework for target speech extraction based on semantic\ninformation, called ConceptBeam. Target speech extraction means extracting the\nspeech of a target speaker in a mixture. Typical approaches have been\nexploiting properties of audio signals, such as harmonic structure and\ndirection of arrival. In contrast, ConceptBeam tackles the problem with\nsemantic clues. Specifically, we extract the speech of speakers speaking about\na concept, i.e., a topic of interest, using a concept specifier such as an\nimage or speech. Solving this novel problem would open the door to innovative\napplications such as listening systems that focus on a particular topic\ndiscussed in a conversation. Unlike keywords, concepts are abstract notions,\nmaking it challenging to directly represent a target concept. In our scheme, a\nconcept is encoded as a semantic embedding by mapping the concept specifier to\na shared embedding space. This modality-independent space can be built by means\nof deep metric learning using paired data consisting of images and their spoken\ncaptions. We use it to bridge modality-dependent information, i.e., the speech\nsegments in the mixture, and the specified, modality-independent concept. As a\nproof of our scheme, we performed experiments using a set of images associated\nwith spoken captions. That is, we generated speech mixtures from these spoken\ncaptions and used the images or speech signals as the concept specifiers. We\nthen extracted the target speech using the acoustic characteristics of the\nidentified segments. We compare ConceptBeam with two methods: one based on\nkeywords obtained from recognition systems and another based on sound source\nseparation. We show that ConceptBeam clearly outperforms the baseline methods\nand effectively extracts speech based on the semantic representation.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.10760", "title": "A Proposal for Foley Sound Synthesis Challenge", "abstract": "  \"Foley\" refers to sound effects that are added to multimedia during\npost-production to enhance its perceived acoustic properties, e.g., by\nsimulating the sounds of footsteps, ambient environmental sounds, or visible\nobjects on the screen. While foley is traditionally produced by foley artists,\nthere is increasing interest in automatic or machine-assisted techniques\nbuilding upon recent advances in sound synthesis and generative models. To\nfoster more participation in this growing research area, we propose a challenge\nfor automatic foley synthesis. Through case studies on successful previous\nchallenges in audio and machine learning, we set the goals of the proposed\nchallenge: rigorous, unified, and efficient evaluation of different foley\nsynthesis systems, with an overarching goal of drawing active participation\nfrom the research community. We outline the details and design considerations\nof a foley sound synthesis challenge, including task definition, dataset\nrequirements, and evaluation criteria.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.10817", "title": "End-to-End and Self-Supervised Learning for ComParE 2022 Stuttering\n  Sub-Challenge", "abstract": "  In this paper, we present end-to-end and speech embedding based systems\ntrained in a self-supervised fashion to participate in the ACM Multimedia 2022\nComParE Challenge, specifically the stuttering sub-challenge. In particular, we\nexploit the embeddings from the pre-trained Wav2Vec2.0 model for stuttering\ndetection (SD) on the KSoF dataset. After embedding extraction, we benchmark\nwith several methods for SD. Our proposed self-supervised based SD system\nachieves a UAR of 36.9% and 41.0% on validation and test sets respectively,\nwhich is 31.32% (validation set) and 1.49% (test set) higher than the best\n(DeepSpectrum) challenge baseline (CBL). Moreover, we show that concatenating\nlayer embeddings with Mel-frequency cepstral coefficients (MFCCs) features\nfurther improves the UAR of 33.81% and 5.45% on validation and test sets\nrespectively over the CBL. Finally, we demonstrate that the summing information\nacross all the layers of Wav2Vec2.0 surpasses the CBL by a relative margin of\n45.91% and 5.69% on validation and test sets respectively. Grand-challenge:\nComputational Paralinguistics ChallengE\n", "date": "2022-07-20", "categories": []}
{"id": "2207.10849", "title": "ASR Error Detection via Audio-Transcript entailment", "abstract": "  Despite improved performances of the latest Automatic Speech Recognition\n(ASR) systems, transcription errors are still unavoidable. These errors can\nhave a considerable impact in critical domains such as healthcare, when used to\nhelp with clinical documentation. Therefore, detecting ASR errors is a critical\nfirst step in preventing further error propagation to downstream applications.\nTo this end, we propose a novel end-to-end approach for ASR error detection\nusing audio-transcript entailment. To the best of our knowledge, we are the\nfirst to frame this problem as an end-to-end entailment task between the audio\nsegment and its corresponding transcript segment. Our intuition is that there\nshould be a bidirectional entailment between audio and transcript when there is\nno recognition error and vice versa. The proposed model utilizes an acoustic\nencoder and a linguistic encoder to model the speech and transcript\nrespectively. The encoded representations of both modalities are fused to\npredict the entailment. Since doctor-patient conversations are used in our\nexperiments, a particular emphasis is placed on medical terms. Our proposed\nmodel achieves classification error rates (CER) of 26.2% on all transcription\nerrors and 23% on medical errors specifically, leading to improvements upon a\nstrong baseline by 12% and 15.4%, respectively.\n", "date": "2022-07-22", "categories": []}
{"id": "2207.11231", "title": "Learning Unsupervised Hierarchies of Audio Concepts", "abstract": "  Music signals are difficult to interpret from their low-level features,\nperhaps even more than images: e.g. highlighting part of a spectrogram or an\nimage is often insufficient to convey high-level ideas that are genuinely\nrelevant to humans. In computer vision, concept learning was therein proposed\nto adjust explanations to the right abstraction level (e.g. detect clinical\nconcepts from radiographs). These methods have yet to be used for MIR.\n  In this paper, we adapt concept learning to the realm of music, with its\nparticularities. For instance, music concepts are typically non-independent and\nof mixed nature (e.g. genre, instruments, mood), unlike previous work that\nassumed disentangled concepts. We propose a method to learn numerous music\nconcepts from audio and then automatically hierarchise them to expose their\nmutual relationships. We conduct experiments on datasets of playlists from a\nmusic streaming service, serving as a few annotated examples for diverse\nconcepts. Evaluations show that the mined hierarchies are aligned with both\nground-truth hierarchies of concepts -- when available -- and with proxy\nsources of concept similarity in the general case.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.10141", "title": "AudioScopeV2: Audio-Visual Attention Architectures for Calibrated\n  Open-Domain On-Screen Sound Separation", "abstract": "  We introduce AudioScopeV2, a state-of-the-art universal audio-visual\non-screen sound separation system which is capable of learning to separate\nsounds and associate them with on-screen objects by looking at in-the-wild\nvideos. We identify several limitations of previous work on audio-visual\non-screen sound separation, including the coarse resolution of spatio-temporal\nattention, poor convergence of the audio separation model, limited variety in\ntraining and evaluation data, and failure to account for the trade off between\npreservation of on-screen sounds and suppression of off-screen sounds. We\nprovide solutions to all of these issues. Our proposed cross-modal and\nself-attention network architectures capture audio-visual dependencies at a\nfiner resolution over time, and we also propose efficient separable variants\nthat are capable of scaling to longer videos without sacrificing much\nperformance. We also find that pre-training the separation model only on audio\ngreatly improves results. For training and evaluation, we collected new human\nannotations of onscreen sounds from a large database of in-the-wild videos\n(YFCC100M). This new dataset is more diverse and challenging. Finally, we\npropose a calibration procedure that allows exact tuning of on-screen\nreconstruction versus off-screen suppression, which greatly simplifies\ncomparing performance between models with different operating points. Overall,\nour experimental results show marked improvements in on-screen separation\nperformance under much more general conditions than previous methods with\nminimal additional computational complexity.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.09732", "title": "Introducing Auxiliary Text Query-modifier to Content-based Audio\n  Retrieval", "abstract": "  The amount of audio data available on public websites is growing rapidly, and\nan efficient mechanism for accessing the desired data is necessary. We propose\na content-based audio retrieval method that can retrieve a target audio that is\nsimilar to but slightly different from the query audio by introducing auxiliary\ntextual information which describes the difference between the query and target\naudio. While the range of conventional content-based audio retrieval is limited\nto audio that is similar to the query audio, the proposed method can adjust the\nretrieval range by adding an embedding of the auxiliary text query-modifier to\nthe embedding of the query sample audio in a shared latent space. To evaluate\nour method, we built a dataset comprising two different audio clips and the\ntext that describes the difference. The experimental results show that the\nproposed method retrieves the paired audio more accurately than the baseline.\nWe also confirmed based on visualization that the proposed method obtains the\nshared latent space in which the audio difference and the corresponding text\nare represented as similar embedding vectors.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.08825", "title": "Contrastive Environmental Sound Representation Learning", "abstract": "  Machine hearing of the environmental sound is one of the important issues in\nthe audio recognition domain. It gives the machine the ability to discriminate\nbetween the different input sounds that guides its decision making. In this\nwork we exploit the self-supervised contrastive technique and a shallow 1D CNN\nto extract the distinctive audio features (audio representations) without using\nany explicit annotations.We generate representations of a given audio using\nboth its raw audio waveform and spectrogram and evaluate if the proposed\nlearner is agnostic to the type of audio input. We further use canonical\ncorrelation analysis (CCA) to fuse representations from the two types of input\nof a given audio and demonstrate that the fused global feature results in\nrobust representation of the audio signal as compared to the individual\nrepresentations. The evaluation of the proposed technique is done on both\nESC-50 and UrbanSound8K. The results show that the proposed technique is able\nto extract most features of the environmental audio and gives an improvement of\n12.8% and 0.9% on the ESC-50 and UrbanSound8K datasets respectively.\n", "date": "2022-07-18", "categories": []}
{"id": "2207.09145", "title": "GAFX: A General Audio Feature eXtractor", "abstract": "  Most machine learning models for audio tasks are dealing with a handcrafted\nfeature, the spectrogram. However, it is still unknown whether the spectrogram\ncould be replaced with deep learning based features. In this paper, we answer\nthis question by comparing the different learnable neural networks extracting\nfeatures with a successful spectrogram model and proposed a General Audio\nFeature eXtractor (GAFX) based on a dual U-Net (GAFX-U), ResNet (GAFX-R), and\nAttention (GAFX-A) modules. We design experiments to evaluate this model on the\nmusic genre classification task on the GTZAN dataset and perform a detailed\nablation study of different configurations of our framework and our model\nGAFX-U, following the Audio Spectrogram Transformer (AST) classifier achieves\ncompetitive performance.\n", "date": "2022-07-19", "categories": []}
{"id": "2207.07429", "title": "Continual Learning For On-Device Environmental Sound Classification", "abstract": "  Continuously learning new classes without catastrophic forgetting is a\nchallenging problem for on-device environmental sound classification given the\nrestrictions on computation resources (e.g., model size, running memory). To\naddress this issue, we propose a simple and efficient continual learning\nmethod. Our method selects the historical data for the training by measuring\nthe per-sample classification uncertainty. Specifically, we measure the\nuncertainty by observing how the classification probability of data fluctuates\nagainst the parallel perturbations added to the classifier embedding. In this\nway, the computation cost can be significantly reduced compared with adding\nperturbation to the raw data. Experimental results on the DCASE 2019 Task 1 and\nESC-50 dataset show that our proposed method outperforms baseline continual\nlearning methods on classification accuracy and computational efficiency,\nindicating our method can efficiently and incrementally learn new classes\nwithout the catastrophic forgetting problem for on-device environmental sound\nclassification.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.07773", "title": "Segment-level Metric Learning for Few-shot Bioacoustic Event Detection", "abstract": "  Few-shot bioacoustic event detection is a task that detects the occurrence\ntime of a novel sound given a few examples. Previous methods employ metric\nlearning to build a latent space with the labeled part of different sound\nclasses, also known as positive events. In this study, we propose a\nsegment-level few-shot learning framework that utilizes both the positive and\nnegative events during model optimization. Training with negative events, which\nare larger in volume than positive events, can increase the generalization\nability of the model. In addition, we use transductive inference on the\nvalidation set during training for better adaptation to novel classes. We\nconduct ablation studies on our proposed method with different setups on input\nfeatures, training data, and hyper-parameters. Our final system achieves an\nF-measure of 62.73 on the DCASE 2022 challenge task 5 (DCASE2022-T5) validation\nset, outperforming the performance of the baseline prototypical network 34.02\nby a large margin. Using the proposed method, our submitted system ranks 2nd in\nDCASE2022-T5. The code of this paper is fully open-sourced at\nhttps://github.com/haoheliu/DCASE_2022_Task_5.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.07911", "title": "Few-shot bioacoustic event detection at the DCASE 2022 challenge", "abstract": "  Few-shot sound event detection is the task of detecting sound events, despite\nhaving only a few labelled examples of the class of interest. This framework is\nparticularly useful in bioacoustics, where often there is a need to annotate\nvery long recordings but the expert annotator time is limited. This paper\npresents an overview of the second edition of the few-shot bioacoustic sound\nevent detection task included in the DCASE 2022 challenge. A detailed\ndescription of the task objectives, dataset, and baselines is presented,\ntogether with the main results obtained and characteristics of the submitted\nsystems. This task received submissions from 15 different teams from which 13\nscored higher than the baselines. The highest F-score was of 60% on the\nevaluation set, which leads to a huge improvement over last year's edition.\nHighly-performing methods made use of prototypical networks, transductive\nlearning, and addressed the variable length of events from all target classes.\nFurthermore, by analysing results on each of the subsets we can identify the\nmain difficulties that the systems face, and conclude that few-show bioacoustic\nsound event detection remains an open challenge.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.07935", "title": "Visually-aware Acoustic Event Detection using Heterogeneous Graphs", "abstract": "  Perception of auditory events is inherently multimodal relying on both audio\nand visual cues. A large number of existing multimodal approaches process each\nmodality using modality-specific models and then fuse the embeddings to encode\nthe joint information. In contrast, we employ heterogeneous graphs to\nexplicitly capture the spatial and temporal relationships between the\nmodalities and represent detailed information about the underlying signal.\nUsing heterogeneous graph approaches to address the task of visually-aware\nacoustic event classification, which serves as a compact, efficient and\nscalable way to represent data in the form of graphs. Through heterogeneous\ngraphs, we show efficiently modelling of intra- and inter-modality\nrelationships both at spatial and temporal scales. Our model can easily be\nadapted to different scales of events through relevant hyperparameters.\nExperiments on AudioSet, a large benchmark, shows that our model achieves\nstate-of-the-art performance.\n", "date": "2022-07-16", "categories": []}
{"id": "2207.04203", "title": "Learning to Separate Voices by Spatial Regions", "abstract": "  We consider the problem of audio voice separation for binaural applications,\nsuch as earphones and hearing aids. While today's neural networks perform\nremarkably well (separating $4+$ sources with 2 microphones) they assume a\nknown or fixed maximum number of sources, K. Moreover, today's models are\ntrained in a supervised manner, using training data synthesized from generic\nsources, environments, and human head shapes.\n  This paper intends to relax both these constraints at the expense of a slight\nalteration in the problem definition. We observe that, when a received mixture\ncontains too many sources, it is still helpful to separate them by region,\ni.e., isolating signal mixtures from each conical sector around the user's\nhead. This requires learning the fine-grained spatial properties of each\nregion, including the signal distortions imposed by a person's head. We propose\na two-stage self-supervised framework in which overheard voices from earphones\nare pre-processed to extract relatively clean personalized signals, which are\nthen used to train a region-wise separation model. Results show promising\nperformance, underscoring the importance of personalization over a generic\nsupervised approach. (audio samples available at our project website:\nhttps://uiuc-earable-computing.github.io/binaural/. We believe this result\ncould help real-world applications in selective hearing, noise cancellation,\nand audio augmented reality.\n", "date": "2022-07-09", "categories": []}
{"id": "2207.07273", "title": "Direction-Aware Joint Adaptation of Neural Speech Enhancement and\n  Recognition in Real Multiparty Conversational Environments", "abstract": "  This paper describes noisy speech recognition for an augmented reality\nheadset that helps verbal communication within real multiparty conversational\nenvironments. A major approach that has actively been studied in simulated\nenvironments is to sequentially perform speech enhancement and automatic speech\nrecognition (ASR) based on deep neural networks (DNNs) trained in a supervised\nmanner. In our task, however, such a pretrained system fails to work due to the\nmismatch between the training and test conditions and the head movements of the\nuser. To enhance only the utterances of a target speaker, we use beamforming\nbased on a DNN-based speech mask estimator that can adaptively extract the\nspeech components corresponding to a head-relative particular direction. We\npropose a semi-supervised adaptation method that jointly updates the mask\nestimator and the ASR model at run-time using clean speech signals with\nground-truth transcriptions and noisy speech signals with highly-confident\nestimated transcriptions. Comparative experiments using the state-of-the-art\ndistant speech recognition system show that the proposed method significantly\nimproves the ASR performance.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.07296", "title": "Direction-Aware Adaptive Online Neural Speech Enhancement with an\n  Augmented Reality Headset in Real Noisy Conversational Environments", "abstract": "  This paper describes the practical response- and performance-aware\ndevelopment of online speech enhancement for an augmented reality (AR) headset\nthat helps a user understand conversations made in real noisy echoic\nenvironments (e.g., cocktail party). One may use a state-of-the-art blind\nsource separation method called fast multichannel nonnegative matrix\nfactorization (FastMNMF) that works well in various environments thanks to its\nunsupervised nature. Its heavy computational cost, however, prevents its\napplication to real-time processing. In contrast, a supervised beamforming\nmethod that uses a deep neural network (DNN) for estimating spatial information\nof speech and noise readily fits real-time processing, but suffers from drastic\nperformance degradation in mismatched conditions. Given such complementary\ncharacteristics, we propose a dual-process robust online speech enhancement\nmethod based on DNN-based beamforming with FastMNMF-guided adaptation. FastMNMF\n(back end) is performed in a mini-batch style and the noisy and enhanced speech\npairs are used together with the original parallel training data for updating\nthe direction-aware DNN (front end) with backpropagation at a\ncomputationally-allowable interval. This method is used with a blind\ndereverberation method called weighted prediction error (WPE) for transcribing\nthe noisy reverberant speech of a speaker, which can be detected from video or\nselected by a user's hand gesture or eye gaze, in a streaming manner and\nspatially showing the transcriptions with an AR technique. Our experiment\nshowed that the word error rate was improved by more than 10 points with the\nrun-time adaptation using only twelve minutes of observation.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.05393", "title": "Western Mediterranean wetlands bird species classification: evaluating\n  small-footprint deep learning approaches on a new annotated dataset", "abstract": "  The deployment of an expert system running over a wireless acoustic sensors\nnetwork made up of bioacoustic monitoring devices that recognise bird species\nfrom their sounds would enable the automation of many tasks of ecological\nvalue, including the analysis of bird population composition or the detection\nof endangered species in areas of environmental interest. Endowing these\ndevices with accurate audio classification capabilities is possible thanks to\nthe latest advances in artificial intelligence, among which deep learning\ntechniques excel. However, a key issue to make bioacoustic devices affordable\nis the use of small footprint deep neural networks that can be embedded in\nresource and battery constrained hardware platforms. For this reason, this work\npresents a critical comparative analysis between two heavy and large footprint\ndeep neural networks (VGG16 and ResNet50) and a lightweight alternative,\nMobileNetV2. Our experimental results reveal that MobileNetV2 achieves an\naverage F1-score less than a 5\\% lower than ResNet50 (0.789 vs. 0.834),\nperforming better than VGG16 with a footprint size nearly 40 times smaller.\nMoreover, to compare the models, we have created and made public the Western\nMediterranean Wetland Birds dataset, consisting of 201.6 minutes and 5,795\naudio excerpts of 20 endemic bird species of the Aiguamolls de l'Empord\\`a\nNatural Park.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.05508", "title": "EfficientLEAF: A Faster LEarnable Audio Frontend of Questionable Use", "abstract": "  In audio classification, differentiable auditory filterbanks with few\nparameters cover the middle ground between hard-coded spectrograms and raw\naudio. LEAF (arXiv:2101.08596), a Gabor-based filterbank combined with\nPer-Channel Energy Normalization (PCEN), has shown promising results, but is\ncomputationally expensive. With inhomogeneous convolution kernel sizes and\nstrides, and by replacing PCEN with better parallelizable operations, we can\nreach similar results more efficiently. In experiments on six audio\nclassification tasks, our frontend matches the accuracy of LEAF at 3% of the\ncost, but both fail to consistently outperform a fixed mel filterbank. The\nquest for learnable audio frontends is not solved.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.04471", "title": "Towards Proper Contrastive Self-supervised Learning Strategies For Music\n  Audio Representation", "abstract": "  The common research goal of self-supervised learning is to extract a general\nrepresentation which an arbitrary downstream task would benefit from. In this\nwork, we investigate music audio representation learned from different\ncontrastive self-supervised learning schemes and empirically evaluate the\nembedded vectors on various music information retrieval (MIR) tasks where\ndifferent levels of the music perception are concerned. We analyze the results\nto discuss the proper direction of contrastive learning strategies for\ndifferent MIR tasks. We show that these representations convey a comprehensive\ninformation about the auditory characteristics of music in general, although\neach of the self-supervision strategies has its own effectiveness in certain\naspect of information.\n", "date": "2022-07-10", "categories": []}
{"id": "2206.11260", "title": "Few-shot Long-Tailed Bird Audio Recognition", "abstract": "  It is easier to hear birds than see them. However, they still play an\nessential role in nature and are excellent indicators of deteriorating\nenvironmental quality and pollution. Recent advances in Deep Neural Networks\nallow us to process audio data to detect and classify birds. This technology\ncan assist researchers in monitoring bird populations and biodiversity. We\npropose a sound detection and classification pipeline to analyze complex\nsoundscape recordings and identify birdcalls in the background. Our method\nlearns from weak labels and few data and acoustically recognizes the bird\nspecies. Our solution achieved 18th place of 807 teams at the BirdCLEF 2022\nChallenge hosted on Kaggle.\n", "date": "2022-06-22", "categories": []}
{"id": "2207.01197", "title": "Multi-Modal Multi-Correlation Learning for Audio-Visual Speech\n  Separation", "abstract": "  In this paper we propose a multi-modal multi-correlation learning framework\ntargeting at the task of audio-visual speech separation. Although previous\nefforts have been extensively put on combining audio and visual modalities,\nmost of them solely adopt a straightforward concatenation of audio and visual\nfeatures. To exploit the real useful information behind these two modalities,\nwe define two key correlations which are: (1) identity correlation (between\ntimbre and facial attributes); (2) phonetic correlation (between phoneme and\nlip motion). These two correlations together comprise the complete information,\nwhich shows a certain superiority in separating target speaker's voice\nespecially in some hard cases, such as the same gender or similar content. For\nimplementation, contrastive learning or adversarial training approach is\napplied to maximize these two correlations. Both of them work well, while\nadversarial training shows its advantage by avoiding some limitations of\ncontrastive learning. Compared with previous research, our solution\ndemonstrates clear improvement on experimental metrics without additional\ncomplexity. Further analysis reveals the validity of the proposed architecture\nand its good potential for future extension.\n", "date": "2022-07-04", "categories": []}
{"id": "2207.01454", "title": "GlowVC: Mel-spectrogram space disentangling model for\n  language-independent text-free voice conversion", "abstract": "  In this paper, we propose GlowVC: a multilingual multi-speaker flow-based\nmodel for language-independent text-free voice conversion. We build on\nGlow-TTS, which provides an architecture that enables use of linguistic\nfeatures during training without the necessity of using them for VC inference.\nWe consider two versions of our model: GlowVC-conditional and GlowVC-explicit.\nGlowVC-conditional models the distribution of mel-spectrograms with\nspeaker-conditioned flow and disentangles the mel-spectrogram space into\ncontent- and pitch-relevant dimensions, while GlowVC-explicit models the\nexplicit distribution with unconditioned flow and disentangles said space into\ncontent-, pitch- and speaker-relevant dimensions. We evaluate our models in\nterms of intelligibility, speaker similarity and naturalness for intra- and\ncross-lingual conversion in seen and unseen languages. GlowVC models greatly\noutperform AutoVC baseline in terms of intelligibility, while achieving just as\nhigh speaker similarity in intra-lingual VC, and slightly worse in the\ncross-lingual setting. Moreover, we demonstrate that GlowVC-explicit surpasses\nboth GlowVC-conditional and AutoVC in terms of naturalness.\n", "date": "2022-07-04", "categories": []}
{"id": "2206.15423", "title": "Implicit Neural Spatial Filtering for Multichannel Source Separation in\n  the Waveform Domain", "abstract": "  We present a single-stage casual waveform-to-waveform multichannel model that\ncan separate moving sound sources based on their broad spatial locations in a\ndynamic acoustic scene. We divide the scene into two spatial regions\ncontaining, respectively, the target and the interfering sound sources. The\nmodel is trained end-to-end and performs spatial processing implicitly, without\nany components based on traditional processing or use of hand-crafted spatial\nfeatures. We evaluate the proposed model on a real-world dataset and show that\nthe model matches the performance of an oracle beamformer followed by a\nstate-of-the-art single-channel enhancement network.\n", "date": "2022-06-30", "categories": []}
{"id": "2206.10805", "title": "Jointist: Joint Learning for Multi-instrument Transcription and Its\n  Applications", "abstract": "  In this paper, we introduce Jointist, an instrument-aware multi-instrument\nframework that is capable of transcribing, recognizing, and separating multiple\nmusical instruments from an audio clip. Jointist consists of the instrument\nrecognition module that conditions the other modules: the transcription module\nthat outputs instrument-specific piano rolls, and the source separation module\nthat utilizes instrument information and transcription results.\n  The instrument conditioning is designed for an explicit multi-instrument\nfunctionality while the connection between the transcription and source\nseparation modules is for better transcription performance. Our challenging\nproblem formulation makes the model highly useful in the real world given that\nmodern popular music typically consists of multiple instruments. However, its\nnovelty necessitates a new perspective on how to evaluate such a model. During\nthe experiment, we assess the model from various aspects, providing a new\nevaluation perspective for multi-instrument transcription. We also argue that\ntranscription models can be utilized as a preprocessing module for other music\nanalysis tasks. In the experiment on several downstream tasks, the symbolic\nrepresentation provided by our transcription model turned out to be helpful to\nspectrograms in solving downbeat detection, chord recognition, and key\nestimation.\n", "date": "2022-06-22", "categories": []}
{"id": "2206.11706", "title": "A Temporal Extension of Latent Dirichlet Allocation for Unsupervised\n  Acoustic Unit Discovery", "abstract": "  Latent Dirichlet allocation (LDA) is widely used for unsupervised topic\nmodelling on sets of documents. No temporal information is used in the model.\nHowever, there is often a relationship between the corresponding topics of\nconsecutive tokens. In this paper, we present an extension to LDA that uses a\nMarkov chain to model temporal information. We use this new model for acoustic\nunit discovery from speech. As input tokens, the model takes a discretised\nencoding of speech from a vector quantised (VQ) neural network with 512 codes.\nThe goal is then to map these 512 VQ codes to 50 phone-like units (topics) in\norder to more closely resemble true phones. In contrast to the base LDA, which\nonly considers how VQ codes co-occur within utterances (documents), the Markov\nchain LDA additionally captures how consecutive codes follow one another. This\nextension leads to an increase in cluster quality and phone segmentation\nresults compared to the base LDA. Compared to a recent vector quantised neural\nnetwork approach that also learns 50 units, the extended LDA model performs\nbetter in phone segmentation but worse in mutual information.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.14659", "title": "Language-Based Audio Retrieval with Converging Tied Layers and\n  Contrastive Loss", "abstract": "  In this paper, we tackle the new Language-Based Audio Retrieval task proposed\nin DCASE 2022. Firstly, we introduce a simple, scalable architecture which ties\nboth the audio and text encoder together. Secondly, we show that using this\narchitecture along with contrastive loss allows the model to significantly beat\nthe performance of the baseline model. Finally, in addition to having an\nextremely low training memory requirement, we are able to use pretrained models\nas it is without needing to finetune them. We test our methods and show that\nusing a combination of our methods beats the baseline scores significantly.\n", "date": "2022-06-29", "categories": []}
{"id": "2206.14723", "title": "DrumGAN VST: A Plugin for Drum Sound Analysis/Synthesis With\n  Autoencoding Generative Adversarial Networks", "abstract": "  In contemporary popular music production, drum sound design is commonly\nperformed by cumbersome browsing and processing of pre-recorded samples in\nsound libraries. One can also use specialized synthesis hardware, typically\ncontrolled through low-level, musically meaningless parameters. Today, the\nfield of Deep Learning offers methods to control the synthesis process via\nlearned high-level features and allows generating a wide variety of sounds. In\nthis paper, we present DrumGAN VST, a plugin for synthesizing drum sounds using\na Generative Adversarial Network. DrumGAN VST operates on 44.1 kHz sample-rate\naudio, offers independent and continuous instrument class controls, and\nfeatures an encoding neural network that maps sounds into the GAN's latent\nspace, enabling resynthesis and manipulation of pre-existing drum sounds. We\nprovide numerous sound examples and a demo of the proposed VST plugin.\n", "date": "2022-06-29", "categories": []}
{"id": "2206.12494", "title": "Multitask vocal burst modeling with ResNets and pre-trained\n  paralinguistic Conformers", "abstract": "  This technical report presents the modeling approaches used in our submission\nto the ICML Expressive Vocalizations Workshop & Competition multitask track\n(ExVo-MultiTask). We first applied image classification models of various sizes\non mel-spectrogram representations of the vocal bursts, as is standard in sound\nevent detection literature. Results from these models show an increase of\n21.24% over the baseline system with respect to the harmonic mean of the task\nmetrics, and comprise our team's main submission to the MultiTask track. We\nthen sought to characterize the headroom in the MultiTask track by applying a\nlarge pre-trained Conformer model that previously achieved state-of-the-art\nresults on paralinguistic tasks like speech emotion recognition and mask\ndetection. We additionally investigated the relationship between the sub-tasks\nof emotional expression, country of origin, and age prediction, and discovered\nthat the best performing models are trained as single-task models, questioning\nwhether the problem truly benefits from a multitask setting.\n", "date": "2022-06-24", "categories": []}
{"id": "2206.12513", "title": "Domain Generalization with Relaxed Instance Frequency-wise Normalization\n  for Multi-device Acoustic Scene Classification", "abstract": "  While using two-dimensional convolutional neural networks (2D-CNNs) in image\nprocessing, it is possible to manipulate domain information using channel\nstatistics, and instance normalization has been a promising way to get\ndomain-invariant features. However, unlike image processing, we analyze that\ndomain-relevant information in an audio feature is dominant in frequency\nstatistics rather than channel statistics. Motivated by our analysis, we\nintroduce Relaxed Instance Frequency-wise Normalization (RFN): a plug-and-play,\nexplicit normalization module along the frequency axis which can eliminate\ninstance-specific domain discrepancy in an audio feature while relaxing\nundesirable loss of useful discriminative information. Empirically, simply\nadding RFN to networks shows clear margins compared to previous domain\ngeneralization approaches on acoustic scene classification and yields improved\nrobustness for multiple audio devices. Especially, the proposed RFN won the\nDCASE2021 challenge TASK1A, low-complexity acoustic scene classification with\nmultiple devices, with a clear margin, and RFN is an extended work of our\ntechnical report.\n", "date": "2022-06-24", "categories": []}
{"id": "2206.12568", "title": "Self-supervision and Learnable STRFs for Age, Emotion, and Country\n  Prediction", "abstract": "  This work presents a multitask approach to the simultaneous estimation of\nage, country of origin, and emotion given vocal burst audio for the 2022 ICML\nExpressive Vocalizations Challenge ExVo-MultiTask track. The method of choice\nutilized a combination of spectro-temporal modulation and self-supervised\nfeatures, followed by an encoder-decoder network organized in a multitask\nparadigm. We evaluate the complementarity between the tasks posed by examining\nindependent task-specific and joint models, and explore the relative strengths\nof different feature sets. We also introduce a simple score fusion mechanism to\nleverage the complementarity of different feature sets for this task.\n  We find that robust data preprocessing in conjunction with score fusion over\nspectro-temporal receptive field and HuBERT models achieved our best\nExVo-MultiTask test score of 0.412.\n", "date": "2022-06-25", "categories": []}
{"id": "2206.13071", "title": "Uncertainty Calibration for Deep Audio Classifiers", "abstract": "  Although deep Neural Networks (DNNs) have achieved tremendous success in\naudio classification tasks, their uncertainty calibration are still\nunder-explored. A well-calibrated model should be accurate when it is certain\nabout its prediction and indicate high uncertainty when it is likely to be\ninaccurate. In this work, we investigate the uncertainty calibration for deep\naudio classifiers. In particular, we empirically study the performance of\npopular calibration methods: (i) Monte Carlo Dropout, (ii) ensemble, (iii)\nfocal loss, and (iv) spectral-normalized Gaussian process (SNGP), on audio\nclassification datasets. To this end, we evaluate (i-iv) for the tasks of\nenvironment sound and music genre classification. Results indicate that\nuncalibrated deep audio classifiers may be over-confident, and SNGP performs\nthe best and is very efficient on the two datasets of this paper.\n", "date": "2022-06-27", "categories": []}
{"id": "2206.13365", "title": "Interpretable Acoustic Representation Learning on Breathing and Speech\n  Signals for COVID-19 Detection", "abstract": "  In this paper, we describe an approach for representation learning of audio\nsignals for the task of COVID-19 detection. The raw audio samples are processed\nwith a bank of 1-D convolutional filters that are parameterized as cosine\nmodulated Gaussian functions. The choice of these kernels allows the\ninterpretation of the filterbanks as smooth band-pass filters. The filtered\noutputs are pooled, log-compressed and used in a self-attention based relevance\nweighting mechanism. The relevance weighting emphasizes the key regions of the\ntime-frequency decomposition that are important for the downstream task. The\nsubsequent layers of the model consist of a recurrent architecture and the\nmodels are trained for a COVID-19 detection task. In our experiments on the\nCoswara data set, we show that the proposed model achieves significant\nperformance improvements over the baseline system as well as other\nrepresentation learning approaches. Further, the approach proposed is shown to\nbe uniformly applicable for speech and breathing signals and for transfer\nlearning from a larger data set.\n", "date": "2022-06-27", "categories": []}
{"id": "2206.13476", "title": "Impact of Acoustic Event Tagging on Scene Classification in a Multi-Task\n  Learning Framework", "abstract": "  Acoustic events are sounds with well-defined spectro-temporal characteristics\nwhich can be associated with the physical objects generating them. Acoustic\nscenes are collections of such acoustic events in no specific temporal order.\nGiven this natural linkage between events and scenes, a common belief is that\nthe ability to classify events must help in the classification of scenes. This\nhas led to several efforts attempting to do well on Acoustic Event Tagging\n(AET) and Acoustic Scene Classification (ASC) using a multi-task network.\nHowever, in these efforts, improvement in one task does not guarantee an\nimprovement in the other, suggesting a tension between ASC and AET. It is\nunclear if improvements in AET translates to improvements in ASC. We explore\nthis conundrum through an extensive empirical study and show that under certain\nconditions, using AET as an auxiliary task in the multi-task network\nconsistently improves ASC performance. Additionally, ASC performance further\nimproves with the AET data-set size and is not sensitive to the choice of\nevents or the number of events in the AET data-set. We conclude that this\nimprovement in ASC performance comes from the regularization effect of using\nAET and not from the network's improved ability to discern between acoustic\nevents.\n", "date": "2022-06-27", "categories": []}
{"id": "2206.12273", "title": "Iterative Sound Source Localization for Unknown Number of Sources", "abstract": "  Sound source localization aims to seek the direction of arrival (DOA) of all\nsound sources from the observed multi-channel audio. For the practical problem\nof unknown number of sources, existing localization algorithms attempt to\npredict a likelihood-based coding (i.e., spatial spectrum) and employ a\npre-determined threshold to detect the source number and corresponding DOA\nvalue. However, these threshold-based algorithms are not stable since they are\nlimited by the careful choice of threshold. To address this problem, we propose\nan iterative sound source localization approach called ISSL, which can\niteratively extract each source's DOA without threshold until the termination\ncriterion is met. Unlike threshold-based algorithms, ISSL designs an active\nsource detector network based on binary classifier to accept residual spatial\nspectrum and decide whether to stop the iteration. By doing so, our ISSL can\ndeal with an arbitrary number of sources, even more than the number of sources\nseen during the training stage. The experimental results show that our ISSL\nachieves significant performance improvements in both DOA estimation and source\nnumber detection compared with the existing threshold-based algorithms.\n", "date": "2022-06-24", "categories": []}
{"id": "2206.05286", "title": "AHD ConvNet for Speech Emotion Classification", "abstract": "  Accomplishments in the field of artificial intelligence are utilized in the\nadvancement of computing and making of intelligent machines for facilitating\nmankind and improving user experience. Emotions are rudimentary for people,\naffecting thinking and ordinary exercises like correspondence, learning and\ndirection. Speech emotion recognition is domain of interest in this regard and\nin this work, we propose a novel mel spectrogram learning approach in which our\nmodel uses the datapoints to learn emotions from the given wav form voice notes\nin the popular CREMA-D dataset. Our model uses log mel-spectrogram as feature\nwith number of mels = 64. It took less training time compared to other\napproaches used to address the problem of emotion speech recognition.\n", "date": "2022-06-10", "categories": []}
{"id": "2206.06680", "title": "Exploring speaker enrolment for few-shot personalisation in emotional\n  vocalisation prediction", "abstract": "  In this work, we explore a novel few-shot personalisation architecture for\nemotional vocalisation prediction. The core contribution is an `enrolment'\nencoder which utilises two unlabelled samples of the target speaker to adjust\nthe output of the emotion encoder; the adjustment is based on dot-product\nattention, thus effectively functioning as a form of `soft' feature selection.\nThe emotion and enrolment encoders are based on two standard audio\narchitectures: CNN14 and CNN10. The two encoders are further guided to forget\nor learn auxiliary emotion and/or speaker information. Our best approach\nachieves a CCC of $.650$ on the ExVo Few-Shot dev set, a $2.5\\%$ increase over\nour baseline CNN14 CCC of $.634$.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.06811", "title": "Adversarial Audio Synthesis with Complex-valued Polynomial Networks", "abstract": "  Time-frequency (TF) representations in audio synthesis have been increasingly\nmodeled with real-valued networks. However, overlooking the complex-valued\nnature of TF representations can result in suboptimal performance and require\nadditional modules (e.g., for modeling the phase). To this end, we introduce\ncomplex-valued polynomial networks, called APOLLO, that integrate such\ncomplex-valued representations in a natural way. Concretely, APOLLO captures\nhigh-order correlations of the input elements using high-order tensors as\nscaling parameters. By leveraging standard tensor decompositions, we derive\ndifferent architectures and enable modeling richer correlations. We outline\nsuch architectures and showcase their performance in audio generation across\nfour benchmarks. As a highlight, APOLLO results in $17.5\\%$ improvement over\nadversarial methods and $8.2\\%$ over the state-of-the-art diffusion models on\nSC09 dataset in audio generation. Our models can encourage the systematic\ndesign of other efficient architectures on the complex field.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.09556", "title": "An Empirical Analysis on the Vulnerabilities of End-to-End Speech\n  Segregation Models", "abstract": "  End-to-end learning models have demonstrated a remarkable capability in\nperforming speech segregation. Despite their wide-scope of real-world\napplications, little is known about the mechanisms they employ to group and\nconsequently segregate individual speakers. Knowing that harmonicity is a\ncritical cue for these networks to group sources, in this work, we perform a\nthorough investigation on ConvTasnet and DPT-Net to analyze how they perform a\nharmonic analysis of the input mixture. We perform ablation studies where we\napply low-pass, high-pass, and band-stop filters of varying pass-bands to\nempirically analyze the harmonics most critical for segregation. We also\ninvestigate how these networks decide which output channel to assign to an\nestimated source by introducing discontinuities in synthetic mixtures. We find\nthat end-to-end networks are highly unstable, and perform poorly when\nconfronted with deformations which are imperceptible to humans. Replacing the\nencoder in these networks with a spectrogram leads to lower overall\nperformance, but much higher stability. This work helps us to understand what\ninformation these network rely on for speech segregation, and exposes two\nsources of generalization-errors. It also pinpoints the encoder as the part of\nthe network responsible for these errors, allowing for a redesign with expert\nknowledge or transfer learning.\n", "date": "2022-06-20", "categories": []}
{"id": "2206.08007", "title": "DCASE 2022: Comparative Analysis Of CNNs For Acoustic Scene\n  Classification Under Low-Complexity Considerations", "abstract": "  Acoustic scene classification is an automatic listening problem that aims to\nassign an audio recording to a pre-defined scene based on its audio data. Over\nthe years (and in past editions of the DCASE) this problem has often been\nsolved with techniques known as ensembles (use of several machine learning\nmodels to combine their predictions in the inference phase). While these\nsolutions can show performance in terms of accuracy, they can be very expensive\nin terms of computational capacity, making it impossible to deploy them in IoT\ndevices. Due to the drift in this field of study, this task has two limitations\nin terms of model complexity. It should be noted that there is also the added\ncomplexity of mismatching devices (the audios provided are recorded by\ndifferent sources of information). This technical report makes a comparative\nstudy of two different network architectures: conventional CNN and Conv-mixer.\nAlthough both networks exceed the baseline required by the competition, the\nconventional CNN shows a higher performance, exceeding the baseline by 8\npercentage points. Solutions based on Conv-mixer architectures show worse\nperformance although they are much lighter solutions.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.07511", "title": "Investigating Multi-Feature Selection and Ensembling for Audio\n  Classification", "abstract": "  Deep Learning (DL) algorithms have shown impressive performance in diverse\ndomains. Among them, audio has attracted many researchers over the last couple\nof decades due to some interesting patterns--particularly in classification of\naudio data. For better performance of audio classification, feature selection\nand combination play a key role as they have the potential to make or break the\nperformance of any DL model. To investigate this role, we conduct an extensive\nevaluation of the performance of several cutting-edge DL models (i.e.,\nConvolutional Neural Network, EfficientNet, MobileNet, Supper Vector Machine\nand Multi-Perceptron) with various state-of-the-art audio features (i.e., Mel\nSpectrogram, Mel Frequency Cepstral Coefficients, and Zero Crossing Rate)\neither independently or as a combination (i.e., through ensembling) on three\ndifferent datasets (i.e., Free Spoken Digits Dataset, Audio Urdu Digits\nDataset, and Audio Gujarati Digits Dataset). Overall, results suggest feature\nselection depends on both the dataset and the model. However, feature\ncombinations should be restricted to the only features that already achieve\ngood performances when used individually (i.e., mostly Mel Spectrogram, Mel\nFrequency Cepstral Coefficients). Such feature combination/ensembling enabled\nus to outperform the previous state-of-the-art results irrespective of our\nchoice of DL model.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.05460", "title": "Hierarchical Conditional Variational Autoencoder Based Acoustic Anomaly\n  Detection", "abstract": "  This paper aims to develop an acoustic signal-based unsupervised anomaly\ndetection method for automatic machine monitoring. Existing approaches such as\ndeep autoencoder (DAE), variational autoencoder (VAE), conditional variational\nautoencoder (CVAE) etc. have limited representation capabilities in the latent\nspace and, hence, poor anomaly detection performance. Different models have to\nbe trained for each different kind of machines to accurately perform the\nanomaly detection task. To solve this issue, we propose a new method named as\nhierarchical conditional variational autoencoder (HCVAE). This method utilizes\navailable taxonomic hierarchical knowledge about industrial facility to refine\nthe latent space representation. This knowledge helps model to improve the\nanomaly detection performance as well. We demonstrated the generalization\ncapability of a single HCVAE model for different types of machines by using\nappropriate conditions. Additionally, to show the practicability of the\nproposed approach, (i) we evaluated HCVAE model on different domain and (ii) we\nchecked the effect of partial hierarchical knowledge. Our results show that\nHCVAE method validates both of these points, and it outperforms the baseline\nsystem on anomaly detection task by utmost 15 % on the AUC score metric.\n", "date": "2022-06-11", "categories": []}
{"id": "2206.05462", "title": "Svadhyaya system for the Second Diagnosing COVID-19 using Acoustics\n  Challenge 2021", "abstract": "  This report describes the system used for detecting COVID-19 positives using\nthree different acoustic modalities, namely speech, breathing, and cough in the\nsecond DiCOVA challenge. The proposed system is based on the combination of 4\ndifferent approaches, each focusing more on one aspect of the problem, and\nreaches the blind test AUCs of 86.41, 77.60, and 84.55, in the breathing,\ncough, and speech tracks, respectively, and the AUC of 85.37 in the fusion of\nthese three tracks.\n", "date": "2022-06-11", "categories": []}
{"id": "2206.06057", "title": "Low-complexity deep learning frameworks for acoustic scene\n  classification", "abstract": "  In this report, we presents low-complexity deep learning frameworks for\nacoustic scene classification (ASC). The proposed frameworks can be separated\ninto four main steps: Front-end spectrogram extraction, online data\naugmentation, back-end classification, and late fusion of predicted\nprobabilities. In particular, we initially transform audio recordings into Mel,\nGammatone, and CQT spectrograms. Next, data augmentation methods of Random\nCropping, Specaugment, and Mixup are then applied to generate augmented\nspectrograms before being fed into deep learning based classifiers. Finally, to\nachieve the best performance, we fuse probabilities which obtained from three\nindividual classifiers, which are independently-trained with three type of\nspectrograms. Our experiments conducted on DCASE 2022 Task 1 Development\ndataset have fullfiled the requirement of low-complexity and achieved the best\nclassification accuracy of 60.1%, improving DCASE baseline by 17.2%.\n", "date": "2022-06-13", "categories": []}
{"id": "2206.04984", "title": "Zero-Shot Audio Classification using Image Embeddings", "abstract": "  Supervised learning methods can solve the given problem in the presence of a\nlarge set of labeled data. However, the acquisition of a dataset covering all\nthe target classes typically requires manual labeling which is expensive and\ntime-consuming. Zero-shot learning models are capable of classifying the unseen\nconcepts by utilizing their semantic information. The present study introduces\nimage embeddings as side information on zero-shot audio classification by using\na nonlinear acoustic-semantic projection. We extract the semantic image\nrepresentations from the Open Images dataset and evaluate the performance of\nthe models on an audio subset of AudioSet using semantic information in\ndifferent domains; image, audio, and textual. We demonstrate that the image\nembeddings can be used as semantic information to perform zero-shot audio\nclassification. The experimental results show that the image and textual\nembeddings display similar performance both individually and together. We\nadditionally calculate the semantic acoustic embeddings from the test samples\nto provide an upper limit to the performance. The results show that the\nclassification performance is highly sensitive to the semantic relation between\ntest and training classes and textual and image embeddings can reach up to the\nsemantic acoustic embeddings when the seen and unseen classes are semantically\nsimilar.\n", "date": "2022-06-10", "categories": []}
{"id": "2206.01918", "title": "Automated Audio Captioning with Epochal Difficult Captions for\n  Curriculum Learning", "abstract": "  In this paper, we propose an algorithm, Epochal Difficult Captions, to\nsupplement the training of any model for the Automated Audio Captioning task.\nEpochal Difficult Captions is an elegant evolution to the keyword estimation\ntask that previous work have used to train the encoder of the AAC model.\nEpochal Difficult Captions modifies the target captions based on a curriculum\nand a difficulty level determined as a function of current epoch. Epochal\nDifficult Captions can be used with any model architecture and is a lightweight\nfunction that does not increase training time. We test our results on three\nsystems and show that using Epochal Difficult Captions consistently improves\nperformance\n", "date": "2022-06-04", "categories": []}
{"id": "2606.21305", "title": "LISE : Listenable Interpretable Speaker Embeddings", "abstract": "Deep neural network-based automatic speaker verification (ASV) systems achieve impressive performance but their embedding representations remain opaque, lacking a structured and perceptually verifiable explanation of the vocal characteristics they encode. Existing approaches either require annotation of speaker attributes or introduce alternative representations whose interpretability is unvalidated with listeners. We propose Listenable Interpretable Speaker Embeddings (LISE), a label-free framework that decomposes pretrained speaker embeddings into a small set of components. This decomposition yields a structured representation that supports the analysis of what information has been encoded by speaker embeddings. LISE preserves ASV performance with negligible EER degradation on x-vector and ECAPA-TDNN. Crucially, the interpretability of these components for human listeners is demonstrated through listening experiments, where participants distinguished speakers with 83.9% accuracy.", "date": "2026-06-19", "categories": []}
{"id": "2606.21335", "title": "Direct Raw Audio Signal Processing via Reservoir Computing: An Investigation into 'Feature-Free' Architectures", "abstract": "This paper evaluates Reservoir Computing (RC) as an autonomous, 'feature-free' framework for audio processing, designed to eliminate traditional, handcrafted feature extraction stages. We investigate whether the high-dimensional temporal dynamics inherent in a reservoir can function as a robust end-to-end processor for the direct classification of raw acoustic signals. By bypassing computationally intensive representations like MFCCs, this approach seeks to mitigate significant intellectual and pre-processing bottlenecks in traditional signal pipelines. Our study evaluates and compares shallow, sequential, and parallel deep reservoir architectures to determine their capacity for hierarchical feature representation. Experimental results demonstrate that the proposed parallel approach consistently outperforms shallow and sequential baselines while maintaining low model complexity. These findings highlight the potential of RC as an efficient and scalable alternative for time-domain audio processing, offering a promising pathway toward deployable, low-power acoustic systems with minimal preprocessing requirements.", "date": "2026-06-19", "categories": []}
{"id": "2606.21365", "title": "LambdaMark: Semantic Audio Watermarking for Robustness and Radioactivity", "abstract": "Recent advances in generative audio have made voice cloning increasingly effortless, enabling voice fraud, impersonation, and other forms of unauthorized use. A common attack finetunes a speech generation model on recordings of a target speaker, allowing the model to synthesize speech in that speaker's voice. Audio watermarking offers a promising defense by embedding detectable signals into audio. A practical watermark must satisfy two key properties: robustness and radioactivity. Existing audio watermarking methods typically embed signals into low-level representations, such as waveforms or spectrograms, which makes them vulnerable to signal-level manipulations and limits their transfer to downstream models. We introduce LambdaMark -- the first generic radioactive watermarking scheme. Unlike all previous approaches, LambdaMark achieves generic radioactivity by embedding multi-bit watermark information into semantic audio latent representations. Our watermarks have semantic interpretation and are thus more likely to be learned by a downstream model through finetuning. LambdaMark includes a lightweight watermark encoder to inject multi-bit message-dependent perturbations into semantic audio representations and a decoder to detect watermark presence and recover the embedded bit information. Encoder and decoder are trained using a custom multi-component loss that preserves fidelity of the watermarked audio, increases bit-level recovery rate, and improves robustness against common distortions and adversarial removal attempts. Experiments show that LambdaMark achieves near-perfect robustness under common distortions. LambdaMark is also the only watermark that is robust against all evaluated removal attacks. Furthermore, LambdaMark exhibits general and robust radioactivity and remains robust to distortions and adversarial removal attacks even on the generated outputs of those finetuned models.", "date": "2026-06-19", "categories": []}
{"id": "2606.21887", "title": "Improving Engine Sound Analysis in Hot-Test Environments via a RAB-U-Net (Residual Attention Block U-Net) Noise Removal Method", "abstract": "During hot tests on a production line, engine-sound analysis is crucial to ensuring product quality and performance. However, background noise often interferes with accurate sound analysis, leading to potential errors in engine diagnostics. Traditionally, skilled technicians listen to engine sounds to assess engine health, but this is prone to significant inaccuracies. This study presents an innovative deep learning-based approach to address this issue by removing background noise from engine sound recordings using a U-Net neural network structure enhanced with Residual Attention Blocks (RAB-U-Net). Our intelligent noise removal system significantly improves the accuracy of engine noise detection, outperforming traditional techniques and providing a robust solution for real-time applications in production line environments. This study proposes a novel system for engine noise detection in production lines, marking a valuable advancement for the automotive industry in applying deep learning methods to improve the quality of engine diagnostics.", "date": "2026-06-20", "categories": []}
{"id": "2606.22218", "title": "An Analysis of Untrained Deep Reservoir Networks for Audio Surveillance", "abstract": "In this paper, we investigate untrained recurrent models from the Reservoir Computing (RC) paradigm for audio surveillance, focusing on bidirectional Echo State Networks with different depths, from shallow to deep configurations, for emergency sound event detection. We evaluate these models on the MIVIA Audio Events dataset in a multiclass setting across different Signal-to-Noise Ratio (SNR) levels, with the goal of assessing the trade-off between depth, recognition performance, and computational efficiency. We compare the proposed architectures against fully trained recurrent and convolutional-recurrent baselines, namely Bidirectional Long Short-Term Memory networks (BiLSTMs) and Convolutional Recurrent Neural Networks (CRNNs). Results show that deep and shallow reservoir-based models achieve competitive recognition rates, with deeper variants being more robust in highly noisy conditions and shallower ones offering the most favorable efficiency profile, particularly on edge devices such as the NVIDIA Orin. In addition, the proposed approach remains robust across different input representations, including log-Mel spectrograms and MFCCs with varying resolutions. These findings highlight untrained reservoir architectures as a promising solution for resource-constrained audio surveillance scenarios.", "date": "2026-06-20", "categories": []}
{"id": "2606.22824", "title": "BranchShine: Compact Raw-Audio-to-IPA Transcription with a RoPE E-Branchformer Encoder", "abstract": "Speech-to-IPA transcription is useful when the desired output is pronunciation rather than orthographic text, but competitive multilingual systems are often large and evaluation is sensitive to normalization choices. This paper presents BranchShine, a 33M-parameter raw-audio CTC recognizer with a lightweight convolutional front end and a 19-block RoPE E-Branchformer encoder. We find that BranchShine provides a compact and competitive operating point for IPA transcription under matched normalization and scoring. On a 16,660-utterance multilingual test set covering 41 language labels, BranchShine obtains 9.19% whitespace-insensitive IPA character error rate, compared with 9.78% for the 575.00M-parameter PhoneticXEUS baseline. A secondary child speech reading analysis shows a complementary operating profile: BranchShine is more conservative on incorrect readings, while Whisper-Medium is stronger on exact acceptance of correct readings. Overall, the results indicate that a compact raw-audio-to-IPA model can approach much larger baselines on character-level IPA transcription.", "date": "2026-06-22", "categories": []}
{"id": "2606.22952", "title": "Domain-incremental audio classification using domain-specific experts and prototype classifier", "abstract": "This technical report presents submission systems for Task 7(domain-incremental audio classification) of the DCASE 2026 Challenge. The main obstacle is that, the system is unable to access to past or future domain's data at once. We approached domain-incremental learning (DIL) as a frozen-feature replay problem. At each incremental stage, one or two compact experts are trained and then kept fixed; at the final stage, the penultimate features from all frozen experts are concatenated and used to train a lightweight per-class prototype classifier solely on cached features. This design prevents catastrophic forgetting by preserving each expert models at inference. To retain earlier-domain knowledge without storing raw audio, some experts were trained with DeepInversion-based generative replay. A cross-stage regression imputer was trained to fill the expert feature slots that did not yet exist at an ealier stage. We submit four fully DIL-compliant systems: three systems based on diverse frozen five-expert backbones and their cross-stack ensemble achieving 78.15% micro / 77.03% macro on the development set, outperforming every individual backbone on both evaluations.", "date": "2026-06-22", "categories": []}
{"id": "2606.23243", "title": "Unlocking In-Context Learning in Audio-Language Models from Decentralized Medical Audio", "abstract": "Clinical audio diagnosis in low-resource settings requires models that identify conditions from minimal examples without large annotated corpora. We propose Federated Self-Contextualization (FSC), a multimodal language model framework for in-context clinical audio diagnosis across federated hospital clients. FSC constructs pseudo-label episodes via unsupervised clustering of audio representations, bypassing scarce real diagnostic labels, and enables contextual reasoning from support-query pairs. Our progressive three-stage pipeline first aligns audio embeddings with the language model via caption-based pretraining, then adapts it for episodic in-context inference through federated optimization. At test time, given a small labeled support set, the model diagnoses an unseen query through multimodal reasoning. On held-out respiratory and cardiac conditions, FSC achieves 71.6% accuracy in 2-way 2-shot evaluation, outperforming audio-language baselines by over 9%.", "date": "2026-06-22", "categories": []}
{"id": "2606.23335", "title": "The Watermark Shortcut: How Provenance Marking Sabotages Audio Deepfake Detection", "abstract": "Provenance watermarking is increasingly treated as a safeguard for synthetic speech, whether built directly into speech-generation models such as Chatterbox, provided through dedicated techniques such as AudioSeal, or deployed by commercial platforms such as ElevenLabs. We identify a previously uncharacterized liability: when synthetic speech is watermarked and human speech is not, detectors trained alongside latch onto the watermark as a spurious \"watermark => fake\" shortcut. This single feature yields three coupled failures: generalization degradation (model performance deteriorates on unseen data), strip-to-evade (a watermarked fake escapes once unwatermarked), and mark-to-frame (watermarking a real voice flags it as fake). In a controlled white-box experiment, a watermark-trained detector shows all three (for example, mark-to-frame lifts Equal Error Rate from 16% to 75%). In a black-box test of a commercial API, we show that adding a watermark to real speech disguises it as fake. However, this shortcut is fixable: retraining with the watermark on both classes decorrelates it and restores clean behavior. We release experiment data as a paired clean-versus-watermarked corpus (WASP).", "date": "2026-06-22", "categories": []}
{"id": "2606.24066", "title": "VieSpeaker: A Large-Scale Vietnamese Speaker Recognition Dataset Beyond Visual Dependency", "abstract": "Speaker recognition has advanced rapidly with large-scale training datasets, yet Vietnamese remains under-resourced, with existing corpora limited in scale and acoustic diversity. Most large-scale datasets rely on facial cues to link speech with speaker identities, restricting data collection to recordings where speakers appear on camera. We propose a face-independent dataset construction pipeline and introduce VieSpeaker, a large-scale Vietnamese speaker recognition dataset. Our approach leverages textual metadata and large language model reasoning to infer speaker identities from transcripts and contextual information. VieSpeaker contains approximately 902 hours of speech from 4,715 speakers. Experiments show that models trained on VieSpeaker achieve improved robustness and generalization compared to existing Vietnamese datasets. This work demonstrates the feasibility of face-independent dataset construction and provides a new direction for building large-scale speech resources.", "date": "2026-06-23", "categories": []}
{"id": "2606.24127", "title": "DTT-BSR+: A Generative-Regression Cascade for Music Source Restoration", "abstract": "Music source restoration (MSR) requires jointly addressing source unmixing and the inversion of non-linear production effects. Current methods struggle to achieve accurate target signal reconstruction while maintaining semantic consistency. To address this limitation, we propose DTT-BSR+, a two-stage cascade MSR system that decouples distribution fitting from signal reconstruction into separate stages. A generative DTT-BSR separator in the first stage produces stems matching the prior of clean sources, and a modified Demucs network in the second stage enhances the first stage output using time-domain and multi-resolution spectral losses. DTT-BSR+ improves multi-mel signal-to-noise ratio (MMSNR) over the single-stage DTT-BSR across all stems, and surpasses the state-of-the-art X-LANCE MSR system on five stems. We also reveal through Fr\\'echet Audio Distance (FAD) decomposition an implicit trade-off between signal reconstruction accuracy and semantic distribution fitting across stems.", "date": "2026-06-23", "categories": []}
{"id": "2606.25328", "title": "Supervised Post-training of Speech Foundation Models for Robust Adaptation in Speech Deepfake Detection", "abstract": "Large speech foundation models have shown strong potential for speech deepfake detection, but direct fine-tuning is limited by a mismatch between self-supervised pre-training objectives and spoof-specific artifacts. To address this, we propose a mix-frame post-training strategy to create localized spoof-oriented perturbations and use frame-level supervision to encourage the SSL model to learn local inconsistencies that are critical for robust spoof detection. On ASVspoof5, we achieve state-of-the-art EER 4.50% for a single model without data augmentation. On ASVspoof2021 LA/DF, it further achieves only 0.16\\% absolute EER gap between LA and DF, indicating strong and balanced robustness across distinct distortion conditions. These results show that supervised post-training provides an effective and practical way to adapt speech foundation models for robust deepfake detection.", "date": "2026-06-24", "categories": []}
{"id": "2606.25391", "title": "From Sounds to Scenes: A Benchmark for Evaluating Context-Aware Auditory Scene Understanding in Large Audio Language Models", "abstract": "Recent Large Audio Language Models (LALMs) have achieved remarkable progress in audio perceptual tasks across individual acoustic layers, including speech, sound, and music. However, existing benchmarks predominantly evaluate these layers in isolation, overlooking the complex contextual relationships that arise when multiple acoustic sources co-occur in real-world auditory scenes. Real-world auditory interpretation requires Context-Aware Auditory Scene Understanding (CASU): the ability to comprehend the holistic scene by integrating sound layers. To evaluate this capability, we introduce the CASU benchmark, which assesses whether Audio LLMs can interpret auditory scenes composed of speech, acoustic events (e.g., announcements), and background environments (e.g., traffic), and reason about the logical relationships between these layers. We propose a scalable pipeline for constructing time-accurate, semi-synthetic audio streams by composing real-world scene sounds with synthetic speech. Building on this data, we design four tasks that probe scene understanding: contextual question answering, entity extraction from the scene, speaker role inference, and counterfactual reasoning where scene is manipulated. Experiments across multiple LALMs demonstrate that effective auditory scene understanding requires integration over all auditory layers, rather than reliance on speech or sound alone, underscoring the necessity of CASU for advancing complex audio understanding in LALMs.", "date": "2026-06-24", "categories": []}
{"id": "2606.27320", "title": "Elastic Time: Dynamic Frame Rate Bottlenecks for Neural Audio Coding", "abstract": "Neural audio autoencoders have become a core component of compression, feature extraction, and generation. However, while existing systems support variable bitrate, the vast majority of models still operate at a fixed latent frame-rate, allocating equal temporal budget to regions with very different information density, which can result in unnecessarily long sequences. We introduce Elastic Time, a dynamic frame-rate bottleneck that converts fixed-frame-rate autoencoders to dynamic ones. Our method learns a lightweight latent predictor used to decide which frames can be skipped and later reconstructed, enabling efficient greedy boundary selection at inference. Experiments show our method enables deployment-time rate control while improving efficiency-quality tradeoffs relative to baselines. Overall, we provide a flexible mechanism for adjusting temporal resolution in audio autoencoders, potentially facilitating more efficient downstream modeling for generation and long-context tasks.", "date": "2026-06-25", "categories": []}
{"id": "2606.27543", "title": "Advancing Speaker-Based Vocal Effort Classification with WavLM and Data Augmentation in Naturalistic Non-Calibrated Speech Recordings", "abstract": "The variations in vocal effort range (e.g. whisper, soft, neutral, loud, shout) alter production and speech acoustics, reducing intelligibility and limiting the robustness of any subsequent speech technology. Classification is challenging since effort lies on a continuum, adjacent categories are easily confused, and labeled data remain scarce. Prior SSL approaches with wav2vec2, HuBERT, and AST improve performance on the AVID corpus but still suffer from boundary errors. In this study, we introduce WavLM for the first time in vocal effort classification and benchmark it against wav2vec2 and HuBERT. To address data scarcity, we conduct a systematic study of augmentation strategies, covering RIR convolution, additive noise, time masking, speed perturbation, band-limiting, MixUp, and CutMix. Augmentation consistently improves WavLM, with gains ranging from +0.6% to +1.8% absolute. We further propose Gaussian-neighbor soft labels, which further reduce near-boundary confusions by modeling the vocal effort continuum. Our best system, WavLM-BASE with gradual unfreezing, augmentation, and Gaussian-neighbor soft labels, achieves 78.2% mean accuracy, establishing a new state-of-the-art on AVID.", "date": "2026-06-25", "categories": []}
{"id": "2606.27751", "title": "From General-Purpose Audio Tagging to Spatially Grounded Sound Event Localization and Detection", "abstract": "This report investigates the extension of pretrained General-Purpose Audio Tagging (GP-AT) models toward spatially grounded Sound Event Localization and Detection (SELD). The proposed AT2SELD framework couples a pretrained AT backbone with compact First-Order Ambisonics (FOA) spatial processing, track-wise SED and Cartesian DOA estimation, permutation aware supervision, and calibration. It characterizes how semantic audio priors support localization-aware scene analysis under data, computation, and deployment constraints. The framework is developed through informed multi-stage Neural Architecture Search (NAS). Stage 1 shows that spectral FOA descriptors, based on magnitude, phase, and Intensity Vectors (IVs), provide the most reliable interface for semantic-to-spatial transfer. Stage 2 identifies early residual spatial encoding as the main capacity-sensitive component, while late track-wise abstraction and recurrent smoothing act mainly as refinement stages. Stage 3 shows that late cross-stitch coupling improves semantic-spatial interaction, whereas early fusion is costlier and less effective. Diagnostic evaluation analyzes the selected architecture under class balancing, focal loss, activity-conditioned DOA supervision, threshold calibration, and transfer across STARSS23, TAU2019, TAU-NIGENS2020, and TAU-NIGENS2021. Focal loss improves the activity point, active-only DOA supervision mitigates inactive target dominance, and validation-selected thresholds recover calibration without replacing spatial learning. Cross-dataset and oracle-activity analyses indicate strong fixed source localization on TAU2019, transferable representations from TAU NIGENS2021, and meaningful but uncertain behavior on STARSS23. Overall, GP-AT priors appear promising for SELD design when embedded in spatial-aware architectures and optimized through integrated calibration and deployment oriented strategies.", "date": "2026-06-26", "categories": []}
{"id": "2606.28728", "title": "Improving Large-Scale Weakly Supervised ASR by Filtering and Selection", "abstract": "Leveraging large-scale weakly supervised datasets is crucial to train robust end-to-end automatic speech recognition (ASR) models. However, such datasets often contain noisy labels and lack domain specificity, limiting their effectiveness. To address these issues and make better use of weakly supervised datasets, we propose a novel training approach incorporating data filtering and selection. Our approach consists of three steps: pretraining on the entire dataset, continued pretraining on a filtered subset based on character error rate (CER), and fine-tuning on a small number of acoustically similar samples to the target domain, selected from the filtered subset. In experiments with a 90,000-hour weakly supervised Japanese dataset, the proposed filtering and selection methods synergistically reduced CER by up to 6.4% and 4.0%, respectively, even though these steps reused training samples already used in the first pretraining step.", "date": "2026-06-27", "categories": []}
{"id": "2606.30682", "title": "ALM2Vec: Learning Audio Embeddings for Universal Audio Retrieval with Large Audio-Language Models", "abstract": "Recent advances in language--audio retrieval have been largely driven by contrastive dual-encoder architectures that align audio and text in a shared embedding space. While effective, existing retrieval embeddings are primarily optimized for audio--caption matching, limiting their ability to support diverse retrieval objectives and controllable retrieval behaviors. We present ALM2Vec, a universal audio embedding framework derived from pretrained large audio--language models (LALMs). By transferring the audio understanding, instruction-following, and reasoning capabilities acquired through large-scale multimodal training, ALM2Vec learns a unified embedding space for retrieval across audio domains and task types. Beyond conventional text--audio retrieval, ALM2Vec incorporates natural-language instructions into the embedding process, enabling instruction-aware retrieval for scenarios such as audio question answering and aspect-conditioned retrieval. Experimental results show that ALM2Vec achieves competitive performance on standard audio and speech retrieval benchmarks while exhibiting promising compositional and controllable retrieval capabilities, highlighting its potential as a unified audio embedding model for retrieval across domains, tasks, and user intents.", "date": "2026-06-27", "categories": []}
{"id": "2606.29544", "title": "Proteus: Automated Adversarial Robustness Testing for Audio Deepfake Detectors", "abstract": "We present Proteus, a framework developed at Resemble AI for automated robustness testing of our audio deepfake detection system. Given a detector, Proteus systematically searches over sequences of everyday audio transformations (codec transcoding, additive noise, reverberation, dynamic-range compression, and VoIP simulation) to find combinations that fool the detector while preserving speech quality. We propose two complementary search strategies: (1) a breadth-first search that exhaustively maps augmentation effectiveness across the parameter space, and (2) a Q-learning agent designed to efficiently discover deeper attack chains by exploiting structural patterns in the BFS data. We report findings from continuous deployment of Proteus against our production detector, showing that specific augmentation chains can reliably flip detection verdicts while preserving speech intelligibility and speaker identity. We discuss how these findings are used to harden the detector through targeted retraining.", "date": "2026-06-28", "categories": []}
{"id": "2606.29901", "title": "Semi-Supervised Sound Event Detection with Conditional Mixup and Embedding-Level Contrastive Loss", "abstract": "Sound event detection (SED) is a core module for acoustic environmental analysis, yet its performance is often limited by scarce labeled data. Recent systems leverage large pretrained audio foundation models, but effective fine-tuning remains challenging because labeled data are limited while unlabeled data are abundant. A previous work, ATST-SED, addressed this problem with a pseudo-label based semi-supervised fine-tuning framework. In this work, we further improve the framework by adopting an embedding-level self-supervised contrastive loss inspired by ATST-Frame pretraining. This contrastive objective better exploits unlabeled data during fine-tuning. One challenge is that mixup serves different roles in the two objectives: pseudo-label learning uses composition mixup, while contrastive learning treats mixup as a perturbation. To resolve this mismatch, we propose conditional mixup, which combines composition mixup and perturbation mixup in one semi-supervised framework and defines the corresponding embedding-level contrastive losses. The resulting model achieves 0.645 PSDS1 and 0.822 PSDS2 on the DESED validation set, establishing a new state of the art.", "date": "2026-06-29", "categories": []}
{"id": "2606.30642", "title": "LeVo 2: Stable and Melodious Song Generation via Hierarchical Representation Modeling and Progressive Post-Training", "abstract": "Full-length song generation must preserve coherence and musicality, render detailed vocal and accompaniment acoustics, and follow lyrics and prompts. Existing language model-based systems face a structural trade-off: mixed-token modeling preserves vocal-instrument coordination but obscures track-specific details, whereas dual-track prediction improves acoustics but requires longer sequences and weakens global planning. We present LeVo 2, a hybrid LLM-Diffusion framework for controllable full-length song generation. LeVo 2 formulates this trade-off as hierarchical modeling: LeLM first predicts mixed tokens for semantic planning, then predicts vocal and accompaniment tokens in parallel for track-specific refinement, while a diffusion-based Music Codec reconstructs full-length waveforms. A central contribution of this extended version is an aesthetics-guided training schedule for alignment. During pre-training, an automated music aesthetic evaluation framework assigns musicality-tier conditions to large-scale data, providing musicality priors before preference alignment. Progressive post-training applies SFT, large-scale offline DPO, and closed-loop semi-online DPO to separately improve generation quality, controllability, and musicality. Modular extension then trains the Track-Specific LM for acoustic refinement while preserving the aligned semantic planner. This schedule separates musicality learning, controllability alignment, and acoustic refinement, mitigating optimization conflict and the limitations of static offline preference pairs. Expert listening tests and objective evaluations show that LeVo 2 outperforms open-source baselines across six subjective dimensions, and approaches leading commercial systems on several listening metrics. Ablations validate the effects of the training strategy, aesthetics guidance, scaling, and hierarchical architecture.", "date": "2026-06-29", "categories": []}
{"id": "2606.30700", "title": "BEST-RQ-2: Contextualize-Then-Predict, a Two-Step Approach for Self-Supervised Audio Representations", "abstract": "Self-supervised learning enables audio representations that transfer across domains and tasks. We present BEST-RQ-2, an evolution of BEST-RQ that retains frozen randomprojection-based discrete targets while introducing a two-step contextualize-then-predict pretraining scheme. A ViT context encoder processes only the unmasked spectrogram regions, and a lightweight predictor infers targets for the masked regions; the predictor is discarded after pretraining. Replacing the original Conformer encoder with a ViT shifts performance across domains, slightly reducing speech performance while improving music and environmental sounds, with comparable average scores. The main improvement comes from decomposing masked prediction into separate contextualization and prediction stages. On the X-ARES and XARES-LLM benchmarks, BEST-RQ-2 consistently outperforms one-stage baselines in overall transfer while keeping inference compute unchanged. Code and model checkpoints are publicly available.", "date": "2026-06-29", "categories": []}
{"id": "2606.30780", "title": "Detecting Audio Deepfakes on the Edge:Lightweight SSL-Based Detection in a Browser Plugin", "abstract": "Audio deepfakes are a growing challenge for the general public, as well as for journalists and fact-checkers. The latter need reliable tools to verify the authenticity of their sources, while at the same time keeping their information private. Commercial deepfake detection solutions rely on cloud-based processing, which raises privacy concerns. To solve this problem, we propose an on-device audio deepfake detection model. We show that a truncated self-supervised backbone with a simple logistic classifier is both very fast and often more accurate than existing solutions. Our solution outperforms the baseline AASIST by 10% and improves inference speed by 40%. We integrate this model into a browser plug-in, which allows journalists and fact-checkers to detect deepfakes easily and securely. Code for the plugin is available at https://github.com/OctavianPascu97/Audio-Deepfakes-Browser-Plugin.", "date": "2026-06-29", "categories": []}
{"id": "2607.00247", "title": "Adaptive Perturbation Selection for Contrastive Audio Decoding", "abstract": "Large audio-language models (LALMs) frequently hallucinate by overriding acoustic evidence with language priors. While contrastive decoding (CD) offers training-free mitigation, existing methods rely on blunt perturbations like masking or noise, leaving structured audio transformations unexplored. We explore this design space by evaluating a diverse library of targeted audio perturbations and adaptively selecting the optimal negative branch for each task and example. First, we improve upon earlier prompt engineering by showing that a simple binary yes/no constraint reduces the model's tendency to falsely confirm absent audio features. Second, evaluating our library across temporal, spectral, frequency, and amplitude domains reveals that optimal transformations are highly task-dependent; for instance, reversing the audio array disrupts temporal coherence, raising accuracy on the temporal order task from 74.7% to 81.4%. Finally, we trained a light-weight perturbation selector on model hidden states to dynamically route negative branches, yielding an additional +4.3% gain on the existence task.", "date": "2026-06-30", "categories": []}