{"title": "Lattice Rescoring Strategies for Long Short Term Memory Language Models\n  in Speech Recognition", "year": 2017, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Shankar Kumar", "Michael Nirschl", "Daniel Holtmann-Rice", "Hank Liao", "Ananda Theertha Suresh", "Felix Yu"], "abstract": "  Recurrent neural network (RNN) language models (LMs) and Long Short Term\nMemory (LSTM) LMs, a variant of RNN LMs, have been shown to outperform\ntraditional N-gram LMs on speech recognition tasks. However, these models are\ncomputationally more expensive than N-gram LMs for decoding, and thus,\nchallenging to integrate into speech recognizers. Recent research has proposed\nthe use of lattice-rescoring algorithms using RNNLMs and LSTMLMs as an\nefficient strategy to integrate these models into a speech recognition system.\nIn this paper, we evaluate existing lattice rescoring algorithms along with new\nvariants on a YouTube speech recognition task. Lattice rescoring using LSTMLMs\nreduces the word error rate (WER) for this task by 8\\% relative to the WER\nobtained using an N-gram LM.\n", "doi": null, "arxiv_id": "1711.05448", "pdf_urls": ["https://arxiv.org/pdf/1711.05448"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "Proceedings of ASRU 2017", "categories": "stat.ML cs.CL cs.LG"}
{"title": "Unsupervised Adaptation with Domain Separation Networks for Robust\n  Speech Recognition", "year": 2017, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Zhong Meng", "Zhuo Chen", "Vadim Mazalov", "Jinyu Li", "Yifan Gong"], "abstract": "  Unsupervised domain adaptation of speech signal aims at adapting a\nwell-trained source-domain acoustic model to the unlabeled data from target\ndomain. This can be achieved by adversarial training of deep neural network\n(DNN) acoustic models to learn an intermediate deep representation that is both\nsenone-discriminative and domain-invariant. Specifically, the DNN is trained to\njointly optimize the primary task of senone classification and the secondary\ntask of domain classification with adversarial objective functions. In this\nwork, instead of only focusing on learning a domain-invariant feature (i.e. the\nshared component between domains), we also characterize the difference between\nthe source and target domain distributions by explicitly modeling the private\ncomponent of each domain through a private component extractor DNN. The private\ncomponent is trained to be orthogonal with the shared component and thus\nimplicitly increases the degree of domain-invariance of the shared component. A\nreconstructor DNN is used to reconstruct the original speech feature from the\nprivate and shared components as a regularization. This domain separation\nframework is applied to the unsupervised environment adaptation task and\nachieved 11.08% relative WER reduction from the gradient reversal layer\ntraining, a representative adversarial training method, for automatic speech\nrecognition on CHiME-3 dataset.\n", "doi": "10.1109/asru.2017.8268938", "arxiv_id": "1711.08010", "pdf_urls": ["https://arxiv.org/pdf/1711.08010"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "2017 IEEE Automatic Speech Recognition and Understanding Workshop\n  (ASRU), Okinawa, 2017, pp. 214-221", "categories": "cs.CL cs.AI cs.SD eess.AS"}
{"title": "The Geometry of Culture: Analyzing Meaning through Word Embeddings", "year": 2018, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Austin C. Kozlowski", "Matt Taddy", "James A. Evans"], "abstract": "  We demonstrate the utility of a new methodological tool, neural-network word\nembedding models, for large-scale text analysis, revealing how these models\nproduce richer insights into cultural associations and categories than possible\nwith prior methods. Word embeddings represent semantic relations between words\nas geometric relationships between vectors in a high-dimensional space,\noperationalizing a relational model of meaning consistent with contemporary\ntheories of identity and culture. We show that dimensions induced by word\ndifferences (e.g. man - woman, rich - poor, black - white, liberal -\nconservative) in these vector spaces closely correspond to dimensions of\ncultural meaning, and the projection of words onto these dimensions reflects\nwidely shared cultural connotations when compared to surveyed responses and\nlabeled historical data. We pilot a method for testing the stability of these\nassociations, then demonstrate applications of word embeddings for\nmacro-cultural investigation with a longitudinal analysis of the coevolution of\ngender and class associations in the United States over the 20th century and a\ncomparative analysis of historic distinctions between markers of gender and\nclass in the U.S. and Britain. We argue that the success of these\nhigh-dimensional models motivates a move towards \"high-dimensional theorizing\"\nof meanings, identities and cultural processes.\n", "doi": "10.1177/0003122419877135", "arxiv_id": "1803.09288", "pdf_urls": ["https://arxiv.org/pdf/1803.09288"], "github": ["https://github.com/UC-MACSS/persp-analysis_A18"], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "American Sociological Review 2019, Vol. 84(5) 905-949", "categories": "cs.CL"}
{"title": "Cosmic rays from supernova remnants and superbubbles", "year": 2018, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Richard E. Lingenfelter"], "abstract": "  Recent high energy gamma-ray observations of both single supernova remnants\nand superbubbles, together with observations of supernovae, star formation\nregions, and local cosmic ray composition, now provide an integrated framework\ntying together the sources, injection, acceleration and propagation of the\ncosmic rays, so that it is possible to determine their relative contributions\nto cosmic ray acceleration for all but the very highest energies.\n", "doi": "10.1016/j.asr.2017.04.006", "arxiv_id": "1807.09726", "pdf_urls": ["https://arxiv.org/pdf/1807.09726"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "Adv. Space Res. (2017),\n  http://dx.doi.org/10.1016/j.asr.2017.04.006", "categories": "astro-ph.HE"}
{"title": "Spoof detection using time-delay shallow neural network and feature\n  switching", "year": 2019, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Mari Ganesh Kumar", "Suvidha Rupesh Kumar", "Saranya M", "B. Bharathi", "Hema\n  A. Murthy"], "abstract": "  Detecting spoofed utterances is a fundamental problem in voice-based\nbiometrics. Spoofing can be performed either by logical accesses like speech\nsynthesis, voice conversion or by physical accesses such as replaying the\npre-recorded utterance. Inspired by the state-of-the-art \\emph{x}-vector based\nspeaker verification approach, this paper proposes a time-delay shallow neural\nnetwork (TD-SNN) for spoof detection for both logical and physical access. The\nnovelty of the proposed TD-SNN system vis-a-vis conventional DNN systems is\nthat it can handle variable length utterances during testing. Performance of\nthe proposed TD-SNN systems and the baseline Gaussian mixture models (GMMs) is\nanalyzed on the ASV-spoof-2019 dataset. The performance of the systems is\nmeasured in terms of the minimum normalized tandem detection cost function\n(min-t-DCF). When studied with individual features, the TD-SNN system\nconsistently outperforms the GMM system for physical access. For logical\naccess, GMM surpasses TD-SNN systems for certain individual features. When\ncombined with the decision-level feature switching (DLFS) paradigm, the best\nTD-SNN system outperforms the best baseline GMM system on evaluation data with\na relative improvement of 48.03\\% and 49.47\\% for both logical and physical\naccess, respectively.\n", "doi": "10.1109/asru46091.2019.9003824", "arxiv_id": "1904.07453", "pdf_urls": ["https://arxiv.org/pdf/1904.07453"], "github": ["https://github.com/mariganeshkumar/TDSNN-spoof-detection"], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "2019 IEEE Automatic Speech Recognition and Understanding Workshop\n  (ASRU), 1011--1017", "categories": "eess.AS cs.CR cs.LG cs.SD"}
{"title": "Spectral fluctuation analysis of ionospheric inhomogeneities over\n  Brazilian territory Part II: E-F valley region plasma instabilities", "year": 2019, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Neelakshi J.", "Reinaldo R. Rosa", "Siomel Savio", "Francisco C. de Meneses", "Stephan Stephany", "Gabriel Fornari", "and Muralikrishna P"], "abstract": "  The turbulent-like process associated with ionospheric instabilities, has\nbeen identified as the main nonlinear process that drives the irregularities\nobserved in the different ionospheric regions. In this complementary study, as\nproposed in the first article of this two-paper series [Fornari et al., Adv.\nSpace Res. 58, 2016], we performed the detrended fluctuation analysis of the\nequatorial E-F valley region electron density fluctuations measured from an in\nsitu experiment performed over the Brazilian territory. The spectral\nconsistency with the K41 turbulent universality class is analyzed for E-F\nvalley region from the DFA spectra for four electron density time series. A\ncomplementary detrended fluctuation analysis for four time series of the\nF-layer electric field is also presented. Consistent with the results obtained\nfor the F region, the analysis for the E-F valley region also shows a very high\nspectral variation ($\\gg50\\%$). Thus, the spectral analysis performed in both\nparts of the series suggest that a process such as the homogeneous turbulence\nK41 ($\\beta =-5/3\\pm 2\\%$) is inappropriate to describe both the fluctuations\nof electron density and the electric field associated with the main ionospheric\ninstabilities.\n", "doi": "10.1016/j.asr.2019.07.015", "arxiv_id": "1906.04316", "pdf_urls": ["https://arxiv.org/pdf/1906.04316"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "https://doi.org/10.1016/j.asr.2019.07.015", "categories": "physics.space-ph physics.geo-ph"}
{"title": "Input-to-state stability for parabolic boundary control: Linear and\n  semi-linear systems", "year": 2019, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Felix Schwenninger"], "abstract": "  Input-to-state stability (ISS) for systems described by partial differential\nequations has seen intensified research activity recently, and in particular\nthe class of boundary control systems, for which truly infinite-dimensional\neffects enter the situation. This note reviews input-to-state stability for\nparabolic equations with respect to general $L^{p}$-input-norms in the linear\ncase and includes extensions of recent results on semilinear equations.\n", "doi": "10.1007/978-3-030-35898-3", "arxiv_id": "1908.08317", "pdf_urls": ["https://arxiv.org/pdf/1908.08317"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "Control Theory of Infinite-Dimensional Systems, J. Kerner, H.\n  Laasri, D. Mugnolo (Eds), Operator Theory: Advances and Applications, vol.\n  277, 2020", "categories": "math.OC math.AP math.FA"}
{"title": "Probing the Information Encoded in X-vectors", "year": 2019, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Desh Raj", "David Snyder", "Daniel Povey", "Sanjeev Khudanpur"], "abstract": "  Deep neural network based speaker embeddings, such as x-vectors, have been\nshown to perform well in text-independent speaker recognition/verification\ntasks. In this paper, we use simple classifiers to investigate the contents\nencoded by x-vector embeddings. We probe these embeddings for information\nrelated to the speaker, channel, transcription (sentence, words, phones), and\nmeta information about the utterance (duration and augmentation type), and\ncompare these with the information encoded by i-vectors across a varying number\nof dimensions. We also study the effect of data augmentation during extractor\ntraining on the information captured by x-vectors. Experiments on the RedDots\ndata set show that x-vectors capture spoken content and channel-related\ninformation, while performing well on speaker verification tasks.\n", "doi": "10.1109/asru46091.2019.9003979", "arxiv_id": "1909.06351", "pdf_urls": ["https://arxiv.org/pdf/1909.06351"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "IEEE Automatic Speech Recognition and Understanding Workshop\n  (ASRU) (2019): 726-733", "categories": "eess.AS cs.CL cs.SD"}
{"title": "Self-Adaptive Soft Voice Activity Detection using Deep Neural Networks\n  for Robust Speaker Verification", "year": 2019, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Youngmoon Jung", "Yeunju Choi", "Hoirin Kim"], "abstract": "  Voice activity detection (VAD), which classifies frames as speech or\nnon-speech, is an important module in many speech applications including\nspeaker verification. In this paper, we propose a novel method, called\nself-adaptive soft VAD, to incorporate a deep neural network (DNN)-based VAD\ninto a deep speaker embedding system. The proposed method is a combination of\nthe following two approaches. The first approach is soft VAD, which performs a\nsoft selection of frame-level features extracted from a speaker feature\nextractor. The frame-level features are weighted by their corresponding speech\nposteriors estimated from the DNN-based VAD, and then aggregated to generate a\nspeaker embedding. The second approach is self-adaptive VAD, which fine-tunes\nthe pre-trained VAD on the speaker verification data to reduce the domain\nmismatch. Here, we introduce two unsupervised domain adaptation (DA) schemes,\nnamely speech posterior-based DA (SP-DA) and joint learning-based DA (JL-DA).\nExperiments on a Korean speech database demonstrate that the verification\nperformance is improved significantly in real-world environments by using\nself-adaptive soft VAD.\n", "doi": "10.1109/asru46091.2019.9003935", "arxiv_id": "1909.11886", "pdf_urls": ["https://arxiv.org/pdf/1909.11886"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "Proc. of ASRU 2019, pp. 365-372", "categories": "eess.AS cs.CL cs.LG cs.SD stat.ML"}
{"title": "Latent space representation for multi-target speaker detection and\n  identification with a sparse dataset using Triplet neural networks", "year": 2019, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Kin Wai Cheuk", "Balamurali B. T.", "Gemma Roig", "Dorien Herremans"], "abstract": "  We present an approach to tackle the speaker recognition problem using\nTriplet Neural Networks. Currently, the $i$-vector representation with\nprobabilistic linear discriminant analysis (PLDA) is the most commonly used\ntechnique to solve this problem, due to high classification accuracy with a\nrelatively short computation time. In this paper, we explore a neural network\napproach, namely Triplet Neural Networks (TNNs), to built a latent space for\ndifferent classifiers to solve the Multi-Target Speaker Detection and\nIdentification Challenge Evaluation 2018 (MCE 2018) dataset. This training set\ncontains $i$-vectors from 3,631 speakers, with only 3 samples for each speaker,\nthus making speaker recognition a challenging task. When using the train and\ndevelopment set for training both the TNN and baseline model (i.e., similarity\nevaluation directly on the $i$-vector representation), our proposed model\noutperforms the baseline by 23%. When reducing the training data to only using\nthe train set, our method results in 309 confusions for the Multi-target\nspeaker identification task, which is 46% better than the baseline model. These\nresults show that the representational power of TNNs is especially evident when\ntraining on small datasets with few instances available per class.\n", "doi": null, "arxiv_id": "1910.01463", "pdf_urls": ["https://arxiv.org/pdf/1910.01463"], "github": ["https://github.com/KinWaiCheuk/MCE2018"], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "Proceedings of IEEE Automatic Speech Recognition and Understanding\n  Workshop (ASRU 2019). Singapore. 2019", "categories": "cs.SD cs.LG eess.AS"}
{"title": "Detecting Deception in Political Debates Using Acoustic and Textual\n  Features", "year": 2019, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Daniel Kopev", "Ahmed Ali", "Ivan Koychev", "Preslav Nakov"], "abstract": "  We present work on deception detection, where, given a spoken claim, we aim\nto predict its factuality. While previous work in the speech community has\nrelied on recordings from staged setups where people were asked to tell the\ntruth or to lie and their statements were recorded, here we use real-world\npolitical debates. Thanks to the efforts of fact-checking organizations, it is\npossible to obtain annotations for statements in the context of a political\ndiscourse as true, half-true, or false. Starting with such data from the\nCLEF-2018 CheckThat! Lab, which was limited to text, we performed alignment to\nthe corresponding videos, thus producing a multimodal dataset. We further\ndeveloped a multimodal deep-learning architecture for the task of deception\ndetection, which yielded sizable improvements over the state of the art for the\nCLEF-2018 Lab task 2. Our experiments show that the use of the acoustic signal\nconsistently helped to improve the performance compared to using textual and\nmetadata features only, based on several different evaluation measures. We\nrelease the new dataset to the research community, hoping to help advance the\noverall field of multimodal deception detection.\n", "doi": null, "arxiv_id": "1910.01990", "pdf_urls": ["https://arxiv.org/pdf/1910.01990"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "ASRU-2019", "categories": "cs.CL cs.AI eess.AS"}
{"title": "Transfer Learning for Context-Aware Spoken Language Understanding", "year": 2020, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Qian Chen", "Zhu Zhuo", "Wen Wang", "Qiuyun Xu"], "abstract": "  Spoken language understanding (SLU) is a key component of task-oriented\ndialogue systems. SLU parses natural language user utterances into semantic\nframes. Previous work has shown that incorporating context information\nsignificantly improves SLU performance for multi-turn dialogues. However,\ncollecting a large-scale human-labeled multi-turn dialogue corpus for the\ntarget domains is complex and costly. To reduce dependency on the collection\nand annotation effort, we propose a Context Encoding Language Transformer\n(CELT) model facilitating exploiting various context information for SLU. We\nexplore different transfer learning approaches to reduce dependency on data\ncollection and annotation. In addition to unsupervised pre-training using\nlarge-scale general purpose unlabeled corpora, such as Wikipedia, we explore\nunsupervised and supervised adaptive training approaches for transfer learning\nto benefit from other in-domain and out-of-domain dialogue corpora.\nExperimental results demonstrate that the proposed model with the proposed\ntransfer learning approaches achieves significant improvement on the SLU\nperformance over state-of-the-art models on two large-scale single-turn\ndialogue benchmarks and one large-scale multi-turn dialogue benchmark.\n", "doi": "10.1109/asru46091.2019.9003902", "arxiv_id": "2003.01305", "pdf_urls": ["https://arxiv.org/pdf/2003.01305"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "2019 IEEE Automatic Speech Recognition and Understanding Workshop\n  (ASRU), SG, Singapore, 2019, pp. 779-786", "categories": "cs.CL"}
{"title": "Novel Memory Structures in QCA Nano Technology", "year": 2020, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Ali H. Majeed", "Esam Alkaldy", "Mohd S. Zainal", "and Danial MD. Nor"], "abstract": "  Quantum-dot Cellular Automata (QCA) is a new emerging technology for\ndesigning electronic circuits in nanoscale. QCA technology comes to overcome\nthe CMOS limitation and to be a good alternative as it can work in\nultra-high-speed. QCA brought researchers attention due to many features such\nas low power consumption, small feature size in addition to high frequency.\nDesigning circuits in QCA technology with minimum costs such as cells count and\nthe area is very important. This paper presents novel structures of D-latch and\nD-Flip Flop with the lower area and cell count. The proposed Flip-Flop has SET\nand RESET ability. The proposed latch and Flip-Flop have lower complexity\ncompared with counterparts in terms of cell counts by 32% and 26% respectively.\nThe proposed circuits are designed and simulated in QCADesigner software.\n", "doi": "10.37917/ijeee.sceeer.3rd.17", "arxiv_id": "2007.01954", "pdf_urls": ["https://arxiv.org/pdf/2007.01954"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "The 3rd Scientific Conference of Electrical and Electronic\n  Engineering Researches (SCEEER) | (15-16) JUNE 2020 | BASRAH / IRAQ", "categories": "cs.ET cond-mat.other"}
{"title": "Reranking Machine Translation Hypotheses with Structured and Web-based\n  Language Models", "year": 2021, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Wen Wang and Andreas Stolcke and Jing Zheng"], "abstract": "  In this paper, we investigate the use of linguistically motivated and\ncomputationally efficient structured language models for reranking N-best\nhypotheses in a statistical machine translation system. These language models,\ndeveloped from Constraint Dependency Grammar parses, tightly integrate\nknowledge of words, morphological and lexical features, and syntactic\ndependency constraints. Two structured language models are applied for N-best\nrescoring, one is an almost-parsing language model, and the other utilizes more\nsyntactic features by explicitly modeling syntactic dependencies between words.\nWe also investigate effective and efficient language modeling methods to use\nN-grams extracted from up to 1 teraword of web documents. We apply all these\nlanguage models for N-best re-ranking on the NIST and DARPA GALE program 2006\nand 2007 machine translation evaluation tasks and find that the combination of\nthese language models increases the BLEU score up to 1.6% absolutely on blind\ntest sets.\n", "doi": "10.1109/asru.2007.4430102", "arxiv_id": "2104.12277", "pdf_urls": ["https://arxiv.org/pdf/2104.12277"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "Proc. 2007 IEEE ASRU Workshop, pp. 159-164", "categories": "cs.CL"}
{"title": "Efficient conformer: Progressive downsampling and grouped attention for\n  automatic speech recognition", "year": 2021, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Maxime Burchi", "Valentin Vielzeuf"], "abstract": "  The recently proposed Conformer architecture has shown state-of-the-art\nperformances in Automatic Speech Recognition by combining convolution with\nattention to model both local and global dependencies. In this paper, we study\nhow to reduce the Conformer architecture complexity with a limited computing\nbudget, leading to a more efficient architecture design that we call Efficient\nConformer. We introduce progressive downsampling to the Conformer encoder and\npropose a novel attention mechanism named grouped attention, allowing us to\nreduce attention complexity from $O(n^{2}d)$ to $O(n^{2}d / g)$ for sequence\nlength $n$, hidden dimension $d$ and group size parameter $g$. We also\nexperiment the use of strided multi-head self-attention as a global\ndownsampling operation. Our experiments are performed on the LibriSpeech\ndataset with CTC and RNN-Transducer losses. We show that within the same\ncomputing budget, the proposed architecture achieves better performances with\nfaster training and decoding compared to the Conformer. Our 13M parameters CTC\nmodel achieves competitive WERs of 3.6%/9.0% without using a language model and\n2.7%/6.7% with an external n-gram language model on the test-clean/test-other\nsets while being 29% faster than our CTC Conformer baseline at inference and\n36% faster to train.\n", "doi": null, "arxiv_id": "2109.01163", "pdf_urls": ["https://arxiv.org/pdf/2109.01163"], "github": ["https://github.com/burchim/efficientconformer"], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "ASRU 2021, Dec 2021, Cartagena, Colombia", "categories": "eess.AS cs.AI cs.CL cs.SD"}
{"title": "Comparison of Self-Supervised Speech Pre-Training Methods on Flemish\n  Dutch", "year": 2021, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Jakob Poncelet", "Hugo Van hamme"], "abstract": "  Recent research in speech processing exhibits a growing interest in\nunsupervised and self-supervised representation learning from unlabelled data\nto alleviate the need for large amounts of annotated data. We investigate\nseveral popular pre-training methods and apply them to Flemish Dutch. We\ncompare off-the-shelf English pre-trained models to models trained on an\nincreasing amount of Flemish data. We find that the most important factors for\npositive transfer to downstream speech recognition tasks include a substantial\namount of data and a matching pre-training domain. Ideally, we also finetune on\nan annotated subset in the target language. All pre-trained models improve\nlinear phone separability in Flemish, but not all methods improve Automatic\nSpeech Recognition. We experience superior performance with wav2vec 2.0 and we\nobtain a 30% WER improvement by finetuning the multilingually pre-trained\nXLSR-53 model on Flemish Dutch, after integration into an HMM-DNN acoustic\nmodel.\n", "doi": "10.1109/asru51503.2021.9688061", "arxiv_id": "2109.14357", "pdf_urls": ["https://arxiv.org/pdf/2109.14357"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "2021 IEEE Automatic Speech Recognition and Understanding Workshop\n  (ASRU), pp. 169-176", "categories": "eess.AS cs.SD"}
{"title": "On the invertibility of a voice privacy system using embedding\n  alignement", "year": 2021, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Pierre Champion (MULTISPEECH", "LIUM)", "Thomas Thebaud (LIUM)", "Ga\\\"el Le\n  Lan", "Anthony Larcher (LIUM)", "Denis Jouvet (MULTISPEECH)"], "abstract": "  This paper explores various attack scenarios on a voice anonymization system\nusing embeddings alignment techniques. We use Wasserstein-Procrustes (an\nalgorithm initially designed for unsupervised translation) or Procrustes\nanalysis to match two sets of x-vectors, before and after voice anonymization,\nto mimic this transformation as a rotation function. We compute the optimal\nrotation and compare the results of this approximation to the official Voice\nPrivacy Challenge results. We show that a complex system like the baseline of\nthe Voice Privacy Challenge can be approximated by a rotation, estimated using\na limited set of x-vectors. This paper studies the space of solutions for voice\nanonymization within the specific scope of rotations. Rotations being\nreversible, the proposed method can recover up to 62% of the speaker identities\nfrom anonymized embeddings.\n", "doi": null, "arxiv_id": "2110.05431", "pdf_urls": ["https://arxiv.org/pdf/2110.05431"], "github": ["https://github.com/deep-privacy/x-vector-procrustes"], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "ASRU 2021 - IEEE Automatic Speech Recognition and Understanding\n  Workshop, Dec 2021, Cartagena, Colombia", "categories": "eess.AS cs.CR cs.LG cs.SD"}
{"title": "Word-level confidence estimation for RNN transducers", "year": 2021, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Mingqiu Wang", "Hagen Soltau", "Laurent El Shafey", "Izhak Shafran"], "abstract": "  Confidence estimate is an often requested feature in applications such as\nmedical transcription where errors can impact patient care and the confidence\nestimate could be used to alert medical professionals to verify potential\nerrors in recognition.\n  In this paper, we present a lightweight neural confidence model tailored for\nAutomatic Speech Recognition (ASR) system with Recurrent Neural Network\nTransducers (RNN-T). Compared to other existing approaches, our model utilizes:\n(a) the time information associated with recognized words, which reduces the\ncomputational complexity, and (b) a simple and elegant trick for mapping\nbetween sub-word and word sequences. The mapping addresses the non-unique\ntokenization and token deletion problems while amplifying differences between\nconfusable words. Through extensive empirical evaluations on two different\nlong-form test sets, we demonstrate that the model achieves a performance of\n0.4 Normalized Cross Entropy (NCE) and 0.05 Expected Calibration Error (ECE).\nIt is robust across different ASR configurations, including target types\n(graphemes vs. morphemes), traffic conditions (streaming vs. non-streaming),\nand encoder types. We further discuss the importance of evaluation metrics to\nreflect practical applications and highlight the need for further work in\nimproving Area Under the Curve (AUC) for Negative Precision Rate (NPV) and True\nNegative Rate (TNR).\n", "doi": null, "arxiv_id": "2110.15222", "pdf_urls": ["https://arxiv.org/pdf/2110.15222"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "Proc. ASRU 2021", "categories": "cs.CL cs.SD eess.AS"}
{"title": "Attentive Contextual Carryover for Multi-Turn End-to-End Spoken Language\n  Understanding", "year": 2021, "venue_abbr": "ASR", "venue_name": "American Sociological Review", "authors": ["Kai Wei", "Thanh Tran", "Feng-Ju Chang", "Kanthashree Mysore Sathyendra", "Thejaswi Muniyappa", "Jing Liu", "Anirudh Raju", "Ross McGowan", "Nathan Susanj", "Ariya Rastrow", "Grant P. Strimel"], "abstract": "  Recent years have seen significant advances in end-to-end (E2E) spoken\nlanguage understanding (SLU) systems, which directly predict intents and slots\nfrom spoken audio. While dialogue history has been exploited to improve\nconventional text-based natural language understanding systems, current E2E SLU\napproaches have not yet incorporated such critical contextual signals in\nmulti-turn and task-oriented dialogues. In this work, we propose a contextual\nE2E SLU model architecture that uses a multi-head attention mechanism over\nencoded previous utterances and dialogue acts (actions taken by the voice\nassistant) of a multi-turn dialogue. We detail alternative methods to integrate\nthese contexts into the state-ofthe-art recurrent and transformer-based models.\nWhen applied to a large de-identified dataset of utterances collected by a\nvoice assistant, our method reduces average word and semantic error rates by\n10.8% and 12.6%, respectively. We also present results on a publicly available\ndataset and show that our method significantly improves performance over a\nnoncontextual baseline\n", "doi": null, "arxiv_id": "2112.06743", "pdf_urls": ["https://arxiv.org/pdf/2112.06743"], "github": [], "is_oa": true, "sources": ["open-arxiv"], "type": "journal", "journal_ref": "ASRU2021", "categories": "cs.CL cs.AI"}