ANC-Spoof: Audio Neural Codec Spoof Dataset

Overview

ANC-Spoof (Audio Neural Codec Spoof Dataset) is a large-scale dataset designed to facilitate research on the robustness of Audio Deepfake Detection (ADD) systems against distortions introduced by neural audio codecs.

Neural audio codecs have recently emerged as important components in modern speech technology, serving both as efficient compression mechanisms and as resynthesis components in emerging audio generation systems. The transformations introduced by neural codecs can substantially alter the acoustic characteristics of both bona fide and spoofed speech, potentially affecting the reliability and generalization of existing audio deepfake detectors.

Unlike conventional post-hoc compression methods such as MP3, AAC, or Opus, neural codecs use learned encoder-decoder architectures to reconstruct speech waveforms. Furthermore, neural codecs increasingly play a dual role: they are used for legitimate speech compression and can also serve as the resynthesis stage in Audio Language Model (ALM)-based speech generation systems.

ANC-Spoof addresses this challenge by applying multiple neural audio codec algorithms to established audio deepfake detection benchmarks. For every original utterance, corresponding neural-codec-compressed versions are generated while preserving the original ground-truth label. This provides a controlled setting for studying whether an ADD system is sensitive to codec-induced artifacts rather than genuine characteristics of spoofed speech.

The dataset is constructed from three publicly available ADD benchmarks:

• ASVspoof 2019 LA
• Fake-or-Real (FoR)
• In-the-Wild

Seven neural audio codecs are used to generate codec-compressed versions of the original audio:

• BigCodec
• DAC
• Higgs Audio V2 Tokenizer
• Mimi
• SNAC
• SpeechTokenizer
• WavTokenizer

The resulting dataset contains both original (uncompressed) and neural codec-compressed speech, enabling systematic evaluation of ADD robustness under neural codec compression.

Motivation

Existing ADD research has primarily considered robustness against signal-level manipulations and conventional post-hoc compression. However, neural audio codecs introduce a fundamentally different type of transformation because their learned encoder-decoder architectures can substantially modify the acoustic representation of speech.

This creates an important robustness question:

Can an audio deepfake detector distinguish genuine spoofing artifacts from artifacts introduced solely by legitimate neural codec compression?

ANC-Spoof provides a controlled benchmark for investigating this question.

In particular, the dataset enables researchers to study whether detectors:

1. Maintain reliable performance on neural-codec-compressed bona fide speech.
2. Maintain reliable performance on neural-codec-compressed spoofed speech.
3. Become overly dependent on codec-induced artifacts as a shortcut for detecting fake speech.
4. Generalize across different neural audio codecs.
5. Remain robust when encountering codecs that were not observed during training.

Dataset Composition

ANC-Spoof is constructed by applying seven neural codecs to three established ADD benchmarks.

For every selected utterance, the original audio is retained and seven codec-compressed counterparts are generated.

Dataset	Split	Original	Compressed	Total
ASVspoof 2019 LA	Training	25,380	177,660	203,040
ASVspoof 2019 LA	Validation	24,986	174,902	199,888
ASVspoof 2019 LA	Evaluation	71,933	503,531	575,464
Fake-or-Real (FoR)	Evaluation	4,634	32,438	37,072
In-the-Wild	Evaluation	31,780	222,460	254,240
Total		158,713	1,110,991	1,269,704

Original refers to the uncompressed source utterances.

Compressed refers to the seven neural-codec-compressed versions generated from the corresponding original utterances.

Because seven codecs are applied to each original utterance, the number of compressed utterances is seven times the number of original utterances.

Dataset Structure

The dataset is organized according to the source benchmark and neural codec:

ANC-Spoof/
├── ASVspoof2019/
│   ├── Original/
│   ├── BigCodec/
│   ├── DAC/
│   ├── HiggsV2/
│   ├── Mimi/
│   ├── SNAC/
│   ├── SpeechTokenizer/
│   ├── WavTokenizer/
│   └── ...
│
├── InTheWild/
│   ├── Original/
│   ├── BigCodec/
│   ├── DAC/
│   ├── HiggsV2/
│   ├── Mimi/
│   ├── SNAC/
│   ├── SpeechTokenizer/
│   ├── WavTokenizer/
│   └── ...
│
└── FoR/
    ├── Original/
    ├── BigCodec/
    ├── DAC/
    ├── HiggsV2/
    ├── Mimi/
    ├── SNAC/
    ├── SpeechTokenizer/
    ├── WavTokenizer/
    └── ...