Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Neural audio codec tokens, especially 3-gram sequences, follow the same statistical laws as natural language, and their language-likeness correlates with better speech recognition and resynthesis.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

NAC token sequences, especially 3-grams, fit Zipf's and Heaps' laws, and their fitted statistics correlate, though weakly and with confounds, with ASR error rates and UTMOS scores.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid descriptive study of NAC token statistics, but the performance-correlation claim is not supported because it never controls for codec capacity. the 3 major comments →

arxiv 2509.01390 v1 pith:6UAEGV7Y submitted 2025-09-01 cs.CL eess.AS

Analysing the Language of Neural Audio Codecs

Joonyong Park , Shinnosuke Takamichi , David M. Chan , Shunsuke Kando , Yuki Saito , Hiroshi Saruwatari This is my paper
classification cs.CL eess.AS
keywords neural audio codecsspeech tokensZipf's lawHeaps' lawn-gram statisticsentropy and redundancyspeech resynthesistokenization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether discrete tokens produced by neural audio codecs behave statistically like words in natural language, and whether that matters for speech systems. Across 15 codec configurations and two languages, it finds that NAC token sequences—most clearly at the 3-gram level—align with Zipf's and Heaps' laws and with entropy-based redundancy measures, just as natural-language word sequences do. It then shows that the codecs whose token statistics look most language-like also tend to have lower speech-recognition error and higher perceived naturalness in resynthesis. The authors argue this suggests language-like statistical regularity is a meaningful signal for designing better generative speech models, not just a curiosity.

Core claim

The central claim is that NAC tokens form a 'language' in a statistical sense: when tokens are flattened across codec codebooks, deduplicated, and analyzed as n-grams, the 3-gram sequences reproduce the rank-frequency decay of Zipf's law, the vocabulary growth of Heaps' law, and the compressibility profile of text more closely than 2-, 4-, or 6-grams. This holds for English and Chinese single-speaker speech. The paper further claims that the fitted exponents—lower Zipf alpha toward 2, Heaps beta toward 1, scaling factor k toward 1, and lower bit reduction rate—correlate with improved WER/CER and UTMOS scores across the 15 codec configurations. Three-grams appear to be the sweet spot: short e

What carries the argument

The analytical machinery is the n-gram token sequence extracted from each codec: per-dimension tokens are offset by codebook size to avoid label collisions across dimensions, concatenated with dimension-start and dimension-end markers, and deduplicated. On these sequences the paper fits Zipf's law and Heaps' law, computes Shannon entropy and the Huffman bit reduction rate, and compares all against word-level natural-language baselines. The same 3-gram statistics are then plotted against speech-recognition error rates and a learned naturalness score to reveal correlations.

Load-bearing premise

The correlational conclusions assume the 15 codec configurations are independent samples and that the association between language-like token statistics and performance is not a side effect of model size—the paper's own figures show that codecs with larger codebooks tend to have both.

What would settle it

Re-run the scatter plots in Figures 7–9 while controlling for codebook dimension and bitrate, either by partial correlation or by matching codecs with equal capacity but different token statistics. If the Zipf, Heaps, and redundancy coefficients vanish or reverse, the reported language-likeness–performance link is a proxy for codec capacity rather than a direct relationship.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • 3-gram token statistics become a compact diagnostic: a single Zipf or Heaps exponent per codec is associated with that codec's relative ASR and resynthesis performance.
  • Token statistics can guide codec design: configurations with very repetitive tokens (low Heaps beta, high bit reduction rate) or very sparse tokens (high Zipf alpha, near-unique 4-grams) sit on the worse end of benchmarks.
  • Language-like statistical regularity is not automatically present in NAC tokens; it emerges at a specific granularity, so generative speech models should be evaluated at the 3-gram level rather than at the token level.
  • Pre-processing choices used in the paper—flattening dimensions with offset codebook IDs and removing consecutive duplicates—preserve the information that correlates with downstream performance, supporting similar pipelines in speech-language modeling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's correlations are computed over 15 codec configurations whose codebook dimensions and bitrates vary together with the token statistics; if the correlations survive within matched pairs (same capacity, different data), that would support a direct role for linguistic regularity rather than model scale.
  • One could test the 3-gram diagnostic as a training-time or early-stopping signal: train two codecs with identical capacity on different data and check that the one with more Zipfian 3-grams achieves the better resynthesis score.
  • The observed language-likeness may partly reflect the residual-vector-quantization structure itself—each dimension reuses tokens in recurring patterns—so the finding may extend to other discrete neural representations, though the paper does not claim this.
  • The paper notes in footnote 3 that exact 2-gram and 4-gram benchmark values were omitted for space; a complete version would let readers directly verify the claimed 3-gram sweet spot.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper analyzes the discrete token sequences produced by 15 configurations of neural audio codecs (NACs). For each configuration, it extracts n-grams (n = 2, 3, 4, 6), fits Zipf's law (exponent α, KS distance) and Heaps' law (k, β), and computes Huffman-based bit-reduction rates. These statistics are compared with word 1-grams of English and Chinese ground-truth transcripts. The authors report that 3-grams are the most 'language-like' level and that codec configurations whose 3-gram statistics are closer to Zipf/Heaps behavior and have lower bit-reduction rates also achieve better WER/CER and UTMOS scores. The conclusion is that NAC tokens, especially 3-grams, exhibit language-like statistical regularities, and that these regularities correlate with improved resynthesis intelligibility and naturalness.

Significance. If the claims were fully supported, the paper would provide useful guidance for generative speech modeling, suggesting that modeling at the 3-gram level and encouraging Zipfian/Heaps-like token distributions could improve downstream performance. The study has genuine strengths: it uses 15 open-source codec configurations with standardized extraction via Codec-SUPERB, analyzes two languages, employs the standard powerlaw library, and makes an explicit effort to handle multi-dimensional codebooks with label offsets. However, the practical takeaway—that language-like statistics correlate with better ASR/UTMOS performance—is currently insecure because the analysis does not control for codec capacity, bitrate, or model family, and because the correlation tests are presented without significance measures. The descriptive finding that 3-grams are closest to word 1-grams also needs a null model to rule out that it is an artifact of n-gram chunking or deduplication.

major comments (3)
  1. [IV-B, Figs. 7–9] The correlational claim is confounded by codec capacity. The paper itself observes that large-dimension (nd) configurations simultaneously show lower α, higher β, lower bit-reduction rate, and better WER/CER/UTMOS (Fig. 7–9). Since the 15 configurations are not independent—e.g., E1–E5 are variants of the same codec trained on related data, F1–F4 all have nd=32—the reported Pearson r values (r=0.21 to -0.78) may simply reflect bitrate/model capacity rather than a direct link between linguistic regularity and quality. No p-values or confidence intervals are given, and no partial correlations or matched comparisons are run. To support the central claim, the authors should report correlations conditional on nd, bitrate, and codec family, or compare configurations matched on nd (e.g., F1–F4, or within-family comparisons). This is a load-bearing issue because the abstract's claim that 'these p
  2. [IV-A, Figs. 2–5] The claim that 3-grams are the most language-like is not tested against a null model. The comparison is between token n-grams and word 1-grams, so the conclusion that 3-gram is special could be an artifact of the n-gram chunking process: as n changes, vocabulary size, frequency distributions, and Heaps parameters change mechanically. The paper should include a null baseline (e.g., shuffled token sequences, randomly grouped tokens, or word n-gram baselines at matched n) to show that the closeness of 3-grams to the word 1-gram reference is meaningful. The dz values in the figure legends are also descriptive only; no significance or confidence intervals are attached, and with only 15 configurations per group, the means are not characterized with error bars.
  3. [III and IV-A (Zipf/Heaps fitting)] The power-law fits are not accompanied by uncertainty quantification or model comparison. The paper uses the powerlaw library, but it does not report the fitted xmin values, bootstrapped confidence intervals for α and β, or log-likelihood ratios against alternative distributions (e.g., log-normal or exponential). Such tests are standard in power-law analysis and are needed to support the statement that NAC tokens 'adhere to' Zipf's law. Additionally, the deduplication of consecutive identical tokens (Section III) directly changes frequency distributions and n-gram counts; its effect on the fitted Zipf/Heaps parameters should be quantified or at least justified, since the paper uses the deduplicated sequences for all statistical analyses.
minor comments (5)
  1. [References] Reference [17] ('Visual instruction tuning') does not match the sentence about NAC token inconsistency and robust instruction tuning; the intended citation appears to be missing or mis-cited.
  2. [IV-B.2] The text says 'Figure 4 shows scatter plots of β values against the benchmark performances,' but the relevant plots are in Figure 8; Figure 4 is the Heaps k–β plot. Please correct.
  3. [IV-A.1, Eq. (6)] The formula for the maximum-likelihood α is malformed: 'α = 1 + nPn i=1 log(xi/xmin)' is missing the division by the sum. Please fix the equation.
  4. [Footnote 3] The statement that exact values and graphs for 2-gram and 4-gram correlations are omitted due to page limitations weakens the robustness claim that 3-gram is uniquely or most strongly correlated with benchmarks. These results should be included in a supplementary appendix rather than only referenced.
  5. [Throughout] There are several grammatical errors and inconsistent captions (e.g., 'It is also showed that model with a configurations of large dimensions'). A careful proofread is needed.

Circularity Check

0 steps flagged

No significant circularity: the paper reports observed correlations between fitted token statistics and independent benchmarks; self-citations are not load-bearing.

full rationale

This paper is a descriptive/empirical study: it computes token-sequence statistics (Zipf α, Heaps β/k, Huffman bit-reduction rate) for 15 NAC configurations and reports simple Pearson correlations of those statistics with WER/CER/UTMOS. There is no derivation chain in which an output is fed back as an input, and no fitted parameter is renamed as a prediction: α, β, k, and R are estimated from token data, while the benchmark scores are measured independently on resynthesized audio. The central '3-gram is most language-like' result is a comparison against natural-language baselines, not a consequence of how the statistics were defined. Self-citations: [15] (Takamichi et al.) and [36] (Chan et al.) appear in related-work/methods contexts; [36] motivates the choice of the powerlaw library, but the fitting itself is performed by the independent Clauset et al. [37] implementation and does not assume the paper's conclusion. The uncontrolled nd/bitrate confound (Section IV-B, Figs. 7–9) is a validity threat to any causal reading of the correlations, but it is not circular: the correlation coefficients are honest summaries of the 15 points, and 'larger codecs do better' is a competing explanation, not an identity between input and output. Footnote 3's omitted 2-gram/4-gram graphs are a completeness limitation, not circularity. Therefore the paper receives a circularity score of 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper's contribution is empirical; it introduces no new theoretical entities. Its results rest on several domain assumptions about token preprocessing and comparability to word statistics, and on fitting four statistical parameters that are then used as independent variables in correlations. The dimension start/end tokens are artificial preprocessing markers, not invented entities with independent evidence.

free parameters (4)
  • Zipf exponent alpha (MLE) = varies by model and n-gram (approx 2.0-3.2 for 3-grams)
    Fit to rank-frequency distributions using the powerlaw library; used as the key independent variable in Figure 7 correlations.
  • Zipf xmin threshold = not reported
    Lower bound of the power-law regime, chosen by the powerlaw library; directly affects the fitted alpha value.
  • Heaps scaling factor k = varies, e.g., 8.11 for mean English 3-gram tokens
    Fitted to vocabulary growth curves; compared with word k in Figure 4.
  • Heaps growth exponent beta = varies, e.g., 0.81 for mean English 3-gram tokens
    Fitted to vocabulary growth curves; used in Figure 8 correlations with benchmarks.
axioms (4)
  • domain assumption Preprocessing (deduplication, dimension offsetting, dimension start/end tokens) preserves the statistical properties relevant to linguistic laws.
    Section III states the procedure mitigates artificial effects across codebooks but does not validate that the resulting flattened sequence is a faithful token stream.
  • domain assumption Word n-grams from manual transcriptions are an appropriate reference for comparing NAC token statistics.
    Token and word vocabularies, segmentation, and lengths differ; similarity in fitted power-law parameters may not indicate genuine linguistic similarity.
  • standard math Power-law (Zipf) and sublinear (Heaps) models are appropriate descriptions; no alternative distributions are tested.
    Section IV-A fits with the powerlaw library and does not compare against log-normal or other heavy-tailed null models.
  • domain assumption The 15 codec configurations are treated as independent samples for correlation analysis.
    Configurations E1-E5 are EnCodec variants and F1-F4 are FunCodec variants sharing architecture and training data, which can inflate correlation significance.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Analysing the Language of Neural Audio Codecs." pith.science (2026). https://pith.science/paper/6UAEGV7Y

@misc{pith2026250901390,
  author       = {Pith},
  title        = {Pith review of: Analysing the Language of Neural Audio Codecs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UAEGV7Y}},
  note         = {Machine review of arXiv:2509.01390}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This study presents a comparative analysis of the statistical and linguistic properties of neural audio codecs (NACs). We investigate discrete speech tokens produced by various NAC models, examining their adherence to linguistic statistical laws such as Zipf's law and Heaps' law, as well as their entropy and redundancy. To assess how these token-level properties relate to semantic and acoustic preservation in synthesized speech, we evaluate intelligibility using error rates of automatic speech recognition, and quality using the UTMOS score. Our results reveal that NAC tokens, particularly 3-grams, exhibit language-like statistical patterns. Moreover, these properties, together with measures of information content, are found to correlate with improved performances in speech recognition and resynthesis tasks. These findings offer insights into the structure of NAC token sequences and inform the design of more effective generative speech models.

Figures

Figures reproduced from arXiv: 2509.01390 by David M. Chan, Hiroshi Saruwatari, Joonyong Park, Shinnosuke Takamichi, Shunsuke Kando, Yuki Saito.

Figure 1
Figure 1. Figure 1: Analysis of NAC tokens conducted in our paper. An NAC model [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relationship between Heap’s law components [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of n-gram token vocabulary growth of the mean NAC token sequences with natural language reference (word 1-gram) for (a) English and (b) Chinese. values tended to exhibit sublinear growth, indicating repeated reuse of common tokens and reduced diversity, as shown in [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of token bit reduction rates across different [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Relationship between the 3-gram Zipf’s law exponent ( [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Relationship between the 3-gram token bit reduction rate and (a) [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 36 canonical work pages

  1. [1]

    Recent advances in discrete speech tokens: A review,

    Y . Guo, Z. Li, H. Wang, B. Li, C. Shao, H. Zhang, C. Du, X. Chen, S. Liu, and K. Yu, “Recent advances in discrete speech tokens: A review,” arXiv, vol. arXiv:2502.06490, 2025

    arXiv 2025

  2. [2]

    AudioLM: A language modeling approach to audio generation,

    Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Shar- ifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasacchi, and N. Zeghi- dour, “AudioLM: A language modeling approach to audio generation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 2523–2533, 2023

    work page 2023

  3. [3]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 29, pp. 3451–3460, 2021

    work page 2021

  4. [4]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Proc. NeurIPS, Vancouver, Canada, Dec. 2020

    work page 2020

  5. [5]

    Self-supervised speech representation learning: A review,

    A. Mohamed, H.-y. Lee, L. Borgholt, J. D. Havtorn, J. Edin, C. Igel, K. Kirchhoff, S.-W. Li, K. Livescu, L. Maaløe, T. N. Sainath, and S. Watanabe, “Self-supervised speech representation learning: A review,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1179–1210, 2022

    work page 2022

  6. [6]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,” Transactions on Machine Learning Research, 2023

    work page 2023

  7. [7]

    SoundStream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “SoundStream: An end-to-end neural audio codec,” IEEE/ACM Transac- tions on Audio, Speech, and Language Processing , vol. 30, pp. 495–507, 2021

    work page 2021

  8. [8]

    Codec-ASR: Training performant automatic speech recog- nition systems with discrete speech representations,

    K. Dhawan, N. R. Koluguri, A. Juki ´c, R. Langman, J. Balam, and B. Ginsburg, “Codec-ASR: Training performant automatic speech recog- nition systems with discrete speech representations,” in Proc. INTER- SPEECH, Kos, Greece, Sep. 2024, pp. 2574–2578

    work page 2024

  9. [9]

    Towards audio codec-based speech separation,

    J. Q. Yip, S. Zhao, D. Ng, E. S. Chng, and B. Ma, “Towards audio codec-based speech separation,” in Proc. INTERSPEECH, Kos, Greece, Sep. 2024, pp. 2190–2194

    work page 2024

  10. [10]

    G. K. Zipf, Human behavior and the principle of least effort. Addison- Wesley Press, 1949

    work page 1949

  11. [11]

    Contribution `a la th ´eorie math ´ematique des jeux de communication,

    B. Mandelbrot, “Contribution `a la th ´eorie math ´ematique des jeux de communication,” in Annales de l’ISUP , vol. 2, 1953, pp. 3–124

    work page 1953

  12. [12]

    Zipf and heaps laws’ coefficients depend on language,

    A. Gelbukh and G. Sidorov, “Zipf and heaps laws’ coefficients depend on language,” in Proc. CICLing, Mexico City, Mexico, 2001, pp. 332– 335

    work page 2001

  13. [13]

    H. S. Heaps, Information retrieval: Computational and theoretical aspects. Academic Press, Inc., 1978

    work page 1978

  14. [14]

    Prediction and entropy of printed english,

    C. E. Shannon, “Prediction and entropy of printed english,” Bell system technical journal , vol. 30, no. 1, pp. 50–64, 1951

    work page 1951

  15. [15]

    Do learned speech symbols follow Zipf’s law?

    S. Takamichi, H. Maeda, J. Park, D. Saito, and H. Saruwatari, “Do learned speech symbols follow Zipf’s law?” in Proc. ICASSP , Seoul, South Korea, Apr. 2024, pp. 12 526–12 530

    work page 2024

  16. [16]

    Analysing discrete self supervised speech representation for spoken language modeling,

    A. Sicherman and Y . Adi, “Analysing discrete self supervised speech representation for spoken language modeling,” inProc. ICASSP, Rhodes, Greece, Jun. 2023

    work page 2023

  17. [17]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” in Proc. NeurIPS, Vancouver, Canada, Dec. 2024

    work page 2024

  18. [18]

    SpeechTokenizer: Unified speech tokenizer for speech large language models,

    X. Zhang, D. Zhang, S. Li, Y . Zhou, and X. Qiu, “SpeechTokenizer: Unified speech tokenizer for speech large language models,” in Proc. ICLR, Vienna, Austria, 2024

    work page 2024

  19. [19]

    HiFi-Codec: Group-residual vector quantization for high fidelity audio codec,

    D. Yang, S. Liu, R. Huang, J. Tian, C. Weng, and Y . Zou, “HiFi-Codec: Group-residual vector quantization for high fidelity audio codec,” arXiv preprint arXiv:2305.02765, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Audiodec: An open-source streaming high-fidelity neural audio codec,

    Y .-C. Wu, I. D. Gebru, D. Markovi ´c, and A. Richard, “Audiodec: An open-source streaming high-fidelity neural audio codec,” in Proc. ICASSP, Jun. 2023

    work page 2023

  21. [21]

    High- fidelity audio compression with improved RVQGAN,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High- fidelity audio compression with improved RVQGAN,” in Proc. NeurIPS, New Orleans, U.S.A., Dec. 2023

    work page 2023

  22. [22]

    FunCodec: A fundamental, reproducible and integrable open-source toolkit for neural speech codec,

    Z. Du, S. Zhang, K. Hu, and S. Zheng, “FunCodec: A fundamental, reproducible and integrable open-source toolkit for neural speech codec,” in Proc. ICASSP, Seoul, South Korea, Apr. 2024

    work page 2024

  23. [23]

    Towards audio language modeling – an overview,

    H. Wu, X. Chen, Y .-C. Lin, K. wei Chang, H.-L. Chung, A. H. Liu, and H. yi Lee, “Towards audio language modeling – an overview,” arXiv preprint 2402.13236, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Librispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An ASR corpus based on public domain audio books,” in Proc. ICASSP , South Brisbane, Australia, Apr. 2015, pp. 5206–5210

    work page 2015

  25. [25]

    LibriTTS: A corpus derived from LibriSpeech for text-to- speech,

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “LibriTTS: A corpus derived from LibriSpeech for text-to- speech,” in Proc. INTERSPEECH, Graz, Austria, Sep. 2019, pp. 1526– 1530

    work page 2019

  26. [26]

    CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit,

    C. Veaux, J. Yamagishi, and K. MacDonald, “CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit,” 2016

    work page 2016

  27. [27]

    AISHELL-1: An open- source Mandarin speech corpus and a speech recognition baseline,

    H. Bu, J. Du, X. Na, B. Wu, and H. Zheng, “AISHELL-1: An open- source Mandarin speech corpus and a speech recognition baseline,” in Proc. O-COCOSDA, Shanghai, China, Oct. 2017

    work page 2017

  28. [28]

    Noisy speech database for training speech en- hancement algorithms and TTS models, 2016 [sound],

    C. Valentini-Botinhao, “Noisy speech database for training speech en- hancement algorithms and TTS models, 2016 [sound],” 2017, university of Edinburgh. School of Informatics. Centre for Speech Technology Research (CSTR)

    work page 2016

  29. [29]

    Common V oice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common V oice: A massively-multilingual speech corpus,” in Proc. LREC , Marseille, France, May 2020, pp. 4218–4222

    work page 2020

  30. [30]

    ICASSP 2022 Deep Noise Suppression Challenge,

    H. Dubey, V . Gopal, R. Cutler, S. Matusevych, S. Braun, E. S. Eskimez, M. Thakker, T. Yoshioka, H. Gamper, and R. Aichner, “ICASSP 2022 Deep Noise Suppression Challenge,” in Proc. ICASSP, Singapore, May 2022

    work page 2022

  31. [31]

    The MTG-Jamendo dataset for automatic music tagging,

    D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra, “The MTG-Jamendo dataset for automatic music tagging,” in Proc. ICML , Long Beach, U.S.A., Jul. 2019

    work page 2019

  32. [32]

    Audio Set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio Set: An ontology and human-labeled dataset for audio events,” in Proc. ICASSP, New Orleans, U.S.A., Mar. 2017, pp. 776–780

    work page 2017

  33. [33]

    FSD50K: an open dataset of human-labeled sound events,

    E. Fonseca, X. Favory, J. Pons, F. Font, and X. Serra, “FSD50K: an open dataset of human-labeled sound events,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 829–852, 2021

    work page 2021

  34. [34]

    Codec-SUPERB: An in- depth analysis of sound codec models,

    H. Wu, H.-L. Chung, Y .-C. Lin, Y .-K. Wu, X. Chen, Y .-C. Pai, H.-H. Wang, K.-W. Chang, A. Liu, and H.-y. Lee, “Codec-SUPERB: An in- depth analysis of sound codec models,” in Findings of ACL , Bangkok, Thailand, Aug. 2024, pp. 10 330–10 348

    work page 2024

  35. [35]

    CSS10: A collection of single speaker speech datasets for 10 languages,

    K. Park and T. Mulc, “CSS10: A collection of single speaker speech datasets for 10 languages,” in Proc. INTERSPEECH, Graz, Austria, Sep. 2019, pp. 1566–1570

    work page 2019

  36. [36]

    Ana- lyzing the language of visual tokens,

    D. M. Chan, R. Corona, J. Park, C. J. Cho, Y . Bai, and T. Darrell, “Ana- lyzing the language of visual tokens,” arXiv preprint arXiv:2411.05001 , 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Power-law distri- butions in empirical data,

    A. Clauset, C. R. Shalizi, and M. E. J. Newman, “Power-law distri- butions in empirical data,” SIAM Review , vol. 51, no. 4, pp. 661–703, 2009

    work page 2009

  38. [38]

    The kolmogorov-smirnov test for goodness of fit,

    F. J. Massey, “The kolmogorov-smirnov test for goodness of fit,” Journal of the American Statistical Association , vol. 46, no. 253, pp. 68–78, 1951

    work page 1951

  39. [39]

    Robust speech recognition via large-scale weak supervision

    A. Radford, K. Jong Wook, X. Tao, B. Greg, M. Christine, and S. Ilya, “Robust speech recognition via large-scale weak supervision.” https:// cdn.openai.com/papers/whisper.pdf, 2022

    work page 2022

  40. [40]

    UTMOS: UTokyo-SaruLab System for V oiceMOS Chal- lenge 2022,

    T. Saeki, D. Xin, W. Nakata, T. Koriyama, S. Takamichi, and H. Saruwatari, “UTMOS: UTokyo-SaruLab System for V oiceMOS Chal- lenge 2022,” in Proc. INTERSPEECH, Incheon, South Korea, Sep. 2022, pp. 4521–4525

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.