Pith. sign in

REVIEW 6 major objections 5 minor 49 references

Probing the Robustness Properties of Neural Speech Codecs

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Neural speech codecs split under noise: DAC leads at high bitrate, EnCodec at 3 kbps, and measured nonlinearity explains part of the gap.

desk verdict A genuinely new empirical comparison of neural codec noise robustness, with a credible ranking and a speculative linearity explanation that needs matched-bitrate testing before it can carry causal weight. read the letter →

arxiv 2505.24248 v1 pith:5DUWU6HV submitted 2025-05-30 eess.AS cs.SD

Wei-Cheng Tseng , David Harwath This is my paper
classification eess.AScs.SD
keywords speechcodingneuralcodecsnoiserobustnesslinearityanalysisfrequencyresponseresidualvectorquantizationtokenizationdownstreamtasks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Neural speech codecs are now used not only for compression but as tokenizers for speech language models, so their behavior in noisy conditions matters for real deployments. This paper claims that codecs differ systematically in noise robustness: at the highest bitrate DAC degrades least across ambient noise, white noise, and reverberation, while at 3 kbps EnCodec outperforms DAC. The differences are partly explained by measured nonlinearity, since codecs that better satisfy additivity and homogeneity introduce fewer artifacts on overlapped and scaled signals, and by frequency-response coloration from time-domain training losses. The upshot is that noise robustness is not a byproduct of clean-signal quality and should be treated as a separate design target.

What carries the argument

The argument is carried by treating each codec as a black-box input-output system and measuring three properties on the same set of codecs. First, a noise-robustness battery mixes clean speech with WHAM ambient noise, generated white noise, and DNS reverberation and scores the reconstructed audio with Whisper-Large for word error rate, ECAPA-TDNN for speaker verification equal error rate, and emotion2vec for emotion recognition accuracy. Second, a linearity probe checks additivity and homogeneity by comparing Mel distance between the codec output of mixtures or scaled inputs and the corresponding combination of individual outputs. Third, a frequency-response probe feeds sine sweeps through each codec and records spectral boosting or attenuation. The correlation between the linearity measurements and the noise-robustness ranking is the mechanism that explains why some codecs fail more gracefully in noise.

What would settle it

A single control experiment, feeding the noisy unencoded speech directly to Whisper-Large, ECAPA-TDNN, and emotion2vec and checking whether their errors stay flat across noise types and SNRs, would settle whether the codec rankings reflect codec behavior or probe sensitivity.

Watch

Extended reading notes

Core claim

The central claim is that neural speech codecs cannot be ranked for real-world use by clean-signal quality alone. Evaluating five codecs under ambient noise, white noise, and reverberation at multiple signal-to-noise ratios, the paper finds that DAC is the most robust codec at its highest bitrate across ASR word error, speaker verification equal error, and emotion recognition accuracy, with EnCodec and FreqCodec competitive; at the matching low bitrate of 3 kbps, EnCodec overtakes DAC. The paper further claims that end-to-end codec linearity, measured by additivity $f(X+Y)=f(X)+f(Y)$ and homogeneity $f(\alpha X)=\alpha f(X)$, tracks robustness: more additive codecs such as DAC distort overlapping speech less, while less homogeneous codecs such as HiFi-Codec distort spectral dynamics. Frequency response measured with sine sweeps shows all codecs boost low frequencies and attenuate high frequencies, with EnCodec's boost most pronounced and HiFi-Codec and SpeechTokenizer showing abrupt high-frequency roll-offs, linking time-domain losses to spectral coloration.

Load-bearing premise

The rankings depend on the three probe models measuring content, speaker identity, and emotion without being themselves degraded by noise; if those probes are noise-sensitive, the reported codec differences could be probe artifacts.

Editorial extensions

If this is right

  • Noise robustness should be reported as a standard axis in future codec evaluations, since high clean-speech quality does not guarantee robustness.
  • The measured linearity properties suggest additivity and homogeneity could be used as explicit training objectives to improve behavior on overlapped speech.
  • Low-bitrate codecs appear to need noise-aware design rather than simply more bits, because EnCodec at 3 kbps beats higher-capacity DAC at the same bitrate.
  • Frequency-response results tie time-domain losses to low-frequency boosting and high-frequency attenuation, so perceptually weighted spectral losses may reduce coloration.
  • Downstream systems built on codec tokens inherit codec-specific noise distortions, so robustness should be assessed at the token level as well as the waveform level.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The ranking might shift if different probe models were used, because the paper does not validate that Whisper-Large, ECAPA-TDNN, and emotion2vec are insensitive to noise on their own.
  • The linearity explanation is based on synthetic sine-sweep and gain tests; extending the same additivity test to real speech-overlap mixtures would test whether the correlation holds outside the lab stimuli.
  • Codec robustness could be improved by augmenting training with noisy speech and by regularizing the end-to-end mapping toward additivity, a testable extension the paper points to but does not run.
  • The oracle baseline itself degrades under heavy noise, so part of the measured codec degradation may actually be a probe-plus-noise interaction; a noise-invariant reference would separate these contributions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 5 minor

Summary. This paper empirically evaluates five neural speech codecs (EnCodec, DAC, SpeechTokenizer, HiFi-Codec, FreqCodec) under ambient noise, white noise, and reverberation at several SNR levels, using PESQ, mel distance, ASR word error rate with Whisper-Large, speaker verification equal error rate with ECAPA-TDNN, and speech emotion recognition accuracy with emotion2vec. The authors report that at the highest available bitrate DAC is the most noise-robust codec, while at 3 kbps EnCodec outperforms DAC, and they interpret this as evidence that noise robustness is not simply inherited from clean-signal fidelity. The paper then measures additivity and homogeneity of the codecs' input-output mappings and characterizes their frequency responses with sine sweeps, arguing that measured non-linearity and spectral shaping partly explain the observed robustness differences.

Significance. If the reported rankings are reliable, the paper provides a useful practical comparison of neural codecs in noisy conditions and introduces two low-cost diagnostic probes (additivity/homogeneity and frequency response) that could inform future codec design. A clear strength is that the evaluation is fully external: it relies on independent noise corpora (WHAM, DNS), external downstream metrics, and an Oracle topline, with no fitted model parameters or circular derivation. However, the absence of error bars and significance tests, the lack of validation that the probe models are noise-invariant rulers, and the untested linearity explanation currently limit the strength of the central claims. The paper is best read as a hypothesis-generating empirical study whose ranking and explanatory conclusions need further statistical grounding.

major comments (6)
  1. [Section 4, Figure 1] The central ranking claim that 'DAC consistently demonstrates the strongest noise robustness' is based on single curves with no error bars, confidence intervals, or significance tests. With only five codecs and, apparently, one noisy-signal realization per condition, differences that appear visually as non-trivial could be within measurement noise. Please report statistics across test utterances and noise draws (for example, bootstrap confidence intervals) and test pairwise codec differences at each SNR and metric.
  2. [Section 3.3] The downstream metrics assume Whisper-Large, ECAPA-TDNN, and emotion2vec are stable rulers that degrade only in response to codec-induced distortion, but the paper provides no validation of that assumption. Because the Oracle topline shows the probes' own performance on noisy unencoded speech, the codec-induced WER/EER/SER-ACC gaps should be reported relative to that topline, or the probes should be explicitly shown to be noise-invariant at the tested SNRs; otherwise the reported codec rankings could reflect probe sensitivity rather than codec behavior.
  3. [Section 5] The explanatory claim that better additivity or homogeneity causes better noise robustness is supported only by visual alignment between Figure 1 and Figure 2. The paper's own 3 kbps result provides the discriminating test: the linearity account predicts that EnCodec should be more additive and homogeneous than DAC at 3 kbps, yet this comparison is never reported. Without a bitrate-matched additivity comparison and, ideally, a quantitative correlation, the robustness-linearity link remains a post hoc story rather than an established explanation.
  4. [Section 5, Figure 2] The additivity and homogeneity metrics are not defined precisely. The y-axis is labeled 'Mel Distance,' but there is no equation stating whether the plotted quantity is, for example, Mel(f(X+Y), f(X)+f(Y)), how time alignment is handled, or how the result is averaged and normalized across test signals. Please specify the exact computation, the number and type of input signals, and whether the same bitrate conditions as in Figure 1 are used; otherwise the curves in Figure 2 are not reproducible.
  5. [Section 6, Figure 3] The frequency-response analysis is qualitative. The text asserts that low-frequency boosting is 'likely due to time-domain L1 loss' and that HiFi-Codec and SpeechTokenizer show 'significant fluctuations,' but no numerical metric (for example, band-specific gain or spectral deviation) or error bars are reported, and the sine-sweep amplitude validation is only mentioned without results. Please quantify the response curves and connect them to the robustness results with a defined measure.
  6. [Section 3.2, Table 1] The evaluated bitrates are not specified precisely. Table 1 lists ranges (for example, EnCodec 0.75–24 kbps), but Section 4 says codecs operate at 'highest available bitrate' without stating the exact bitrate, checkpoint version, or sampling rate for each model. Given the paper's own finding that robustness is highly bitrate-sensitive, exact operating points must be reported for reproducibility and for the 3 kbps comparison to be interpretable.
minor comments (5)
  1. [Section 3.1] The dataset name is typeset with an unintended space as 'RA VDESS' and should be 'RAVDESS'.
  2. [Throughout] The model name is inconsistent: 'EnCodec' and 'Encodec' are both used, and the reference [10] is 'Funcodec' while the text calls it 'FreqCodec.' Please standardize the notation and clarify the relationship.
  3. [Section 2.2, Eq. (1)] In the definition of bitrate, the symbol N is used for both the number of frames and in the expression f_N = N·f_s/T; please clarify the notation and state the units of each quantity.
  4. [Figure 1] Each subplot repeats the same legend entries; consider simplifying the legend to one panel and adding explicit bitrate annotations (for example, 'EnCodec@24kbps') to each curve.
  5. [References] There are minor typos in the reference list: '[15] Y o. Guo' and '[21] V oicecraft' should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the robustness evaluation is an empirical comparison against external metrics and datasets, and the linearity/frequency analyses are independent probes rather than fitted inputs.

full rationale

The paper's central claims are derived from direct empirical measurements: clean speech is mixed with external noise corpora (WHAM, DNS, Gaussian white noise), passed through publicly available codecs, and scored with external metrics (PESQ, Whisper-Large WER, ECAPA-TDNN EER, emotion2vec accuracy). No parameter is fitted to the target comparisons, and no claimed 'prediction' is constructed from the data it is said to explain. The linearity analysis in Section 5 measures additivity and homogeneity as separate input-output properties, and the stated connection to noise robustness is presented as an observed alignment ('These trends, in fact, well align with our noise robustness observations'), not as a derivation that reduces to the same fitted values. The frequency response analysis uses sine sweeps and is again an independent probe. Citations to prior codec papers are standard references to the evaluated systems, and the only self-citation (VoiceCraft, [21]) is a related-work mention that is not load-bearing for the robustness or linearity arguments. The inclusion of an Oracle topline partly controls for probe sensitivity, and the 3 kbps comparison (EnCodec vs DAC) shows the authors are not forcing a single monotonic explanation. Thus the evaluation is self-contained against external benchmarks and no self-definitional, fitted-input, or self-citation-circular step is present.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the validity of the chosen metrics and noise models, and on the assumption that the measurement instruments do not bias the ranking. No free parameters are fitted and no new entities are introduced.

assumptions (4)
  • domain assumption The objective metrics PESQ, ASR-WER, ASV-EER, and SER-ACC are accepted proxies for perceptual quality and information preservation.
    Section 3.3 relies on these metrics to rank codecs; if any probe model (Whisper-Large, ECAPA-TDNN, emotion2vec) is itself noise-sensitive, the measured 'codec robustness' would be confounded.
  • domain assumption The noise conditions (WHAM ambient, white noise, DNS reverberation) approximate real-world degradation of speech.
    Section 4 uses these three noise types to draw general conclusions about real-world robustness; they do not cover all possible distortions such as clipping, codec tandeming, or babble noise.
  • domain assumption Operating each codec at its highest available bitrate is a fair basis for cross-codec comparison.
    Section 4 states 'Each codec is operated at its highest available bitrate unless otherwise specified'; codecs have different maximum bitrates and architecture-specific rate-distortion trade-offs.
  • domain assumption The mel-distance-based measurements in Section 5 capture practical linearity (additivity and homogeneity) of the end-to-end codec mapping.
    Section 5 defines ideal linearity with f(X+Y)=f(X)+f(Y), but Figure 2 plots an unspecified mel-distance error; this assumes mel distance is a valid deviation metric for linearity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probing the Robustness Properties of Neural Speech Codecs." pith.science (2026). https://pith.science/paper/5DUWU6HV

@misc{pith2026250524248,
  author       = {Pith},
  title        = {Pith review of: Probing the Robustness Properties of Neural Speech Codecs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DUWU6HV}},
  note         = {Machine review of arXiv:2505.24248}
}
read the original abstract

Neural speech codecs have revolutionized speech coding, achieving higher compression while preserving audio fidelity. Beyond compression, they have emerged as tokenization strategies, enabling language modeling on speech and driving paradigm shifts across various speech processing tasks. Despite these advancements, their robustness in noisy environments remains underexplored, raising concerns about their generalization to real-world scenarios. In this work, we systematically evaluate neural speech codecs under various noise conditions, revealing non-trivial differences in their robustness. We further examine their linearity properties, uncovering non-linear distortions which partly explain observed variations in robustness. Lastly, we analyze their frequency response to identify factors affecting audio fidelity. Our findings provide critical insights into codec behavior and future codec design, as well as emphasizing the importance of noise robustness for their real-world integration.

Figures

Figures reproduced from arXiv: 2505.24248 by the authors.

Figure 1
Figure 1. Evaluation results of neural audio codecs under different noise conditions. Each row corresponds to a specific noise type (from top to bottom): ambient noise, white noise, and reverberation . 4. Noise Robustness Evaluation Neural audio codecs have proven useful not only for compres￾sion but also as tokenizers for language modeling on speech signals. However, their deployment in real-world environ￾ments—where backgro… view at source ↗
Figure 3
Figure 3. Frequency response of neural speech codecs. 6. Frequency Response Characterization Speech datasets vary in bandwidth due to differences in record￾ing conditions and applications. While compression inevitably introduces distortions, a robust neural codec should maintain spectral integrity. Evaluating frequency response provides in￾sight into codec behavior and informs potential avenues for im￾provement. In our analys… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 22 canonical work pages

  1. [1]

    Introduction Compression of speech signals, formally known as speech cod- ing, has been an active research area for decades. Early ef- forts focused on minimizing bitrates to reduce storage require- ments while preserving intelligibility, playing a crucial role in digital communication systems [1, 2, 3, 4]. Recently, neural speech codecs [5, 6, 7, 8, 9, 1...

    work page Pith review arXiv 2025

  2. [2]

    Preliminaries 2.1. Formulation of Compression and Reconstruction In this work, we consider neural speech codecs composed of three main components: an encoderE(·), a RVQ module R VQ(·), and a DecoderD(·). LetX∈R T be an audio waveform sampled at ratef s. The encoder first mapsXinto a sequence of latent representationsZ=E(X)∈R N×d , whereNis the number of f...

    work page

  3. [3]

    3.1), the evaluated neu- ral speech codecs(sec

    Experimental Setups Here, we describe the dataset used(sec. 3.1), the evaluated neu- ral speech codecs(sec. 3.2), and the evaluation metrics(sec. 3.3). 3.1. Datasets In this work, we aim to systematically analyze the noise robust- ness and behavior of neural speech codecs. To ensure that the observed results are attributable to the codec models themselves...

    work page

  4. [4]

    However, their deployment in real-world environ- ments—where background noise is common—raises concerns about their robustness

    Noise Robustness Evaluation Neural audio codecs have proven useful not only for compres- sion but also as tokenizers for language modeling on speech signals. However, their deployment in real-world environ- ments—where background noise is common—raises concerns about their robustness. Understanding how these models per- form under noisy conditions is esse...

    work page

  5. [5]

    Linearity Analysis An ideal speech codec for real-world applications should be- have approximately linear to ensure signal integrity when speech overlaps with other sources or when amplitude varies. Although the underlying neural networks necessarily employ non-linear components, we treat each codec as a black-box sys- tem and empirically ask whether its ...

    work page

  6. [6]

    While compression inevitably introduces distortions, a robust neural codec should maintain spectral integrity

    Frequency Response Characterization Speech datasets vary in bandwidth due to differences in record- ing conditions and applications. While compression inevitably introduces distortions, a robust neural codec should maintain spectral integrity. Evaluating frequency response provides in- sight into codec behavior and informs potential avenues for im- provem...

    work page

  7. [7]

    Our findings reveal sig- nificant variations in their noise robustness, influenced by fac- tors such as training data diversity, operating bitrate, and quanti- zation strategies

    Conclusions In this work, we present a systematic evaluation of neural speech codecs under noisy conditions. Our findings reveal sig- nificant variations in their noise robustness, influenced by fac- tors such as training data diversity, operating bitrate, and quanti- zation strategies. Additionally, our linearity analysis highlights the presence of non-l...

    work page

  8. [8]

    7 khz audio coding within 64 kbit/s,

    X. Maitre, “7 khz audio coding within 64 kbit/s,”IEEE Journal on Selected Areas in Communications, vol. 6, no. 2, pp. 283–298, 1988

    work page 1988

Show all 49 references
  1. [9]

    Mpeg-4 low delay audio coding based on the aac codec,

    E. Allamanche, R. Geiger, J. Herre, and T. Sporer, “Mpeg-4 low delay audio coding based on the aac codec,” inAudio Engineering Society Convention 106. Audio Engineering Society, 1999

    1999

  2. [10]

    High- quality, low-delay music coding in the opus codec,

    J.-M. Valin, G. Maxwell, T. B. Terriberry, and K. V os, “High- quality, low-delay music coding in the opus codec,”arXiv preprint arXiv:1602.04845, 2016

    2016 arXiv

  3. [11]

    Review of methods for coding of speech sig- nals,

    D. O’Shaughnessy, “Review of methods for coding of speech sig- nals,”EURASIP Journal on Audio, Speech, and Music Processing, vol. 2023, no. 1, p. 8, 2023

    2023

  4. [12]

    Soundstream: An end-to-end neural au- dio codec,

    N. Zeghidouret al., “Soundstream: An end-to-end neural au- dio codec,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 30, pp. 495–507, 2021

    2021

  5. [13]

    High fidelity neural audio compression,

    A. D ´efossezet al., “High fidelity neural audio compression,” arXiv preprint arXiv:2210.13438, 2022

    2022 arXiv

  6. [14]

    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,”Ad- vances in Neural Information Processing Systems, vol. 36, 2024

    2024

  7. [15]

    Speechtokenizer: Unified speech tokenizer for speech language models,

    X. Zhanget al., “Speechtokenizer: Unified speech tokenizer for speech language models,” 2023

    2023

  8. [16]

    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

    2023 arXiv

  9. [17]

    Funcodec: A funda- mental, reproducible and integrable open-source toolkit for neural speech codec,

    Z. Du, S. Zhang, K. Hu, and S. Zheng, “Funcodec: A funda- mental, reproducible and integrable open-source toolkit for neural speech codec,” inICASSP 2024-2024 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 591–595

    2024

  10. [18]

    Apcodec: A neural audio codec with parallel amplitude and phase spectrum encoding and decoding,

    Y . Ai, X.-H. Jiang, Y .-X. Lu, H.-P. Du, and Z.-H. Ling, “Apcodec: A neural audio codec with parallel amplitude and phase spectrum encoding and decoding,”arXiv preprint arXiv:2402.10533, 2024

    2024 arXiv

  11. [19]

    Scaling transformers for low-bitrate high-quality speech coding,

    J. D. Parker, A. Smirnov, J. Pons, C. Carr, Z. Zukowski, Z. Evans, and X. Liu, “Scaling transformers for low-bitrate high-quality speech coding,”arXiv preprint arXiv:2411.19842, 2024

    2024 arXiv

  12. [20]

    Semanticodec: An ultra low bitrate semantic audio codec for general sound,

    H. Liu, X. Xu, Y . Yuan, M. Wu, W. Wang, and M. D. Plumb- ley, “Semanticodec: An ultra low bitrate semantic audio codec for general sound,”arXiv preprint arXiv:2405.00233, 2024

    2024 arXiv

  13. [21]

    Autoregres- sive image generation using residual quantization,

    D. Lee, C. Kim, S. Kim, M. Cho, and W.-S. Han, “Autoregres- sive image generation using residual quantization,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 2022, pp. 11 523–11 532

    2022

  14. [22]

    Recent advances in discrete speech tokens: A review,

    Y . o. Guo, “Recent advances in discrete speech tokens: A review,” arXiv preprint arXiv:2502.06490, 2025

    2025

  15. [23]

    Sparks of large audio models: A survey and out- look,

    S. Latifet al., “Sparks of large audio models: A survey and out- look,”arXiv preprint arXiv:2308.12792, 2023

    2023 arXiv

  16. [24]

    A survey on speech large language models,

    J. Penget al., “A survey on speech large language models,”arXiv preprint arXiv:2410.18908, 2024

    2024

  17. [25]

    Neural codec language models are zero-shot text to speech synthesizers,

    C. Wanget al., “Neural codec language models are zero-shot text to speech synthesizers,”arXiv preprint arXiv:2301.02111, 2023

    2023 arXiv

  18. [26]

    Viola: Unified codec language models for speech recognition, synthesis, and translation,

    T. Wang, L. Zhou, Z.-H. Zhang, Y . Wu, S. Liu, Y . Gaur, Z. Chen, J. Li, and F. Wei, “Viola: Unified codec language models for speech recognition, synthesis, and translation,”arXiv preprint arXiv:2305.16107, 2023

    2023 arXiv

  19. [27]

    Speechx: Neural codec language model as a versatile speech transformer,

    X. Wanget al., “Speechx: Neural codec language model as a versatile speech transformer,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  20. [28]

    V oicecraft: Zero-shot speech editing and text-to-speech in the wild,

    P. Peng, P.-Y . Huang, D. Li, A. Mohamed, and D. Harwath, “V oicecraft: Zero-shot speech editing and text-to-speech in the wild,”arXiv preprint arXiv:2403.16973, 2024

    2024 arXiv

  21. [29]

    Moshi: a speech-text foundation model for real-time dialogue,

    A. D ´efossezet al., “Moshi: a speech-text foundation model for real-time dialogue,”arXiv preprint arXiv:2410.00037, 2024

    2024 arXiv

  22. [30]

    Anygpt: Unified multimodal llm with discrete se- quence modeling,

    J. Zhanet al., “Anygpt: Unified multimodal llm with discrete se- quence modeling,”arXiv preprint arXiv:2402.12226, 2024

    2024 arXiv

  23. [31]

    Discrete audio representation as an alter- native to mel-spectrograms for speaker and speech recognition,

    K. C. Puvvadaet al., “Discrete audio representation as an alter- native to mel-spectrograms for speaker and speech recognition,” inICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12 111–12 115

    2024

  24. [32]

    Codec-asr: Training performant automatic speech recognition systems with discrete speech representations,

    K. Dhawanet al., “Codec-asr: Training performant automatic speech recognition systems with discrete speech representations,” arXiv preprint arXiv:2407.03495, 2024

    2024 arXiv

  25. [33]

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

    H. Wu, H.-L. Chunget al., “Codec-superb: An in-depth analysis of sound codec models,”arXiv preprint arXiv:2402.13071, 2024

    2024 arXiv

  26. [34]

    Dasb–discrete audio and speech benchmark,

    P. Mousavi, L. Della Libera, J. Duret, A. Ploujnikov, C. Subakan, and M. Ravanelli, “Dasb–discrete audio and speech benchmark,” arXiv preprint arXiv:2406.14294, 2024

    2024 arXiv

  27. [35]

    Espnet-codec: Comprehensive training and evalua- tion of neural codecs for audio, music, and speech,

    J. Shiet al., “Espnet-codec: Comprehensive training and evalua- tion of neural codecs for audio, music, and speech,” inIEEE SLT

  28. [36]

    Perceptual evaluation of speech quality (pesq)- a new method for speech quality assessment of telephone net- works and codecs,

    A. W. Rixet al., “Perceptual evaluation of speech quality (pesq)- a new method for speech quality assessment of telephone net- works and codecs,” in2001 IEEE international conference on acoustics, speech, and signal processing. Proceedings (Cat. No. 01CH37221), vol. 2. IEEE, 2...

    2001

  29. [37]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  30. [38]

    The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english,

    S. R. Livingstone and F. A. Russo, “The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english,”PloS one, vol. 13, no. 5, p. e0196391, 2018

    2018

  31. [39]

    Seanet: A multi-modal speech enhancement network,

    M. Tagliasacchi, Y . Li, K. Misiunas, and D. Roblek, “Seanet: A multi-modal speech enhancement network,”arXiv preprint arXiv:2009.02095, 2020

    2009 arXiv

  32. [40]

    Long short-term memory,

    A. Graves and A. Graves, “Long short-term memory,”Supervised sequence labelling with recurrent neural networks, pp. 37–45, 2012

    2012

  33. [41]

    Wavlm: Large-scale self- supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chenet al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,”IEEE JSTSP, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  34. [42]

    Libritts: A corpus derived from librispeech for text-to-speech,

    H. Zenet al., “Libritts: A corpus derived from librispeech for text-to-speech,”arXiv preprint arXiv:1904.02882, 2019

    1904 arXiv

  35. [43]

    Mel-cepstral distance measure for objective speech quality assessment,

    R. Kubichek, “Mel-cepstral distance measure for objective speech quality assessment,” inProceedings of IEEE pacific rim confer- ence on communications computers and signal processing, vol. 1. IEEE, 1993, pp. 125–128

    1993

  36. [45]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  37. [46]

    Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,”arXiv preprint arXiv:2005.07143, 2020

    2005 arXiv

  38. [47]

    emotion2vec: Self-supervised pre-training for speech emotion representation,

    Z. Ma, Z. Zheng, J. Ye, J. Li, Z. Gao, S. Zhang, and X. Chen, “emotion2vec: Self-supervised pre-training for speech emotion representation,”arXiv preprint arXiv:2312.15185, 2023

    2023 arXiv

  39. [48]

    Wham!: Extend- ing speech separation to noisy environments,

    G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, D. Crow, E. Manilow, and J. L. Roux, “Wham!: Extend- ing speech separation to noisy environments,”arXiv preprint arXiv:1907.01160, 2019

    1907 arXiv

  40. [49]

    The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,

    C. K. Reddy, V . Gopal, R. Cutler, E. Beyrami, R. Cheng, H. Dubey, S. Matusevych, R. Aichner, A. Aazami, S. Braun et al., “The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” arXiv preprint arXiv:2005.13981, 2020

    2020 arXiv

  41. [2024]

    IEEE, 2024, pp. 562–569

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.