Pith. sign in

REVIEW 4 major objections 7 minor 43 references

Distillation and Pruning for Scalable Self-Supervised Representation-Based Speech Quality Assessment

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read This paper claims that distilling an XLS-R-based speech quality teacher into a 4.3-million-parameter student on unlabeled degraded speech halves the gap to the teacher while shrinking the model by two orders of magnitude, and that…

desk verdict Useful engineering comparison of distillation vs. pruning for speech quality assessment, but the 'halve the gap' headline rests on weighted averaging and an unexplained model-variant choice. read the letter →

arxiv 2502.05356 v1 pith:LQ5A4IYF submitted 2025-02-07 eess.AS cs.SD

classification eess.AScs.SD
keywords speechqualityassessmentmeanopinionscoreknowledgedistillationmodelpruningself-supervisedrepresentationsXLS-Rnon-intrusivepredictionpseudo-labels
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

This paper asks whether the quality-assessment power of a huge self-supervised speech model can be packed into a small model that runs in practice. It retrains an XLS-R-based teacher on more than 100,000 labeled speech clips, then distills it into convolutional-transformer students using pseudo-labels on unlabeled degraded speech. The central result is that a 4.3-million-parameter student reaches a weighted mean Pearson correlation of 0.76 with ground-truth MOS labels, compared with 0.72 for a same-size baseline trained on labels alone and 0.81 for the teacher. That closes roughly half the gap between baseline and teacher while cutting model size by two orders of magnitude. The paper also shows that data-driven pruning of the teacher is preferable when larger models are acceptable.

What carries the argument

The load-bearing device is teacher-student distillation through pseudo-labels: the frozen XLS-R(2B)-L10+T(BA) teacher scores a large unlabeled corpus of degraded speech (VoIP-style degradations, TTS output, codecs), and the student is trained to match those scores, with 20% labeled clips mixed in and per-dataset bias transforms applied to ground-truth labels. The student is a convolutional transformer fed with real and imaginary complex compressed spectrogram features, so it does not touch XLS-R features at inference. For the pruning arm, the mechanism is Taylor-importance pruning: weights are removed in order of estimated squared loss change, with periodic fine-tuning and recursive averaging of importance scores.

What would settle it

Take the released 4.3-million-parameter distilled model and run it on the Blizzard 2023 French text-to-speech test set. The paper reports a Pearson correlation of 0.22, essentially the labeled-only baseline; a result clearly above that would contradict the claimed boundary that distillation transfers only where the teacher already correlates. The complementary check is to measure the student on a held-out VoIP-type test set where teacher correlation is high and ask whether the student closes the reported half-gap.

Watch

Extended reading notes

Core claim

Trained only on labeled clips, small models plateau, but a student trained to imitate the XLS-R(2B)-L10+T(BA) teacher on unlabeled degraded and synthetic speech improves substantially, especially on VoIP-typical degradations. The best distilled student has 4.3 million parameters, about 1/100 the teacher's size, and improves weighted mean correlation from 0.72 to 0.76, closing about half of the 0.09 gap to the teacher's 0.81. A pruned teacher retaining 29% of its parameters reaches 0.80, showing that pruning preserves teacher-level accuracy at larger sizes. The method does not transfer to out-of-domain synthetic speech: on the Blizzard 2023 French TTS data the distilled model's correlation (0.22) is no better than the labeled-only baseline (0.24) and far below the teacher (0.58).

Load-bearing premise

The whole distillation pipeline assumes the teacher's pseudo-labels on unlabeled speech are accurate enough to learn from; the paper reports that where the teacher's own correlation is low, such as Blizzard 2023 French TTS, the distilled student inherits the failure and gains nothing over the labeled baseline.

Editorial extensions

If this is right

  • A 4.3-million-parameter model can run reference-free speech quality assessment at practical cost while still tracking human ratings on VoIP and enhancement data.
  • Distillation transfers the teacher's quality judgments to a compact architecture using unlabeled data, so new degradation types can be covered without collecting new MOS labels, provided the teacher itself rates them reliably.
  • Data-driven pruning preserves more correlation than naive magnitude pruning at the same budget, and at around 29% parameters the pruned teacher nearly matches the full teacher's weighted mean correlation.
  • Combining distillation for small sizes and pruning for larger sizes gives a scalable family of quality-assessment models spanning more than three orders of magnitude in parameter count.
  • The poor results on out-of-domain TTS data show that neither approach adds generalization the teacher does not already have.

Reading between the lines

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

  • A natural extension would be to filter unlabeled training samples by teacher confidence, so pseudo-labels from low-certainty regions are down-weighted or discarded; that could extend distillation to TTS domains where the teacher currently fails.
  • The per-dataset scale-and-shift calibration applied to the teacher's logits may itself carry transferable information about dataset bias, and a student trained to also predict those calibration parameters might generalize better across datasets.
  • The pruning and distillation results suggest a combined recipe: first prune the teacher to a moderate size, then distill the pruned teacher into a small student, which could yield small models that inherit improved robustness from the pruning fine-tuning.
  • Because the distilled student uses only spectrogram features, the same training recipe could be applied to other audio quality tasks, such as music or conference-room audio assessment, wherever a strong SSL teacher is available.
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

4 major / 7 minor

Summary. The paper addresses model-size reduction for non-intrusive speech quality assessment. It retrains an XLS-R-based teacher model (XLS-R(2B)-L10+T(BA)) on a large compilation of MOS-labeled datasets, then investigates two compression routes: (i) distillation of the teacher into compact convolutional-transformer student models using pseudo-labels on unlabeled degraded speech with a 20% ground-truth mix-in, and (ii) importance-based pruning of the XLS-R embedding model with iterative fine-tuning. The main reported result is that a 4.3M-parameter distilled student reaches a clip-count-weighted mean Pearson correlation of 0.76 versus 0.72 for the labeled-only baseline and 0.81 for the teacher, roughly halving the gap, while a pruned teacher at 29% of its parameters reaches 0.80. The paper also reports that the method fails on out-of-domain synthetic French TTS data (Blizzard 2023), where the teacher itself has low correlation.

Significance. If the central claims hold, the paper provides a practical route to a roughly 100x smaller speech quality model with a modest loss in accuracy, plus a pruning-based alternative for larger model sizes. The empirical study is unusually broad: it uses over 100k labeled clips for training, 46k clips across 16 test sets, and compares against DNSMOS, NISQA, TorchaudioSQUIM, UTMOS, and the original XLS-R-SQA. The authors also release model weights and describe the unlabeled degradation simulation recipe in detail, which are concrete reproducibility strengths. However, the headline quantitative claim rests on a single model variant chosen from ten candidates, on a single aggregate metric, with no uncertainty estimates; these issues need to be resolved before the claim can be taken at face value.

major comments (4)
  1. [Section IV-B and Section VI, Table I and Fig. 3] The paper does not state how student variant v7 was selected from the ten variants described in Section IV-B. Figure 3 plots all variants on the test sets, and Table I reports only v7. If v7 was chosen after inspecting test-set curves, the reported 0.76 weighted mean is a selected maximum, not an expected result, and the 'halves the gap' claim is optimistic. Please state the selection rule (e.g., best validation MSE among variants) or, if v7 was selected on test data, provide an honest estimate of the selection bias, for example by reporting all variants and a validation-based selection protocol.
  2. [Section VI, Table I and Abstract] The 'halve the gap' headline is metric-dependent: under the clip-count-weighted mean the distilled v7 improves from 0.72 to 0.76 against a teacher gap of 0.09 (0.72 to 0.81), but under the unweighted mean the improvement is 0.70 to 0.73 against a teacher gap of 0.12, i.e., roughly a quarter, not a half. The weighted mean is dominated by the largest challenge test sets. Please report both weighted and unweighted means throughout, justify the choice of weighting, and adjust the abstract and conclusion so that the strength of the claim matches the evidence.
  3. [Section IV-B, Section VI, Table I] No multiple-seed experiments, error bars, or statistical tests are reported. Given that the reported gains are of size 0.03-0.04 in mean correlation and that model variants were selected, seed-to-seed variability could be of the same order as the reported improvement. Please report mean and variance over at least three training runs for the baseline v7, the distilled v7, and the pruned checkpoint, or otherwise justify why selection noise cannot explain the headline difference.
  4. [Section VI, Blizzard 2023 rows in Table I] The paper itself states that 'distillation only works well for data on which the teacher model already achieves high correlation,' and the Blizzard TTS row (0.22 distilled vs. 0.24 baseline vs. 0.58 teacher) confirms this limitation. This is a candid and important disclosure, but the abstract and title currently make an unqualified claim about distillation. The scope of the claim should be narrowed to in-domain degraded speech (VoIP, noise suppression, and similar degradations), and the implications for using pseudo-labels on unlabeled data in practice should be stated explicitly.
minor comments (7)
  1. [Section III-B] The dataset description contains a typo: 'ConferncingSpeech' should be 'ConferencingSpeech'.
  2. [Section III] The text contains a typo: 'Additonally' should be 'Additionally'.
  3. [Section III-B, NISQA P.501 row] The phrase 'degradations from transmission unsing V oIP services' contains a typo: 'unsing' should be 'using'.
  4. [Section III and Table I] The number of clips for DNS 3 FB is given as 4200 in the text but 4500 in Table I; please reconcile these numbers.
  5. [Section IV-A] The paper says the teacher is selected by 'lowest weighted mean squared error during validation,' but the weighting across the heterogeneous validation datasets is not defined; please specify the weights.
  6. [Section IV-B] The architecture search for the ten student variants is shown only graphically in Fig. 2; for reproducibility, please provide a table listing the exact hyperparameters of each variant, especially v7.
  7. [Section VII] The conclusion says 'combining data-driven pruning and the proposed distillation technique,' but the experiments treat pruning and distillation separately and never prune a distilled student; please rephrase to avoid implying a joint method was evaluated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the student is trained on teacher pseudo-labels plus labeled data and evaluated on held-out ground-truth MOS, while pruning is evaluated against ground-truth labels.

full rationale

The paper's central claim is an empirical result, not a derivation. The teacher model is retrained on labeled MOS data, the student is trained on teacher pseudo-labels mixed with 20% ground-truth labels, and all reported correlations are computed against held-out ground-truth MOS on test datasets not used for training. The pseudo-labels are training targets only; the evaluation metric is independent of them. The pruning method uses a first-order Taylor approximation of loss change to remove low-importance weights, and the pruned models are evaluated on the same held-out ground-truth test sets; there is no fitted parameter that is later renamed as a prediction. Self-citations in the paper, such as the compressed spectrogram features from [39], are architectural input choices rather than load-bearing justifications of the claimed result. Concerns about the selection of variant 7 and the difference between weighted and unweighted means are statistical reporting issues, not circularity, because the paper does not define its headline metric in terms of the selected variant's test performance by construction. No equation or definition in the paper reduces a predicted quantity to its own input.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on supervised training with teacher-generated pseudo-labels, a grid-selected label mix-in, and a specific data synthesis pipeline, plus several domain assumptions about shared quality scales and evaluation metrics. No new entities are postulated. The main free parameters are hyperparameters of the distillation and pruning schedules; none is a fitted physical constant.

free parameters (5)
  • labeled mix-in percentage p = 20%
    Grid search over 0%, 10%, 20%, 40%, 80%; 20% chosen as optimal for distillation.
  • universal logits scale and shift = grid-searched after training
    Per-dataset scale/shift parameters are learned, then universal values are selected by grid search; this is a fitted calibration of the teacher's output.
  • student architecture variants = 10 variants, v7 selected
    The paper varies conv channels, transformer dim, heads, and layers; variant 7 is used for the headline results, but the selection criterion is not documented.
  • pruning rate and fine-tuning schedule = 0.5% per step, 30 update steps, smoothing 0.9
    These hyperparameters are chosen without sensitivity analysis and affect the pruning curve.
  • training step counts = 72,000 teacher; 250,000 student
    Set by early stopping on validation; standard but manually chosen.
assumptions (5)
  • standard math Taylor expansion importance approximation (Eq. 1)
    Importance score approximates loss change by (dL/dw)^2; standard in pruning literature.
  • domain assumption MOS datasets share a common quality construct
    The paper pools labels from different procedures with per-dataset scale/shift, assuming a shared underlying quality notion.
  • domain assumption Teacher pseudo-labels are valid supervision
    Student is trained on teacher outputs plus 20% ground truth; if teacher errors are systematic, student inherits them.
  • domain assumption Pearson correlation is the appropriate metric
    Evaluation relies on per-dataset PCC and its weighted mean, ignoring calibration and scale of predictions.
  • ad hoc to paper Unlabeled data generation covers target degradations
    The 85% real speech pipeline (DNS4 noise, RIRs, codecs, vocoders) is designed in-house; Blizzard results show it does not cover synthetic speech.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Distillation and Pruning for Scalable Self-Supervised Representation-Based Speech Quality Assessment." pith.science (2026). https://pith.science/paper/LQ5A4IYF

@misc{pith2026250205356,
  author       = {Pith},
  title        = {Pith review of: Distillation and Pruning for Scalable Self-Supervised Representation-Based Speech Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQ5A4IYF}},
  note         = {Machine review of arXiv:2502.05356}
}
read the original abstract

In this paper, we investigate distillation and pruning methods to reduce model size for non-intrusive speech quality assessment based on self-supervised representations. Our experiments build on XLS-R-SQA, a speech quality assessment model using wav2vec 2.0 XLS-R embeddings. We retrain this model on a large compilation of mean opinion score datasets, encompassing over 100,000 labeled clips. For distillation, using this model as a teacher, we generate pseudo-labels on unlabeled degraded speech signals and train student models of varying sizes. For pruning, we use a data-driven strategy. While data-driven pruning performs better at larger model sizes, distillation on unlabeled data is more effective for smaller model sizes. Distillation can halve the gap between the baseline's correlation with ground-truth MOS labels and that of the XLS-R-based teacher model, while reducing model size by two orders of magnitude compared to the teacher model.

Figures

Figures reproduced from arXiv: 2502.05356 by the authors.

Figure 1
Figure 1. XLS-R-based speech quality assessment and its usage [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Resulting model sizes for student model variants. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Weighted average Pearson correlation coefficient as a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [1]

    Non-intrusive speech quality assessment using neural networks,

    A. R. Avila et al., “Non-intrusive speech quality assessment using neural networks,” in Proc. 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2019, pp. 631–635

  2. [2]

    Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. K. A. Reddy et al., “Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in Proc. 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6493–6497

  3. [3]

    NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,

    G. Mittag et al. , “NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,” in Proc. Interspeech 2021 , 2021, pp. 2127–2131

  4. [4]

    Utilizing Self-Supervised Representations for MOS Prediction,

    W.-C. Tseng et al., “Utilizing Self-Supervised Representations for MOS Prediction,” in Proc. Interspeech 2021 , 2021, pp. 2781–2785

  5. [5]

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

    A. Baevski et al., “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Information Processing Systems, H. Larochelle et al., Eds., vol. 33. Curran Associates, Inc., 2020, pp. 12 449–12 460

  6. [6]

    Deep learning-based non-intrusive multi-objective speech assessment model with cross-domain features,

    R. E. Zezario et al., “Deep learning-based non-intrusive multi-objective speech assessment model with cross-domain features,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 54–70, 2023

  7. [7]

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

    W.-N. Hsu et al., “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

  8. [8]

    UTMOS: UTokyo-SaruLab System for V oiceMOS Challenge 2022,

    T. Saeki et al. , “UTMOS: UTokyo-SaruLab System for V oiceMOS Challenge 2022,” in Proc. Interspeech 2022 , 2022, pp. 4521–4525

Show all 43 references
  1. [9]

    The voicemos challenge 2022,

    W. C. Huang et al., “The voicemos challenge 2022,” in Interspeech 2022, 2022, pp. 4536–4540

  2. [10]

    Analysis of XLS-R for speech quality assessment,

    B. Tamm et al., “Analysis of XLS-R for speech quality assessment,” in 2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2023, pp. 1–5

  3. [11]

    XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,

    A. Babu et al. , “XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,” in Proc. Interspeech 2022 , 2022, pp. 2278–2282

  4. [12]

    ConferencingSpeech 2022 Challenge: Non-intrusive Objective Speech Quality Assessment (NISQA) Challenge for Online Conferencing Applications,

    G. Yi et al. , “ConferencingSpeech 2022 Challenge: Non-intrusive Objective Speech Quality Assessment (NISQA) Challenge for Online Conferencing Applications,” in Proc. Interspeech, 2022, pp. 3308–3312

  5. [13]

    PAM: Prompting audio-language models for audio quality assessment,

    S. Deshmukh et al., “PAM: Prompting audio-language models for audio quality assessment,” arXiv preprint arXiv:2402.00282 , 2023

  6. [14]

    CLAP: Learning audio concepts from natural language supervision,

    B. Elizalde et al., “CLAP: Learning audio concepts from natural language supervision,” in Proc. 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2023, pp. 1–5

  7. [15]

    DNN No-Reference PSTN Speech Quality Prediction,

    G. Mittag et al., “DNN No-Reference PSTN Speech Quality Prediction,” in Proc. Interspeech 2020 , 2020, pp. 2867–2871

  8. [16]

    ICASSP 2021 deep noise suppression challenge,

    C. K. A. Reddy et al., “ICASSP 2021 deep noise suppression challenge,” in Proc. 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2021, pp. 6623–6627

  9. [17]

    Interspeech 2022 audio deep packet loss concealment challenge,

    L. Diener et al., “Interspeech 2022 audio deep packet loss concealment challenge,” in Interspeech 2022, 2022, pp. 580–584

  10. [18]

    ICASSP 2023 speech signal improvement challenge,

    R. Cutler et al., “ICASSP 2023 speech signal improvement challenge,” IEEE Open Journal of Signal Processing , vol. 5, pp. 662–674, 2024

  11. [19]

    Protocol for the collection of databases of recordings for forensic-voice-comparison research and practice,

    G. Morrison et al., “Protocol for the collection of databases of recordings for forensic-voice-comparison research and practice,” Australian Journal of Forensic Sciences, vol. 44, pp. 1–13, 06 2012

  12. [20]

    Tcd-voip, a research database of degraded speech for assessing quality in voip applications,

    N. Harte et al. , “Tcd-voip, a research database of degraded speech for assessing quality in voip applications,” in Proc. 2015 Seventh International Workshop on Quality of Multimedia Experience (QoMEX) , 2015, pp. 1–6

  13. [21]

    Speech quality factors for traditional and neural- based low bit rate vocoders,

    W. A. Jassim et al., “Speech quality factors for traditional and neural- based low bit rate vocoders,” in Proc. 2020 Twelfth International Conference on Quality of Multimedia Experience (QoMEX) , 2020, pp. 1–6

  14. [22]

    The blizzard challenge 2023,

    O. Perrotin et al. , “The blizzard challenge 2023,” in 18th Blizzard Challenge Workshop, 2023, pp. 1–27

  15. [23]

    Interspeech 2021 deep noise suppression challenge,

    C. K. Reddy et al., “Interspeech 2021 deep noise suppression challenge,” in Interspeech 2021, 2021, pp. 2796–2800

  16. [24]

    The ICASSP 2024 audio deep packet loss concealment grand challenge,

    L. Diener et al., “The ICASSP 2024 audio deep packet loss concealment grand challenge,” in Proc. 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSP) , 2024, pp. 41–42

  17. [25]

    ICASSP 2024 speech signal improvement challenge,

    N.-C. Ristea et al., “ICASSP 2024 speech signal improvement challenge,” in Proc. 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSP) , 2024, pp. 15–16

  18. [26]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in International Conference on Learning Representations , 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

  19. [27]

    Bias-aware loss for training image and speech quality prediction models from multiple datasets,

    G. Mittag et al. , “Bias-aware loss for training image and speech quality prediction models from multiple datasets,” in Proc. 2021 13th International Conference on Quality of Multimedia Experience (QoMEX) , 2021, pp. 97–102

  20. [28]

    Coqui TTS,

    G. Eren and the Coqui TTS Team, “Coqui TTS,” Jan. 2021. [Online]. Available: https://github.com/coqui-ai/TTS

  21. [29]

    MultiSubs: A large-scale multimodal and multilingual dataset,

    J. Wang et al., “MultiSubs: A large-scale multimodal and multilingual dataset,” in Proceedings of the Thirteenth Language Resources and Evaluation Conference , N. Calzolari et al. , Eds. Marseille, France: European Language Resources Association, Jun. 2022, pp. 6776–6785. [Onl...

  22. [30]

    ICASSP 2022 deep noise suppression challenge,

    H. Dubey et al., “ICASSP 2022 deep noise suppression challenge,” in Proc. 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2022, pp. 9271–9275

  23. [31]

    Deepfilternet: Perceptually motivated real-time speech enhancement,

    H. Schr¨oter et al., “Deepfilternet: Perceptually motivated real-time speech enhancement,” in INTERSPEECH 2023, 2023, pp. 2008–2009

  24. [32]

    Data augmentation and loss normalization for deep noise suppression,

    S. Braun and I. Tashev, “Data augmentation and loss normalization for deep noise suppression,” in Speech and Computer , A. Karpov and R. Potapova, Eds. Cham: Springer International Publishing, 2020, pp. 79–86

  25. [33]

    Real time speech enhancement in the waveform domain,

    A. Defossez et al. , “Real time speech enhancement in the waveform domain,” in Interspeech, 2020

  26. [34]

    timsainb/noisereduce: v1.0,

    T. Sainburg, “timsainb/noisereduce: v1.0,” 2019. [Online]. Available: https://doi.org/10.5281/zenodo.3243139

  27. [35]

    Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,

    J. Kong et al. , “Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,” in Advances in Neural Information Processing Systems , H. Larochelle et al. , Eds., vol. 33. Curran Associates, Inc., 2020, pp. 17 022–17 033. [Online]. Available: ht...

  28. [36]

    LPCNET: Improving neural speech synthesis through linear prediction,

    J.-M. Valin and J. Skoglund, “LPCNET: Improving neural speech synthesis through linear prediction,” in Proc. 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2019, pp. 5891–5895

  29. [37]

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

    J.-M. Valin et al., “High-quality, low-delay music coding in the Opus codec,” in Proc. 135th Audio Engineering Society Convention , 2013, pp. 1–10, URL: http://www.aes.org/e-lib/browse.cfm?elib=16992

  30. [38]

    High fidelity neural audio compression,

    A. D ´efossez et al. , “High fidelity neural audio compression,” Transactions on Machine Learning Research, 2023, featured Certification, Reproducibility Certification. [Online]. Available: https://openreview.net/ forum?id=ivCd8z8zR2

  31. [39]

    Effect of noise suppression losses on speech distortion and asr performance,

    S. Braun and H. Gamper, “Effect of noise suppression losses on speech distortion and asr performance,” in Proc. 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2022, pp. 996–1000

  32. [40]

    Importance estimation for neural network pruning,

    P. Molchanov et al., “Importance estimation for neural network pruning,” in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

  33. [41]

    Dnsmos p.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. Reddy et al., “Dnsmos p.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in Proc. 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 886–890

  34. [42]

    Torchaudio-squim: Reference-less speech quality and intelligibility measures in torchaudio,

    A. Kumar et al., “Torchaudio-squim: Reference-less speech quality and intelligibility measures in torchaudio,” in Proc. 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2023, pp. 1–5

  35. [43]

    The state of sparsity in deep neural networks,

    T. Gale et al. , “The state of sparsity in deep neural networks,” 2019. [Online]. Available: https://arxiv.org/abs/1902.09574

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.