Pith. sign in

REVIEW 4 major objections 5 minor 36 references

LL-SDR claims that a non-autoregressive, token-based speech enhancer—using a variance-ordered residual quantizer and a HuBERT semantic discriminator—can match the audible quality of large autoregressive models while running about thirty tim

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 ↗

LL-SDR, a non-autoregressive token-based enhancer with variance-ordered residual vector quantization and a HuBERT semantic discriminator, matches autoregressive systems on DNSMOS at roughly one-tenth the GPU real-time factor.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A practical NAR enhancer with a clever variance-ordered quantizer; the AR-parity claim needs a shared re-run, but the internal ablations make it a solid contribution. the 4 major comments →

arxiv 2603.20242 v2 pith:UZCVX232 submitted 2026-03-10 cs.SD eess.AS

LL-SDR: Low-Latency Speech enhancement through Discrete Representations

Jingyi Li , Luca Della Libera , Mirco Ravanelli , Mingkun Xu , Cem Subakan This is my paper
classification cs.SD eess.AS
keywords speech enhancementdiscrete audio tokensresidual vector quantizationvariance-ordered representationsemantic discriminatorlow-latency inferenceDNSMOSnon-autoregressive generation
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

LL-SDR aims to show that discrete representations themselves improve speech enhancement, not just enable autoregressive generation. The paper claims that a residual quantizer whose codebooks are ordered by variance—speech in the first four, noise in the last—cleanly separates the two, and that a HuBERT-based discriminator keeps the enhanced tokens semantically faithful. If right, a lightweight 74M-parameter non-autoregressive model can match the perceptual quality of much larger autoregressive systems like GenSE and LLaSE-G1 while running roughly 30x faster on GPU. That would make high-quality speech enhancement practical for real-time telephony, hearing aids, and ASR front-ends.

Core claim

The central claim is that a non-autoregressive speech enhancement model can match the perceptual quality of autoregressive token-based systems—GenSE and LLaSE-G1—by exploiting discretization rather than generation. The mechanism is VO-RVQ, a residual vector quantizer modified with an order-inducing loss and triangular masking so that the first four codebooks capture high-variance speech structure and the final codebook absorbs residual noise. On DNSMOS, LL-SDR reaches 3.27 OVRL (reverb) and 3.39 (no-reverb) with 74M parameters, 40G MACs, and a GPU real-time factor of 0.0108, about 30x faster than LLaSE-G1. Ablations show each component contributes: discrete tokens beat a continuous baseline,

What carries the argument

Variance-Ordered Residual Vector Quantizer (VO-RVQ): at each of N residual stages the latent is projected to a shared full-dimensional space but only the first d_i dimensions are quantized (triangular mask), with an order-inducing loss that forces early codebooks to encode high-variance content. With a 4:1 split of N_e speech codebooks to N_n noise codebooks, the paper assumes speech occupies high-variance dimensions and noise is pushed into later residuals. A HuBERT-based semantic discriminator (L2 regression plus InfoNCE contrastive loss) aligns the enhanced embeddings with clean-speech semantic features.

Load-bearing premise

The whole architecture rests on the assumption that speech energy always lands in high-variance latent dimensions and noise in low-variance ones, enforced by a fixed 4:1 codebook split calibrated to one training-set ratio; if a test noise type has variance comparable to speech, the disentanglement and the quality gains weaken.

What would settle it

Take a test set with non-stationary noise such as babble or music, whose variance is comparable to speech, retrain or fine-tune LL-SDR with the same 4:1 split, and measure DNSMOS plus the spectral-clustering accuracy from Table 3. If clustering accuracy falls toward the RVQ baseline (~50%) or DNSMOS drops below the continuous baseline, the variance-ordering assumption fails.

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

If this is right

  • Discretization can be a benefit in itself for non-autoregressive speech enhancement, not merely a prerequisite for autoregressive modeling.
  • Variance-ordered quantization gives codec designers a reusable recipe for disentangling signal and noise beyond this specific task.
  • A semantic discriminator outside the tokenizer can supply the grounding that internal-supervision methods get, making it a drop-in for other non-autoregressive audio models.
  • Perceptual quality near 3.3 OVRL at RTF 0.01 suggests real-time deployment in hearing aids and telephony is feasible with a 74M-parameter model.

Where Pith is reading between the lines

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

  • The 4:1 codebook split is calibrated to a single training-set variance ratio; a sensitivity sweep over Ne:Nn splits (e.g., 3:2, 5:0) would show whether the architecture is robust to mismatch, and whether the gains come from ordering or from the extra capacity of the first codebooks.
  • A direct probe of the disentanglement claim is to decode the enhanced signal from only the four speech codebooks: if the output still contains noise, the variance-ordering assumption is violated for that input.
  • Since noise variance is time- and class-dependent, a dynamic split that adapts Ne:Nn per utterance could extend the idea to non-stationary noise, which the current fixed split does not address.
  • The clustering experiment (71% accuracy) is evidence of separability in the latent space, but a human listening test or an ASR word-error-rate evaluation on enhanced audio would test whether the disentanglement yields downstream benefits beyond DNSMOS.
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

4 major / 5 minor

Summary. The paper proposes LL-SDR, a non-autoregressive speech enhancement model that operates on discrete codec tokens. The main contributions are a Variance-Ordered Residual Vector Quantizer (VO-RVQ) with triangular masking intended to separate speech from noise across codebooks, and a HuBERT-based semantic discriminator that aligns enhanced embeddings with clean-speech semantic features. The model is trained with reconstruction, adversarial, order-inducing, and semantic losses. Experiments on DNS-style test sets report DNSMOS scores, model complexity, and RTF, and claim performance comparable to autoregressive token-based systems such as GenSE and LLaSE-G1 at much lower latency. Ablations show incremental gains from discretization, VO-RVQ, and semantic alignment.

Significance. If the empirical claims are supported, the paper makes a useful contribution: it challenges the assumption that autoregressive generation is required for competitive token-based speech enhancement, and its efficiency figures (74M parameters, 40G MACs, RTF 0.0108 on H100) are attractive for latency-sensitive applications. The VO-RVQ idea is interesting and the release of code and demos is a strength. However, the central comparison to autoregressive baselines currently rests on DNSMOS scores taken from a single prior paper, with no shared evaluation or uncertainty quantification. The disentanglement mechanism is also under-specified, and one free structural parameter (the 4:1 codebook split) is not tested for sensitivity. The contribution is therefore promising but not yet demonstrated at the level required by the main claim.

major comments (4)
  1. [Sec. 3.4, Table 1] The headline claim that LL-SDR "achieves performance comparable to recent autoregressive generative methods such as GenSE and LLaSE-G1" is supported only by DNSMOS values taken from prior work [6]. DNSMOS scores are sensitive to model version, resampling, and preprocessing, and the observed gaps are small (reverb OVRL 3.27 vs 3.33; no-reverb 3.39 vs 3.42). The compared models were also trained on different data. Without re-running the baselines in the authors' own pipeline, or at least providing error bars and significance tests over multiple test splits, the comparison is not controlled. This is load-bearing because it is the paper's main empirical contribution. Please re-evaluate the baselines under a common setup, or add uncertainty quantification and temper the claim to "similar to published scores."
  2. [Sec. 2.2, Algorithm 1] The pseudocode for VO-RVQ appears internally inconsistent with the text. The loop computes y_{q,i} for every stage, but the line "if i <= N_e, y_q <- y_q + y_{q,i}" means the outputs of codebooks i > N_e are never added to the final representation. The residual update r_i <- r_{i-1} - y_{q,i} also has no effect on the returned y_q after the loop ends, because no later stage contributes to y_q. The text, however, states that "these subsequent codebooks (the last N_n codebooks) effectively represent noise components." As written, the noise codebooks cannot influence the output, making the disentanglement mechanism unclear. Please correct the algorithm or explain how the later codebooks interact with the decoder; if they are discarded by design, justify why they are trained at all.
  3. [Sec. 2.4 / Table 4] The 4:1 split of codebooks between speech and noise is a fitted structural parameter, chosen from an "empirical variance ratio between speech and noise estimated from the training dataset." Neither the estimation procedure nor any sensitivity analysis is reported. Since VO-RVQ's advantage over plain RVQ in Table 4 is the main evidence for the ordering mechanism, the paper should describe how the variance ratio is computed and ablate over Ne:Nn (e.g., 3:2, 5:0, 4:1) and ideally over the total number of codebooks. Without this, the single 4:1 choice is a free parameter that carries a core part of the contribution.
  4. [Sec. 3.6, Table 4] The "Continuous" ablation baseline is not specified. It is unclear whether it uses the same DAC encoder/decoder, the same training loss, the same number of parameters, and no quantization, or a different architecture. Because this baseline is the basis for the claim that "discretization ... consistently outperforms continuous baselines," the architecture and training details (and ideally parameter count and MACs) must be provided.
minor comments (5)
  1. [Sec. 2.2, Algorithm 1] The masking dimension is referred to as "d" without defining d_i for each stage. The input also lists {C_n}_{n=1}^{N_e} but the loop runs to N; please clarify that there are N codebooks.
  2. [Sec. 3.5, Table 3] The clustering experiment would benefit from more detail: which embeddings are clustered (which codebook outputs or layers), how many utterances, and variance over seeds. A single 71.33% accuracy number is hard to interpret.
  3. [Sec. 3.4, Table 2] CPU model and thread count are not specified, and RTF can vary substantially with hardware. Please also clarify whether RTF includes feature extraction and HuBERT scoring or only the enhancement forward pass.
  4. [Sec. 3.4] The statement "Qualitative results (see demo page) further suggest ..." should be supported by an in-paper figure or a quantitative measure of speaker similarity; demos are not part of the manuscript.
  5. [Sec. 2.3, Eq. (3)] Please clarify whether the HuBERT features of the clean target are frozen during training and whether HuBERT is used only as a train-time teacher or also at inference.

Circularity Check

0 steps flagged

No significant circularity: the claims are supported by the paper's own training/ablation experiments and external anchors, with borrowed baseline scores being an evaluation-validity concern rather than a circular-derivation concern.

full rationale

The paper's central derivation chain is empirical rather than definitional. VO-RVQ is defined by a loss (Eq. 2) and algorithm (Alg. 1) that do not presuppose the outcome; the 4:1 speech/noise codebook split is a fitted hyperparameter based on a training-set variance ratio, but the paper never 'predicts' this ratio from the model—it simply adopts it as a design choice and then tests the resulting system on held-out DNSMOS (Table 1), including an ablation (Table 4) showing VO-RVQ improves over RVQ and over the continuous baseline. The disentanglement analysis (Table 3) tests the hypothesis empirically with spectral clustering on held-out embeddings, which is a genuine validation rather than a tautology. The semantic discriminator uses HuBERT as an external teacher, not as a fitted parameter of the enhancement target. The only self-citations ([28], [29]) justify not using PESQ/SI-SNR/STOI for generative models, and this decision is peripheral rather than load-bearing to the claimed contribution. The fact that baseline DNSMOS numbers are 'taken from prior work [6]' (Section 3.3) is an evaluation-consistency concern about comparison validity, not a circular-derivation concern under the review rules, because it does not make the model's output equal to its input or make the headline claim true by construction. Overall, the paper is self-contained against external benchmarks and no circular step can be exhibited.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claims rest on a small number of architectural and data assumptions: a specific 4:1 codebook split fit to the training-set speech/noise variance ratio; an unproven transfer of the ordered-autoencoding objective to multi-stage RVQ; and the use of HuBERT clean-speech features as a semantic target. No novel physical entities are introduced. Main free parameters (loss weights, InfoNCE temperature) are not reported, limiting exact reproduction.

free parameters (3)
  • Ne:Nn codebook allocation ratio = 4:1
    Set in §2.4 to the "empirical variance ratio between speech and noise estimated from the training dataset"; determines how many codebooks model speech vs noise and is central to the disentanglement claim.
  • Loss weights (lambda_ord, lambda_stft, lambda_adv, lambda_align, alpha) = not reported
    Overall objective in Eq. 6; values are not specified, so the contribution of each loss term cannot be reproduced exactly.
  • InfoNCE temperature tau = not reported
    Used in Eq. 5 to control the sharpness of the contrastive similarity distribution; the default value is not stated.
axioms (4)
  • domain assumption Speech dominates the high-variance components of the codec latent; noise is primarily pushed into the residual, so the first 4 codebooks carry speech and the last 1 carries noise.
    Stated in §2.2 and used to justify VO-RVQ's triangle masking and the 4:1 codebook split; not proven or tested via sensitivity analysis.
  • domain assumption The ordered-representation objective from [15] induces PCA-like variance ordering when applied to an RVQ residual stack with triangular masking.
    The paper transfers this result to multi-stage RVQ plus adversarial losses without a proof or empirical check of the ordering property per codebook.
  • domain assumption HuBERT features computed on clean target speech are a valid semantic teacher for enhanced embeddings.
    Used in §2.3; assumes clean-speech HuBERT features are a stable, discriminative semantic target for the discriminator.
  • domain assumption DAC-style training and architecture are suitable for enhancement when all reconstruction terms are against clean target speech.
    The paper follows the DAC recipe (§2.4) but does not state whether the codec is initialized from pretrained weights or trained from scratch, nor how the clean-target objective changes codec behavior.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LL-SDR: Low-Latency Speech enhancement through Discrete Representations." pith.science (2026). https://pith.science/paper/UZCVX232

@misc{pith2026260320242,
  author       = {Pith},
  title        = {Pith review of: LL-SDR: Low-Latency Speech enhancement through Discrete Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZCVX232}},
  note         = {Machine review of arXiv:2603.20242}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Many speech enhancement (SE) methods rely on continuous representations. Recently, discrete audio tokens have been explored to enable autoregressive generation for SE. However, it remains unclear whether discretization itself consistently improves SE performance. In this paper, we introduce LL-SDR, a token-based speech enhancement framework that explicitly leverages discretization to better separate speech and noise. Our first contribution is a Variance-Ordered Residual Vector Quantizer (VO-RVQ), designed to disentangle speech and noise distributions during tokenization. Second, we propose a latent-space discriminator to better align enhanced embeddings with semantic embeddings. Experiments show that LL-SDR outperforms continuous baselines and matches the performance of autoregressive token-based approaches. Despite its strong enhancement performance, LL-SDR remains lightweight and efficient, requiring only 40G MACs for a single forward pass on a 10-second 16 kHz speech segment and achieving low-latency inference with an RTF of 0.01 on GPU and 0.24 on CPU. Demos and source code are available at our project websites.

Figures

Figures reproduced from arXiv: 2603.20242 by Cem Subakan, Jingyi Li, Luca Della Libera, Mingkun Xu, Mirco Ravanelli.

Figure 1
Figure 1. Figure 1: LL-SDR speech enhancement framework: a noisy waveform is encoded and enhanced in the discrete space via our variance￾ordered quantizer where Ne is four and Nn is one, then decoded back to waveform. The triangle structure here illustrates that different quantizers have different dimensional sizes; a HuBERT-based semantic discriminator to align the enhanced representation with clean speech semantics. yc = DA… 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

36 extracted references · 2 linked inside Pith

  1. [1]

    Most existing speech enhancement approaches are built upon continuous acoustic features

    Introduction Speech enhancement (SE) aims to improve the quality and intel- ligibility of speech signals by suppressing noise, reverberation, or other distortions while preserving the underlying speech con- tent. Most existing speech enhancement approaches are built upon continuous acoustic features. These primarily include methods formulated in the time-...

  2. [2]

    Framework Overview As shown in Figure 1, the generator of our SE architecture consists of a codec encoder, a quantizer, and a decoder

    Method 2.1. Framework Overview As shown in Figure 1, the generator of our SE architecture consists of a codec encoder, a quantizer, and a decoder. Formally, arXiv:2603.20242v1 [cs.SD] 10 Mar 2026 Decoder Quantizer HuBERT Semantic Discriminator Discriminator Noisy Speech Xin Clean Speech Enhanced Speech Xout yq1 yq3yq2 yq yq4 yc zq1 zq2 zq3 zq4 r1 r2 r3 r4...

    Pith/arXiv arXiv 2026

  3. [3]

    Dataset For the ablation study, we train our model on LibriSpeech- 100 [24] and the DNS Challenge [25] datasets

    Experiments 3.1. Dataset For the ablation study, we train our model on LibriSpeech- 100 [24] and the DNS Challenge [25] datasets. Room impulse responses (RIRs) are sourced from the DNS Challenge dataset, and noise signals are also drawn from DNS. For comparisons against other methods, we additionally train a larger variant us- ing LibriSpeech-360 together...

    arXiv 2020

  4. [4]

    Conclusion In this paper, we introduced LL-SDR, a low-latency non- autoregressive speech enhancement model that achieves per- formance comparable to autoregressive approaches. Our method leverages a variance-ordered quantizer with a triangular mask- ing structure to disentangle speech and noise components in the latent space, prioritizing high-information...

  5. [5]

    Generative AI tools were used solely for minor editing and language polishing to improve clarity and fluency of the manuscript

    Generative AI Use Disclosure All authors take full responsibility and accountability for the con- tent of this paper and consent to its submission. Generative AI tools were used solely for minor editing and language polishing to improve clarity and fluency of the manuscript

  6. [6]

    FRCRN: Boosting feature representation using frequency recurrence for monaural speech enhancement,

    S. Zhao, B. Ma, K. N. Watcharasupat, and W. S. Gan, “FRCRN: Boosting feature representation using frequency recurrence for monaural speech enhancement,”IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 9281– 9285, 2022

    2022

  7. [7]

    Demucs: Deep extractor for music sources with extra unlabeled data remixed,

    A. D´efossez, N. Usunier, L. Bottou, and F. Bach, “Demucs: Deep extractor for music sources with extra unlabeled data remixed,” arXiv preprint arXiv:1909.01174, 2019

    Pith/arXiv arXiv 1909

  8. [8]

    Conv-TasNet: Surpassing ideal time–frequency magnitude masking for speech separation,

    Y . Luo and N. Mesgarani, “Conv-TasNet: Surpassing ideal time–frequency magnitude masking for speech separation,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing (TASLP), vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  9. [9]

    Genhancer: High-fidelity speech enhancement via generative modeling on discrete codec tokens,

    H. Yang, J. Su, M. Kim, and Z. Jin, “Genhancer: High-fidelity speech enhancement via generative modeling on discrete codec tokens,” inInterspeech, 2024, pp. 1170–1174

    2024

  10. [10]

    SELM: Speech enhancement using discrete tokens and language models,

    Z. Wang, X. Zhu, Z. Zhang, Y . Lv, N. Jiang, G. Zhao, and L. Xie, “SELM: Speech enhancement using discrete tokens and language models,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 11 561–11 565

    2024

  11. [11]

    LLaSE-g1: In- centivizing generalization capability for LLaMA-based speech enhancement,

    B. Kang, X. Zhu, Z. Zhang, Z. Ye, M. Liu, Z. Wang, Y . Zhu, G. Ma, J. Chen, L. Xiao, C. Weng, W. Xue, and L. Xie, “LLaSE-g1: In- centivizing generalization capability for LLaMA-based speech enhancement,” inAnnual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), 2025, pp. 13 292– 13 305

    2025

  12. [12]

    GenSE: Generative speech enhancement via language models using hi- erarchical modeling,

    J. Yao, H. Liu, C. Chen, Y . Hu, E. Chng, and L. Xie, “GenSE: Generative speech enhancement via language models using hi- erarchical modeling,” inInternational Conference on Learning Representations (ICLR), 2025

    2025

  13. [13]

    An EM approach to non-autoregressive conditional sequence generation,

    Z. Sun and Y . Yang, “An EM approach to non-autoregressive conditional sequence generation,” inInternational Conference on Machine Learning (ICML), 2020, pp. 9249–9258

    2020

  14. [14]

    Real time speech en- hancement in the waveform domain,

    A. D ´efossez, G. Synnaeve, and Y . Adi, “Real time speech en- hancement in the waveform domain,” inInterspeech, 2020, pp. 3291–3295

    2020

  15. [15]

    A scalable noisy speech dataset and online subjective test framework,

    C. K. Reddy, E. Beyrami, J. Pool, R. Cutler, S. Srinivasan, and J. Gehrke, “A scalable noisy speech dataset and online subjective test framework,” inInterspeech, 2019, pp. 1816–1820

    2019

  16. [16]

    An individualized super-Gaussian single microphone speech enhancement for hearing aid users with smart- phone as an assistive device,

    C. Karadagur Ananda Reddy, N. Shankar, G. Shreedhar Bhat, R. Charan, and I. Panahi, “An individualized super-Gaussian single microphone speech enhancement for hearing aid users with smart- phone as an assistive device,”IEEE Signal Processing Letters, vol. 24, no. 11, pp. 1601–1605, 2017

    2017

  17. [17]

    An investigation into the effectiveness of enhancement in ASR training and test for Chime-5 dinner party transcription,

    C. Zorila, C. Boeddeker, R. S. Doddipatla, and R. H¨ab-Umbach, “An investigation into the effectiveness of enhancement in ASR training and test for Chime-5 dinner party transcription,”IEEE Au- tomatic Speech Recognition and Understanding Workshop (ASRU), pp. 47–53, 2019

    2019

  18. [18]

    Speech enhancement using continuous embeddings of neural audio codec,

    H. Li, J. Q. Yip, T. Fan, and E. S. Chng, “Speech enhancement using continuous embeddings of neural audio codec,” inIEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), Apr. 2025, pp. 1–5

    2025

  19. [19]

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

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasac- chi, “SoundStream: An end-to-end neural audio codec,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing (TASLP), pp. 495–507, 2021

    2021

  20. [20]

    Anytime sampling for autoregressive models via ordered autoen- coding,

    Y . Xu, Y . Song, S. Garg, L. Gong, R. Shu, A. Grover, and S. Ermon, “Anytime sampling for autoregressive models via ordered autoen- coding,” inInternational Conference on Learning Representations (ICLR), 2021

    2021

  21. [21]

    Learning ordered repre- sentations with nested dropout,

    O. Rippel, M. Gelbart, and R. Adams, “Learning ordered repre- sentations with nested dropout,” inInternational Conference on Machine Learning (ICML), 2014, pp. 1746–1754

    2014

  22. [22]

    SoCodec: A semantic-ordered multi-stream speech codec for ef- ficient language model based text-to-speech synthesis,

    H. Guo, F. Xie, K. Xie, D. Yang, D. Guo, X. Wu, and H. Meng, “SoCodec: A semantic-ordered multi-stream speech codec for ef- ficient language model based text-to-speech synthesis,” inIEEE Spoken Language Technology Workshop (SLT), 2024, pp. 645–651

    2024

  23. [23]

    HuBERT: How much can a bad teacher benefit ASR pre-training?

    W.-N. Hsu, Y .-H. Tsai, B. Bolte, R. Salakhutdinov, and A. Mo- hamed, “HuBERT: How much can a bad teacher benefit ASR pre-training?” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6533–6537

    2021

  24. [24]

    SpeechTok- enizer: Unified speech tokenizer for speech language models,

    X. Zhang, D. Zhang, S. Li, Y . Zhou, and X. Qiu, “SpeechTok- enizer: Unified speech tokenizer for speech language models,” in International Conference on Learning Representations (ICLR), 2024

    2024

  25. [25]

    InfoNCE: Identifying the gap between theory and practice,

    E. Rusak, P. Reizinger, A. Juhos, O. Bringmann, R. S. Zimmer- mann, and W. Brendel, “InfoNCE: Identifying the gap between theory and practice,” inInternational Conference on Artificial Intelligence and Statistics, 2025, pp. 4159–4167

    2025

  26. [26]

    Revisiting over- smoothness in text to speech,

    Y . Ren, X. Tan, T. Qin, Z. Zhao, and T.-Y . Liu, “Revisiting over- smoothness in text to speech,” inAnnual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), May 2022, pp. 8197–8213

    2022

  27. [27]

    Noise contrastive estimation and negative sampling for conditional models: Consistency and statistical effi- ciency,

    Z. Ma and M. Collins, “Noise contrastive estimation and negative sampling for conditional models: Consistency and statistical effi- ciency,” inConference on Empirical Methods in Natural Language Processing (EMNLP), 2018, pp. 3698–3707

    2018

  28. [28]

    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,” inIn- ternational Conference on Neural Information Processing Systems (NeurIPS), 2023

    2023

  29. [29]

    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,” inIEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2015, pp. 5206–5210

    2015

  30. [30]

    The Interspeech 2020 deep noise suppression challenge: Datasets, sub- jective 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. Braunet al., “The Interspeech 2020 deep noise suppression challenge: Datasets, sub- jective testing framework, and challenge results,” inInterspeech, 2020

    2020

  31. [31]

    The diverse environments multi-channel acoustic noise database (DEMAND): A database of multichannel environmental noise recordings,

    J. Thiemann, N. Ito, and E. Vincent, “The diverse environments multi-channel acoustic noise database (DEMAND): A database of multichannel environmental noise recordings,”Proceedings of Meetings on Acoustics, vol. 19, no. 1, 2013

    2013

  32. [32]

    DNSMOS: A non-intrusive perceptual objective speech quality metric to evaluate noise sup- pressors,

    C. K. Reddy, V . Gopal, and R. Cutler, “DNSMOS: A non-intrusive perceptual objective speech quality metric to evaluate noise sup- pressors,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6493–6497

    2021

  33. [33]

    Focal- Codec: Low-bitrate speech coding via focal modulation networks,

    L. Della Libera, F. Paissan, C. Subakan, and M. Ravanelli, “Focal- Codec: Low-bitrate speech coding via focal modulation networks,” inInternational Conference on Neural Information Processing Systems (NeurIPS), 2025

    2025

  34. [34]

    FocalCodec- Stream: Streaming low-bitrate speech coding via causal distilla- tion,

    L. Della Libera, C. Subakan, and M. Ravanelli, “FocalCodec- Stream: Streaming low-bitrate speech coding via causal distilla- tion,”arXiv preprint arXiv:2509.16195, 2025

    arXiv 2025

  35. [35]

    MaskSR: Masked language model for full-band speech restoration,

    X. Li, Q. Wang, and X. Liu, “MaskSR: Masked language model for full-band speech restoration,” inInterspeech, 2024, pp. 2275–2279

    2024

  36. [36]

    AnyEnhance: A unified generative model with prompt- guidance and self-critic for voice enhancement,

    J. Zhang, J. Yang, Z. Fang, Y . Wang, Z. Zhang, Z. Wang, F. Fan, and Z. Wu, “AnyEnhance: A unified generative model with prompt- guidance and self-critic for voice enhancement,”IEEE Transac- tions on Audio, Speech and Language Processing (TASLP), vol. 33, pp. 3085–3098, 2025

    2025

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.