Pith. sign in

REVIEW 3 major objections 3 minor 44 references

Rethinking Language Model-Based Generative Speech Enhancement in the Latent Space of a Neural Audio Codec

T0 review · 3 major / 3 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Language-model speech enhancement is better when it models continuous codec features rather than discrete tokens, and the best variant is a continuous non-autoregressive model.

desk verdict A useful six-way benchmark of LM-based SE paradigms in codec latent space, with a plausible but not yet bulletproof continuous-over-discrete ranking. read the letter →

arxiv 2608.12082 v1 pith:UMFYPHPD submitted 2026-08-12 eess.AS

Yihui Fu , Zhengyang Li , Tim Fingscheidt This is my paper · ORCID
classification eess.AS
keywords language-modelspeechenhancementneuralaudiocodeccontinuouslatentfeaturesdiscretetokensnon-autoregressiveflowmatchingauxiliarylossfine-tuningqualitymetrics
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 argues that, for language-model-based generative speech enhancement, modeling the continuous latent features of a neural audio codec is better than modeling its discrete tokens. It unifies six popular paradigms—autoregressive and non-autoregressive, discrete and continuous, plus discrete diffusion and continuous flow matching—under one decoder-only transformer trained from scratch, and evaluates them on a common challenge corpus with both intrusive and non-intrusive quality metrics. The central result is that every continuous-domain variant outperforms its discrete counterpart, and the continuous non-autoregressive (CNAR) model is best on nearly every metric. The paper also claims that a fine-tuning stage with auxiliary losses on the reconstructed waveform improves DNSMOS, NISQA, PESQ, and POLQA consistently across all six paradigms.

What carries the argument

The central object is a unified decoder-only language-model architecture over neural-audio-codec latent features. A frozen codec and a frozen self-supervised speech encoder supply the inputs: either discrete tokens from a 12-codebook residual vector quantizer, or continuous dequantized features with 1024 dimensions per frame, plus frame-aligned semantic condition features. A 24-layer transformer with hidden size 1024 is trained either with causal attention for autoregressive variants or full attention for non-autoregressive variants to predict clean latents from noisy latents. The six paradigms are instantiations of this same container: cross-entropy over the 12 codebooks for discrete methods, L1 regression over the continuous features for continuous methods, a masked-token generation schedule for discrete diffusion, and a flow-matching velocity field for continuous flow matching. This shared substrate is what makes the comparison fair.

What would settle it

Train the discrete non-autoregressive model with an alternative token-flattening scheme that interleaves the 12 codebooks per frame, with a comparable parameter count and a doubled training budget; if it then matches or beats the continuous non-autoregressive model on PESQ and DNSMOS, the paper's central continuous-over-discrete conclusion would be falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a ranking: when all six LM-based generative enhancement paradigms share the same codec, the same frozen condition encoder, the same transformer size, the same training budget, and the same inference-time search over only a few settings, continuous-domain methods consistently beat discrete-token methods. The winning paradigm, CNAR, predicts the 1024-dimensional unquantized codec features in a single parallel pass, then quantizes and decodes them; it is top-ranked in all but one metric among the six. The paper attributes the gap to fault tolerance: regressing continuous features is a softer target than classifying 12 discrete tokens per frame, and quantization at the output reduces error accumulation. A further claim is that fine-tuning the LM with auxiliary time- and frequency-domain losses on the decoded waveform improves both non-intrusive and intrusive quality scores in every paradigm.

Load-bearing premise

The ranking assumes all six implementations are equally well tuned: same transformer size, same 150K training steps, same learning rate, and only inference settings tuned, so the result could be an artifact of settings that suit continuous paradigms more than discrete ones.

Editorial extensions

If this is right

  • New LM-based speech enhancers should start from continuous codec latents with a non-autoregressive decoder, since this combination is the strongest and also the fastest, requiring only a single parallel forward pass.
  • Discrete-token approaches should be compared against a continuous regression baseline with matched compute and architecture before claiming gains.
  • Auxiliary-loss fine-tuning that leaves the codec frozen is a cheap and general recipe that transfers across all six paradigms, so it can likely be appended to future generative speech-enhancement models.
  • Intrusive and non-intrusive metrics can move together in this setting: PESQ and POLQA improve alongside DNSMOS and NISQA, contrary to the common worry that generative models only help perceptual scores.

Reading between the lines

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

  • A testable extension would vary the number of codebooks in the discrete branch: if the continuous advantage comes from fault tolerance, increasing codebook count or using a flattened token sequence should narrow the gap.
  • Because the fine-tuning objective includes differentiable PESQ and STOI losses, the same recipe could be applied to other codec latent domains, such as semantic tokens, where the auxiliary losses may matter even more.
  • The non-autoregressive advantage suggests that for long-form audio, continuous non-autoregressive models may scale better than autoregressive token predictors, since latency does not grow linearly with sequence length.
  • If the fault-tolerance explanation is right, continuous models should degrade gracefully under reduced codec bitrate, whereas discrete models should show step-like collapses; that is a direct, checkable consequence.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. This manuscript presents a unified decoder-only LM framework for six generative speech enhancement (SE) paradigms built on DAC latent features: discrete/continuous autoregressive (D/ CAR), discrete/continuous non-autoregressive (D/CNAR), discrete diffusion (DDiff), and continuous flow matching (CFM). Using the URGENT 2025 challenge data splits and a fixed backbone, the authors compare the paradigms on intrusive and non-intrusive metrics, propose an auxiliary-loss fine-tuning stage with straight-through estimation for discrete models, and report that continuous-domain paradigms outperform their discrete counterparts, with CNAR as the overall best method. The paper also includes an ablation on codec choice and a small inference-hyperparameter study. Code is made publicly available.

Significance. The paper provides a useful synopsis of six LM-based generative SE paradigms under a single experimental framework, which the community has lacked. The use of an official held-out test set, training all models from scratch, and the control of training non-fine-tuned models for the same extra 50K steps are good experimental practices. If the ranking is robust, the design rule — continuous NAC latents, non-autoregressive decoding, and auxiliary-loss fine-tuning — would be directly actionable. The main significance is conditional on two unresolved points: whether the six implementations are comparably tuned, and whether the reported margins are statistically meaningful despite coming from single runs.

major comments (3)
  1. [Sec. 2.3(i), Sec. 3.1, Sec. 3.3] The central claim that continuous-domain paradigms consistently outperform discrete-domain paradigms in Table 2 rests on an unvalidated assumption that the discrete implementations are comparably tuned. Section 2.3(i) deliberately replaces the token-flattening strategy of [6] with a parallel sum of all C=12 codebook embeddings, and Sections 3.1 and 3.3 fix the same 24-layer/H=1024 transformer, 150K training steps, and one learning rate for every variant, tuning only inference settings such as beam size and N' on the validation set. The continuous-over-discrete gaps in Table 2 could therefore reflect this fixed discrete token-fusion and training budget rather than an intrinsic advantage of continuous latents. Please add ablations (e.g., flattened token sequences, alternative codebook fusion, or longer training for discrete variants) or otherwise demonstrate that the discrete baselines are not handicapped by these choices.
  2. [Table 2, Sec. 4.2] Every reported number appears to come from a single training run with no confidence intervals or significance tests. The claim that CNAR is top in all but one metric and second in the other relies on small margins in several cases (e.g., DNSMOS 3.01 vs. 2.95 for CAR; NISQA 3.39 vs. 3.46 for CFM). Likewise, some fine-tuning improvements are tiny (e.g., CFM-FT vs. CFM DNSMOS 2.96 vs. 2.95; CNAR-FT vs. CNAR DNSMOS 3.03 vs. 3.01), so the statement that fine-tuning 'consistently improves' DNSMOS, NISQA, PESQ, and POLQA is not statistically grounded. Please report variance across seeds or paired bootstrap confidence intervals, at least for the close comparisons.
  3. [Eqs. (4)–(5), Sec. 2.3(vi)] The text defines the velocity estimator as v_theta(X1, X0, C, t), but Eq. (4) writes v_theta(X1, X0, Y_c, t) and Eq. (5) uses the same argument. This is an internal notational inconsistency: it is unclear whether the flow-matching model is conditioned on the WavLM features C or on the discrete noisy tokens Y_c. Please correct the notation and make the conditioning of the CFM model explicit, as this matters for reproducing the results.
minor comments (3)
  1. [Table 1, Sec. 4.1] The sentence 'with a significant margin, DAC is the clear winner' is only justified for the intrusive metrics PESQ and ESTOI; on DNSMOS, NISQA, and UTMOS, BiCodec and the X-Codec variants score higher. Please state explicitly that DAC was selected for its intrusive-metric fidelity rather than overall perceptual quality.
  2. [Sec. 2.4, Table 2] Because the fine-tuning loss includes a differentiable PESQ term (delta * J_PESQ) and PESQ is among the reported metrics, the PESQ improvements are partly by construction. The support for the fine-tuning claim is stronger on the non-optimized metrics (DNSMOS, NISQA, POLQA), all of which do improve in Table 2; the paper could make this point more explicitly.
  3. [Sec. 2.3(ii)] The notation in the CAR description would benefit from distinguishing more clearly between the predicted unquantized features \hX and the quantized/dequantized version \h\tildeX used as the next autoregressive step's input; the current text uses \hX for both the training target and the inference-time prediction.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the six-paradigm comparison is empirical and self-contained; the only overlap is that fine-tuning optimizes a PESQ surrogate while PESQ is also reported, but independent metrics and non-load-bearing self-citation keep this minor.

full rationale

The paper makes no mathematical derivation claim: its principal assertions are empirical rankings from a shared experimental pipeline. All six paradigms use the same frozen DAC and WavLM front-ends, the same transformer scale, training budget, and data splits, so the continuous-vs-discrete comparison is an experimental observation rather than a quantity forced by construction. The only overlap between training objective and evaluation metric is that the fine-tuning loss (Eq. 7) includes differentiable PESQ and STOI terms, while Table 2 reports PESQ (and ESTOI). This makes the reported PESQ gains unsurprising, but it is not a fitted-parameter-as-prediction step: no test-set metric value enters training, and the same fine-tuning also consistently improves DNSMOS, NISQA, and POLQA, which are not optimized. The reference [10] to the authors' earlier DisContSE paper supplies only additional DDiff inference details; the current text already specifies the MaskGIT-style iterative remasking procedure, and DDiff is not the winning paradigm, so the self-citation is not load-bearing. No uniqueness theorem, ansatz smuggled in via citation, or renamed known result carries the argument. Concerns about equal tuning potential of the discrete variants are about experimental fairness, not circularity, and belong to correctness risk rather than to a circular step. Accordingly, no step in the paper's reasoning reduces by construction to its own inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's claims are empirical, so the ledger contains no invented entities and no mathematical derivation. The free parameters are hand-set fine-tuning weights and a validation-selected inference count. The main unstated costs are the assumptions of frozen, sufficient representations and comparably tuned implementations.

free parameters (5)
  • Alpha (spectral auxiliary loss weight) = 10.0
    Chosen by hand in Section 3.3 for the fine-tuning objective in Eq. (7); controls the contribution of Braun's STFT loss.
  • Gamma (time-domain L1 auxiliary loss weight) = 20.0
    Chosen by hand in Section 3.3; weights the L1 waveform loss in Eq. (7).
  • Delta (differentiable PESQ loss weight) = 0.5
    Chosen by hand in Section 3.3; adds metric-oriented pressure during fine-tuning.
  • Eta (differentiable STOI loss weight) = 1.0
    Chosen by hand in Section 3.3; weights the intelligibility-oriented training signal.
  • Inference iterations N' for DDiff and CFM = 3
    Selected on the validation set in Table 1; the reported D_test scores for diffusion and flow matching depend on this choice.
assumptions (4)
  • domain assumption The pretrained DAC and WavLM representations carry sufficient information for all six enhancement paradigms.
    Section 3.1 freezes both extractors; if either representation favors one paradigm, the ranking is biased.
  • domain assumption The seven objective metrics approximate speech quality and intelligibility well enough to rank systems.
    Section 3.2 uses only objective metrics, with no listening test; the conclusions inherit the metrics' biases.
  • ad hoc to paper Each of the six models is comparably tuned.
    Section 3.3 uses the same 24-layer transformer size, the same training budget, and single runs; only inference hyperparameters are tuned on validation, so the comparison assumes equal implementation quality.
  • domain assumption The URGENT 2025 data splits, minus the stated exclusions, cover the intended operating conditions.
    Section 3.2 excludes CommonVoice 19.0 and bandwidth limitation, so results may not transfer to those conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Language Model-Based Generative Speech Enhancement in the Latent Space of a Neural Audio Codec." pith.science (2026). https://pith.science/paper/UMFYPHPD

@misc{pith2026260812082,
  author       = {Pith},
  title        = {Pith review of: Rethinking Language Model-Based Generative Speech Enhancement in the Latent Space of a Neural Audio Codec},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UMFYPHPD}},
  note         = {Machine review of arXiv:2608.12082}
}
read the original abstract

Language model (LM)-based speech enhancement (SE) has recently emerged rapidly using latent space features of neural audio codecs (NACs). In this paper, first, we present a unified framework covering six popular LM-based generative SE modeling paradigms based on discrete/continuous latent NAC features: discrete or continuous autoregressive (D/CAR) SE, discrete or continuous non-autoregressive (D/CNAR) SE, discrete diffusion (DDiff) SE, and continuous flow matching (CFM) SE. Second, we are the first to compare their performance in a unified experimental setup and synopsis with diverse intrusive and non-intrusive metrics, enabling a fair and comprehensive evaluation. Third, we propose a fine-tuning strategy with auxiliary losses on reconstructed speech to improve both intrusive and non-intrusive metrics. Trained and evaluated on URGENT 2025 Speech Enhancement Challenge data splits, all continuous-domain paradigms excel their discrete-domain counterparts. The overall best approach turns out to be CNAR. We further show that our proposed auxiliary loss fine-tuning strategy helps to improve DNSMOS, NISQA, PESQ, and POLQA consistently in all six paradigms.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 40 canonical work pages

  1. [6]

    ACKNOWLEDGMENT Computational resources were provided by the German AI Service Center WestAI

    work page

  2. [1]

    INTRODUCTION Language model (LM)-based generative speech enhancement (SE) approaches have gained immense attention. Employing LMs in the enhancement process in a generative manner can harness their pow- erful semantic modeling capabilities, enabling better speech recon- struction beyond simple noise removal. LM-based SE typically takes discrete tokens or ...

    work page

  3. [2]

    METHODS 2.1. Neural Audio Codecs Neural audio codecs (NACs) [13, 14, 15, 16] learn to map a time- domain signal into compact sequences of discrete tokens via a resid- ual vector quantization (RVQ) scheme [17]. Lettingyandxrepre- sent the time-domain noisy and clean speech vectors, respectively, arXiv:2608.12082v1 [eess.AS] 12 Aug 2026 WavLMDACWavLM S N C ...

    work page Pith review arXiv 2026

  4. [3]

    Networks In the proposed SE models, the pretrained 16 kHzDAC[25] and WavLM[26] weights are adopted

    EXPERIMENTAL SETUP 3.1. Networks In the proposed SE models, the pretrained 16 kHzDAC[25] and WavLM[26] weights are adopted. ForDAC, the encoder output di- mension isD=1024, the codebook size isN=1024, and the number of codebooks isC=12, respectively. ForWavLM, the 6th encoder layer’s output is adopted [2] with dimensionD=1024. For the AR and NAR tasks, we...

    work page 2025

  5. [4]

    C... ”) are consistently outperforming their discrete-domain counterparts (methods named “D

    RESULTS AND DISCUSSION 4.1. Ablation on the Validation Set In Table 1, we perform an ablation study onD val. We first apply the recent popular NACs to the clean speech, includingDAC[14] with 12 codebooks, the single-codebookBiCodec[15],X-Codecwith 1 and 8 codebooks (X-Codec-1 andX-Codec-8), respectively. Concerning the non-intrusive metrics (DNSMOS, NISQA...

    work page

  6. [5]

    We are the first to de- liver a comprehensive and fair synopsis of the paradigms by employ- ing both non-intrusive and intrusive metrics

    CONCLUSIONS In this paper, we present a unified framework for six popular decoder-only LM-based generative SE modeling paradigms based on discrete/continuous latent NAC features. We are the first to de- liver a comprehensive and fair synopsis of the paradigms by employ- ing both non-intrusive and intrusive metrics. Continuous-domain paradigms excel discre...

    work page

  7. [7]

    Speech Enhancement Using Continuous Embeddings of Neural Audio Codec,

    Haoyang Li et al., “Speech Enhancement Using Continuous Embeddings of Neural Audio Codec,” inProc. of ICASSP, Hyderabad, India, Apr. 2025, pp. 1–5

    work page 2025

  8. [8]

    AnyEnhance: A Unified Generative Model with Prompt-Guidance and Self-Critic for V oice En- hancement,

    Junan Zhang et al., “AnyEnhance: A Unified Generative Model with Prompt-Guidance and Self-Critic for V oice En- hancement,”arXiv:2501.15417, pp. 1–5, Jan. 2025

    arXiv 2025

Show all 44 references
  1. [9]

    Llama: Open and Efficient Foundation Language Models,

    Hugo Touvron et al., “Llama: Open and Efficient Foundation Language Models,”arXiv:2302.13971, pp. 1–27, Feb. 2023

    2023 arXiv

  2. [10]

    GenSE: Generative Speech Enhancement via Language Models using Hierarchical Modeling,

    Jixun Yao et al., “GenSE: Generative Speech Enhancement via Language Models using Hierarchical Modeling,” inProc. of ICLR, Singapore, Apr. 2025, pp. 1–21

    2025

  3. [11]

    LLaSE-G1: Incentivizing Generalization Capability for Llama-based Speech Enhancement,

    Boyi Kang et al., “LLaSE-G1: Incentivizing Generalization Capability for Llama-based Speech Enhancement,” inProc. of ACL, Vienna, Austria, Jul. 2025, pp. 13292–13305

    2025

  4. [12]

    UniSE: A Unified Framework for Decoder-only Autoregressive LM-based Speech Enhance- ment,

    Haoyin Yan et al., “UniSE: A Unified Framework for Decoder-only Autoregressive LM-based Speech Enhance- ment,”arXiv:2510.20441, pp. 1–5, Oct. 2025

    2025 arXiv

  5. [13]

    Modeling Strategies for Speech En- hancement in the Latent Space of a Neural Audio Codec,

    Sofiene Kammoun et al., “Modeling Strategies for Speech En- hancement in the Latent Space of a Neural Audio Codec,” pp. 17407–17411, May 2026

    2026

  6. [14]

    as the NAC model, with 20 ms non-overlapping frames on a 16 kHz waveform, which is further justified in Section 4.1. 2.2. Condition Features Extractor We adopt the pre-trainedWavLM[18] encoder as the condition fea- ture extractor to provide auxiliary semantic information.WavLM...

  7. [15]

    High-Fidelity Speech Enhance- ment via Discrete Audio Tokens,

    Luca A. Lanzend ¨orfer et al., “High-Fidelity Speech Enhance- ment via Discrete Audio Tokens,” pp. 17397–17401, May 2026

    2026

  8. [16]

    Codec Does Matter: Exploring the Seman- tic Shortcoming of Codec for Audio Language Model,

    Zhen Ye et al., “Codec Does Matter: Exploring the Seman- tic Shortcoming of Codec for Audio Language Model,” in Proc. of Annual AAAI Conference on Artificial Intelligence, Kos, Greece, Feb. 2025, pp. 25697–25705

    2025

  9. [17]

    Genhancer: High-Fidelity Speech Enhance- ment via Generative Modeling on Discrete Codec Tokens,

    Haici Yang et al., “Genhancer: High-Fidelity Speech Enhance- ment via Generative Modeling on Discrete Codec Tokens,” in Proc. of Interspeech, Kos, Greece, Sep. 2024, pp. 1170–1174

    2024

  10. [18]

    DisContSE: Single-Step Diffusion Speech Enhancement Based on Joint Discrete and Continuous Embed- dings,

    Yihui Fu et al., “DisContSE: Single-Step Diffusion Speech Enhancement Based on Joint Discrete and Continuous Embed- dings,” inProc. of ICASSP, Barcelona, Spain, May 2026, pp. 17392–17396

    2026

  11. [19]

    Flow Matching for Generative Model- ing,

    Yaron Lipman et al., “Flow Matching for Generative Model- ing,” inProc. of ICLR, Kigali, Rwanda, May 2023, pp. 1–28

    2023

  12. [20]

    FlowDec: A Flow-Based Full-Band General Audio Codec with High Perceptual Quality,

    Simon Welker et al., “FlowDec: A Flow-Based Full-Band General Audio Codec with High Perceptual Quality,” inProc. of ICLR, Singapore, Apr. 2025, pp. 1–28

    2025

  13. [21]

    High Fidelity Neural Audio Com- pression,

    Alexandre D ´efossez et al., “High Fidelity Neural Audio Com- pression,”Transactions on Machine Learning Research, vol. 2023, pp. 1–20, Sep. 2023

    2023

  14. [22]

    High-Fidelity Audio Compression with Improved RVQGAN,

    Rithesh Kumar et al., “High-Fidelity Audio Compression with Improved RVQGAN,” inProc. of NeurIPS, New Orleans, LA, Dec. 2023, pp. 27980–27993

    2023

  15. [23]

    Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens,

    Xinsheng Wang et al., “Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens,”arXiv:2503.01710, pp. 1–22, Mar. 2025

    2025 arXiv

  16. [24]

    Understanding Straight-Through Esti- mator in Training Activation Quantized Neural Nets,

    Penghang Yin et al., “Understanding Straight-Through Esti- mator in Training Activation Quantized Neural Nets,” inProc. of ICLR, New Orleans, LA, May 2019, pp. 1–30

    2019

  17. [25]

    Residual Quantization with Implicit Neural Codebooks,

    Iris A. M. Huijben et al., “Residual Quantization with Implicit Neural Codebooks,” inProc. of ICML, Vienna, Austria, Jul. 2024, pp. 20682–20699

    2024

  18. [26]

    WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing,

    Sanyuan Chen et al., “WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing,”IEEE Jour- nal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, Oct. 2022

    2022

  19. [27]

    MaskGIT: Masked Generative Image Transformer,

    Huiwen Chang et al., “MaskGIT: Masked Generative Image Transformer,” inProc. of CVPR, New Orleans, LA, Jun. 2022, pp. 11315–11325

    2022

  20. [28]

    A Consolidated View of Loss Func- tions for Supervised Deep Learning-Based Speech Enhance- ment,

    Sebastian Braun et al., “A Consolidated View of Loss Func- tions for Supervised Deep Learning-Based Speech Enhance- ment,” inProc. of TSP, Brno, Czech Republic, Jul 2021, pp. 72–76

    2021

  21. [29]

    due to its occasional background noise. For training dataD train, we mainly follow the strategies in the challenge [28], in which six kinds of distortions are adopted: additive noise, reverberation, clip- ping, codec loss (MP3 and OGG), packet loss, and wind noise. We do not a...

  22. [30]

    EffCRN: An Efficient Convolutional Re- current Network for High-Performance Speech Enhancement,

    Marvin Sach et al., “EffCRN: An Efficient Convolutional Re- current Network for High-Performance Speech Enhancement,” inProc. of ICASSP, Dublin, Ireland, Aug. 2023, pp. 2508– 2512

    2023

  23. [31]

    Loss Function Inspired by the PESQ Score,

    Lorenz Schmidt et al., “Loss Function Inspired by the PESQ Score,” 2022–,https://github.com/audiolabs/ torch-pesq

    2022

  24. [32]

    PyTorch Implementation of STOI ,

    Manuel Pariente et al., “PyTorch Implementation of STOI ,” 2020–,https://github.com/mpariente/ pytorch_stoi

    2020

  25. [33]

    Descript Audio Codec (.dac): High-Fidelity Audio Compression with Improved RVQ- GAN,

    Rithesh Kumar et al., “Descript Audio Codec (.dac): High-Fidelity Audio Compression with Improved RVQ- GAN,” 2023–,https://github.com/descriptinc/ descript-audio-codec

    2023

  26. [34]

    WavLM-Large,

    Sanyuan Chen et al., “WavLM-Large,” 2023–,https:// huggingface.co/microsoft/wavlm-large

    2023

  27. [35]

    Attention Is All You Need,

    Ashish Vaswani et al., “Attention Is All You Need,” inProc. of NeurIPS, Long Beach, CA, Jan. 2017, pp. 1–11

    2017

  28. [36]

    Interspeech 2025 URGENT Speech En- hancement Challenge,

    Kohei Saijo et al., “Interspeech 2025 URGENT Speech En- hancement Challenge,” inProc. of Interspeech, Rotterdam, The Netherlands, Aug. 2025, pp. 858–862

    2025

  29. [37]

    Common V oice: A Massively- Multilingual Speech Corpus,

    Rosana Ardila et al., “Common V oice: A Massively- Multilingual Speech Corpus,” inProc. of LREC, Marseille, France, May. 2020, pp. 4218–4222

    2020

  30. [38]

    P .56: Objective Measurement of Active Speech Level, International Telecommunication Union, Telecommu- nication Standardization Sector (ITU-T), Dec

    ITU,Rec. P .56: Objective Measurement of Active Speech Level, International Telecommunication Union, Telecommu- nication Standardization Sector (ITU-T), Dec. 2011

    2011

  31. [39]

    Perceptual Evaluation of Speech Quality (PESQ)—A New Method for Speech Quality Assessment of Telephone Networks and Codecs,

    Antony W. Rix et al., “Perceptual Evaluation of Speech Quality (PESQ)—A New Method for Speech Quality Assessment of Telephone Networks and Codecs,” inProc. of ICASSP, Salt Lake City, UT, USA, May 2001, pp. 749–752

    2001

  32. [40]

    An Algorithm for Predicting the Intel- ligibility of Speech Masked by Modulated Noise Maskers,

    Jesper Jensen et al., “An Algorithm for Predicting the Intel- ligibility of Speech Masked by Modulated Noise Maskers,” IEEE/ACM T-ASLP, vol. 24, no. 11, pp. 2009–2022, Nov. 2016

    2009

  33. [41]

    DNSMOS P. 835: A Non- Intrusive Perceptual Objective Speech Quality Metric to Eval- uate Noise Suppressors,

    Chandan K. A. Reddy et al., “DNSMOS P. 835: A Non- Intrusive Perceptual Objective Speech Quality Metric to Eval- uate Noise Suppressors,” inProc. of ICASSP, Singapore, May 2022, pp. 886–890

    2022

  34. [42]

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

    Gabriel Mittag et al., “NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,” inProc. of Interspeech, Brno, Czech Republic, Aug. 2021, pp. 2127–2131

    2021

  35. [43]

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

    Takaaki Saeki et al., “UTMOS: UTokyo-SaruLab System for V oiceMOS Challenge 2022,” inProc. of Interspeech, Incheon, Korea, Sep. 2022, pp. 4521–4525

    2022

  36. [44]

    Evaluation Metrics for Generative Speech Enhancement Methods: Issues and Perspectives,

    Jan Pirklbauer et al., “Evaluation Metrics for Generative Speech Enhancement Methods: Issues and Perspectives,” in Proc. of ITG Speech Communication, Aachen, Germany, Sep. 2023, p. 265–269

    2023

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.