Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Improving Noise Robustness of LLM-based Zero-shot TTS via Discrete Acoustic Token Denoising

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Token-domain denoising restores LLM-based zero-shot TTS quality from noisy prompts.

desk verdict A compact token-domain denoiser for LLM-based zero-shot TTS that largely delivers on the TTS side, though the 'outperforms SOTA SE' claim is not established by the evidence. read the letter →

arxiv 2505.13830 v2 pith:ZHJGOSJM submitted 2025-05-20 eess.AS cs.SD

Ye-Xin Lu , Hui-Peng Du , Fei Liu , Yang Ai , Zhen-Hua Ling This is my paper · ORCID
classification eess.AScs.SD
keywords zero-shotTTSnoise-robustspeechenhancementcodecdenoisingdiscreteacoustictokensresidualvectorquantizationLLM-basedembeddingrefiner
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

LLM-based zero-shot text-to-speech systems preserve the acoustic environment of the supplied prompt, so a noisy recording of a few seconds degrades the synthesized voice and intelligibility. This paper argues that the cleanest place to attack that problem is the discrete acoustic-token domain, not the waveform. It proposes a codec-based denoiser whose token denoiser predicts the first two groups of clean acoustic tokens from all groups of noisy tokens, with an embedding refiner that reconstructs the summed clean embedding for waveform recovery. Paired with an LLM-based TTS system, the enhanced tokens restore synthesis to the level of a clean prompt across subjective and objective metrics, while also beating two speech enhancement baselines in perceptual quality at roughly a third of the compute. The central claim is that the first two residual-vector-quantizer groups of the codec carry enough speaker and prosodic information to serve both denoising and TTS prompting.

What carries the argument

The carrier of the argument is the codec denoiser, built around a pre-trained audio codec with $K$ residual vector quantizers (RVQs). It has three learned components: a token denoiser, which sums the embeddings of all $K$ noisy token groups and passes them through $N_{\mathrm{TD}}$ Conformer blocks to predict the probability distribution of the first two clean token groups, followed by an argmax; an embedding refiner, which concatenates the predicted first-two-group embedding with the summed noisy embedding and uses $N_{\mathrm{ER}}$ Conformer blocks to predict the summed clean embedding $\sum_{k=1}^K E(\hat{y})_k$; and the codec decoder, which turns that embedding into the enhanced waveform. Training jointly optimizes a cross-entropy loss on the two predicted token groups and an embedding-refinement loss combining $\ell^1$ and Frobenius terms, with teacher forcing that randomly feeds clean tokens. The load-bearing design choice is restricting prediction to the first two RVQ groups, which keeps the prediction task easy and, the paper reports, is also the point of highest denoising quality.

What would settle it

Take a pre-trained codec denoiser and feed it prompts corrupted by a noise source absent from its training data, such as background music or room reverberation at a signal-to-noise ratio of -5 dB, then synthesize speech and compare speaker similarity and word error rate against the TTS system given the clean prompt. A significant drop would indicate that the two-group token representation is insufficient under distribution shift; a more direct test is to measure, via information-theoretic probing, how much speaker identity is retained in the first two token groups relative to later groups.

Watch

Extended reading notes

Core claim

The paper's central discovery is that residual vector quantization itself acts as a noise bottleneck: the codec's first two token groups retain the speaker and prosodic essentials, and later groups add reconstruction detail that is both harder to predict and unnecessary for the TTS prompt. The authors train a Conformer token denoiser to map every group of noisy tokens to the first two groups of clean tokens, then an embedding refiner to predict the summed clean embedding from those tokens plus the noisy full embedding, and finally the codec decoder restores the clean waveform. In their comparisons, predicting exactly two token groups gives the best perceptual scores; predicting more groups lowers quality and raises FLOPs, and the two-group design outperforms two strong signal-domain speech enhancement baselines. Integrated into the TTS system, the denoised tokens yield synthesized speech that matches the clean-prompt reference on naturalness, speaker similarity, and intelligibility, while a waveform-enhancement front end leaves a speaker-similarity gap.

Load-bearing premise

The whole method rests on the premise that the first two residual vector quantization groups of the codec already contain enough clean acoustic information to preserve speaker identity and essential prosody, for both waveform reconstruction and TTS prompting; if more groups are actually needed, the denoiser and the conditioned TTS both degrade.

Editorial extensions

If this is right

  • The integrated TTS system, given only a noisy prompt, produces speech whose naturalness, speaker similarity, and intelligibility match the same system given a clean prompt, closing the gap left by waveform-level enhancement.
  • The codec denoiser scores higher on a perceptual speech-quality metric than two signal-domain speech enhancement baselines while needing roughly a third of their FLOPs.
  • Predicting only the first two codec token groups yields better denoised quality than predicting 4, 8, 16, or 32 groups, making the extra groups pure cost in this design.
  • Because the enhanced tokens are used directly as the TTS prompt, the approach avoids the artifacts that signal-domain enhancement injects into the prompt, which is why speaker similarity stays high.
  • The full noise-robust pipeline adds 1.10 GFLOPs over the TTS model, far below the 38.93 GFLOPs charged by the waveform-enhancement front end.

Reading between the lines

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

  • If the two-group bottleneck is the real source of the gain, the same token-denoising front end should transfer to other codec-conditioned generative models, such as voice conversion or speech-to-speech translation, with little re-tuning.
  • The method's noise coverage in training and testing comes from distinct noise corpora and a limited SNR range, so its real-world robustness depends on the noise types seen; adversarial or non-stationary noise would be a stress test the paper does not run.
  • The claim that the first two groups contain essential acoustic information is inferred from an ablation, not directly measured; a probe of token-group information content would tell whether the design generalizes to codecs with different VQ counts or codebook sizes.
  • Because the embedding refiner receives the noisy summed embedding as a condition, the model may learn to suppress noise rather than reconstruct clean content; comparing its behavior on clean inputs would reveal whether it introduces unnecessary alterations.
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

4 major / 5 minor

Summary. The paper proposes a noise-robust zero-shot TTS system, NR-LauraTTS, in which a codec-based denoiser operates in the discrete acoustic token domain before the prompt is fed to the LLM-based TTS backbone LauraTTS. The codec denoiser uses an Encodec encoder to obtain all RVQ token groups, a Conformer token denoiser that predicts only the first two clean token groups, and an embedding refiner that estimates the summed clean acoustic embedding, which is then decoded into a waveform or used directly as the acoustic prompt for LauraTTS. Experiments compare the codec denoiser with CMGAN and MP-SENet using DNSMOS P.835 and FLOPs, and compare NR-LauraTTS with LauraTTS under clean, noisy, and MP-SENet-enhanced prompts using MOS, SMOS, SECS, WER, CER, and additional FLOPs. The authors report that the codec denoiser outperforms state-of-the-art SE methods and that NR-LauraTTS approaches clean-prompt performance with lower complexity than a signal-domain SE front-end.

Significance. If the reported results are robust, the paper makes a useful contribution: it demonstrates that denoising in the token domain is a viable and efficient alternative to signal-domain enhancement for LLM-based zero-shot TTS, and the two-group prediction design is well aligned with LauraTTS's prompting mechanism. The paper has several concrete strengths: the architecture is clearly described, the comparison includes a clean-prompt upper bound and a strong signal-domain SE baseline, the evaluation is performed on a held-out test set with no circularity in the loss or test procedure, and a public demo page is provided. The main weaknesses are that the closest token-domain SE baselines are not evaluated, the SE results are reported without confidence intervals or conventional metrics, and the wording of the TTS claims is stronger than the small differences in Table 3 support. These issues are addressable and do not invalidate the central idea, but they currently limit the reliability of the headline claims.

major comments (4)
  1. [§4.1, Table 1] The claim that the codec denoiser 'outperforms state-of-the-art SE methods' is not established because the closest competitors, the token-domain SE models MaskSR [18] and Genhancer [19], are cited in Section 2.2 but never evaluated. These are the models whose design decision (predicting tokens from all RVQ groups) the authors explicitly argue against, so a comparison with them is necessary to support both the SOTA claim and the two-group design. Please add these baselines, or restrict the claim to 'outperforms CMGAN and MP-SENet'.
  2. [§4.1, Table 1] The SE evaluation reports only DNSMOS P.835 scores with no confidence intervals, significance tests, or conventional objective metrics such as PESQ, STOI, or SI-SNR. The reported margins over MP-SENet are small (SIG 3.62 vs. 3.56, BAK 4.11 vs. 4.09, OVRL 3.36 vs. 3.30), and DNSMOS is known to vary across utterances and noise conditions. Please report error bars or statistical tests and, ideally, corroborating objective metrics before claiming superiority.
  3. [§4.2, Table 3] The abstract and conclusion state that NR-LauraTTS 'surpasses' the approach using additional SE models, but Table 3 shows MOS 4.02 vs. 4.01 with identical 95% confidence intervals, WER 2.44 vs. 2.54, and CER 1.27 vs. 1.41; only SMOS and SECS are numerically higher. No paired significance tests are reported, so the 'surpasses' wording is too strong. Either provide significance testing or rewrite the claim to say 'comparable to the SE-based approach with improved speaker similarity'.
  4. [§2.1, Table 2] The design premise that the first two RVQ groups contain sufficient acoustic information (speaker identity and essential prosody) is supported only by the DNSMOS ablation in Table 2, whose differences are small (OVRL 3.36 for two groups vs. 3.35 for one and four groups, 3.33 for eight) and are reported without variance. This does not establish that two groups are sufficient for faithful voice cloning or for the embedding refiner. Please add a direct analysis of token-group information content, evaluate downstream TTS speaker similarity when different numbers of groups are predicted, or present the two-group choice as an empirical design choice with appropriate uncertainty rather than as a validated property of Encodec.
minor comments (5)
  1. [§3.1] The sentence 'we utilized the 580-hour training set of the 15-hour development set' appears to contain a grammatical error; presumably the 580-hour training set and the 15-hour development set were both used.
  2. [§2.2.2 and §3.1] The number of embedding-refiner Conformer blocks is denoted N_ER in the methodology but reported as 'NCV' in Section 3.1; please make the notation consistent.
  3. [Tables 1 and 2] The DNSMOS results are given without indicating the number of test utterances or the SNR breakdown of the test set; please report the test-set size and, if possible, per-condition results.
  4. [§4.1] The statement that 'the VQs of the codec act as information bottlenecks, allowing the quantized discrete acoustic tokens to filter out some noise' is presented as an explanation but is not directly tested; consider presenting it as a hypothesis rather than a confirmed mechanism.
  5. [Figures 1 and 2] The figures are dense and the two-group output is not immediately easy to identify; adding explicit labels in the captions for the RVQ dimension and the 'first two groups' path would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proposed codec denoiser is a supervised model trained on noisy-clean pairs and evaluated on a held-out test set; the first-two-token-groups design is an empirically tested choice, not an input re-labeled as output.

full rationale

I walked the derivation chain. The codec denoiser's objective (Eq. 1-3) is a standard cross-entropy/embedding loss against clean Encodec tokens and embeddings obtained from the target clean speech; this is a supervised mapping, not a quantity defined in terms of its own prediction. The claim that the first two RVQ groups suffice is stated in Sec. 2.1 as a property of Encodec RVQ and validated by the ablation in Table 2, not assumed into existence by the method. The TTS evaluation compares the full system against LauraTTS with clean/noisy prompts and LauraTTS+MP-SENet on held-out utterances; none of these comparisons are inputs to the training objective. The only same-author comparator, MP-SENet [14], is a baseline, not a load-bearing premise; citing it does not make the result circular. Weaknesses such as DNSMOS without error bars, no significance tests, and the omission of token-domain baselines MaskSR and Genhancer in Table 1 are robustness/correctness concerns, not examples of a prediction reducing to a fit or to a self-citation chain. I therefore find no specific circular step and score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central design rests on an empirical property of RVQ codecs (initial token groups carry speaker and most acoustic content) and on the adequacy of DNSMOS as a quality metric; these are plausible domain assumptions but not verified in this paper. The free parameters are hyperparameters and the chosen number of predicted token groups (2), which was selected from an ablation. No new entities are introduced.

free parameters (3)
  • Loss weights (lambda1, lambda2) = 1.0, 0.5
    Chosen by hand; weighted sum of CE and ER losses in Eq. (3); not justified by a grid search or validation protocol in the paper.
  • Number of predicted token groups = 2
    Selected based on DNSMOS results across models predicting 1,2,4,8,16,32 groups (Table 2); used for the main NR-LauraTTS, so this is a hand-chosen value affecting a central design.
  • Conformer block counts (N_TD, N_ER) = 12, 6
    Architecture hyperparameters set by the authors; no search described.
assumptions (4)
  • domain assumption In Encodec's RVQ, the first two quantizer groups capture enough acoustic information, including speaker identity, to reconstruct clean speech and support TTS.
    Invoked in Sec. 2.1 to motivate predicting only the first two groups; no direct evidence is provided beyond the ablation in Table 2, which itself is the basis for choosing this number.
  • domain assumption The VQs of the codec act as information bottlenecks that filter out some noise, making token-domain denoising easier than signal-level enhancement.
    Stated in Sec. 4.1 analysis as the reason for superiority; not directly tested.
  • domain assumption DNSMOS P.835 is a valid proxy for subjective speech quality in this comparison.
    Used as the sole objective SE metric in Table 1; its reliability for token-reconstructed audio is not discussed.
  • domain assumption The authors' re-implementation of LauraTTS is faithful to the published LauraTTS and its one-step codec vocoder.
    The paper says 'we pre-trained LauraTTS using LibriLight' (Sec. 3.1) without releasing it, so results depend on an unreleased implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Noise Robustness of LLM-based Zero-shot TTS via Discrete Acoustic Token Denoising." pith.science (2026). https://pith.science/paper/ZHJGOSJM

@misc{pith2026250513830,
  author       = {Pith},
  title        = {Pith review of: Improving Noise Robustness of LLM-based Zero-shot TTS via Discrete Acoustic Token Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZHJGOSJM}},
  note         = {Machine review of arXiv:2505.13830}
}
read the original abstract

Large language model (LLM) based zero-shot text-to-speech (TTS) methods tend to preserve the acoustic environment of the audio prompt, leading to degradation in synthesized speech quality when the audio prompt contains noise. In this paper, we propose a novel neural codec-based speech denoiser and integrate it with the advanced LLM-based TTS model, LauraTTS, to achieve noise-robust zero-shot TTS. The proposed codec denoiser consists of an audio codec, a token denoiser, and an embedding refiner. The token denoiser predicts the first two groups of clean acoustic tokens from the noisy ones, which can serve as the acoustic prompt for LauraTTS to synthesize high-quality personalized speech or be converted to clean speech waveforms through the embedding refiner and codec decoder. Experimental results show that our proposed codec denoiser outperforms state-of-the-art speech enhancement (SE) methods, and the proposed noise-robust LauraTTS surpasses the approach using additional SE models.

Figures

Figures reproduced from arXiv: 2505.13830 by the authors.

Figure 1
Figure 1. Noise-robust zero-shot TTS synthesis process of the proposed NR-LauraTTS, where ⃝S , ⃝T , and ⃝E denote the “start of sequence”, “turn of speech”, and “end of sequence” tokens. proaches. Research on how to improve the noise robustness of LLM-based zero-shot TTS is still limited. A common ap￾proach involves employing a speech enhancement (SE) model to denoise the noisy audio prompt before feeding it to the TTS models… view at source ↗
Figure 2
Figure 2. Overall structure of the proposed codec denoiser. waveform. Additionally, we integrate the codec denoiser into LauraTTS to construct a noise-robust zero-shot TTS system, NR-LauraTTS, which is capable of synthesizing high-quality personalized speech from noisy audio prompts. Experimental results demonstrate that our proposed codec denoiser surpasses SOTA SE methods in enhanced speech quality and model’s computational… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 23 canonical work pages

  1. [18]

    Speech enhancement of noisy and reverberant speech for text-to-speech,

    C. Valentini-Botinhao and J. Yamagishi, “Speech enhancement of noisy and reverberant speech for text-to-speech,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 26, no. 8, pp. 1420–1433, 2018

    2018

  2. [19]

    CMGAN: Conformer-based metric GAN for speech enhancement,

    R. Cao, S. Abdulatif, and B. Yang, “CMGAN: Conformer-based metric GAN for speech enhancement,” inProc. Interspeech, 2022, pp. 936–940

    work page 2022

  3. [1]

    Improving Noise Robustness of LLM-based Zero-shot TTS via Discrete Acoustic Token Denoising

    Introduction Zero-shot text-to-speech (TTS) synthesis [1–5] aims to syn- thesize any speaker’s voice based on few seconds of audio prompt of the speaker. Typically trained on extensive high- quality speech data, these models face challenges in real-life scenarios where obtaining clean audio prompts is challenging. Recorded audio prompts are often distorte...

    work page Pith review arXiv 2025

  4. [2]

    Methodology In this section, we first elaborate on the noise-robust zero-shot TTS synthesis process of the NR-LauraTTS model, and then describe in detail the model structure and training criteria of the proposed codec denoiser. 2.1. Noise-Robust LauraTTS The overall noise-robust zero-shot TTS synthesis process of the proposed NR-LauraTTS is depicted in Fi...

    work page

  5. [3]

    test-clean

    Experiments 3.1. Dataset and Experimental Setup Due to the high training cost of LLM-based TTS models, we pre-trained LauraTTS using the LibriLight dataset [21], which contains about 60,000 hours of English speech data from over 7,000 speakers, and constructed a noisy-clean dataset based on the LibriTTS-R dataset [22] to train the codec denoiser. For trai...

    work page 2022

  6. [4]

    Results on the SE Task The experimental results of the SE task are depicted in Table 1

    Results and Analysis 4.1. Results on the SE Task The experimental results of the SE task are depicted in Table 1. Overall, the proposed codec denoiser outperformed the SOTA SE methods CMGAN and MP-SENet among all the DNSMOS metrics, demonstrating that the speech enhanced by the pro- posed codec denoiser surpassed those of CMGAN and MP- SENet in speech qua...

    work page

  7. [5]

    Conclusion In this paper, we proposed a codec denoiser for high-quality SE in the discrete acoustic token domain, and integrated it into the LauraTTS model to construct a noise-robust zero-shot TTS system, NR-LauraTTS. The codec denoiser predicted the first two groups of enhanced acoustic tokens through a token de- noiser, and used an embedding refiner to...

    work page

  8. [6]

    Acknowledgements This work was funded by the National Nature Science Foun- dation of China under Grant U23B2053 and 62301521, and the Anhui Provincial Natural Science Foundation under Grant 2308085QF200

    work page

Show all 34 references
  1. [7]

    SC- GlowTTS: An efficient zero-shot multi-speaker text-to-speech model,

    E. Casanova, C. Shulby, E. G ¨olge, N. M. M¨uller, F. S. De Oliveira, A. C. Junior, A. d. S. Soares, S. M. Aluisio, and M. A. Ponti, “SC- GlowTTS: An efficient zero-shot multi-speaker text-to-speech model,” in Proc. Interspeech, 2021, pp. 3645–3649

    2021

  2. [8]

    YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone,

    E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G ¨olge, and M. A. Ponti, “YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone,” in Proc. ICML , 2022, pp. 2709–2720

    2022

  3. [9]

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

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al. , “Neural codec language mod- els are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111, 2023

    2023 arXiv

  4. [10]

    LauraGPT: Listen, attend, understand, and regenerate audio with GPT,

    Z. Du, J. Wang, Q. Chen, Y . Chu, Z. Gao, Z. Li, K. Hu, X. Zhou, J. Xu, Z. Ma et al., “LauraGPT: Listen, attend, understand, and regenerate audio with GPT,” arXiv preprint arXiv:2310.04673 , 2023

    2023 arXiv

  5. [11]

    Naturalspeech 3: Zero-shot speech syn- thesis with factorized codec and diffusion models,

    Z. Ju, Y . Wang, K. Shen, X. Tan, D. Xin, D. Yang, Y . Liu, Y . Leng, K. Song, S. Tang et al., “Naturalspeech 3: Zero-shot speech syn- thesis with factorized codec and diffusion models,”arXiv preprint arXiv:2403.03100, 2024

    2024 arXiv

  6. [12]

    Noise-robust zero-shot text-to-speech synthesis conditioned on self-supervised speech-representation model with adapters,

    K. Fujita, H. Sato, T. Ashihara, H. Kanagawa, M. Delcroix, T. Moriya, and Y . Ijima, “Noise-robust zero-shot text-to-speech synthesis conditioned on self-supervised speech-representation model with adapters,” in Proc. ICASSP, 2024, pp. 11 471–11 475

    2024

  7. [13]

    DINO-VITS: Data-efficient zero-shot TTS with self-supervised speaker verification loss for noise robustness,

    V . Pankov, V . Pronina, A. Kuzmin, M. Borisov, N. Usolt- sev, X. Zeng, A. Golubkov, N. Ermolenko, A. Shirshova, and Y . Matveeva, “DINO-VITS: Data-efficient zero-shot TTS with self-supervised speaker verification loss for noise robustness,” in Proc. Interspeech, 2024, pp. 697–701

    2024

  8. [14]

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

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

    2021

  9. [15]

    High fidelity neural audio compression,

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

    2023

  10. [16]

    Speech enhancement for a noise-robust text-to-speech synthesis system using deep recurrent neural networks,

    C. V . Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “Speech enhancement for a noise-robust text-to-speech synthesis system using deep recurrent neural networks,” inProc. Interspeech, 2016, pp. 352–356

    2016

  11. [17]

    In- vestigating RNN-based speech enhancement methods for noise- robust text-to-speech

    C. Valentini-Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “In- vestigating RNN-based speech enhancement methods for noise- robust text-to-speech.” in Proc. SSW, 2016, pp. 146–152

    2016

  12. [20]

    MP-SENet: A speech enhance- ment model with parallel denoising of magnitude and phase spec- tra,

    Y .-X. Lu, Y . Ai, and Z.-H. Ling, “MP-SENet: A speech enhance- ment model with parallel denoising of magnitude and phase spec- tra,” in Proc. Interspeech, 2023, pp. 3834–3838

    2023

  13. [21]

    Explicit estimation of magnitude and phase spectra in parallel for high-quality speech enhancement,

    ——, “Explicit estimation of magnitude and phase spectra in parallel for high-quality speech enhancement,” arXiv preprint arXiv:2308.08926, 2023

    2023 arXiv

  14. [22]

    How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,

    K. Iwamoto, T. Ochiai, M. Delcroix, R. Ikeshita, H. Sato, S. Araki, and S. Katagiri, “How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,” in Proc. Inter- speech, 2022, pp. 5418–5422

    2022

  15. [23]

    Learning to enhance or not: Neural network-based switching of enhanced and observed signals for overlapping speech recognition,

    H. Sato, T. Ochiai, M. Delcroix, K. Kinoshita, N. Kamo, and T. Moriya, “Learning to enhance or not: Neural network-based switching of enhanced and observed signals for overlapping speech recognition,” in Proc. ICASSP, 2022, pp. 6287–6291

    2022

  16. [24]

    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,” in Proc. Interspeech, 2024, pp. 2275–2279

    2024

  17. [25]

    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,” in Proc. Interspeech, 2024, pp. 1170–1174

    2024

  18. [26]

    Conformer: Convolution- augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu et al. , “Conformer: Convolution- augmented transformer for speech recognition,” in Proc. Inter- speech, 2020, pp. 5036–5040

    2020

  19. [27]

    Libri-light: A benchmark for ASR with limited or no su- pervision,

    J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu, P.-E. Mazar´e, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegen et al., “Libri-light: A benchmark for ASR with limited or no su- pervision,” in Proc. ICASSP, 2020, pp. 7669–7673

    2020

  20. [28]

    Libritts-R: A re- stored multi-speaker text-to-speech corpus,

    Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, M. Bacchiani, Y . Zhang, W. Han, and A. Bapna, “Libritts-R: A re- stored multi-speaker text-to-speech corpus,” in Proc. Interspeech, 2023, pp. 5496–5500

    2023

  21. [29]

    ICASSP 2022 deep noise suppression challenge,

    H. Dubey, V . Gopal, R. Cutler, A. Aazami, S. Matusevych, S. Braun, S. E. Eskimez, M. Thakker, T. Yoshioka, H. Gamper et al., “ICASSP 2022 deep noise suppression challenge,” in Proc. ICASSP, 2022, pp. 9271–9275

    2022

  22. [30]

    WHAM!: Extending 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!: Extending speech separation to noisy environments,” in Proc. Interspeech, 2019, pp. 1368–1372

    2019

  23. [31]

    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,” in Proc. ICASSP, 2024, pp. 591–595

    2024

  24. [32]

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

    C. K. Reddy, V . Gopal, and R. Cutler, “DNSMOS P. 835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in Proc. ICASSP, 2022, pp. 886–890

    2022

  25. [33]

    Resemblyzer,

    G. Louppe, “Resemblyzer,” https://github.com/resemble-ai/ Resemblyzer, 2019

    2019

  26. [34]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” in Proc. ICML, 2023, pp. 28 492–28 518

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.