Pith. sign in

REVIEW 3 major objections 3 minor 35 references

Generative Data Augmentation Challenge: Zero-Shot Speech Synthesis for Personalized Speech Enhancement

T0 review · 3 major / 3 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that zero-shot TTS synthesized speech, even from low-quality voice clones, enables personalized speech enhancement models to outperform generalist models, while true clean speech remains the best training target.

desk verdict Useful challenge baseline with a real confound in the headline claim: TTS-augmented models beat the generalist, but the design also adapts to speaker-specific noises, so the effect isn't isolated. read the letter →

arxiv 2501.13372 v1 pith:KV6PU2E3 submitted 2025-01-23 eess.AS cs.AI

Jae-Sung Bae , Anastasia Kuznetsova , Dinesh Manocha , John Hershey , Trausti Kristjansson , Minje Kim This is my paper · ORCID
classification eess.AScs.AI
keywords zero-shottext-to-speechpersonalizedspeechenhancementgenerativedataaugmentationspeakersimilarityvoicecloningsyntheticquality
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 launches a two-stage challenge: use a zero-shot text-to-speech (TTS) system, which clones a speaker's voice from a short enrollment clip, to generate unlimited clean training speech for a personalized speech enhancement (PSE) model, then measure whether the quality of the synthetic speech determines how well the PSE model cleans noisy audio for that speaker in its own noise environment. The baseline experiments establish the central empirical claim: every PSE model fine-tuned on synthetic cloned speech outperformed the generalist model on every metric across all three model sizes, even when the TTS system had the lowest measured quality. They also show that data quality beats quantity: a model fine-tuned on only six minutes of real clean speech beat every model trained on thirty minutes of synthetic speech. Among the synthetic systems, the one with the highest speaker similarity produced the best enhancement and intelligibility scores, while the one with the highest perceived naturalness produced the best perceptual quality score. The paper's conclusion is that zero-shot TTS is a viable route around the privacy and recording problems of personalization, but the ceiling is set by how faithfully synthetic data reproduces the target speaker.

What carries the argument

The mechanism is a two-phase pipeline. In phase one, a zero-shot TTS model takes a short enrollment utterance from a target speaker plus a text sentence and produces a clean synthetic utterance in that speaker's voice; the synthetic utterance becomes the denoising target. In phase two, a ConvTasNet-based PSE model is first trained as a generalist on LibriSpeech and FSD50K, then fine-tuned per speaker by mixing the synthetic clean targets with that speaker's five assigned noise types from MUSAN and training with a negative SDR loss. The reported correlation chain is what carries the argument: SECS and WER of the TTS output track SDRI, SDR, and eSTOI of the PSE model, while UTMOS tracks PESQ, so the downstream task is a functional test of generative data quality.

What would settle it

Take a target speaker, generate personalized PSE training data with a zero-shot TTS system using a correctly matched enrollment clip, and train the PSE model. Then repeat the identical fine-tuning recipe using an enrollment clip from a different speaker, or from the same speaker but with the speaker embedding deliberately scrambled, while keeping text, noise, and count identical. If the mismatched-identity model still beats the generalist by the same margin, the benefit of augmentation is not speaker personalization but exposure to synthetic speech patterns. The same experiment can be run in the other direction: if matched enrollment outperforms scrambled enrollment, speaker similarity carries the effect.

Watch

Extended reading notes

Core claim

The discovery the paper argues for is that speaker similarity and intelligibility of synthetic speech, rather than sheer volume, are the factors that drive downstream personalized speech enhancement. Across the medium, small, and tiny PSE models, the generalist model had the lowest scores on SDRI, SDR, eSTOI, and PESQ, and every model fine-tuned on TTS-generated utterances did better. The oracle model trained on 40 real utterances (about six minutes) per speaker outperformed all 30-minute synthetic-data models, showing the quality ceiling. Among zero-shot TTS baselines, SpeechT5 was best at speaker similarity and word error rate and produced the best SDRI, SDR, and eSTOI in the 30-minute setting, while XTTS, with the best perceptual quality, produced the best PESQ. These results are presented as evidence for the challenge's hypothesis: higher-quality augmented speech makes better personalized enhancers, and TTS metrics are useful but do not all predict downstream gains equally.

Load-bearing premise

The load-bearing premise is that a synthetic utterance cloned from a short enrollment clip is a faithful enough stand-in for the target speaker's real clean speech that a PSE model trained on synthetic targets will transfer to the speaker's actual voice in real test mixtures; if synthetic artifacts or identity errors go unnoticed by the SECS, UTMOS, and WER metrics, the measured gains could come from learning synthetic-speech patterns rather than true personalization.

Editorial extensions

If this is right

  • Even the weakest zero-shot TTS baseline, with the worst measured naturalness or intelligibility, yields a PSE model that beats the speaker- and noise-agnostic generalist on all four metrics, so voice cloning is a practical alternative to collecting private user recordings.
  • Adding five times more synthetic utterances (30 minutes vs. 6 minutes) produces only marginal gains, so the return on generating more synthetic data is small once quality is held fixed.
  • Speaker similarity and intelligibility are the TTS attributes that most strongly predict enhancement quality (SDRI, SDR, eSTOI); perceived naturalness specifically predicts PESQ, so no single TTS metric should be used to rank augmentation systems.
  • Training on six minutes of real clean speech outperforms all 30-minute synthetic datasets, establishing the quality ceiling that future TTS augmentation systems would need to close.
  • Virtual speakers show the same ranking as real speakers, which suggests privacy-preserving synthetic personas can stand in for real target speakers in the personalization pipeline.

Reading between the lines

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

  • A natural extension the paper does not run is to filter synthetic utterances by their SECS or UTMOS scores before fine-tuning; if quality is the binding constraint, discarding low-score clones should move the 30-minute models closer to the GT-6min oracle at no extra recording cost.
  • The cross-model metric dissociation (best similarity system winning SDRI/SDR/eSTOI while best naturalness system wins PESQ) suggests that a downstream task like PSE could be used as a task-specific TTS benchmark, and that an ensemble or multi-objective TTS optimizing similarity and naturalness jointly might dominate both.
  • The results imply that voice-cloned augmentation may be most useful not to replace real data when it exists, but to bootstrap personalization for speakers with no available recordings, reducing the privacy burden to a single short enrollment clip.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. This paper describes a challenge, accompanying the Generative Data Augmentation workshop at ICASSP 2025, in which participants build zero-shot TTS systems to synthesize personalized speech and then use that synthetic speech to fine-tune personalized speech enhancement (PSE) models. The authors provide baseline experiments using three open-source zero-shot TTS systems (YourTTS, SpeechT5, XTTS), evaluating the synthesized speech with SECS, UTMOS, and WER, and the resulting PSE models with SDRI, SDR, eSTOI, and PESQ on real and virtual speakers. The central empirical claims are that all TTS-augmented PSE models outperform a generalist model across model sizes, that fine-tuning on ground-truth clean speech (GT-6min) outperforms all synthetic-data models, and that better TTS speaker similarity and intelligibility are associated with better PSE performance.

Significance. If the central claims hold, the paper provides a useful benchmark and baseline for an emerging application of generative data augmentation, and the challenge itself may catalyze community progress. The release of baseline code and checkpoints, the inclusion of virtual speakers as a privacy-preserving option, and the use of multiple TTS and PSE metrics are concrete strengths. However, the empirical evidence for the main claims currently rests on a confounded comparison and on single-run evaluations with no statistical significance assessment, which limits the conclusions that can be drawn from the baseline experiments.

major comments (3)
  1. [Section V-C, Tables IV and V] The claim that every TTS-augmented PSE model outperforms the generalist model does not isolate the effect of personalized synthetic speech, because each fine-tuned model is trained and evaluated on a speaker-specific set of five MUSAN noise types (Section III-C), whereas the generalist model is evaluated on the same test mixtures without any adaptation to those noise types. As the authors acknowledge in Section V-C ('the adaptation to the noise sources could have contributed to better PSE performance'), the improvement over the generalist could be driven by noise-environment adaptation alone. To support the paper's central conclusion, the authors should add a control condition in which the generalist model is fine-tuned on the same speaker-specific noise sets but with non-personalized (e.g., arbitrary-speaker) clean targets, or otherwise demonstrate that the observed gains are not attributable to noise adaptation.
  2. [Section V-C, Tables IV and V] All PSE results are reported as single numbers with no variance estimates, repeated seeds, or statistical tests. Many of the comparative statements in this section rely on small differences; for example, real-world medium SpeechT5-30min achieves SDRI 12.519 versus 12.302 for XTTS-6min and 12.341 for XTTS-30min. These gaps may be within run-to-run variation. The authors should provide means and standard deviations across multiple fine-tuning seeds and, ideally, per-speaker paired comparisons or significance tests before claiming that one TTS system is better than another on a given PSE metric.
  3. [Section V-C, Discussion] The claim that 'speaker similarity and intelligibility emerged as the most relevant factors' for PSE performance is supported only by an informal ranking of three TTS models: SpeechT5 has the highest SECS and lowest WER, and its 30min models achieve the best SDRI/SDR/eSTOI in several configurations. With only three TTS systems, this correlation is anecdotal, and the relationship is not quantified (e.g., no correlation coefficient across conditions or speakers). A more cautious interpretation, or an explicit correlational analysis over speakers and TTS conditions, is needed before this factor-based claim can stand.
minor comments (3)
  1. [Table I] The column header 'Descriptoin' is a typo and should read 'Description'.
  2. [Section V-C, paragraph on PESQ] The phrase 'PESQ focuses on perpetual quality' should read 'perceptual quality'.
  3. [Reference [1]] The reference for Whisper lists the venue as 'Proc. Int. Conf. on Machine Learning (ICLR)'; the paper by Radford et al. was published at ICML 2023, not ICLR.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; the paper's claims are empirically evaluated against held-out clean speech.

full rationale

This is an empirical challenge-baseline paper rather than a derivation or prediction from first principles. Its central claims—that zero-shot TTS augmented fine-tuning outperforms a generalist PSE model and that higher-quality TTS data (GT-6min) yields the best PSE results—are supported by direct measurements on held-out ground-truth clean speech using standard external metrics (SDRI, SDR, eSTOI, PESQ) and open-source TTS models (YourTTS, SpeechT5, XTTS). The baseline PSE architecture and fine-tuning recipe are adopted from the authors' own prior work [9], [32], but this self-citation is transparent, non-load-bearing, and not used to justify the central result. Section V-C explicitly acknowledges the noise-adaptation confound ('the adaptation to the noise sources could have contributed to better PSE performance'), which is an experimental limitation, not a definitional or self-citational circularity. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported, and no known result is renamed. Therefore no circular step is identified.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central findings rest on several domain assumptions about metric validity, synthetic-to-real transfer, and the chosen fine-tuning recipe. There are no invented physical or theoretical entities; the free parameters are experimental design choices (dataset sizes, hyperparameters) rather than fitted constants. The most important unverified premise is that synthetic speech is a faithful training surrogate for real target-speaker speech.

free parameters (5)
  • PSE fine-tuning corpus sizes (40 and 220 utterances per speaker) = 40 and 220 utterances (about 6 and 30 minutes)
    The comparison between 6min and 30min conditions, and the conclusion that extra synthetic data gives only marginal gains, depends on these chosen sizes.
  • Fine-tuning hyperparameters (learning rate, batch size, early-stopping patience, SNR range) = lr=1e-6, batch=8, patience=20, SNR in [-5,5] dB
    Chosen following [9] and [32]; not swept, and may affect which TTS-generated data helps the downstream PSE model.
  • Number of enrollment utterances per speaker (1) and enrollment length (3-14 s) = 1 utterance, 3-14 seconds
    Zero-shot TTS cloning quality and hence PSE augmentation quality depend on this design choice.
  • Number of text prompts per speaker for TTS evaluation and training (50, with 40 used for GT-6min) = 50 prompts total; 40 for GT-6min
    The training and evaluation splits are determined by these choices.
  • Number of validation utterances for early stopping (10 for 6min, 30 for 30min) = 10 and 30
    Early stopping behavior and final model selection depend on these quantities.
assumptions (6)
  • domain assumption SECS (x-vector cosine similarity), WER from Whisper, and UTMOS are valid proxies for speaker similarity, intelligibility, and perceptual quality when choosing data for downstream PSE training.
    The paper defines TTS quality with these metrics in Section IV-A and uses the resulting scores to interpret PSE outcomes in Section VI.
  • domain assumption Synthetic clean speech is a faithful training target for the target speaker's real speech, so a PSE model fine-tuned on synthetic targets will improve on real test utterances.
    Used throughout Section V-B and V-C; this is the main unverified premise behind TTS data augmentation for PSE.
  • domain assumption The ConvTasNet architecture and the generalist-then-fine-tune recipe from [9], [32] are a valid testbed for comparing data augmentation methods.
    The challenge fixes this model; conclusions about TTS utility may not transfer to other PSE architectures.
  • domain assumption LibriTTS test-clean speakers and MUSAN noise selections are representative of real personalization scenarios.
    The challenge and baselines use these corpora to define target speakers and noise environments (Section III).
  • domain assumption Virtual speakers generated with Meta's unspecified TTS system are a meaningful proxy for diverse human speakers.
    Used for the privacy-preserving setting in Section III-B; no details on the model are given.
  • domain assumption SDRI, SDR, eSTOI, and PESQ against held-out clean speech measure PSE performance adequately for comparing data augmentation methods.
    These metrics define the challenge evaluation in Section IV-B.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Data Augmentation Challenge: Zero-Shot Speech Synthesis for Personalized Speech Enhancement." pith.science (2026). https://pith.science/paper/KV6PU2E3

@misc{pith2026250113372,
  author       = {Pith},
  title        = {Pith review of: Generative Data Augmentation Challenge: Zero-Shot Speech Synthesis for Personalized Speech Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KV6PU2E3}},
  note         = {Machine review of arXiv:2501.13372}
}
read the original abstract

This paper presents a new challenge that calls for zero-shot text-to-speech (TTS) systems to augment speech data for the downstream task, personalized speech enhancement (PSE), as part of the Generative Data Augmentation workshop at ICASSP 2025. Collecting high-quality personalized data is challenging due to privacy concerns and technical difficulties in recording audio from the test scene. To address these issues, synthetic data generation using generative models has gained significant attention. In this challenge, participants are tasked first with building zero-shot TTS systems to augment personalized data. Subsequently, PSE systems are asked to be trained with this augmented personalized dataset. Through this challenge, we aim to investigate how the quality of augmented data generated by zero-shot TTS models affects PSE model performance. We also provide baseline experiments using open-source zero-shot TTS models to encourage participation and benchmark advancements. Our baseline code implementation and checkpoints are available online.

Figures

Figures reproduced from arXiv: 2501.13372 by the authors.

Figure 1
Figure 1. Overall flow of our challenge. (a) First, participants augment the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 26 canonical work pages

  1. [1]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. Mcleavey et al., “Robust speech recognition via large-scale weak supervision,” in Proc. Int. Conf. on Machine Learning (ICLR) , 2023

    work page 2023

  2. [2]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” in Proc. Int. Conf. on Machine Learning (ICLR) , 2021

    work page 2021

  3. [3]

    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 et al. , “YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone,” in Proc. Int. Conf. on Machine Learning (ICML), 2022

    work page 2022

  4. [4]

    XTTS: A massively multilingual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea et al. , “XTTS: A massively multilingual zero-shot text-to-speech model,” in Proc. Interspeech, 2024

    work page 2024

  5. [5]

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

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu et al. , “Neural codec language models are zero-shot text to speech synthesizers,” CoRR, vol. abs/2301.02111, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2301.02111

    arXiv doi:10.48550/arxiv.2301.02111 2023

  6. [6]

    Metric- GAN+: An improved version of MetricGAN for speech enhancement,

    S.-W. Fu, C. Yu, T.-A. Hsieh, P. Plantinga, M. Ravanelli et al., “Metric- GAN+: An improved version of MetricGAN for speech enhancement,” in Proc. Interspeech, 2021

    work page 2021

  7. [7]

    MANNER: Multi-view attention netwfork for noise erasure,

    H. J. Park, B. H. Kang, W. Shin, J. S. Kim, and S. W. Han, “MANNER: Multi-view attention netwfork for noise erasure,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , 2022

    work page 2022

  8. [8]

    Text is all you need: Personalizing ASR models using controllable speech synthesis,

    K. Yang, T.-Y . Hu, J.-H. R. Chang, H. Swetha Koppula, and O. Tuzel, “Text is all you need: Personalizing ASR models using controllable speech synthesis,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , 2023

    work page 2023

Show all 35 references
  1. [9]

    The potential of neural speech synthesis-based data augmentation for personalized speech en- hancement,

    A. Kuznetsova, A. Sivaraman, and M. Kim, “The potential of neural speech synthesis-based data augmentation for personalized speech en- hancement,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), 2023

    2023

  2. [10]

    A survey on image data augmen- tation for deep learning,

    C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmen- tation for deep learning,” Journal of Big Data , vol. 6, p. 60, 2019

    2019

  3. [11]

    Effec- tive data augmentation with diffusion models,

    B. Trabucco, K. Doherty, M. A. Gurinas, and R. Salakhutdinov, “Effec- tive data augmentation with diffusion models,” in Proc. Int. Conf. on Learning Representations (ICLR) , 2024

    2024

  4. [12]

    SpecAug- ment: A simple data augmentation method for automatic speech recog- nition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph et al. , “SpecAug- ment: A simple data augmentation method for automatic speech recog- nition,” in Proc. Interspeech, 2019

    2019

  5. [13]

    Latent filling: Latent space data augmentation for zero-shot speech synthesis,

    J.-S. Bae, J. Y . Lee, J.-H. Lee, S. Mun, T. Kang et al. , “Latent filling: Latent space data augmentation for zero-shot speech synthesis,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , 2024

    2024

  6. [14]

    Improving few-shot learning for talking face system with TTS data augmentation,

    Q. Chen, Z. Ma, T. Liu, X. Tan, Q. Lu et al. , “Improving few-shot learning for talking face system with TTS data augmentation,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , 2023

    2023

  7. [15]

    Utilizing TTS synthesized data for efficient development of keyword spotting model,

    H. J. Park, D. Agarwal, N. Chen, R. Sun, K. Partridge et al. , “Utilizing TTS synthesized data for efficient development of keyword spotting model,” CoRR, vol. abs/2407.18879, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2407.18879

    2024 doi

  8. [16]

    Zero shot text to speech augmentation for automatic speech recognition on low-resource accented speech corpora,

    F. Nespoli, D. Barreda, and P. A. Naylor, “Zero shot text to speech augmentation for automatic speech recognition on low-resource accented speech corpora,” in 57th Asilomar Conference on Signals, Systems, and Computers, 2023

    2023

  9. [17]

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

    H. Zen, R. Clark, R. J. Weiss, V . Dang, Y . Jiaet al., “LibriTTS: A corpus derived from librispeech for text-to-speech,” in Proc. Interspeech, 2019

    2019

  10. [18]

    MUSAN: A music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “MUSAN: A music, speech, and noise corpus,” CoRR, vol. abs/1510.08484, 2015. [Online]. Available: http://arxiv.org/abs/1510.08484

    2015 arXiv

  11. [19]

    SpeechBrain: A general-purpose speech toolkit,

    M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell et al. , “SpeechBrain: A general-purpose speech toolkit,” CoRR, vol. abs/2106.04624, 2021. [Online]. Available: https://arxiv.org/abs/2106. 04624

    2021 arXiv

  12. [20]

    Deep MOS predictor for synthetic speech using cluster-based modeling,

    Y . Choi, Y . Jung, and H. Kim, “Deep MOS predictor for synthetic speech using cluster-based modeling,” in Proc. Interspeech, 2020

    2020

  13. [21]

    NISQA: A deep CNN- self-attention model for multidimensional speech quality prediction with crowdsourced datasets,

    G. Mittag, B. Naderi, A. Chehadi, and S. Moller, “NISQA: A deep CNN- self-attention model for multidimensional speech quality prediction with crowdsourced datasets,” in Proc. Interspeech, 2021

    2021

  14. [22]

    The singing voice conversion challenge 2023,

    W.-C. Huang, L. P. Violeta, S. Liu, J. Shi, and T. Toda, “The singing voice conversion challenge 2023,” in Proc. IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) , 2023

    2023

  15. [23]

    NaturalSpeech 3: Zero- shot speech synthesis with factorized codec and diffusion models,

    Z. Ju, Y . Wang, K. Shen, X. Tan, D. Xin et al., “NaturalSpeech 3: Zero- shot speech synthesis with factorized codec and diffusion models,” in Proc. Int. Conf. on Machine Learning (ICML) , 2024

    2024

  16. [24]

    The T05 system for the V oiceMOS challenge 2024: Transfer learning from deep image classifier to naturalness MOS prediction of high-quality synthetic speech,

    K. Baba, W. Nakata, Y . Saito, and H. Saruwatari, “The T05 system for the V oiceMOS challenge 2024: Transfer learning from deep image classifier to naturalness MOS prediction of high-quality synthetic speech,” CoRR, vol. abs/2409.09305, 2024. [Online]. Available: https://doi.o...

    2024 doi arXiv

  17. [25]

    The V oiceMOS challenge 2024: Beyond speech quality prediction,

    W. Huang, S. Fu, E. Cooper, R. E. Zezario, T. Toda, H. Wang, J. Yamagishi, and Y . Tsao, “The V oiceMOS challenge 2024: Beyond speech quality prediction,” CoRR, vol. abs/2409.07001, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2409.07001

    2024 doi arXiv

  18. [26]

    Performance measurement in blind audio source separation,

    E. Vincent, R. Gribonval, and C. Fevotte, “Performance measurement in blind audio source separation,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 14, no. 4, pp. 1462–1469, 2006

    2006

  19. [27]

    A short- time objective intelligibility measure for time-frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “A short- time objective intelligibility measure for time-frequency weighted noisy speech,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), 2010

    2010

  20. [28]

    Perceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,

    A. Rix, J. Beerends, M. Hollier, and A. Hekstra, “Perceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,” in Proc. IEEE Int. Conf. on Acoustics, Speech, and Signal Processing. (ICASSP) , 2001

    2001

  21. [29]

    SpeechT5: Unified- modal encoder-decoder pre-training for spoken language processing,

    J. Ao, R. Wang, L. Zhou, C. Wang, S. Ren et al. , “SpeechT5: Unified- modal encoder-decoder pre-training for spoken language processing,” in Proc. of the 60th Annual Meeting of the Association for Computational Linguistics, 2022

    2022

  22. [30]

    Better speech synthesis through scaling,

    J. Betker, “Better speech synthesis through scaling,” CoRR, vol. abs/2305.07243, 2023. [Online]. Available: https://doi.org/10.48550/ arXiv.2305.07243

    2023 doi arXiv

  23. [31]

    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 Processing , vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  24. [32]

    Efficient personalized speech enhancement through self-supervised learning,

    A. Sivaraman and M. Kim, “Efficient personalized speech enhancement through self-supervised learning,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1342–1356, 2022

    2022

  25. [33]

    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,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , 2015

    2015

  26. [34]

    FSD50K: An open dataset of human-labeled sound events,

    E. Fonseca, X. Favory, J. Pons, F. Font, and X. Serra, “FSD50K: An open dataset of human-labeled sound events,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 829–852, 2022

    2022

  27. [35]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. Int. Conf. on Learning Representations (ICLR) , 2015. [Online]. Available: http://arxiv.org/abs/1412.6980

    2015 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.