Pith. sign in

REVIEW 3 major objections 6 minor 35 references

Advanced Zero-Shot Text-to-Speech for Background Removal and Preservation with Controllable Masked Speech Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A one-bit control signal lets a zero-shot TTS model either erase or preserve background audio.

desk verdict Solid incremental methods paper with fair baselines, but the 'strong generalization' claim outruns the single out-of-domain noise test. read the letter →

arxiv 2502.07345 v1 pith:WBQWCC76 submitted 2025-02-11 cs.SD eess.AS

Leying Zhang , Wangyou Zhang , Zhengyang Chen , Yanmin Qian This is my paper · ORCID
classification cs.SDeess.AS
keywords zero-shottext-to-speechflowmatchingbackgroundremovalpreservationcontrollablemaskedspeechpredictiondualspeakerencodernoise-robustTTSacousticcontrol
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

The paper aims to give zero-shot text-to-speech a single switch: given a speech prompt polluted by background noise, reverberation, or another talker, the same model can either strip the background and speak in the prompt speaker's voice, or keep the background and speak inside that same acoustic scene. It proposes Controllable Masked Speech Prediction, where a flow-matching acoustic model is trained with two complementary masked-prediction targets—one predicting clean speech, the other predicting the augmented speech—and a binary control signal selects which target is active. A dual speaker encoder, with one branch per task, is added so that timbre extraction and background modeling are trained separately rather than tangled together. If the claims hold, this makes background handling a user-level choice rather than a fixed pre-processing step, which matters for applications ranging from audiobook production to conversational agents that need environmental context.

What carries the argument

The central object is the Controllable Masked Speech Prediction (CMSP) loss, which merges background removal and preservation into one masked-prediction objective with two conditional terms. Both terms use the flow-matching training objective of VoiceBox, but one predicts the clean target from the augmented prompt while the other predicts the augmented target from the same prompt, with a binary control signal $c$ deciding which term is active. The dual speaker encoder is the supporting mechanism: two small transformer encoders process the noisy prompt, and the control signal routes the model to the encoder for the active task, preventing timbre and background information from interfering during multi-task training.

What would settle it

A concrete test: present the trained model with prompts containing musical background or multi-speaker babble—two classes absent from training—and measure whether the one-bit control still produces a continuous preserved background (B-MOS) and a clean removal (MOS). If the control collapses or the background is only active during voiced segments, the paper's generalization claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the contribution is a training strategy rather than a new architecture: replace the single masked speech prediction loss of the VoiceBox-style flow-matching model with the Controllable Masked Speech Prediction loss. The model is trained on clean speech and on the same speech augmented with environmental noise, reverberation, or an interfering speaker; the removal branch predicts the clean target from the noisy prompt under control signal c0, and the preservation branch predicts the augmented target under c1. At test time, flipping the control signal flips the behavior. With this setup the authors report that their model beats the VoiceBox, VoiceBox plus speech enhancement, and VoiceBox plus masked speech denoising baselines on speaker similarity and quality for removal, and beats VoiceBox on background similarity, including out-of-domain VCTK-TUT noise; it also produces backgrounds that persist through silences, unlike the baseline.

Load-bearing premise

The load-bearing premise is that the simulated training backgrounds (WHAM! noise, RIR-NOISES reverberation, and LibriTTS interference at the stated SNRs) are representative enough of real acoustic backgrounds that a one-bit control learned on them transfers to unseen scenarios.

Editorial extensions

If this is right

  • With one binary flag at inference, a user can request either clean speech or the same speech embedded in the prompt's original acoustic scene, with no retraining or model switching.
  • The preserved background remains audible through silent pauses in the generated speech, fixing an inconsistency the authors identify in VoiceBox.
  • The single model covers three degradation types—environmental noise, reverberation, and interfering speakers—and reports higher speaker similarity than the SE-augmented and MSD baselines on the removal task.
  • On out-of-domain noise, the model keeps some removal and preservation ability, and prepending a universal speech enhancer further improves OOD speaker similarity.
  • For interfering speakers, the model partially clones the second talker's timbre (0.20 SIM after separation), pointing toward unified multi-talker generation.

Reading between the lines

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

  • We infer that the binary control mechanism may serve as a general audio foreground/background switch beyond speech, for example letting a singing-voice or music model choose between dry and room-matched output, because the two branches impose a task-level disentanglement.
  • Because the model can emit both clean and environment-matched versions of the same prompt, it could become a cheap paired-data generator for training speech enhancement or separation models, with perfectly matched content.
  • A natural test of the architecture's inductive bias is to probe the two encoder branches: if the disentanglement is real, the c0 branch should encode little background energy and the c1 branch should encode the background spectrum, which is checkable with standard linear probes.
  • The authors defer music backgrounds to future work; until then, generalization to non-stationary or babble-like backgrounds remains unverified, so the 'strong generalization' claim should be read as covering the tested additive-noise and reverberation conditions.
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 / 6 minor

Summary. The paper proposes a zero-shot text-to-speech system that jointly handles background removal and background preservation from a noisy prompt, controlled by a binary control signal c. The method, Controllable Masked Speech Prediction (CMSP), extends the VoiceBox flow-matching objective with two target losses: one predicts the clean mel-spectrogram from an augmented prompt with c0, the other predicts the augmented mel-spectrogram with c1 (Eq. 3). A dual speaker encoder, with branches selected by c, is added to improve speaker timbre robustness. Experiments on LibriTTS with WHAM noise, RIR-NOISES reverberation, and LibriTTS interfering speakers show improvements over reimplemented VoiceBox, VoiceBox+SE, and VoiceBox+MSD baselines on SIM, MCD, MOS, and B-MOS, plus one out-of-domain test (VCTK-TUT noise). The paper claims precise control and strong generalization.

Significance. If the results hold, the paper makes a useful contribution: a single flow-matching TTS model whose output background can be switched between removed and preserved via a one-bit control signal, with publicly comparable baselines (all reimplemented with the same backbone and data, Sec. IV-C). The controlled baseline setup is a strength, as is the inclusion of both objective and subjective metrics. However, the significance of the claims is limited by the lack of statistical validation, the absence of a direct control-signal toggle experiment, and the narrow scope of the out-of-domain evaluation. With those addressed, the work could be a solid advance in controllable speech synthesis.

major comments (3)
  1. [Section IV-D, Tables I–II] No confidence intervals, error bars, or significance tests are reported for any objective metric (SIM, MCD) or subjective MOS/B-MOS. The MOS test uses only 15 listeners rating 10 utterances per scenario (Sec. IV-D), and several reported differences are small (e.g., Table I clean MOS: 3.96 vs 3.90; clean SIM: 0.59 vs 0.59). Without variance estimates or significance statements, the claim of "obviously higher speech quality" over baselines (Sec. V-A) is not statistically supported. The paper should report per-condition variances or confidence intervals and, where appropriate, significance tests for the key comparisons.
  2. [Section III-A and Section V] The central claim of "precise control" is not directly validated: no experiment toggles the control signal c0↔c1 while keeping the prompt, text, and initial noise fixed. Tables I and II compare different systems (Ours vs baselines) under either removal or preservation, but they do not show that flipping c in the trained model changes the output in the intended direction. A within-model comparison with identical inputs and only c varied, measured by background-preservation metrics or listening, is required to establish that the binary signal is causally controlling the output rather than the CMSP loss merely improving both tasks jointly.
  3. [Section V-C] The out-of-domain (OOD) evaluation is limited to one additive-noise condition (VCTK-TUT). No OOD test is reported for reverberation or interfering speakers, although the abstract and contribution list claim "strong generalization capabilities in unseen scenarios" and "various acoustic conditions." The in-domain reverb and interfering-speaker tests use the same RIR-NOISES and LibriTTS interference distributions as training, so they do not test generalization beyond the training manifold. The "strong generalization" claim is therefore established only for a noise type similar to the training noise, and the experiments should be expanded or the claim appropriately qualified.
minor comments (6)
  1. [Eq. (3)] The labels "Backgroud Removal Loss" and "Backgroud Preservation Loss" contain a typo: they should read "Background Removal Loss" and "Background Preservation Loss." The equation label "LCM SP" also has an extra space.
  2. [Eq. (4)] The unconditional term in the classifier-free guidance is written as "˜v_t(w, c; θ)" but, following Eq. (2), all conditioning variables (including c) should be dropped, so the term should likely be "˜v_t(w; θ)" or the notation should be clarified.
  3. [Section I and Section II-A] There are typos: "controll background preservation" should be "control background preservation" (in Section I), and "these models are sensible for speech prompts" should likely be "these models are sensitive to speech prompts" (in Section II-A).
  4. [Section III-B] The phrase "minimize computational cost in the model interference between the removal and preservation tasks" is awkward; it should probably be "minimize interference between the removal and preservation tasks" or similar.
  5. [Table II] The table header appears misaligned: the mapping between the condition columns (Clean, Noise, Reverb, Interference, VCTK-TUT) and the metric columns (SIM, MCD, BMOS) is unclear, and the "Noisy GT" row includes values for Clean that are not well-defined for the preservation task on clean prompts. The table should be reformatted for readability.
  6. [Section IV-D] The speaker embedding extraction cites both [32] and [33] for WavLM; verify that [32] is the correct reference or restrict the citation to the WavLM paper [33].

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the dual losses define distinct targets, and overlapping-author citations are not load-bearing.

full rationale

This is an experimental systems paper whose central mechanism is the Controllable Masked Speech Prediction loss in Eq. 3. The two terms define genuinely different regression targets: the removal term predicts the clean sample x from the augmented prompt x_aug_ctx, and the preservation term predicts the augmented sample x_aug itself. Neither target is defined in terms of the model's own output, a fitted evaluation quantity, or the final claim, so the core 'prediction' is not equivalent to its input by construction. The binary control signal c is an input that selects the task, and no parameter is fitted to the reported test sets in a way that would make the results forced. The overlapping-author references are used for hyperparameters ([4]), a baseline SE module ([7]), and a feature extractor for evaluation ([32]); none supplies the paper's main claim as an unexamined premise. In particular, the SE baseline [7] is used as a comparison system, and the paper explicitly reports that this baseline achieves higher OOD speaker similarity than the proposed method, which is not the behavior of a stacked circular argument. The out-of-domain generalization claim rests on only one VCTK-TUT noise condition, but this is an external-validity limitation rather than a circularity, because the reported improvement is not reduced to the training augmentations by construction. No circular step can be exhibited from the paper's equations or citation chain.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities are postulated. The free parameters are training-distribution and inference choices that are not fitted to the evaluation metrics in a circular way, but they are hand-set and could influence the conclusions. The main assumptions are domain-level: the simulation proxy, the metric validity, the control-gate efficacy, and the small listening tests.

free parameters (3)
  • Augmentation probabilities P_N, P_R, P_IS, P_C = 0.2, 0.2, 0.2, 0.4
    Set by hand in Section IV-B to balance clean and degraded training samples; no sensitivity analysis, so the reported behavior may depend on this choice.
  • SNR ranges for simulated noise and interference = Noise: -5 to 10 dB; interference: 1 to 10 dB
    Chosen in Section IV-A; these ranges fix the difficulty of the training and evaluation conditions.
  • Classifier-free guidance strength alpha = Not stated; follows prior work [4]
    Equation 4 uses alpha; the paper defers hyperparameters to CoVoMix [4], so the exact value used in inference is not in this paper, which affects output quality and the reported metrics.
assumptions (5)
  • standard math Flow matching and ODE solvers produce mel-spectrograms that the HiFiGAN vocoder can convert to intelligible speech
    Assumed from VoiceBox framework [21]; Equations 1-2 in Section II-A are taken from prior work without re-derivation.
  • domain assumption The simulated augmentations with WHAM noise, RIR-NOISES reverberation, and single-speaker LibriTTS interference are a valid proxy for real acoustic backgrounds
    Section IV-A defines the training data; the generalization claim in Section V-C rests on this proxy, but only one OOD noise set (VCTK-TUT) is tested.
  • ad hoc to paper A one-bit binary control signal can gate two generation targets in a single model without cross-task interference
    This is the core design of CMSP and the dual speaker encoder (Section III); it is supported only by ablations, not by a mechanism or theoretical argument.
  • domain assumption WavLM embeddings measure speaker similarity robustly for both clean and noisy generated audio
    Section IV-D uses WavLM cosine similarity as SIM for both tasks; if WavLM is biased by background, SIM does not isolate timbre.
  • domain assumption MOS and B-MOS from 15 listeners rating 10 utterances per scenario provide stable quality measurements
    Section IV-D describes the listening tests; no confidence intervals or significance testing are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advanced Zero-Shot Text-to-Speech for Background Removal and Preservation with Controllable Masked Speech Prediction." pith.science (2026). https://pith.science/paper/WBQWCC76

@misc{pith2026250207345,
  author       = {Pith},
  title        = {Pith review of: Advanced Zero-Shot Text-to-Speech for Background Removal and Preservation with Controllable Masked Speech Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBQWCC76}},
  note         = {Machine review of arXiv:2502.07345}
}
read the original abstract

The acoustic background plays a crucial role in natural conversation. It provides context and helps listeners understand the environment, but a strong background makes it difficult for listeners to understand spoken words. The appropriate handling of these backgrounds is situation-dependent: Although it may be necessary to remove background to ensure speech clarity, preserving the background is sometimes crucial to maintaining the contextual integrity of the speech. Despite recent advancements in zero-shot Text-to-Speech technologies, current systems often struggle with speech prompts containing backgrounds. To address these challenges, we propose a Controllable Masked Speech Prediction strategy coupled with a dual-speaker encoder, utilizing a task-related control signal to guide the prediction of dual background removal and preservation targets. Experimental results demonstrate that our approach enables precise control over the removal or preservation of background across various acoustic conditions and exhibits strong generalization capabilities in unseen scenarios.

Figures

Figures reproduced from arXiv: 2502.07345 by the authors.

Figure 1
Figure 1. Model Architecture i.e. x aug = x is denoted as P C . A random portion of the speech is masked and to be predicted, noted as x˜ and x˜ aug for samples before and after augmentation. The remaining unmasked segments are the prompt x aug ctx . We design a multi-task learning loss as shown in Eq.3 based on the conventional flow-matching loss in Eq.1, with notations described in Section.II-A. By introducing binary contro… view at source ↗
Figure 2
Figure 2. Mean Speaker Similarity Comparison by SNRs for Each Model [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Mel-spectrogram Comparison for Generated Speech with Preserved [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 29 canonical work pages

  1. [1]

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

    Z. Ju, Y . Wang, K. Shen, X. Tan, D. Xin, D. Yang, E. Liu, Y . Leng, K. Song, S. Tang et al. , “Naturalspeech 3: Zero-shot speech synthesis with factorized codec and diffusion models,” in Forty-first International Conference on Machine Learning , 2024

    work page 2024

  2. [2]

    Neural codec language models 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 models are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111 , 2023

    arXiv 2023

  3. [3]

    NaturalSpeech 2: Latent diffusion models are natural and zero- shot speech and singing synthesizers,

    K. Shen, Z. Ju, X. Tan, E. Liu, Y . Leng, L. He, T. Qin, J. Bian et al., “NaturalSpeech 2: Latent diffusion models are natural and zero- shot speech and singing synthesizers,” in The Twelfth International Conference on Learning Representations , 2023

    work page 2023

  4. [4]

    CoV oMix: Advancing zero-shot speech generation for human-like multi-talker conversations,

    L. Zhang, Y . Qian, L. Zhou, S. Liu, D. Wang, X. Wang, M. Yousefi, Y . Qian, J. Li, L. He et al. , “CoV oMix: Advancing zero-shot speech generation for human-like multi-talker conversations,” Proceedings of the 38th International Conference on Neural Information Processing Systems, 2024

    work page 2024

  5. [5]

    An investigation of noise robustness for flow- matching-based zero-shot tts,

    X. Wang, S. E. Eskimez, M. Thakker, H. Yang, Z. Zhu, M. Tang, Y . Xia, J. Li, S. Zhao, J. Li et al., “An investigation of noise robustness for flow- matching-based zero-shot tts,” arXiv preprint arXiv:2406.05699 , 2024

    arXiv 2024

  6. [6]

    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

    work page 2018

  7. [7]

    Toward universal speech enhancement for diverse input conditions,

    W. Zhang, K. Saijo, Z.-Q. Wang, S. Watanabe, and Y . Qian, “Toward universal speech enhancement for diverse input conditions,” in 2023 IEEE Automatic Speech Recognition and Understanding Workshop . IEEE, 2023, pp. 1–6

    work page 2023

  8. [8]

    DDTSE: Discriminative diffusion model for target speech extraction,

    L. Zhang, L. Y . Qian, L. Yu, H. Wang, H. Yang, S. Liu, L. Zhou, and Y . Qian, “DDTSE: Discriminative diffusion model for target speech extraction,” in IEEE Spoken Language Technology Workshop , 2024

    work page 2024

Show all 35 references
  1. [9]

    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,” arXiv preprint arXiv:2201.06685 , 2022

    2022 arXiv

  2. [10]

    DRSpeech: Degradation- Robust Text-to-Speech Synthesis with Frame-Level and Utterance-Level Acoustic Representation Learning,

    T. Saeki, K. Tachibana, and R. Yamamoto, “DRSpeech: Degradation- Robust Text-to-Speech Synthesis with Frame-Level and Utterance-Level Acoustic Representation Learning,” in Proc. Interspeech 2022, 2022, pp. 793–797

    2022

  3. [11]

    Denoispeech: Denoising text to speech with frame-level noise modeling,

    C. Zhang, Y . Ren, X. Tan, J. Liu, K. Zhang, T. Qin, S. Zhao, and T.- Y . Liu, “Denoispeech: Denoising text to speech with frame-level noise modeling,” in IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2021, pp. 7063–7067

    2021

  4. [12]

    Noise-robust zero-shot text-to-speech synthesis condi- tioned 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 condi- tioned on self-supervised speech-representation model with adapters,” in IEEE International Conference on Acoustics, Speech and Signal P...

    2024

  5. [13]

    An experimental comparison of noise- robust text-to-speech synthesis systems based on self-supervised repre- sentation,

    X. Zhao, Q. Zhu, and Y . Hu, “An experimental comparison of noise- robust text-to-speech synthesis systems based on self-supervised repre- sentation,” in IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2024, pp. 11 441–11 445

    2024

  6. [14]

    V oiceLDM: Text-to- speech with environmental context,

    Y . Lee, I. Yeon, J. Nam, and J. S. Chung, “V oiceLDM: Text-to- speech with environmental context,” in IEEE International Conference on Acoustics, Speech and Signal Processing . IEEE, 2024, pp. 12 566– 12 571

    2024

  7. [15]

    Audiobox: Unified audio gener- ation with natural language prompts,

    A. Vyas, B. Shi, M. Le, A. Tjandra, Y .-C. Wu, B. Guo, J. Zhang, X. Zhang, R. Adkins, W. Ngan et al., “Audiobox: Unified audio gener- ation with natural language prompts,” arXiv preprint arXiv:2312.15821, 2023

    2023 arXiv

  8. [16]

    Environment Aware Text-to-Speech Synthesis,

    D. Tan, G. Zhang, and T. Lee, “Environment Aware Text-to-Speech Synthesis,” in Proc. Interspeech 2022 , 2022, pp. 481–485

    2022

  9. [17]

    Preserving background sound in noise-robust voice conversion via multi-task learning,

    J. Yao, Y . Lei, Q. Wang, P. Guo, Z. Ning, L. Xie, H. Li, J. Liu, and D. Xie, “Preserving background sound in noise-robust voice conversion via multi-task learning,” in IEEE International Conference on Acoustics, Speech and Signal Processing . IEEE, 2023, pp. 1–5

    2023

  10. [18]

    Direct noisy speech modeling for noisy-to-noisy voice conversion,

    C. Xie, Y .-C. Wu, P. L. Tobing, W.-C. Huang, and T. Toda, “Direct noisy speech modeling for noisy-to-noisy voice conversion,” in ICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2022, pp. 6787–6791

    2022

  11. [19]

    Noisy-to-noisy voice conversion under variations of noisy condition,

    C. Xie and T. Toda, “Noisy-to-noisy voice conversion under variations of noisy condition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

    2023

  12. [20]

    Speechx: Neural codec lan- guage model as a versatile speech transformer,

    X. Wang, M. Thakker, Z. Chen, N. Kanda, S. E. Eskimez, S. Chen, M. Tang, S. Liu, J. Li, and T. Yoshioka, “Speechx: Neural codec lan- guage model as a versatile speech transformer,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , 2024

    2024

  13. [21]

    V oicebox: Text- guided multilingual universal speech generation at scale,

    M. Le, A. Vyas, B. Shi, B. Karrer, L. Sari, R. Moritz, M. Williamson, V . Manohar, Y . Adi, J. Mahadeokar, and W.-N. Hsu, “V oicebox: Text- guided multilingual universal speech generation at scale,” in Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Gl...

    2023

  14. [22]

    Neural ordinary differential equations,

    R. T. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural ordinary differential equations,” Advances in Neural Information Pro- cessing Systems, vol. 31, 2018

    2018

  15. [23]

    Maximum likelihood training of score-based diffusion models,

    Y . Song, C. Durkan, I. Murray, and S. Ermon, “Maximum likelihood training of score-based diffusion models,” Advances in Neural Informa- tion Processing Systems , vol. 34, pp. 1415–1428, 2021

    2021

  16. [24]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  17. [25]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  18. [26]

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

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “LibriTTS: A corpus derived from librispeech for text-to-speech,” in Interspeech 2019, 2019, pp. 1526–1530

    2019

  19. [27]

    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 Interspeech 2019, 2019, pp. 1368–1372

    2019

  20. [28]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing . IEEE, 2017, pp. 5220–5224

    2017

  21. [29]

    CSTR VCTK Corpus: English multi-speaker corpus for cstr voice cloning toolkit,

    C. Veaux, J. Yamagishi, and K. MacDonald, “CSTR VCTK Corpus: English multi-speaker corpus for cstr voice cloning toolkit,” 2017

    2017

  22. [30]

    TUT database for acoustic scene classification and sound event detection,

    A. Mesaros, T. Heittola, and T. Virtanen, “TUT database for acoustic scene classification and sound event detection,” in 2016 24th European Signal Processing Conference (EUSIPCO). IEEE, 2016, pp. 1128–1132

    2016

  23. [31]

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

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,” Advances in Neural Information Processing Systems , vol. 33, pp. 17 022–17 033, 2020

    2020

  24. [32]

    Large-scale self-supervised speech representation learning for automatic speaker verification,

    Z. Chen, S. Chen, Y . Wu, Y . Qian, C. Wang, S. Liu, Y . Qian, and M. Zeng, “Large-scale self-supervised speech representation learning for automatic speaker verification,” in IEEE International Conference on Acoustics, Speech and Signal Processing . IEEE, 2022, pp. 6147–6151

    2022

  25. [33]

    Wavlm: Large-scale self-supervised pre- training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao et al. , “Wavlm: Large-scale self-supervised pre- training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  26. [34]

    I. T. U. T. S. Sector, Methods for subjective determination of transmis- sion quality. International Telecommunication Union, 1996

    1996

  27. [35]

    Music source separation with band-split rnn,

    Y . Luo and J. Yu, “Music source separation with band-split rnn,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 1893–1901, 2023

    1901

Pith tools

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