Pith. sign in

REVIEW 5 minor 38 references

Technical Report for MERL's Real-TSE Challenge Submission

T0 review · 0 major / 5 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Careful multi-stage data preparation, not a new architecture, won first place in real-world target speech extraction; DNSMOS and speaker similarity can be adversarially maxed without harming speech recognition or VAD scores.

desk verdict Solid first-place Real-TSE report: data curriculum + clean metric-fragility demo, not a new architecture. read the letter →

arxiv 2607.09043 v1 pith:MXJIEVZ4 submitted 2026-07-10 eess.AS

Dominik Klement , Yoshiki Masuyama , Christoph Boeddeker , Kohei Saijo , Julius Richter , Gordon Wichern , Jonathan Le Roux This is my paper · ORCID
classification eess.AS
keywords targetspeechextractionrealfar-fieldmixturesdatapreparationcurriculumlearningDNSMOSspeakersimilarityadversarialattackpseudo-targets
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

Target speech extraction systems have mostly been trained and scored on clean, fully overlapped synthetic mixtures, so they often collapse on real far-field conversations that contain noise, reverb, and partial overlap. This technical report shows that a modest enlargement of the challenge baseline, trained in four carefully staged curricula, can reach first place on the Real-TSE second track when the training data itself is aggressively cleaned and adapted. The decisive step is not a new network but a pipeline that turns real close-talk and far-field pairs into usable pseudo-targets via enhancement, synchronization, and causal Wiener projection, then mixes those real mixtures with simulated data. Along the way the authors notice that two popular non-intrusive metrics—DNSMOS and speaker similarity—can be driven to near-perfect values by tiny adversarial perturbations that leave token error rate and VAD F1 unchanged. The practical claim is therefore double: high-quality data work still beats architectural novelty for real-world TSE, and the community’s favorite automatic quality scores are too brittle to trust once they become optimization targets.

What carries the argument

Four-stage curriculum that ends with real far-field mixtures whose supervised targets are produced by close-talk enhancement, max-peak synchronization, causal Wiener projection onto the distant channel, and a second enhancement pass; metric-aware losses and a subsequent per-utterance adversarial attack that regularizes spectrogram magnitude while maximizing DNSMOS and cosine speaker similarity.

What would settle it

Train an identical architecture on the same real far-field mixtures but replace the Wiener-projected pseudo-targets with either pure silence or with unprocessed close-talk signals; if token error rate and VAD F1 on the Real-TSE evaluation set collapse, the pseudo-target quality claim is falsified.

Watch

Extended reading notes

Core claim

A baseline Band-split RNN with a larger speaker encoder, trained through four progressive stages that move from fully overlapped clean mixtures to real far-field multi-talker conversations whose targets are pseudo-labels obtained from processed close-talk microphones, achieves first place on the Real-TSE Challenge second track; the same work also demonstrates that DNSMOS and speaker-similarity scores can be pushed to extreme values by imperceptible adversarial noise while token error rate and VAD-based F1 remain essentially unchanged.

Load-bearing premise

The multi-step close-talk-to-far-field projection pipeline produces pseudo-targets accurate and unbiased enough to serve as reliable supervised labels for real multi-talker mixtures.

Editorial extensions

If this is right

  • Future Real-TSE or similar challenges can expect large gains from data-cleaning and curriculum design even when model architecture is held fixed.
  • Leaderboards that include DNSMOS or speaker-similarity scores can be gamed by small adversarial perturbations that leave ASR and VAD metrics intact.
  • Challenge organizers are given concrete evidence to drop or replace those two metrics when computing official rankings.
  • Pseudo-label pipelines that project close-talk onto far-field channels become a reusable recipe for adapting separation models to other real multi-microphone corpora.

Reading between the lines

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

  • Any non-intrusive neural quality metric that is differentiable and has been seen during training is likely to admit similar low-energy adversarial attacks, so the problem is not limited to DNSMOS.
  • Once pseudo-target quality is accepted as the bottleneck, the next practical research question becomes how little close-talk supervision is still sufficient for the projection stage.
  • The same curriculum-plus-projection recipe could be tested on streaming or multi-channel TSE without changing the core network, offering a low-risk path to on-device systems.
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

0 major / 5 minor

Summary. This technical report describes MERL’s first-place submission to the Real-TSE Challenge. The authors retain a strengthened baseline BSRNN with multi-level speaker conditioning (EcapaTDNN-1024, 10 blocks) and concentrate on data preparation and a four-stage curriculum: (1) fully-overlapped Libri2Mix-style pre-training, (2) noisy/reverberant simulated conversations, (3) on-the-fly far-field mixtures from single-talker segments of CHiME-6/AMI/AISHELL, and (4) real multi-talker mixtures supervised by pseudo-targets obtained via close-talk enhancement, max-peak synchronization, causal Wiener projection and a second enhancement pass. Table I documents monotonic gains in TER and F1 across stages; a pure-reconstruction ablation matches the final TER/DNSMOS. An additional per-utterance adversarial attack (Algorithm 1, Table II) drives DNSMOS and speaker-similarity scores to extremes while leaving TER and VAD-based F1 unchanged, illustrating metric fragility.

Significance. If the reported ranking and ablation hold, the work supplies concrete evidence that careful multi-source data cleaning, heavy mixture/enrollment augmentation and staged adaptation to real far-field conditions can outperform architectural novelty on a difficult real-world TSE benchmark. The controlled metric attack cleanly isolates the brittleness of DNSMOS and speaker similarity (Goodhart’s law), offering a practical warning for challenge organizers. Strengths that raise the contribution above a pure system description include the external leaderboard validation, the pure-reconstruction ablation in Table I, and the fully specified attack procedure that preserves the non-attacked metrics.

minor comments (5)
  1. Throughout the manuscript (abstract, §I, §VII) the challenge name appears inconsistently as “Real-TSE”, “RealTSE” and “Real-T”. Standardize to one form.
  2. Algorithm 1 contains incomplete expressions (“loss = loss.pow().mean()”) and uses undefined helpers (stft, dnsmos, spksim). A short clarifying sentence or corrected pseudocode would improve reproducibility.
  3. §III.A–C repeatedly write “ClearerV oice”, “V oxCeleb2”, “V AD”. These appear to be PDF-extraction artifacts; restore the correct spellings (ClearerVoice, VoxCeleb2, VAD).
  4. Eq. (5) weights are given without justification or sensitivity analysis. A one-sentence remark on how the coefficients 10 / 0.01 were chosen would help readers.
  5. Table I caption uses “SPKFOR SPEAKER” (missing space) and the final pure-reconstruction row is labeled only by a star; a clearer legend would avoid ambiguity.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity; mild self-training bootstrap on pseudo-targets is secondary and does not force the external ranking or attack results.

  1. other [Section III.C (Real Far-field Mixture Training) and Section IV.D (fourth training stage)]
    "Afterwards, we use the best-performing TSE model trained on far-field mixtures to reduce cross-talk on CHiME-6 and AMI close-talk microphones. ... Lastly, the fourth stage contains real noisy mixture training against targets that were extracted by our previous TSE model."

    A prior version of the same model family produces the pseudo-targets used to train the submitted model. This is iterative self-training rather than a definitional identity, and evaluation (TER, F1, external ranking) remains independent of the pseudo-targets; it is therefore only a mild, non-load-bearing circularity.

full rationale

This is an empirical challenge technical report, not a first-principles derivation. The central claims (first-place ranking on the Real-TSE second track; DNSMOS/spk-sim fragility under adversarial attack) rest on external leaderboard evaluation and on a controlled post-hoc per-utterance attack (Algorithm 1, Table II) that leaves TER and VAD-F1 unchanged while driving the attacked metrics to extremes. Table I further shows that the progressive TER/F1 gains across the four curriculum stages, and the pure-reconstruction ablation (last row), do not require the optional metric-aware losses. The only mild circular element is the fourth-stage use of a previous TSE model to clean close-talk signals that then become pseudo-targets for the next model; this is ordinary iterative self-training / bootstrapping, not a definitional reduction of a claimed prediction to its own inputs, and the final metrics remain independent of those pseudo-targets. No self-definitional equations, fitted-parameter-as-prediction, load-bearing self-citation uniqueness theorems, or renamed known results appear. Score 1 reflects only that secondary bootstrap; the paper is otherwise self-contained against external benchmarks.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claims rest on standard speech-processing assumptions plus a large collection of hand-chosen thresholds, loss weights and sampling probabilities that control data cleaning and curriculum. No new physical entities are postulated; the main modeling leap is the trustworthiness of the engineered pseudo-targets.

free parameters (8)
  • speaker-similarity filter threshold = 0.85
    Recordings kept only if ResNet34 cosine similarity after enhancement exceeds 0.85; chosen by inspection of outliers.
  • DNSMOS pre-filter threshold = 3.0
    Only utterances with DNSMOS > 3.0 retained for pre-training.
  • WER filtering thresholds = 30 % / 50 %
    30 % (AMI/AISHELL) or 50 % (CHiME-6) used to discard low-quality close-talk segments.
  • loss weights (Lstft, LMOS, Lspk) = 10 / 0.01 / 1
    Final loss = 10 Lstft + Ltime + 0.01 LMOS + Lspk; coefficients chosen to balance metrics.
  • RIR and noise application probability = 0.8
    Both applied independently with probability 0.8; SNR range [-5,15] dB.
  • non-target enrollment probability = 0.05
    5 % of training examples force zero output by pairing enrollment with absent speaker.
  • learning rates and schedule = 5e-4 / 1e-4
    AdamW 5e-4 then 1e-4 with linear warmup and cosine decay; ReduceLROnPlateau on validation SNR.
  • model capacity knobs = 10 blocks / 1024-dim
    BSRNN depth raised from 6 to 10 blocks; ECAPA-TDNN embedding size raised from 512 to 1024.
assumptions (4)
  • domain assumption Processed close-talk signals after enhancement, synchronization and causal Wiener projection constitute valid supervised targets for far-field multi-talker mixtures.
    Invoked throughout Section III.C and the fourth training stage; no independent clean reference exists for real data.
  • domain assumption Curriculum learning from fully-overlapped synthetic data through simulated conversations to real far-field mixtures improves final performance (Bengio et al., 2009).
    Stated as motivation for the four-stage schedule in Section IV.
  • domain assumption Token error rate and VAD-based F1 are reliable proxies for extraction quality that cannot be gamed by the same adversarial perturbations that affect DNSMOS and speaker similarity.
    Used to interpret Table II as evidence that the metrics are brittle while 'true' performance is preserved.
  • standard math Standard optimizers (AdamW), STFT magnitude losses and speaker-embedding cosine losses behave as expected for TSE.
    Background machinery taken from the literature without re-derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Technical Report for MERL's Real-TSE Challenge Submission." pith.science (2026). https://pith.science/paper/MXJIEVZ4

@misc{pith2026260709043,
  author       = {Pith},
  title        = {Pith review of: Technical Report for MERL's Real-TSE Challenge Submission},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXJIEVZ4}},
  note         = {Machine review of arXiv:2607.09043}
}
read the original abstract

Target speech extraction (TSE) has largely been dominated by neural network-based approaches trained and evaluated on synthetic fully overlapped data. The Real-TSE Challenge aims to advance performance on real-world far-field noisy and reverberant recordings. This technical report describes MERL's submission to the Real-TSE Challenge. Rather than proposing a novel model architecture, we built upon the baseline model and focused primarily on data preparation and cleaning. Our system was trained in four stages, beginning with pre-training on fully overlapped mixtures and simulated multi-talker conversations with noise and reverberation applied to both the mixture and the enrollment utterances. We then adapted the model to real-world conditions using noisy far-field recordings with pseudo-targets derived from processed close-talk microphone signals. Our submission achieved first place in the second track, demonstrating the critical importance of high-quality data preparation. Furthermore, we observed that DNSMOS and speaker similarity are susceptible to over-optimization, motivating an investigation of their robustness using adversarial attacks. The results show that both metrics can be driven to extreme values without degrading the token error rate or the VAD-based F1 score.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 5 linked inside Pith

  1. [1]

    LibriMix: An open-source dataset for generalizable speech separation,

    J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vincent, “LibriMix: An open-source dataset for generalizable speech separation,” arXiv preprint arXiv:2005.11262, 2020

    arXiv 2005

  2. [2]

    CHiME-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings,

    S. Watanabe, M. Mandel, J. Barker, E. Vincent, A. Arora, X. Chang, S. Khudanpur, V . Manohar, D. Povey, D. Raj, D. Snyder, A. S. Subrama- nian, J. Trmal, B. B. Yair, C. Boeddeker, Z. Ni, Y . Fujita, S. Horiguchi, N. Kanda, T. Yoshioka, and N. Ryant, “CHiME-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings,” inProc. 6th Inter...

    2020

  3. [3]

    DiPCo – dinner party corpus,

    M. Van Segbroeck, A. Zaid, K. Kutsenko, C. Huerta, T. Nguyen, X. Luo, B. Hoffmeister, J. Trmal, M. Omologo, and R. Maas, “DiPCo – dinner party corpus,” inProc. Interspeech, 2020, pp. 434–436

    2020

  4. [4]

    REAL-T: Real conversational mixtures for target speaker extraction,

    S. Li, S. Wang, J. Han, K. Zhang, W. Wang, and H. Li, “REAL-T: Real conversational mixtures for target speaker extraction,” inProc. Interspeech, 2025, pp. 1923–1927

    2025

  5. [5]

    Multi-level speaker representation for target speaker extraction,

    K. Zhang, J. Li, S. Wang, Y . Wei, Y . Wang, Y . Wang, and H. Li, “Multi-level speaker representation for target speaker extraction,” in Proc. ICASSP, 2025

    2025

  6. [6]

    ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,” inProc. Interspeech, 2020, pp. 3830–3834

    2020

  7. [7]

    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,” inProc. ICASSP, 2015, pp. 5206–5210

    2015

  8. [8]

    V oxCeleb2: Deep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxCeleb2: Deep speaker recognition,” inProc. Interspeech, 2018, pp. 1086–1090

    2018

Show all 38 references
  1. [9]

    Emilia: A large- scale, extensive, multilingual, and diverse dataset for speech generation,

    H. He, Z. Shang, C. Wang, X. Li, Y . Gu, H. Hua, L. Liu, C. Yang, J. Li, P. Shi, Y . Wang, K. Chen, P. Zhang, and Z. Wu, “Emilia: A large- scale, extensive, multilingual, and diverse dataset for speech generation,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 33, pp. 40...

    2025

  2. [10]

    CSTR VCTK Corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

    J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK Corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),” 2019

    2019

  3. [11]

    EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,

    J. Richter, Y .-C. Wu, S. Krenn, S. Welker, B. Lay, S. Watanabe, A. Richard, and T. Gerkmann, “EARS: An anechoic fullband speech dataset benchmarked for speech enhancement and dereverberation,” in Proc. Interspeech, 2024, pp. 4873–4877

    2024

  4. [12]

    ClearerV oice-Studio: Bridging advanced speech processing research and practical deployment,

    S. Zhao, Z. Pan, and B. Ma, “ClearerV oice-Studio: Bridging advanced speech processing research and practical deployment,” inProc. Inter- speech, 2025, pp. 2980–2984

    2025

  5. [13]

    MossFormer: Pushing the performance limit of monaural speech separation using gated single-head transformer with convolution-augmented joint self-attentions,

    S. Zhao and B. Ma, “MossFormer: Pushing the performance limit of monaural speech separation using gated single-head transformer with convolution-augmented joint self-attentions,” inProc. ICASSP, 2023

    2023

  6. [14]

    WeSpeaker: A research and production oriented speaker embedding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y . Deng, and Y . Qian, “WeSpeaker: A research and production oriented speaker embedding learning toolkit,” inProc. ICASSP, 2023

    2023

  7. [15]

    Montreal forced aligner: Trainable text-speech alignment using Kaldi,

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Sonderegger, “Montreal forced aligner: Trainable text-speech alignment using Kaldi,” inProc. Interspeech, 2017, pp. 498–502

    2017

  8. [16]

    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, and R. Aichner, “ICASSP 2022 deep noise suppression challenge,” inProc. ICASSP, 2022, pp. 9271–9275

    2022

  9. [17]

    Pyroomacoustics: A python package for audio room simulation and array processing algorithms,

    R. Scheibler, E. Bezzam, and I. Dokmani ´c, “Pyroomacoustics: A python package for audio room simulation and array processing algorithms,” in Proc. ICASSP, 2018, pp. 351–355

    2018

  10. [18]

    The third CHiME speech separation and recognition challenge: Dataset, task and base- lines,

    J. Barker, R. Marxer, E. Vincent, and S. Watanabe, “The third CHiME speech separation and recognition challenge: Dataset, task and base- lines,” inProc. ASRU, Dec. 2015, pp. 504–511

    2015

  11. [19]

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

    J. Thiemann, N. Ito, and E. Vincent, “The diverse environments multi- channel acoustic noise database (DEMAND): A database of multichan- nel environmental noise recordings,” inProceedings of Meetings on Acoustics, vol. 19, no. 1. Acoustical Society of America, 2013, p. 035081

    2013

  12. [20]

    FMA: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “FMA: A dataset for music analysis,” inProc. ISMIR, 2017, pp. 316–323

    2017

  13. [21]

    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 Trans. Audio, Speech, Lang. Process., vol. 30, pp. 829–852, 2022

    2022

  14. [22]

    MUSAN: A music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “MUSAN: A music, speech, and noise corpus,”arXiv preprint arXiv:1510.08484, 2015

    2015 arXiv

  15. [23]

    WHAM!: Extending speech separation to noisy environments,

    G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, D. Crow, E. Manilow, and J. Le Roux, “WHAM!: Extending speech separation to noisy environments,” inInterspeech 2019, 2019

    2019

  16. [24]

    URGENT challenge: Universality, robustness, and generalizability for speech en- hancement,

    W. Zhang, R. Scheibler, K. Saijo, S. Cornell, C. Li, Z. Ni, J. Pirkl- bauer, M. Sach, S. Watanabe, T. Fingscheidt, and Y . Qian, “URGENT challenge: Universality, robustness, and generalizability for speech en- hancement,” inProc. Interspeech, 2024, pp. 4868–4872

    2024

  17. [25]

    The AMI meeting corpus: A pre-announcement,

    J. Carletta, S. Ashby, S. Bourban, M. Flynn, M. Guillemot, T. Hain, J. Kadlec, V . Karaiskos, W. Kraaij, M. Kronenthalet al., “The AMI meeting corpus: A pre-announcement,” inInternational workshop on machine learning for multimodal interaction. Springer, 2005, pp. 28– 39

    2005

  18. [26]

    AISHELL-4: An open source dataset for speech enhancement, separation, recognition and speaker diarization in conference scenario,

    Y . Fu, L. Cheng, S. Lv, Y . Jv, Y . Kong, Z. Chen, Y . Hu, L. Xie, J. Wu, H. Bu, X. Xu, J. Du, and J. Chen, “AISHELL-4: An open source dataset for speech enhancement, separation, recognition and speaker diarization in conference scenario,” inProc. Interspeech, 2021, pp. 3665–3669

    2021

  19. [27]

    AISHELL-5: The first open-source in-car multi-channel multi-speaker speech dataset for automatic speech diarization and recognition,

    Y . Dai, H. Wang, X. Li, Z. Zhang, S. Wang, L. Xie, X. Xu, H. Guo, S. Zhang, H. Bu, and W. Chen, “AISHELL-5: The first open-source in-car multi-channel multi-speaker speech dataset for automatic speech diarization and recognition,” inProc. Interspeech, 2025, pp. 5493–5497

    2025

  20. [28]

    Front-end processing for the CHiME-5 dinner party scenario,

    C. Boeddecker, J. Heitkaemper, J. Schmalenstroeer, L. Drude, J. Hey- mann, and R. Haeb-Umbach, “Front-end processing for the CHiME-5 dinner party scenario,” inProc. 5th International Workshop on Speech Processing in Everyday Environments (CHiME 2018), 2018, pp. 35–40

    2018

  21. [29]

    Gen- erating training targets for real-world speech enhancement via close- to-distant microphone projection,

    T. Nakatani, R. Ikeshita, N. Kamo, M. Delcroix, and S. Araki, “Gen- erating training targets for real-world speech enhancement via close- to-distant microphone projection,” inProc. ICASSP, 2026, pp. 18 912– 18 916

    2026

  22. [30]

    Canary-1B-V2 & Parakeet- TDT-0.6B-V3: Efficient and high-performance models for multilingual ASR and AST,

    M. Sekoyan, N. R. Koluguri, N. Tadevosyan, P. Zelasko, T. Bartley, N. Karpov, J. Balam, and B. Ginsburg, “Canary-1B-V2 & Parakeet- TDT-0.6B-V3: Efficient and high-performance models for multilingual ASR and AST,”arXiv preprint arXiv:2509.14128, 2025

    2025

  23. [31]

    Curriculum learning,

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” inProc. ICML, ser. ACM International Conference Proceeding Series, vol. 382. ACM, 2009, pp. 41–48

    2009

  24. [32]

    Mind the gap: Impact of synthetic conversational data on multi-talker ASR and speaker diarization,

    A. Polok, I. Medennikov, J. ˇCernock`y, S. Watanabe, L. Burget, and S. Cornell, “Mind the gap: Impact of synthetic conversational data on multi-talker ASR and speaker diarization,”arXiv preprint arXiv:2605.15442, 2026

    2026 arXiv

  25. [33]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inProc. ICLR, 2019

    2019

  26. [34]

    SGDR: Stochastic gradient descent with warm restarts,

    ——, “SGDR: Stochastic gradient descent with warm restarts,” in International Conference on Learning Representations, 2017. [Online]. Available: https://openreview.net/forum?id=Skq89Scxx

    2017

  27. [35]

    The text-to-speech in the wild (TITW) database,

    J.-W. Jung, W. Zhang, S. Maiti, Y . Wu, X. Wang, J.-H. Kim, Y . Mat- sunaga, S. Um, J. Tian, H.-J. Shim, N. Evans, J. S. Chung, S. Takamichi, and S. Watanabe, “The text-to-speech in the wild (TITW) database,” in Interspeech 2025. ISCA: ISCA, Aug. 2025, pp. 4798–4802

    2025

  28. [36]

    Multilayer feedforward networks are universal approximators,

    K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,”Neural Networks, vol. 2, no. 5, pp. 359–366, 1989. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/0893608089900208

    1989

  29. [37]

    Beyond waveform robustness: Robust feature-vocoder adversarial attacks on automatic speech recognition,

    Y . Liao, Z. Zhang, Z. Sun, Y . Sun, X. Zheng, and X. He, “Beyond waveform robustness: Robust feature-vocoder adversarial attacks on automatic speech recognition,”arXiv preprint arXiv:2606.05678, 2026

    2026 arXiv

  30. [38]

    Attacking UTMOS: Probing the robustness of a speech quality assessment model,

    W.-C. Huang and T. Toda, “Attacking UTMOS: Probing the robustness of a speech quality assessment model,” 2026. [Online]. Available: https://arxiv.org/abs/2606.31105

    2026 arXiv

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.