Pith. sign in

REVIEW 2 major objections 2 minor 39 references

Zero-shot voice cloning produces synthetic dysarthric speech that trains ASR nearly as well as real recordings.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-26 16:03 UTC pith:P4EYEEYS

load-bearing objection Zero-shot cloning gets competitive WER on dysarthric ASR but the paper leaves the key assumption about preserving dysarthric traits untested. the 2 major comments →

arxiv 2606.19823 v1 pith:P4EYEEYS submitted 2026-06-18 eess.AS cs.LG

Low-Burden Data Augmentation for Dysarthric ASR via Zero-Shot Voice Cloning

Satwinder Singh , Qianli Wang , Zihan Zhong , Clarion Mendes , Hasegawa-Johnson , Waleed Abdulla , Seyed Reza Shahamiri This is my paper
classification eess.AS cs.LG
keywords dysarthric ASRzero-shot voice cloningdata augmentationTORGO datasetWhisper fine-tuningsynthetic speech dataspeech recognition
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that cloning existing dysarthric voices from the TORGO dataset with a zero-shot model creates training data that lets Whisper-medium reach 26 percent word error rate on held-out real speech. This sits close to the 24.44 percent from real data and 25.12 percent from mixed data, and it beats real-data training for moderate-to-severe speakers while also leading on a separate corpus. The approach removes the need to record new speakers for each case of dysarthria, addressing the main barrier to building reliable recognizers for this speech.

Core claim

Fine-tuning Whisper-medium on speech cloned from TORGO speakers via Higgs Audio V2 yields 26.00 percent WER on held-out real dysarthric test speech, versus 31.62 percent without fine-tuning and 24.44 percent with real-data fine-tuning; the cloned and hybrid sets outperform real data for moderate-severe speakers, and cloned fine-tuning gives the largest gain (11.45 percent relative) on the SAP-1102 corpus.

What carries the argument

Zero-shot voice cloning with Higgs Audio V2 to generate synthetic dysarthric utterances from TORGO speakers for ASR fine-tuning.

Load-bearing premise

Cloned speech must preserve the acoustic and articulatory traits of real dysarthria that actually affect recognition accuracy.

What would settle it

A test showing that models trained only on the cloned data produce markedly higher word error rates than real-data models across multiple held-out dysarthric corpora.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Training sets can be scaled for dysarthric ASR without recording additional speakers.
  • Cloned data works especially well for moderate and severe dysarthria cases.
  • Hybrid real-plus-cloned sets deliver performance close to all-real sets.
  • Cross-corpus generalization improves when cloned data is included in training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same cloning step could supply data for other disordered or accented speech if the model captures the relevant distortions.
  • If cloning fidelity rises, entirely synthetic training pipelines become feasible for some low-resource speech tasks.
  • Pairing cloned data with existing augmentation methods may yield further gains without new recordings.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript claims that zero-shot voice cloning via Higgs Audio V2 offers a low-burden augmentation strategy for dysarthric ASR. Using the TORGO dataset, the authors clone speakers and fine-tune Whisper-medium on cloned, real, and hybrid data; they report that Clone FT reaches 26.00% WER on held-out real speech (vs. 24.44% Real FT and 25.12% Hybrid FT), outperforms Real FT on moderate-severe speakers, and yields the best cross-corpus result (11.45% relative) on SAP-1102.

Significance. If the cloned utterances preserve the acoustic and articulatory deviations characteristic of dysarthria, the approach would materially lower the cost of obtaining usable training data for a population where collection is expensive and variable. The direct comparison of WER on held-out real speech and the cross-corpus result constitute concrete, falsifiable evidence supporting the scalability claim.

major comments (2)
  1. [Results] Results section (WER table): the central claim that Clone FT is 'competitive' with Real FT rests on point estimates (26.00% vs 24.44%) without reported error bars, bootstrap intervals, or paired statistical tests. The 1.56-point absolute gap is small enough that sampling variability could reverse the ranking; this directly affects the strength of the 'nearly matching' conclusion.
  2. [Experimental setup] Experimental setup / §3 (cloning procedure): no acoustic or perceptual verification is provided that the zero-shot clones retain dysarthria-specific traits (imprecise consonants, altered formants, reduced speaking rate) rather than generic speaker variability. Without similarity metrics, formant comparisons, or dysarthria-severity ratings on the clones, the WER gains could be explained by ordinary data augmentation; this assumption is load-bearing for the interpretation that zero-shot cloning circumvents the dysarthric data bottleneck.
minor comments (2)
  1. [Abstract] Abstract: the phrase '11.45% relative' improvement on SAP-1102 does not state the baseline against which the relative reduction is computed.
  2. [Experimental setup] The manuscript does not specify the exact train/validation/test speaker splits or the number of utterances per severity class, hindering reproducibility of the reported WER figures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below, indicating planned revisions where appropriate.

read point-by-point responses

  1. Referee: [Results] Results section (WER table): the central claim that Clone FT is 'competitive' with Real FT rests on point estimates (26.00% vs 24.44%) without reported error bars, bootstrap intervals, or paired statistical tests. The 1.56-point absolute gap is small enough that sampling variability could reverse the ranking; this directly affects the strength of the 'nearly matching' conclusion.

    Authors: We agree that uncertainty quantification is necessary to support claims of competitiveness. In the revised manuscript we will report bootstrap confidence intervals for all WER estimates and include paired statistical tests (e.g., McNemar’s test on per-utterance errors) to evaluate whether observed differences are statistically significant. revision: yes

  2. Referee: [Experimental setup] Experimental setup / §3 (cloning procedure): no acoustic or perceptual verification is provided that the zero-shot clones retain dysarthria-specific traits (imprecise consonants, altered formants, reduced speaking rate) rather than generic speaker variability. Without similarity metrics, formant comparisons, or dysarthria-severity ratings on the clones, the WER gains could be explained by ordinary data augmentation; this assumption is load-bearing for the interpretation that zero-shot cloning circumvents the dysarthric data bottleneck.

    Authors: This point is well taken; downstream WER alone does not directly confirm retention of dysarthric characteristics. We will add acoustic verification in the revision by computing formant frequency statistics and speaking-rate measurements on a subset of cloned versus original utterances. We will also explicitly discuss the reliance on task performance as an indirect indicator and note the absence of perceptual ratings as a limitation. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical WER comparisons on held-out data

full rationale

The paper reports direct experimental results from fine-tuning Whisper on real, zero-shot cloned (via Higgs Audio V2), and hybrid data from TORGO, then measuring WER on held-out real speech and cross-corpus SAP-1102. No equations, normalizations, fitted parameters renamed as predictions, or self-citation chains appear in the derivation. All reported numbers (e.g., 26.00% Clone FT WER vs 24.44% Real FT) are independent measurements, not reductions to prior inputs by construction. The central claim rests on external empirical benchmarks rather than self-referential definitions or ansatzes.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work is empirical and relies on standard machine-learning evaluation assumptions rather than new theoretical constructs; no free parameters are explicitly fitted in the abstract, and no new entities are postulated.

axioms (1)
  • domain assumption Held-out real speech from TORGO is representative of dysarthric variability for reliable WER measurement.
    Invoked by the evaluation protocol comparing models on held-out real speech.

pith-pipeline@v0.9.1-grok · 5738 in / 1418 out tokens · 53311 ms · 2026-06-26T16:03:52.659081+00:00 · methodology

0 comments
read the original abstract

Automatic speech recognition remains unreliable for dysarthric speech due to data scarcity and high inter-speaker variability. While synthetic data can address these gaps, traditional methods often require extensive speaker-specific data, reintroducing the collection bottleneck. We investigate zero-shot voice cloning as a low-burden augmentation strategy, using Higgs Audio V2 to clone speakers in the TORGO dataset. We fine-tune (FT) Whisper-medium on cloned, real, and hybrid data and evaluate on held-out real speech. Compared to the zero-shot (31.62%), Clone FT achieved a competitive 26.00% WER, nearly matching the 24.44% and 25.12% seen with Real and Hybrid FT, respectively. Notably, Clone and Hybrid FT outperform Real FT for moderate-severe speakers. Clone FT achieves the best results (11.45% relative) in cross-corpus evaluation on the SAP-1102. These results suggest that zero-shot cloning provides scalable training data that circumvents the costly data collection bottleneck.

Figures

Figures reproduced from arXiv: 2606.19823 by Clarion Mendes, Hasegawa-Johnson, Qianli Wang, Satwinder Singh, Seyed Reza Shahamiri, Waleed Abdulla, Zihan Zhong.

Figure 1
Figure 1. Figure 1: Overview of voice cloning and ASR fine-tuning and evaluation pipeline [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Duration distribution of cloned utterances in TORGO-Synth. words (e.g., “yes”, “no”), phonetically balanced restricted sen￾tences, and unrestricted spontaneous speech, offering rich pho￾netic and prosodic coverage. Dysarthric speakers are clinically categorized by severity: severe, moderate-severe, moderate, and mild. For this study, we focus exclusively on restricted speech utterances to ensure linguistic… view at source ↗
Figure 4
Figure 4. Figure 4: Cross-corpus generalization results (WER %) on SAP-1102 dataset across three etiologies. 6. Conclusion This study demonstrates that zero-shot voice cloning from a single reference can provide a robust training signal for dysarthric ASR, significantly reducing the data-collection bur￾den on speech-impaired individuals. Our results reveal that per￾formance gains are non-monotonic, with an optimal synthetic v… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Index Terms: Whisper, dysarthria, automatic speech recogni- tion, zero-shot voice cloning

    These results suggest that zero-shot cloning provides scalable training data that circumvents the costly data collec- tion bottleneck. Index Terms: Whisper, dysarthria, automatic speech recogni- tion, zero-shot voice cloning

  2. [2]

    sweet spot

    Introduction Automatic speech recognition (ASR) has achieved strong ac- curacy on typical speech, driven by large-scale datasets and transformer-based models such as Whisper [1] and wav2vec 2.0 [2]. However, these gains do not extend reliably to dysarthric speech, which arises from neurological conditions such as Parkinson’s disease (PD), cerebral palsy (...

  3. [3]

    Low-Burden Data Augmentation for Dysarthric ASR via Zero-Shot Voice Cloning

    Related Work 2.1. Synthetic Data for Dysarthric ASR Data scarcity has motivated a range of augmentation strate- gies for dysarthric ASR. Conventional signal-level perturba- tions (e.g., speed or noise) can increase acoustic diversity but do not create new lexical content or explicitly model dysarthria- specific timing and articulation patterns [21–23]. Co...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  4. [4]

    The quick brown fox jumps over the lazy dog

    Method 3.1. Voice Cloning Setup In this study, we adopted the Higgs Audio V2 voice cloning model [18] to generate synthetic dysarthric speech samples. It is a large-scale (5B parameters), open-source state-of-the- art audio foundation model trained on over 10 million hours of diverse audio-text pairs. The model performs zero-shot TTS synthesis without req...

  5. [5]

    yes”, “no

    Experimental Setup 4.1. Datasets TORGO dataset[3], developed by the University of Toronto, provides clinically annotated speech recordings from individu- als with dysarthria due to CP and ALS. The dataset contains approximately 23 hours of speech from 8 dysarthric speakers (5 male, 3 female; 7 with CP, 1 with ALS) and 7 neurotypical con- trols. Recordings...

    2000

  6. [6]

    sweet spot

    Results and Discussion 5.1. Speaker Similarity Analysis Figure 3 shows 2D t-SNE projections of TitaNet speaker em- beddings [31] for cloned utterances, with the single enrollment reference per speaker marked by a star. Several speakers (e.g., F03, M03, F01, and F04) form tight, well-separated clusters, with the reference embedded within the same region, i...

    work page arXiv

  7. [7]

    Our results reveal that per- formance gains are non-monotonic, with an optimal synthetic volume near 15 hours, beyond which synthesis artifacts may impede generalization

    Conclusion This study demonstrates that zero-shot voice cloning from a single reference can provide a robust training signal for dysarthric ASR, significantly reducing the data-collection bur- den on speech-impaired individuals. Our results reveal that per- formance gains are non-monotonic, with an optimal synthetic volume near 15 hours, beyond which synt...

  8. [8]

    Generative AI Use Disclosure The authors did not use any generative AI in the conceptualiza- tion, writing, or preparation of this manuscript

  9. [9]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning (ICML). PMLR, 2023, pp. 28 492–28 518

    2023

  10. [10]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  11. [11]

    The TORGO database of acoustic and articulatory speech from speakers with dysarthria,

    F. Rudzicz, A. K. Namasivayam, and T. Wolff, “The TORGO database of acoustic and articulatory speech from speakers with dysarthria,”Language resources and evaluation, vol. 46, pp. 523– 541, 2012

    2012

  12. [12]

    Fre- quency of consonant articulation errors in dysarthric speech,

    H. Kim, K. Martin, M. Hasegawa-Johnson, and A. Perlman, “Fre- quency of consonant articulation errors in dysarthric speech,” Clinical linguistics & phonetics, vol. 24, no. 10, pp. 759–770, 2010

    2010

  13. [13]

    Disorders of communication: Dysarthria,

    P. Enderby, “Disorders of communication: Dysarthria,”Hand- book of clinical neurology, vol. 110, pp. 273–281, 2013

    2013

  14. [14]

    A Comprehensive Performance Evaluation of Whisper Models in Dysarthric Speech Recogni- tion,

    S. Singh, Z. Zhong, Q. Wang, C. Mendes, M. Hasegawa-Johnson, W. Abdulla, and S. R. Shahamiri, “A Comprehensive Performance Evaluation of Whisper Models in Dysarthric Speech Recogni- tion,” inInternational Conference on Neural Information Process- ing. Springer, 2024, pp. 75–90

    2024

  15. [15]

    Robust Cross-Etiology and Speaker-Independent Dysarthric Speech Recognition,

    S. Singh, Q. Wang, Z. Zhong, C. Mendes, M. Hasegawa-Johnson, W. Abdulla, and S. R. Shahamiri, “Robust Cross-Etiology and Speaker-Independent Dysarthric Speech Recognition,” inIEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  16. [16]

    Convolution-augmented transformers for enhanced speaker-independent dysarthric speech recognition,

    Z. Zhong, Q. Wang, S. Singh, C. Mendes, M. Hasegawa-Johnson, W. Abdulla, and S. R. Shahamiri, “Convolution-augmented transformers for enhanced speaker-independent dysarthric speech recognition,”IEEE Transactions on Neural Systems and Rehabil- itation Engineering, 2025

    2025

  17. [17]

    Improving the efficiency of dysarthria voice conversion system based on data augmentation,

    W.-Z. Zheng, J.-Y . Han, C.-Y . Chen, Y .-J. Chang, and Y .-H. Lai, “Improving the efficiency of dysarthria voice conversion system based on data augmentation,”IEEE Transactions on Neural Sys- tems and Rehabilitation Engineering, vol. 31, pp. 4613–4623, 2023

    2023

  18. [18]

    Dysarthric speech conformer: Adaptation for sequence-to-sequence dysarthric speech recogni- tion,

    Q. Wang, Z. Zhong, S. Singh, C. Mendes, M. Hasegawa-Johnson, W. Abdulla, and S. R. Shahamiri, “Dysarthric speech conformer: Adaptation for sequence-to-sequence dysarthric speech recogni- tion,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  19. [19]

    Chal- lenges and practical guidelines for atypical speech data collection, annotation, usage and sharing: A multi-project perspective,

    Z. Yue, M. Barberis, T. Patel, J. Dineley, W. Doedens, L. Stip- donk, Y . Zhang, E. d. Witte, E. Loweimi, D. Satoeret al., “Chal- lenges and practical guidelines for atypical speech data collection, annotation, usage and sharing: A multi-project perspective,” in Proc. Interspeech 2025, 2025, pp. 3943–3947

    2025

  20. [20]

    The re- liability and validity of speech-language pathologists’ estimations of intelligibility in dysarthria,

    M. E. Hirsch, A. Thompson, Y . Kim, and K. L. Lansford, “The re- liability and validity of speech-language pathologists’ estimations of intelligibility in dysarthria,”Brain Sciences, vol. 12, no. 8, p. 1011, 2022

    2022

  21. [21]

    Dysarthric speech database for universal access research,

    H. Kim, M. Hasegawa-Johnson, A. Perlman, J. R. Gunderson, T. S. Huang, K. L. Watkin, and S. Frame, “Dysarthric speech database for universal access research,” inProc. Interspeech, 2008, pp. 1741–1744

    2008

  22. [22]

    Community-supported shared infrastructure in support of speech accessibility,

    M. Hasegawa-Johnson, X. Zheng, H. Kim, C. Mendes, M. Dickin- son, E. Hege, C. Zwilling, M. M. Channell, L. Mattie, H. Hodges et al., “Community-supported shared infrastructure in support of speech accessibility,”Journal of Speech, Language, and Hearing Research, vol. 67, no. 11, pp. 4162–4175, 2024

    2024

  23. [23]

    Personalized Fine-Tuning with Controllable Synthetic Speech from LLM-Generated Transcripts for Dysarthric Speech Recognition,

    D. Wagner, I. Baumann, N. Engert, S. Lee, E. N ¨oth, K. Riedham- mer, and T. Bocklet, “Personalized Fine-Tuning with Controllable Synthetic Speech from LLM-Generated Transcripts for Dysarthric Speech Recognition,” inInterspeech, 2025, pp. 3294–3298

    2025

  24. [24]

    VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers

    S. Chen, S. Liu, L. Zhou, Y . Liu, X. Tan, J. Li, S. Zhao, Y . Qian, and F. Wei, “Vall-E 2: Neural codec language models are hu- man parity zero-shot text to speech synthesizers,”arXiv preprint arXiv:2406.05370, 2024

    work page Pith review arXiv 2024

  25. [25]

    F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. JianZhao, K. Yu, and X. Chen, “F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025, pp. 6255–6271

    2025

  26. [26]

    Higgs Audio V2: Redefining Expressiveness in Au- dio Generation,

    Boson AI, “Higgs Audio V2: Redefining Expressiveness in Au- dio Generation,” https://github.com/boson-ai/higgs-audio, 2025, GitHub repository. Release blog available at https://www.boson. ai/blog/higgs-audio-v2

    2025

  27. [27]

    Accurate synthesis of dysarthric speech for asr data augmenta- tion,

    M. Soleymanpour, M. T. Johnson, R. Soleymanpour, and J. Berry, “Accurate synthesis of dysarthric speech for asr data augmenta- tion,”Speech Communication, vol. 164, p. 103112, 2024

    2024

  28. [28]

    Towards identity preserving normal to dysarthric voice conversion,

    W.-C. Huang, B. M. Halpern, L. P. Violeta, O. Scharenborg, and T. Toda, “Towards identity preserving normal to dysarthric voice conversion,” inIEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6672–6676

    2022

  29. [29]

    SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition,”Proc. Interspeech, p. 2613, 2019

    2019

  30. [30]

    Audio augmen- tation for speech recognition

    T. Ko, V . Peddinti, D. Povey, and S. Khudanpur, “Audio augmen- tation for speech recognition.” inProc. Interspeech, vol. 2015, 2015, p. 3586

    2015

  31. [31]

    PhasePerturbation: Speech data augmentation via phase perturbation for automatic speech recognition,

    C. Lei, S. Singh, F. Hou, X. Jia, and R. Wang, “PhasePerturbation: Speech data augmentation via phase perturbation for automatic speech recognition,” inProceedings of the 5th ACM International Conference on Multimedia in Asia Workshops, 2023, pp. 1–6

    2023

  32. [32]

    Training data augmentation for dysarthric automatic speech recognition by text- to-dysarthric-speech synthesis,

    W.-Z. Leung, M. Cross, A. Ragni, and S. Goetze, “Training data augmentation for dysarthric automatic speech recognition by text- to-dysarthric-speech synthesis,” inProc. Interspeech 2024, 2024, pp. 2494–2498

    2024

  33. [33]

    Few-shot dysarthric speech recog- nition with text-to-speech data augmentation,

    E. Hermann and M. M. Doss, “Few-shot dysarthric speech recog- nition with text-to-speech data augmentation,” inProc. Inter- speech 2023, 2023, pp. 156–160

    2023

  34. [34]

    Data Augmentation Using Healthy Speech for Dysarthric Speech Recognition,

    B. Vachhani, C. Bhat, and S. K. Kopparapu, “Data Augmentation Using Healthy Speech for Dysarthric Speech Recognition,” inIn- terspeech, 2018, pp. 471–475

    2018

  35. [35]

    Simulating dysarthric speech for training data augmentation in clinical speech applica- tions,

    Y . Jiao, M. Tu, V . Berisha, and J. Liss, “Simulating dysarthric speech for training data augmentation in clinical speech applica- tions,” inIEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 6009–6013

    2018

  36. [36]

    YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot V oice 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 V oice Conversion for Everyone,” inInternational conference on machine learning. PMLR, 2022, pp. 2709–2720

    2022

  37. [37]

    Synthe- sis of new words for improved dysarthric speech recognition on an expanded vocabulary,

    J. Harvill, D. Issa, M. Hasegawa-Johnson, and C. Yoo, “Synthe- sis of new words for improved dysarthric speech recognition on an expanded vocabulary,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6428–6432

    2021

  38. [38]

    Lib- riSpeech: an ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- riSpeech: an ASR corpus based on public domain audio books,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  39. [39]

    TitaNet: Neural model for speaker representation with 1D depth-wise separable convo- lutions and global context,

    N. R. Koluguri, T. Park, and B. Ginsburg, “TitaNet: Neural model for speaker representation with 1D depth-wise separable convo- lutions and global context,” inIEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2022, pp. 8102–8106

    2022