Pith. sign in

REVIEW 4 major objections 7 minor 43 references

Towards Pretraining Robust ASR Foundation Model with Acoustic-Aware Data Augmentation

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Transcription generalization in ASR is driven mainly by how speech sounds, not by what the words are; with just 960 hours of Librispeech, acoustic augmentations cut word-error rates by up to 19.24% on unseen accented and children's speech.

desk verdict A useful acoustic augmentation recipe with an overreaching causal claim; the recipe is worth trying, the claim needs a cleaner test. read the letter →

arxiv 2505.20606 v1 pith:EC43MMLB submitted 2025-05-27 cs.CL cs.MM

Dancheng Liu , Amir Nassereldine , Chenhui Xu , Jinjun Xiong This is my paper · ORCID
classification cs.CLcs.MM
keywords automaticspeechrecognitiondataaugmentationacousticdiversitylinguisticsynthesisWhisperworderrorrateASRpretraining
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 asks why Whisper is so robust on accented and children's speech and whether that robustness can be bought without Whisper's 680,000 hours of data. Its answer, argued through experiments, is that a pretrained ASR model's transcription generalization is carried by acoustic diversity — how speech sounds — rather than linguistic diversity, what the words are. A model trained on an 11-million-sentence corpus spoken by a single synthetic voice collapses on real audio (WER above 100%), whereas augmenting only 960 hours of Librispeech with pitch, amplitude, duration, and vowel manipulations cuts word-error rates by up to 19.24% on unseen accented and children's speech. If these findings hold, robust foundation ASR models could be built from modest human recordings plus acoustically focused augmentation, instead of data collections on the 680k-hour scale.

What carries the argument

The load-bearing mechanism is the paper's acoustic-centric augmentation recipe. On the waveform side, pitch is shifted with gender-specific ranges (lowering pitch to simulate elderly speakers, raising it to simulate children) and amplitude is scaled by a uniform factor of 0.5–1.5. On the spectrogram side, the magnitude is normalized to 0–1, columns corresponding to vowel pronunciations are found by a 0.3 threshold and grouped by adjacency (each group is one vowel), then each group is randomly stretched or compressed in duration, has its columns randomly swapped, and is scaled in intensity by a factor in (0.5, 2.0), before denormalizing. This preserves the spectrogram-to-transcription mapping while enlarging the acoustic space, in contrast to SpecAugment's masking, which the authors argue extends linguistic diversity. The supporting probe is the purely synthetic pretraining run: 11 million WikiDialog sentences voiced by a single Kokoro TTS voice, which isolates the effect of linguistic diversity when acoustic diversity is nearly absent.

What would settle it

Train the same Whisper-base recipe on the same 11 million WikiDialog sentences rendered by a deliberately diverse set of synthetic voices (many accents, age ranges, and speaking rates), then evaluate on ENNI, MyST, and L2-Arctic: if real-speech WER stays above 100%, acoustic diversity is not the driver the paper claims, while a substantial WER drop would show that the single-voice setup caused the original failure and confirm the thesis. The converse check is equally decisive: train on a tiny text set, say one hundred sentences, recorded from hundreds of real speakers, and ask whether out-of-distribution WER drops.

Watch

Extended reading notes

Core claim

The paper's central claim is that an ASR foundation model's ability to transcribe out-of-distribution speech is set mainly by the acoustic variation in its training data, not by the variety of sentences it was trained on. The authors decompose human speech into an acoustic space (how a sound is produced: pitch, amplitude, duration, articulation) and a linguistic space (what the words are), and argue that the two spaces play different roles in pretraining. In the linguistic-diversity experiment, a Whisper-base model trained on purely synthetic data — 11 million WikiDialog sentences rendered by a single Kokoro TTS voice — reaches near-zero WER on synthetic audio but fails completely on real audio with WER above 100%. In the acoustic-diversity experiment, the same architecture trained on the 960-hour Librispeech set with acoustic augmentations (pitch shifting calibrated to simulate elderly and child speakers, amplitude scaling, and spectrogram-level manipulation of vowel groups) outperforms no augmentation, SpecAugment, Mixup, and SpecMix on the unseen ENNI, MyST, and L2-Arctic test sets, with up to 19.24% WER reduction over the unaugmented baseline. The paper also argues that SpecAugment's masking is chiefly a linguistic-diversity mechanism, because it forces the model to guess words from partial spectrograms, and is therefore a suboptimal investment for acoustic robustness.

Load-bearing premise

The central claim rests on the Section 4 synthetic-data experiment, which assumes that a corpus of 11 million WikiDialog sentences spoken by a single Kokoro TTS voice fairly represents 'linguistically rich, acoustically poor' training data, and that the resulting WER above 100% on real audio is caused by missing acoustic diversity rather than by the synthetic-to-real domain gap, training instability, or stopping after less than 2% of the corpus.

Editorial extensions

If this is right

  • A 960-hour Librispeech pretraining run with acoustic-centric augmentation beats the same run without augmentation on accented and children's speech by up to 19.24% WER, and also edges out Mixup and SpecMix baselines.
  • SpecAugment, the de facto augmentation used in Whisper-v2 pretraining, is a suboptimal choice for out-of-distribution robustness because masking primarily exercises linguistic rather than acoustic diversity.
  • Scaling linguistic diversity alone — an 11-million-sentence corpus spoken by one TTS voice — produces a model with near-zero WER on synthetic audio but WER above 100% on real human speech, so text-only synthetic pretraining cannot substitute for acoustic coverage.
  • Adding acoustic variation to synthetic data (54 voices plus timbre, speed, pitch, and amplitude changes) pulls real-speech WER down from above 100% to roughly 70–80%, indicating that synthetic-plus-acoustic pipelines are a viable middle path.
  • The augmentation gains continue to grow with additional training passes (up to 60k hours of effective seen data), so the method converts compute, not new data collection, into robustness.

Reading between the lines

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

  • Extending the paper's logic, the relevant scaling law for robust ASR would be speakers-per-hour rather than hours: a few hundred hours spread over thousands of distinct voices could rival corpora many times larger that come from a single accent, a trade-off the paper does not directly measure.
  • A cheap diagnostic follows from the thesis: before commissioning a large data collection, augment pitch and duration of existing audio and measure out-of-distribution word-error rate; if it does not move, acoustic diversity is not the binding constraint for the target domain.
  • The vowel-group manipulation is a simplified stand-in for articulatory variation; a phonetically graded successor that warps formant trajectories or perturbs specific vowel classes would test whether the mechanism is formant geometry rather than generic spectral perturbation.
  • The results also imply a data-collection budget rule for TTS-driven pipelines: invest in many-voice synthesis rather than larger text corpora, since Section 5.4 suggests multi-voice synthetic data transfers partially to real speech while single-voice data transfers not at all.
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

4 major / 7 minor

Summary. The paper asks whether ASR robustness in pretraining is driven more by acoustic or by linguistic diversity in the training data. It argues that SpecAugment is suboptimal because it augments linguistic rather than acoustic diversity, and it proposes acoustic-centric augmentations (pitch shifting, amplitude scaling, vowel-duration and vowel-intensity perturbation) for Whisper-base pretraining on Librispeech-960h. The experiments report that (i) a model pretrained on single-voice Kokoro TTS renderings of WikiDialog reaches near-zero WER on synthetic audio but fails (>100% WER) on real audio; (ii) the proposed augmentations improve out-of-distribution WER relative to baselines in Table 1; and (iii) adding 54 TTS voices with acoustic variation brings real-audio WER down to 70-80% before training collapses.

Significance. If the causal claim were established, the paper would be practically important: it would suggest that small real corpora plus acoustic augmentation can substitute for massive data collection, a claim with direct implications for foundation-model pretraining. The manuscript is clearly written and transparent about resource constraints, and the concrete augmentation recipe, together with the matched-budget rows in Table 1 (Ours 20k vs No Augmentation 20k), constitutes a useful preliminary empirical contribution. However, the headline conceptual claim that acoustic diversity, rather than linguistic richness, drives robustness is not supported by the present experimental design because the key comparisons are confounded. The paper does not provide code, which limits reproducibility of the augmentation pipeline.

major comments (4)
  1. [Section 4] The experiment that motivates the central claim compares a single-voice Kokoro TTS rendering of WikiDialog with real Librispeech audio. This joint variation changes the TTS domain gap, the number of voices (one), the text corpus, and the training budget (less than 2% of the corpus), so the WER >100% on real audio cannot be attributed specifically to the absence of acoustic diversity. A factorial control is needed: hold the text corpus fixed while varying the TTS voice count (one vs many), and hold the voice count fixed while varying the text corpus, then evaluate on the same real-audio test sets. Without such conditions, the paper's claim that linguistic richness is not the driver is not established.
  2. [Table 1 / Section 5.3] The headline 'up to 19.24% WER reduction' appears to come from comparing MyST WER 77.28 (No Augmentation, 20k hours) with 58.04 (Ours, 60k hours), a difference of 19.24 percentage points. Because the rows differ in total training hours, this gain conflates the augmentation method with a longer training budget. The matched-budget comparison (Ours 20k vs No Augmentation 20k) gives a smaller gain (77.28 vs 63.08, or 14.20 points), and at 20k hours the proposed method is not uniformly better than SpecMix: SpecMix achieves 22.30 vs 25.01 on Librispeech-other and 44.14 vs 44.90 on L2-Arctic. The abstract and Section 5.3 should either report the matched-budget comparison as the primary result or explicitly attribute the larger gain to augmentation plus additional training data.
  3. [Section 5.1] The claim that SpecAugment's time and frequency masking 'augments linguistic diversity' rather than acoustic diversity is asserted without a direct test. The lower out-of-distribution performance of SpecAugment in Table 1 could equally be explained by the destruction of salient spectrogram information (the explanation offered by SpecMix [6]) or by hyperparameter choices. To make this mechanism claim load-bearing, the authors need an experiment that varies masking type and measures whether time versus frequency masking affects linguistic versus acoustic robustness separately, or they should soften the claim to a conjecture.
  4. [Section 5.4] The synthetic-data ablation reports only a range ('around 70-80% WER') and mentions that further training collapses WER above 100%, without a table, the number of steps, or training curves. As reported, this result is consistent with a persistent synthetic-to-real domain gap and training instability, not specifically with the hypothesis that acoustic variation improves generalization. Please report the exact WER at the stable point, the step count, and the loss curves, and discuss why additional training degrades performance.
minor comments (7)
  1. [Section 5.3] Typo: 'Whipser-base' should be 'Whisper-base'.
  2. [Section 5.2] Typo: 'Liborsa' should be 'librosa'.
  3. [Table 1] The table formatting is ambiguous: the '20k' value appears only on the No Augmentation row, and it is unclear whether SpecAugment, Mixup, SpecMix, and Ours 20k also use 20k training hours; please make the Max Training Data column explicit for every row.
  4. [Abstract] The phrase 'up to 19.24% WER reduction' should state whether this is relative or absolute WER reduction and identify the exact dataset and training-hour pair used for the comparison.
  5. [Section 4] The statement that training ran 'until convergence' on less than 2% of the corpus should be accompanied by the actual number of optimization steps, batch size, and a validation-loss curve, since the convergence criterion affects the interpretation of the result.
  6. [Section 5.2] The vowel-identification threshold (0.3) and the intensity multiplication range (0.5, 2) are heuristics; a sensitivity analysis or a reference justifying these values would improve reproducibility.
  7. [References] Reference [28] points to the authors' own alignment tool rather than to a description of the ENNI corpus; please cite the dataset source or describe its collection and access conditions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical and the reported reductions are not forced by definition or fitted parameters.

full rationale

The paper does not present a derivation chain; it reports empirical comparisons. The central claim that acoustic diversity matters more than linguistic diversity is supported by a synthetic-data experiment (Section 4) and by augmentation comparisons (Table 1), not by a definition or a fitted parameter renamed as a prediction. The 19.24% WER reduction is an observed result on held-out benchmarks (L2-Arctic, MyST, ENNI), and none of those numbers is used as a training objective or as an input to the method. The two self-citations ([2] for background on ASR challenges and [28] for the ENNI evaluation dataset) are not load-bearing for the main claim: ENNI is an external speech corpus used for evaluation, and the paper's conclusion does not reduce to the existence of that dataset. The Section 4 experiment is confounded (single TTS voice, synthetic-to-real domain gap, stopping before 2% of the corpus), but confounding is a validity concern, not circularity: the experimental outcome is not logically equivalent to the setup by construction. No equation is reused as its own prediction, and no fitted parameter is presented as a prediction. The paper is self-contained against external benchmarks and baselines, so the honest finding is no significant circularity.

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

The empirical claims rest on a set of domain assumptions that are stated but not independently validated: the linguistic/acoustic decomposition of speech, the use of a single TTS voice to represent acoustic poverty, the claim that SpecAugment acts mainly linguistically, and heuristic vowel-boundary detection. All augmentation hyperparameters are hand-chosen. No new entities are postulated.

free parameters (5)
  • pitch shift distribution mapping = Table 2: probability 0.2 to 0.3 per gender, shift bounds -4 to +6 semitones
    Hand-chosen to simulate elderly and children voices; affects out-of-distribution WER.
  • amplitude scale range = 0.5 to 1.5
    Hand-chosen random gain augmentation applied to all samples.
  • vowel column threshold = 0.3 normalized magnitude
    Used to identify vowel columns in the mel spectrogram; hand-chosen.
  • vowel intensity multiplication range = 0.5 to 2
    Hand-chosen factor applied to each vowel group during augmentation.
  • training budget in hours = 20k, 40k, 60k hours
    The paper reports Ours at three budgets; comparing Ours 60k to No Augmentation at 20k confounds augmentation with training duration.
assumptions (6)
  • domain assumption Human speech can be decomposed into independent linguistic and acoustic feature spaces.
    Section 3 introduces this decomposition and uses it to interpret all subsequent experiments; no evidence is given that the two feature spaces are independent.
  • ad hoc to paper A single-voice Kokoro TTS corpus represents linguistically rich but acoustically poor training data.
    Section 4 relies on this equivalence to attribute the synthetic model's failure to missing acoustic diversity.
  • ad hoc to paper SpecAugment's time and frequency masking augments linguistic diversity rather than acoustic diversity.
    Section 5.1 argues this without measuring what SpecAugment changes in the trained representations.
  • ad hoc to paper Vowel pronunciations can be identified by a fixed 0.3 threshold on the normalized mel spectrogram and modified without breaking the transcription mapping.
    Section 5.2 describes this heuristic augmentation pipeline; no validation is provided that the modifications preserve phonetic content.
  • domain assumption Whisper-base trained from scratch on 960h Librispeech is representative of foundation ASR pretraining.
    Sections 4 and 5 base all conclusions on this single architecture and corpus.
  • domain assumption ENNI, MyST, and L2-Arctic are valid out-of-distribution evaluation sets for robustness.
    Section 5.3 uses these datasets to measure generalization; ENNI is semi-proprietary and is tied to the authors' own reference [28].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Pretraining Robust ASR Foundation Model with Acoustic-Aware Data Augmentation." pith.science (2026). https://pith.science/paper/EC43MMLB

@misc{pith2026250520606,
  author       = {Pith},
  title        = {Pith review of: Towards Pretraining Robust ASR Foundation Model with Acoustic-Aware Data Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EC43MMLB}},
  note         = {Machine review of arXiv:2505.20606}
}
read the original abstract

Whisper's robust performance in automatic speech recognition (ASR) is often attributed to its massive 680k-hour training set, an impractical scale for most researchers. In this work, we examine how linguistic and acoustic diversity in training data affect the robustness of the ASR model and reveal that transcription generalization is primarily driven by acoustic variation rather than linguistic richness. We find that targeted acoustic augmentation methods could significantly improve the generalization ability of ASR models, reducing word-error rates by up to 19.24 percent on unseen datasets when training on the 960-hour Librispeech dataset. These findings highlight strategic acoustically focused data augmentation as a promising alternative to massive datasets for building robust ASR models, offering a potential solution to future foundation ASR models when massive human speech data is lacking.

Figures

Figures reproduced from arXiv: 2505.20606 by the authors.

Figure 1
Figure 1. High-level pipeline of pre-training and ASR model with synthetic data and with data augmentation strategies on the real data. gram. SpecMix [6] combines the idea of SpecAumgnet [5] and Mixup [13] to create time and frequency masks using other speech samples instead of zeroing them out. Later in Section 5, we will show that SpecAugment [5] is actually a subopti￾mal augmentation strategy due to its indirect relationsh… view at source ↗
Figure 2
Figure 2. Normalized log mel spectrogram before (top) and af￾ter (bottom) acoustic augmentation. columns representing vowel pronunciations by a threshold (0.3 in the implementation) and group those columns by adjacency (each group then represents one complete vowel pronunciation); 3. randomly change the duration of vowel pronunciations by re￾peating or removing some columns of vowel pronunciations; 4. randomly swap vowel colu… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 30 canonical work pages

  1. [6]

    Analysis of Data Augmentation from a Human Speech Perspective As mentioned earlier, human speech is conceptually a combina- tion of how people speak (acoustics) and what they speak (lin- guistics). In this section, we provide a high-level discussion of how these linguistic and acoustic features affect the diversity of the ASR model’s training set, which i...

    work page

  2. [1]

    Introduction Automatic speech recognition (ASR) technology empowers nu- merous real-world applications from interactive voice assistants to educational and medical systems. Among existing ASR mod- els, Whisper [1] stands out for its robust performance across di- verse audio inputs, particularly accented and children’s speech, where most ASR models face si...

    work page

  3. [2]

    We discuss the difference between linguistic and acoustic variations in human speech and their connections to the ro- bustness of foundation ASR models

    work page

  4. [3]

    We show that ASR models benefit mainly from the acoustic diversity in the training data than the linguistic diversity

    work page

  5. [4]

    We demonstrate that acoustic-centric data augmentation strategies can significantly improve the robustness of ASR models, especially when the training data is not massive

    work page

  6. [5]

    Towards Pretraining Robust ASR Foundation Model with Acoustic-Aware Data Augmentation

    Related Works 2.1. Speech Data Augmentation Techniques Speech data augmentation has been a well-established field in the research community. From the early works leveraging sim- ple transformations such as adding noise [8], changing playback speed [9], applying reverberation [10], and concatenation [11], to more modern approaches like SpecAugment [5], Spe...

    work page Pith review arXiv 2025

  7. [7]

    Pre-training with Linguistic Diversity Results from prior works using synthetically generated data seem to indicate that synthetic data could lead to model robust- ness [3, 4]. Indeed, synthetic data could increase the diversity of transcription in the training data, which will fill in the gaps of missing inter-word dependencies in natural languages when ...

    work page

  8. [8]

    Deep speech: Scaling up end-to-end speech recognition,

    A. Hannun, C. Case, J. Casper, B. Catanzaro, G. Diamos, E. Elsen, R. Prenger, S. Satheesh, S. Sengupta, A. Coates, and A. Y . Ng, “Deep speech: Scaling up end-to-end speech recognition,” 2014. [Online]. Available: https://arxiv.org/abs/ 1412.5567

    arXiv 2014

Show all 43 references
  1. [9]

    In this section, we focus on acoustic data augmentation techniques and investigate their ef- fects on the robustness of the pre-trained ASR models

    Pre-training with Acoustic Diversity In Section 4, empirical results show that pre-training with only synthetic diversity limits the model’s ability to generalize be- yond the synthetic distribution. In this section, we focus on acoustic data augmentation techniques and invest...

  2. [10]

    When data sources are lim- ited, acoustic augmentations can significantly outperform exist- ing data augmentation methods on unseen speech

    Conclusion In conclusion, through analysis and experiments, this work re- veals that robust ASR performance existing in foundation ASR models relies primarily on acoustic diversity as opposed to lin- guistic diversity in the training data. When data sources are lim- ited, acou...

  3. [11]

    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,” 2022. [Online]. Available: https: //arxiv.org/abs/2212.04356

    2022 arXiv

  4. [12]

    Automatic screening for children with speech disorder using automatic speech recognition: Opportunities and challenges,

    D. Liu, J. Yang, I. Albrecht-Buehler, H. Qin, S. Li, Y . Hu, A. Nassereldine, and J. Xiong, “Automatic screening for children with speech disorder using automatic speech recognition: Opportunities and challenges,” 2024. [Online]. Available: https://arxiv.org/abs/2410.11865

    2024 arXiv

  5. [13]

    Synthasr: Unlocking synthetic data for speech recognition,

    A. Fazel, W. Yang, Y . Liu, R. Barra-Chicote, Y . Meng, R. Maas, and J. Droppo, “Synthasr: Unlocking synthetic data for speech recognition,” inInterspeech 2021, 2021, pp. 896–900

    2021

  6. [14]

    On the effect of purely synthetic training data for different automatic speech recognition architectures,

    B. Hilmes, N. Rossenbach, and R. Schl ¨uter, “On the effect of purely synthetic training data for different automatic speech recognition architectures,” inSynthetic Data’s Transformative Role in F oundational Speech Models, 2024, pp. 46–50

    2024

  7. [15]

    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,” inInterspeech 2019, 2019, pp. 2613–2617

    2019

  8. [16]

    Specmix : A mixed sample data augmentation method for training with time-frequency do- main features,

    G. Kim, D. K. Han, and H. Ko, “Specmix : A mixed sample data augmentation method for training with time-frequency do- main features,” inInterspeech 2021, 2021, pp. 546–550

    2021

  9. [17]

    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), 2015, pp. 5206–5210

    2015

  10. [18]

    Adversarial audio synthesis,

    C. Donahue, J. McAuley, and M. Puckette, “Adversarial audio synthesis,” inICLR, 2019

    2019

  11. [19]

    Audio aug- mentation for speech recognition,

    T. Ko, V . Peddinti, D. Povey, and S. Khudanpur, “Audio aug- mentation for speech recognition,” inInterspeech 2015, 2015, pp. 3586–3589

    2015

  12. [20]

    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,” in2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 5220–5224

    2017

  13. [21]

    Make more of your data: Minimal effort data augmentation for automatic speech recog- nition and translation,

    T. K. Lam, S. Schamoni, and S. Riezler, “Make more of your data: Minimal effort data augmentation for automatic speech recog- nition and translation,” inICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  14. [22]

    Specaugment++: A hidden space data augmentation method for acoustic scene classifica- tion,

    H. Wang, Y . Zou, and W. Wang, “Specaugment++: A hidden space data augmentation method for acoustic scene classifica- tion,” inInterspeech 2021, 2021, pp. 551–555

    2021

  15. [23]

    mixup: Beyond empirical risk minimization,

    H. Zhang, “mixup: Beyond empirical risk minimization,”arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  16. [24]

    Sapaugment: Learning a sample adaptive policy for data augmentation,

    T.-Y . Hu, A. Shrivastava, J.-H. R. Chang, H. Koppula, S. Braun, K. Hwang, O. Kalinli, and O. Tuzel, “Sapaugment: Learning a sample adaptive policy for data augmentation,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2...

    2021

  17. [25]

    G- augment: Searching for the meta-structure of data augmentation policies for asr,

    G. Wang, E. D. Cubuk, A. Rosenberg, S. Cheng, R. J. Weiss, B. Ramabhadran, P. J. Moreno, Q. V . Le, and D. S. Park, “G- augment: Searching for the meta-structure of data augmentation policies for asr,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 23–30

    2023

  18. [26]

    Sample adaptive data augmentation with progressive scheduling,

    H. Lu and B. Li, “Sample adaptive data augmentation with progressive scheduling,” 2024. [Online]. Available: https: //arxiv.org/abs/2412.00415

    2024 arXiv

  19. [27]

    Natural tts synthesis by condi- tioning wavenet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, R. Skerrv-Ryan, R. A. Saurous, Y . Agiomvrgiannakis, and Y . Wu, “Natural tts synthesis by condi- tioning wavenet on mel spectrogram predictions,” in2018 IEEE International Conference...

    2018

  20. [28]

    Fasa: a flexible and automatic speech aligner for extracting high-quality aligned children speech data,

    D. Liu and J. Xiong, “Fasa: a flexible and automatic speech aligner for extracting high-quality aligned children speech data,”

  21. [29]

    Kokoro-82m (revision d8b4fc7),

    Hexgrad, “Kokoro-82m (revision d8b4fc7),” 2025. [Online]. Available: https://huggingface.co/hexgrad/Kokoro-82M

    2025

  22. [30]

    Improving accented speech recognition using data augmentation based on unsupervised text-to-speech synthesis,

    C.-T. Do, S. Imai, R. Doddipatla, and T. Hain, “Improving accented speech recognition using data augmentation based on unsupervised text-to-speech synthesis,” 2024. [Online]. Available: https://arxiv.org/abs/2407.04047

    2024 arXiv

  23. [31]

    Investigating the use of syn- thetic speech data for the analysis of spanish-accented english pronunciation patterns in asr,

    M. Masson and J. Carson-Berndsen, “Investigating the use of syn- thetic speech data for the analysis of spanish-accented english pronunciation patterns in asr,” inSynthetic Data’s Transformative Role in F oundational Speech Models, 2024, pp. 81–85

    2024

  24. [32]

    Asr data augmentation in low-resource settings using cross-lingual multi- speaker tts and cross-lingual voice conversion,

    E. Casanova, C. Shulby, A. Korolev, A. C. Junior, A. da Silva Soares, S. Alu ´ısio, and M. A. Ponti, “Asr data augmentation in low-resource settings using cross-lingual multi- speaker tts and cross-lingual voice conversion,” inInterspeech 2023, 2023, pp. 1244–1248

    2023

  25. [33]

    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,” 2024. [Online]. Available: https://arxiv.org/abs/2406.08568

    2024 arXiv

  26. [34]

    Dialog inpainting: Turning documents to dialogs,

    Z. Dai, A. T. Chaganty, V . Zhao, A. Amini, M. Green, Q. Rashid, and K. Guu, “Dialog inpainting: Turning documents to dialogs,” inInternational Conference on Machine Learning (ICML). PMLR, 2022

    2022

  27. [35]

    istftnet: Fast and lightweight mel-spectrogram vocoder incorporating inverse short-time fourier transform,

    T. Kaneko, K. Tanaka, H. Kameoka, and S. Seki, “istftnet: Fast and lightweight mel-spectrogram vocoder incorporating inverse short-time fourier transform,” 2022. [Online]. Available: https://arxiv.org/abs/2203.02395

    2022 arXiv

  28. [36]

    Styletts 2: Towards human-level text-to- speech through style diffusion and adversarial training with large speech language models,

    Y . A. Li, C. Han, V . S. Raghavan, G. Mischler, and N. Mesgarani, “Styletts 2: Towards human-level text-to- speech through style diffusion and adversarial training with large speech language models,” 2023. [Online]. Available: https://arxiv.org/abs/2306.07691

    2023 arXiv

  29. [37]

    librosa/librosa: 0.10.2,

    B. McFee and et. al., “librosa/librosa: 0.10.2,” May 2024. [Online]. Available: https://doi.org/10.5281/zenodo.4923181

    2024 doi

  30. [40]

    My science tutor (myst) – a large corpus of children’s conversational speech,

    S. S. Pradhan, R. A. Cole, and W. H. Ward, “My science tutor (myst) – a large corpus of children’s conversational speech,”

  31. [42]

    L2-arctic: A non- native english speech corpus,

    G. Zhao, S. Sonsaat, A. Silpachai, I. Lucic, E. Chukharev- Hudilainen, J. Levis, and R. Gutierrez-Osuna, “L2-arctic: A non- native english speech corpus,” inInterspeech 2018, 2018, pp. 2783–2787

    2018

  32. [43]

    Phonological differences between received pronunciation and standard scottish english,

    R. Cao and S. Jin, “Phonological differences between received pronunciation and standard scottish english,” inProceedings of the 2017 International Seminar on Social Science and Humanities Research (SSHR 2017). Atlantis Press, 2017/12, pp. 121–130. [Online]. Available: https:/...

    2017 doi

  33. [64]

    All experiments are conducted on a server with 4 A6000 GPUs

    whenever experiments are conducted. All experiments are conducted on a server with 4 A6000 GPUs. We train our model until convergence, which surprisingly takes less than 2% of the training corpus. We then evaluate the pre-trained ASR model on both synthetic data and real-world...

  34. [2023]

    Available: https://arxiv.org/abs/2309.13347

    [Online]. Available: https://arxiv.org/abs/2309.13347

    arXiv

  35. [2024]

    Available: https://arxiv.org/abs/2406.17926

    [Online]. Available: https://arxiv.org/abs/2406.17926

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.