Adaptability of ASR Models on Low-Resource Language: A Comparative Study of Whisper and Wav2Vec-BERT on Bangla

Md. Aminur Rahman; Md Sazzadul Islam Ridoy; Sumi Akter

REVIEW 5 major objections 6 minor 2 cited by

Adaptability of ASR Models on Low-Resource Language: A Comparative Study of Whisper and Wav2Vec-BERT on Bangla

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read On Bangla, Wav2Vec-BERT fine-tuned on 70,000 utterances reaches a word error rate of 14.42%, beating Whisper Large-v2's 28.86% while training in less time and fitting on a 12 GB GPU.

desk verdict Useful Bangla ASR benchmark, but the headline accuracy gap rests on an unablated 8 kHz round-trip and an underspecified t-test. read the letter →

arxiv 2507.01931 v1 pith:KDDCQ67T submitted 2025-07-02 cs.CL cs.AIcs.SDeess.AS

Md Sazzadul Islam Ridoy , Sumi Akter , Md. Aminur Rahman This is my paper

classification cs.CLcs.AIcs.SDeess.AS

keywords automaticspeechrecognitionbanglaasrwav2vec-bertwhisperrepresentationmodelspretrainedtransformerlow-resourcelanguage

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 which pretrained speech-recognition model adapts better to Bangla, a language with scarce annotated speech. By fine-tuning Whisper (Small and Large-v2) and Wav2Vec-BERT on Bangla subsets that grow from 2,000 to 70,000 utterances, the study reports that Wav2Vec-BERT wins on every measured axis: word error rate 14.42% versus Whisper Large-v2's 28.86%, character error rate 2.67% versus 7.47%, shorter training time (13:26 versus 21:52), and the ability to complete fine-tuning on a 12 GB GPU at all dataset sizes, where Whisper Large-v2 failed and Whisper Small hit memory limits beyond 20k samples. The authors conclude that self-supervised Wav2Vec-BERT is the more practical choice for resource-constrained Bangla ASR, and that Whisper's extra data and compute do not translate into better accuracy here.

What carries the argument

Wav2Vec-BERT carries the argument: a convolutional encoder converts raw audio into latent speech representations, a Conformer adapter (a transformer variant with convolutional modules) and a bidirectional BERT-style transformer are pretrained to predict masked speech segments, and fine-tuning then needs only modest labeled data. Whisper is the contrast case: an encoder-decoder transformer pretrained on roughly 680,000 hours of weakly supervised labeled audio, consuming 80-channel log-Mel spectrograms and requiring much more memory. The comparison mechanism is a controlled fine-tuning pipeline using two public Bangla corpora (about 86 hours total), five dataset subsets from 2k to 70k samples, three hyperparameter configurations per model, and two GPU tiers. The claimed superiority of Wav2Vec-BERT rests on its lower WER and CER at every scale, faster training, and memory headroom on a 12 GB GPU.

What would settle it

Fine-tune Whisper Large-v2 on the same 70,000-utterance Bangla subset with a substantially wider hyperparameter search, including effective batch size matched to Wav2Vec-BERT via gradient accumulation, learning rates spanning 1e-6 to 1e-4, and longer epoch ranges; if any Whisper configuration reaches WER at or below Wav2Vec-BERT's 14.42% on the same test set, the paper's claim that Wav2Vec-BERT outperforms Whisper across all key metrics is overturned.

Watch

Extended reading notes

Core claim

On its own datasets and tuning protocol, the paper's central finding is that Wav2Vec-BERT is the more adaptable model for Bangla ASR. At the largest training scale of 70,000 samples with 8 epochs and learning rate 1e-5, Wav2Vec-BERT reaches WER 14.42% and CER 2.67%, while Whisper Large-v2's best configuration (70k samples, 10 epochs, 3e-5) reaches WER 28.86% and CER 7.47%. Wav2Vec-BERT trains in 13:26 versus 21:52 for Whisper Large-v2, and only Wav2Vec-BERT completes all dataset sizes on the lower-memory GPU. Paired t-tests on the 70k setting give p = 0.0041 for WER and p = 0.0037 for CER, which the paper takes as evidence the gap is not random. The authors interpret the overall pattern as showing that self-supervised speech representations are particularly valuable when labeled data are limited.

Load-bearing premise

The comparison is fair: both models received equivalent fine-tuning effort, so the chosen hyperparameter grids, batch sizes, learning rates, and checkpoint selections do not systematically favor Wav2Vec-BERT over Whisper.

Editorial extensions

If this is right

A Bangla ASR system on a single consumer-grade GPU should start with Wav2Vec-BERT, since it delivers the lower error rates, trains faster, and fits in 12 GB of VRAM at all dataset sizes tested.
Whisper's much larger pretraining corpus and parameter count do not buy accuracy on Bangla; Whisper Large-v2 ends with roughly double Wav2Vec-BERT's WER despite requiring the high-end GPU.
Wav2Vec-BERT's WER flattens after about 40k samples, so for this test distribution roughly 40k-70k labeled utterances is enough to reach its practical floor.
Whisper Small, not Large-v2, is the feasible Whisper option when high-end hardware is unavailable, with best WER 32.61% at 40k samples.
The two models make different phoneme-level errors (nasal and retroflex confusions for Wav2Vec-BERT; fricative, aspiration, and numeral errors for Whisper), so targeted augmentation or an ensemble could push accuracy beyond either model.

Reading between the lines

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

The preprocessing step that downsamples audio from 16 kHz to 8 kHz and back could help Wav2Vec-BERT, a raw-waveform learner, more than Whisper, which expects 16 kHz log-Mel spectrograms; a follow-up could fine-tune Whisper on native 16 kHz audio to test whether part of the gap is a preprocessing artifact.
The tuning protocol allowed Wav2Vec-BERT larger batch sizes and forced Whisper into gradient accumulation to fit memory, so some of the advantage may come from optimization conditions rather than architecture alone; matching effective batch sizes or using parameter-efficient tuning for Whisper would separate these factors.
Because Wav2Vec-BERT's error curve saturates around 40k samples, the paper's results suggest that for Bangla, collecting more labeled data may matter less than improving text normalization and coverage of rare phoneme contrasts.
If the pattern holds, self-supervised speech models may be the default choice for other low-resource Indo-Aryan languages whose sound systems resemble Bangla, although the paper only evaluates Bangla.

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.

Desk Editor's Note

read the letter

Honest take: this is a useful benchmark for anyone choosing between Whisper and Wav2Vec-BERT on Bangla, and the first direct comparison I know of on that language with dataset scaling and hardware variation. The authors report real training time, memory failures, and a phoneme-level error analysis, all on public datasets. That is worth having.

The main accuracy claim, though, is not as secure as the abstract suggests. All audio is resampled 16 kHz -> 8 kHz -> 16 kHz before training and evaluation, and there is no baseline without that round-trip. Whisper's front end is an 80-bin log-Mel over the 0-8 kHz band, so a 4 kHz cutoff zeros out a large chunk of its input features. Wav2Vec-BERT consumes raw waveforms and may be less sensitive to the same band limit. The paper says the round-trip is a form of augmentation, but it is actually a confound: the reported WER gap (14.42 vs. 28.86 at 70k) could be partly an artifact of preprocessing that disproportionately hurts Whisper. The resource-use difference is probably real, but the headline claim is accuracy plus efficiency, so the accuracy gap needs to survive a standard 16 kHz run. This is not a manufactured flaw; the stress-test note lands.

The other issues are softer but real. The paired t-test in Section IV-C is not reproducible: no number of paired observations, no definition of what is paired across models, no error bars or confidence intervals anywhere. The hyperparameter search is three configs per model, with different learning rates and batch sizes that favor Wav2Vec-BERT (larger batches, different LR), and the "best" result is selected post hoc. That does not sink the comparison, but it does mean "outperforms" is conditional on tuning protocol. There are also small reporting inconsistencies: Section IV-B says Whisper Small best WER is 32.61% on 40k, Table II says 32.17; those need cleanup.

Who is this for? Practitioners building Bangla ASR systems who want a starting point on model choice. It is not a methodological advance. I would send it to peer review rather than desk reject, because the comparison is useful and the data are real, but a serious referee should demand the 16 kHz ablation, a proper statistical test, and corrected tables. If the authors deliver those, the paper becomes a solid within-field benchmark.

Referee Report

5 major / 6 minor

Summary. This paper presents a comparative empirical study of Wav2Vec-BERT, Whisper Small, and Whisper Large-v2 for Bangla automatic speech recognition. The authors fine-tune the models on subsets of Mozilla Common Voice-17 and OpenSLR ranging from 2,000 to 70,000 samples, and they evaluate WER, CER, training time, and computational efficiency on two GPU setups. They report that Wav2Vec-BERT achieves the best accuracy (WER 14.42%, CER 2.67% at 70k samples), requires less training time than Whisper Large-v2 (13:26 vs. 21:52), and runs on hardware where Whisper Large-v2 fails. They also report paired t-tests indicating significant differences and provide a qualitative phoneme-level error analysis. The central claim is that Wav2Vec-BERT outperforms Whisper across all key evaluation metrics for Bangla ASR.

Significance. If the result holds, the paper offers a useful empirical data point for low-resource Bangla ASR: Wav2Vec-BERT appears to provide a better accuracy-efficiency trade-off than Whisper, which is relevant to practitioners building ASR systems under resource constraints. The scaling analysis across five dataset sizes and two hardware configurations is a meaningful contribution, as is the qualitative error analysis that identifies distinct phoneme-level failure patterns. However, the strength of the claim is currently conditional because of three load-bearing issues: the unablated 8 kHz round-trip preprocessing, the absence of any variance or confidence-interval reporting, and the incomplete description of the paired t-test. The lack of code and random seeds further prevents reproduction. These issues are fixable in principle, so the paper's contribution is not yet fully established but warrants a revision.

major comments (5)

[§III-B] The audio preprocessing step resamples every file from 16 kHz to 8 kHz and back to 16 kHz for both training and evaluation, with no ablation or native-16 kHz baseline reported. Whisper's input is an 80-bin log-Mel spectrogram spanning 0–8 kHz, so a 4 kHz band limit zeros a large fraction of its input features; Wav2Vec-BERT's raw-waveform CNN may have different sensitivity to the same band limiting. Since the headline result (Table II: WER 14.42% vs. 28.86% at 70k samples) depends on this unablated pipeline, the reported accuracy gap could reflect a preprocessing artifact rather than a robust architectural advantage. The authors should add a 16 kHz baseline for both model families and show that the ranking is invariant to this round-trip resampling.
[§IV-C] The paired t-test is not reproducible and, as described, cannot be computed. The authors report p-values for '70k dataset size (8 epochs)' comparing Whisper Large-v2 and Wav2Vec-BERT, but with a single WER and CER value per model at that configuration, a paired t-test requires at least two paired observations; no number of pairs, no seeds, no test-set partitions, and no within-configuration repeats are reported. The p = 0.0041 and p = 0.0037 claims are therefore unsupported. The authors must state the number of paired observations, define what constitutes a pair, and report variance estimates or confidence intervals.
[§III-D(2) and Table I] The model comparison is not controlled. Wav2Vec-BERT was allowed larger batch sizes than Whisper, and gradient accumulation is mentioned only for Whisper Large-v2, leaving the effective batch sizes unequal and unclear. In addition, each model's 'best' configuration in Table II is selected post hoc from a small three-configuration grid, so the claim that Wav2Vec-BERT 'outperformed Whisper across all key evaluation metrics' reflects favorable choices rather than a controlled comparison. Table I shows strong sensitivity to the grid (e.g., Whisper Small WER ranges from 32.28% to 33.91% at 70k samples), so the ranking at the headline row could shrink or reverse under a different, equally reasonable tuning protocol. The paper should report the full grid results and/or use an equal tuning budget (same number of runs, same effective batch size, same early-stopping criterion) for both models.
[§IV-B and Table II] There is a numerical inconsistency in the reported best Whisper Small result. The text states 'Whisper Small achieved the best WER of 32.61% and CER of 18.17% on the 40k dataset,' but Table II lists WER 32.17% for the same configuration (40k samples, 15 epochs, lr 1e-5), and Table I does not include a 40k row. The paper does not explain this discrepancy, and the 'best' configuration selection is therefore not reproducible. Please correct the numbers and ensure the text, tables, and figures agree.
[§IV-A and Table I] All WER and CER results are reported as single values without standard deviations, confidence intervals, or multiple runs across seeds. Given that the paper's central claim is a comparative advantage (e.g., 14.42% vs. 28.86% WER at 70k samples), the absence of any variance estimate means the reader cannot assess whether the gap is larger than run-to-run noise. At minimum, the authors should provide repeated fine-tuning runs or bootstrap intervals over test utterances for the headline configurations in Table II.

minor comments (6)

[§II-A and Table III] The Bangla script examples are garbled in the manuscript (e.g., 'Worbo¯No', 'ibPÐ'), making these passages unreadable; please use Unicode with a font that renders the script correctly.
[§III-B] The phrase 'as a form of augmentation [13]' is not supported by the cited reference, which concerns Wav2Vec2 for child speech recognition; please provide a relevant reference or clarify how the round-trip resampling is expected to improve robustness.
[§I] The claim 'To the best of our knowledge, this is the first comprehensive analysis...' is unverifiable without a more thorough related-work comparison; please temper the claim or support it with a literature review.
[Table III] The table's layout is unclear: the first row labels the error type as '(Wav2Vec-BERT)' while the subsequent rows label errors as Whisper-related, and the 'True Text' column appears to contain the same string as one of the model outputs in some rows; please restructure the table so each row clearly shows the true transcription, each model's output, and the model(s) that made the error.
[§IV-C] The sentence 'A p-value measures the likelihood that the results occurred under the null hypothesis' is imprecise; the p-value is the probability of observing data as extreme as the observed data under the null hypothesis, not the probability that the results occurred.
[Figures 4 and 5] The figure captions state 'on the OpenSLR and Common Voice-17 Test Set,' but the text does not specify how the test set is constructed from the two datasets; please clarify the test-set composition and split.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark; the claimed Wav2Vec-BERT advantage is a measured result under a specified pipeline, not a quantity derived from its own inputs.

full rationale

This paper does not present a derivation chain that could be circular. The central claim, that Wav2Vec-BERT outperforms Whisper on Bangla ASR, is an empirical measurement reported in Tables I and II, based on fine-tuning both models on public datasets and evaluating WER/CER on a test set. No fitted parameter is renamed as a prediction: the best hyperparameter configuration is selected from a small grid and then reported, which is a post-hoc selection issue, not a case of predicting a quantity that was used to fit the model. The statistical test is applied to the selected configuration, but the WER/CER values are direct evaluations, not outputs reconstructed from the training objective. The only notable methodological concern is the unablated 16 kHz -> 8 kHz -> 16 kHz resampling described in Section III-B, which could differentially affect the two model families and thus weaken external validity. That is a confounding-variable concern, not a circularity concern: the result does not reduce by construction to its inputs, and there is no self-citation chain carrying the argument. The authors do not invoke their own prior results as evidence, and none of the load-bearing claims depend on a self-citation. Accordingly, no circular step meeting the required evidentiary standard can be identified, and the appropriate score is 0.

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

The central claim rests on standard model checkpoints, public datasets, and hand-selected hyperparameters. There are no new physical or algorithmic entities introduced. The main ledger items are the unablated preprocessing choices and the unspecified statistical test design, which a reader would need to understand fully to reproduce the comparison.

free parameters (4)

Learning rate values = 1e-5, 3e-5, 5e-5
Three learning rates were tested per model and the best was chosen post hoc. The central comparison depends on these arbitrary choices; no sensitivity analysis is reported.
Number of epochs = 8, 10, 15 depending on dataset size
Epoch counts were selected by hand: 10/15 for small subsets and 8/10 for large subsets. The best result is dependent on this choice.
Batch size / gradient accumulation = Not reported numerically
Batch sizes were adjusted to GPU memory and gradient accumulation was used every four steps for Whisper Large-v2. Exact values are missing, which affects both the accuracy and the training time comparison.
Audio resampling schedule = 16 kHz -> 8 kHz -> 16 kHz
The authors deliberately resampled audio down and back up as augmentation. This is a nonstandard preprocessing choice that could affect both models unevenly and is not ablated.

assumptions (4)

domain assumption The Common Voice-17 and OpenSLR Bangla datasets provide correctly transcribed training and test audio, and the WER/CER computation is standard.
The paper assumes the public data labels are accurate and that the tokenizer used for WER/CER matches the models' output. No manual verification or tokenizer details are given.
standard math The paired t-test conditions hold, including paired observations and independent differences, and the test is correctly computed from the available WER/CER values.
Section IV-C reports p-values but does not state the number of paired samples, the grouping variable, or whether the differences are normally distributed. The test cannot be verified from the paper.
domain assumption Whisper's Bangla language token <|bn|> and Wav2Vec-BERT's default tokenizer are correctly configured for Bangla text.
No details are given for how text is tokenized or normalized for each model beyond the general description in Section III-B, so an incorrect language token or vocabulary mapping would change the results.
domain assumption The unique-prompt filtering used to build dataset subsets does not introduce a bias that favors one model over the other.
The paper filters duplicates to prioritize unique prompts but does not analyze whether this changes the difficulty distribution for either model.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptability of ASR Models on Low-Resource Language: A Comparative Study of Whisper and Wav2Vec-BERT on Bangla." pith.science (2026). https://pith.science/paper/KDDCQ67T

@misc{pith2026250701931,
  author       = {Pith},
  title        = {Pith review of: Adaptability of ASR Models on Low-Resource Language: A Comparative Study of Whisper and Wav2Vec-BERT on Bangla},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KDDCQ67T}},
  note         = {Machine review of arXiv:2507.01931}
}

read the original abstract

In recent years, neural models trained on large multilingual text and speech datasets have shown great potential for supporting low-resource languages. This study investigates the performances of two state-of-the-art Automatic Speech Recognition (ASR) models, OpenAI's Whisper (Small & Large-V2) and Facebook's Wav2Vec-BERT on Bangla, a low-resource language. We have conducted experiments using two publicly available datasets: Mozilla Common Voice-17 and OpenSLR to evaluate model performances. Through systematic fine-tuning and hyperparameter optimization, including learning rate, epochs, and model checkpoint selection, we have compared the models based on Word Error Rate (WER), Character Error Rate (CER), Training Time, and Computational Efficiency. The Wav2Vec-BERT model outperformed Whisper across all key evaluation metrics, demonstrated superior performance while requiring fewer computational resources, and offered valuable insights to develop robust speech recognition systems in low-resource linguistic settings.

Figures

Figures reproduced from arXiv: 2507.01931 by the authors.

**Figure 3.** Architecture of the Fine-Tuning Process for Bangla ASR using [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 5.** Wav2Vec-BERT and Whisper CER Results on the OpenSLR and [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Wav2Vec-BERT and Whisper WER Results on the OpenSLR and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

When Surveys Become Conversations: Adaptive Matrix Validation for AI-Assisted Interviews
stat.ME 2026-06 unverdicted novelty 6.0 of 10

Adaptive Matrix Validation calibrates AI-mapped survey responses using sparse randomized validation questions from other respondents then corrects with the target's own answers, with estimators and planning formulas f...
A Unified Denoising and Adaptation Framework for Self-Supervised Bengali Dialectal ASR
cs.SD 2025-08 conditional novelty 5.0 of 10

A two-stage, noise-augmented fine-tuning of WavLM achieves state-of-the-art WER/CER on Bengali dialectal ASR under clean and noisy conditions.

Reference graph

Works this paper leans on

21 extracted references · 14 canonical work pages · cited by 2 Pith papers

[1]

A comparison of sequence-to-sequence models for speech recognition

R. Prabhavalkar, K. Rao, T. N. Sainath, B. Li, L. Johnson, and N. Jaitly, “A comparison of sequence-to-sequence models for speech recognition.” in Interspeech, 2017, pp. 939–943

work page 2017
[2]

Transformers: State- of-the-art natural language processing,

T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y . Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. Rush, “Transformers: State- of-the-art natural language processing,” in Proceedings of the 2020 Conference on Empirical Meth...

work page 2020
[3]

A survey on contrastive self-supervised learning,

A. Jaiswal, A. R. Babu, M. Z. Zadeh, D. Banerjee, and F. Makedon, “A survey on contrastive self-supervised learning,” Technologies, vol. 9, no. 1, p. 2, 2020

2020
[4]

Bangla speech recognition system using lpc and ann,

A. K. Paul, D. Das, and M. M. Kamal, “Bangla speech recognition system using lpc and ann,” in 2009 seventh international conference on advances in pattern recognition . IEEE, 2009, pp. 171–174

work page 2009
[5]

Bangladeshi bangla speech corpus for automatic speech recognition research,

S. Kibria, A. M. Samin, M. H. Kobir, M. S. Rahman, M. R. Selim, and M. Z. Iqbal, “Bangladeshi bangla speech corpus for automatic speech recognition research,” Speech Communication, vol. 136, pp. 84– 97, 2022

work page 2022
[6]

w2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,

Y .-A. Chung, Y . Zhang, W. Han, C.-C. Chiu, J. Qin, R. Pang, and Y . Wu, “w2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,” 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) , pp. 244–250, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:237048255

work page 2021
[7]

Evaluating openai’s whisper asr: Performance analysis across diverse accents and speaker traits,

C. Graham and N. Roll, “Evaluating openai’s whisper asr: Performance analysis across diverse accents and speaker traits,” JASA Express Letters, vol. 4, no. 2, 2024

work page 2024
[8]

Common voice: A massively-multilingual speech corpus,

R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), 2020, pp. 4211–4215

2020

Show all 21 references

[9]

Crowd-Sourced Speech Corpora for Javanese, Sundanese, Sinhala, Nepali, and Bangladeshi Bengali,

O. Kjartansson, S. Sarin, K. Pipatsrisawat, M. Jansche, and L. Ha, “Crowd-Sourced Speech Corpora for Javanese, Sundanese, Sinhala, Nepali, and Bangladeshi Bengali,” in Proc. The 6th Intl. Workshop on Spoken Language Technologies for Under-Resourced Languages (SLTU), Gurugram, ...

2018 doi
[10]

A Step-by-Step Process for Building TTS V oices Using Open Source Data and Framework for Bangla, Javanese, Khmer, Nepali, Sinhala, and Sundanese,

K. Sodimana, K. Pipatsrisawat, L. Ha, M. Jansche, O. Kjartansson, P. D. Silva, and S. Sarin, “A Step-by-Step Process for Building TTS V oices Using Open Source Data and Framework for Bangla, Javanese, Khmer, Nepali, Sinhala, and Sundanese,” in Proc. The 6th Intl. Workshop on S...

2018 doi
[11]

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

A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,”
[12]

The cnn: The architecture behind artificial intelligence develop- ment,

Y . Lee, “The cnn: The architecture behind artificial intelligence develop- ment,” Journal of Student Research, vol. 12, no. 4, Nov. 2023. [Online]. Available: https://www.jsr.org/hs/index.php/path/article/view/5579

2023
[13]

A wav2vec2-based experimental study on self-supervised learning methods to improve child speech recognition,

R. Jain, A. Barcovschi, M. Y . Yiwere, D. Bigioi, P. Corcoran, and H. Cucu, “A wav2vec2-based experimental study on self-supervised learning methods to improve child speech recognition,” IEEE Access , vol. 11, pp. 46 938–46 948, 2023

2023
[14]

Modular domain adaptation for conformer-based streaming asr,

Q. Li, B. Li, D. Hwang, T. Sainath, and P. Mengibar, “Modular domain adaptation for conformer-based streaming asr,” 08 2023, pp. 3357–3361

2023
[15]

Bert: Pre-training of deep bidirectional transformers for language understanding,

J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” 2019. [Online]. Available: https://arxiv.org/abs/1810.04805

2019 arXiv
[16]

Kazakh speech recognition: Wav2vec2. 0 vs. whisper,

Z. Kozhirbayev, “Kazakh speech recognition: Wav2vec2. 0 vs. whisper,” Journal of Advances in Information Technology, vol. 14, no. 6, pp. 1382– 1389, 2023

2023
[17]

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
[18]

Model adaptation for asr in low-resource indian languages,

A. Singh, A. S. Mehta, J. Nanavati, J. Bandekar, K. Basumatary, S. Badiger, S. Udupa, S. Kumar, P. K. Ghosh, P. Pai et al. , “Model adaptation for asr in low-resource indian languages,” arXiv preprint arXiv:2307.07948, 2023

2023 arXiv
[19]

Whisper-at: Noise- robust automatic speech recognizers are also strong general audio event taggers,

Y . Gong, S. Khurana, L. Karlinsky, and J. Glass, “Whisper-at: Noise- robust automatic speech recognizers are also strong general audio event taggers,” in INTERSPEECH 2023, ser. interspeech_2023. ISCA, Aug

2023
[2020]

Available: https://arxiv.org/abs/2006.11477

[Online]. Available: https://arxiv.org/abs/2006.11477

2006 arXiv
[2023]

Available: http://dx.doi.org/10.21437/Interspeech.2023- 2193

[Online]. Available: http://dx.doi.org/10.21437/Interspeech.2023- 2193

2023 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A comparison of sequence-to-sequence models for speech recognition

R. Prabhavalkar, K. Rao, T. N. Sainath, B. Li, L. Johnson, and N. Jaitly, “A comparison of sequence-to-sequence models for speech recognition.” in Interspeech, 2017, pp. 939–943

work page 2017

[2] [2]

Transformers: State- of-the-art natural language processing,

T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y . Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. Rush, “Transformers: State- of-the-art natural language processing,” in Proceedings of the 2020 Conference on Empirical Meth...

work page 2020

[3] [3]

A survey on contrastive self-supervised learning,

A. Jaiswal, A. R. Babu, M. Z. Zadeh, D. Banerjee, and F. Makedon, “A survey on contrastive self-supervised learning,” Technologies, vol. 9, no. 1, p. 2, 2020

2020

[4] [4]

Bangla speech recognition system using lpc and ann,

A. K. Paul, D. Das, and M. M. Kamal, “Bangla speech recognition system using lpc and ann,” in 2009 seventh international conference on advances in pattern recognition . IEEE, 2009, pp. 171–174

work page 2009

[5] [5]

Bangladeshi bangla speech corpus for automatic speech recognition research,

S. Kibria, A. M. Samin, M. H. Kobir, M. S. Rahman, M. R. Selim, and M. Z. Iqbal, “Bangladeshi bangla speech corpus for automatic speech recognition research,” Speech Communication, vol. 136, pp. 84– 97, 2022

work page 2022

[6] [6]

w2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,

Y .-A. Chung, Y . Zhang, W. Han, C.-C. Chiu, J. Qin, R. Pang, and Y . Wu, “w2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,” 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) , pp. 244–250, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:237048255

work page 2021

[7] [7]

Evaluating openai’s whisper asr: Performance analysis across diverse accents and speaker traits,

C. Graham and N. Roll, “Evaluating openai’s whisper asr: Performance analysis across diverse accents and speaker traits,” JASA Express Letters, vol. 4, no. 2, 2024

work page 2024

[8] [8]

Common voice: A massively-multilingual speech corpus,

R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), 2020, pp. 4211–4215

2020