Pith. sign in

REVIEW 3 major objections 6 minor 39 references

Whale: Large-Scale multilingual ASR model with w2v-BERT and E-Branchformer with large speech data

T0 review · 3 major / 6 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 Whale, a 1.87B-parameter multilingual ASR model, achieves a 2.4% WER on LibriSpeech test-clean and a 3.4% CER on CSJ eval3, outperforming Whisper large-v3 and OWSM v3.1.

desk verdict Whale is a serious engineering effort, but the headline numbers are uninterpretable until the authors state that the official test splits of CSJ/LibriSpeech/CommonVoice/FLEURS were excluded from the 250k-hour training corpus. read the letter →

arxiv 2506.01439 v1 pith:AGPWXYVB submitted 2025-06-02 cs.CL eess.AS

Yosuke Kashiwagi , Hayato Futami , Emiru Tsunoo , Satoshi Asakawa This is my paper
classification cs.CLeess.AS
keywords multilingualASRw2v-BERTE-BranchformerjointCTC-attentioncurriculumlearningself-conditionedCTCWhalelarge-scalespeechrecognition
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 introduces Whale, a 1.87-billion-parameter multilingual speech recognizer that combines a w2v-BERT self-supervised front end, a 24-layer E-Branchformer encoder, a 6-layer Transformer decoder, and a joint CTC-attention decoding strategy. It is trained on 250,000 hours across 144 languages, with a deliberately large Japanese component, using a seven-stage curriculum that ends by unfreezing the w2v-BERT parameters. The authors claim that Whale reaches a word error rate of 2.4% on LibriSpeech test-clean and a character error rate of 3.4% on CSJ eval3, surpassing Whisper large-v3 and OWSM v3.1. If correct, this shows that a transparent, mostly public-data model can match or beat proprietary systems on English and Japanese spontaneous speech, and that language-specific data collection is a decisive factor.

What carries the argument

The central mechanism is the stacked architecture of a w2v-BERT front end (24 Conformer layers, pre-trained with masked prediction and contrastive loss), a 24-layer E-Branchformer encoder with self-conditioned CTC branches at layers 8 and 16, and a 6-layer Transformer decoder trained with a weighted CTC plus attention loss (0.3/0.7) and decoded by joint CTC-attention beam search. This carries the argument by showing that no new block is needed: scaling existing SSL plus an E-Branchformer encoder-decoder on a carefully mixed corpus—especially the 30k hours of Japanese—is enough to beat Whisper and OWSM on the claimed benchmarks.

What would settle it

Compute the overlap between the exact audio files (by file path, speaker ID, or utterance fingerprint) in the training data and those in the CSJ eval1-3 and FLEURS test partitions; any nonzero overlap would invalidate the generalization reading of the headline numbers. Alternatively, evaluating the released model on a fresh, never-trained Japanese spontaneous-speech corpus and checking whether the CER stays near 3-5% would settle the claim.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a large-scale multilingual ASR model built entirely from known components—w2v-BERT feature extraction, E-Branchformer encoding, and joint CTC-attention decoding—can achieve 2.4% WER on LibriSpeech test-clean and 3.4% CER on CSJ eval3, both better than the corresponding Whisper large-v3 and OWSM v3.1 numbers (Whisper: 2.5%/16.5%; OWSM: 2.4%/21.9%). The improvement on CSJ is especially large (3.4% vs 16.5% for Whisper), which the authors attribute to collecting about 30k hours of Japanese in-house data, roughly four times Whisper's Japanese data. They further report that the final curriculum stage, which updates previously frozen w2v-BERT parameters, is critical: it cuts LibriSpeech other WER from 5.8% to 4.5% while preserving clean-set performance. The paper also reports that the self-conditioned CTC language-adaptation mechanism did not help on most FLEURS languages, which they interpret as the SSL front end already absorbing the linguistic variability.

Load-bearing premise

The reported benchmark scores assume that the evaluation splits of LibriSpeech, CSJ, CommonVoice, and FLEURS were excluded from the 250k-hour training corpus; the paper lists these corpora as training sources without ever stating that the test partitions were held out.

Editorial extensions

If this is right

  • If the reported numbers hold, a transparent recipe of public, purchased, and in-house data can yield a model that beats Whisper large-v3 on clean English and on Japanese spontaneous speech, suggesting that Whisper's web-scale data advantage is not absolute.
  • The 30k hours of Japanese in-house data, nearly all read speech, appears to be the decisive factor behind the CSJ result (3.4% CER vs Whisper's 16.5%), implying that targeted language-specific data collection can outperform generic web-scale sampling for a given language.
  • The stage-7 improvement on LibriSpeech other (5.8% to 4.5% WER) indicates that unfreezing and jointly fine-tuning the SSL front end is a cheap, high-impact step after large-scale curriculum training.
  • The model's weaker FLEURS performance relative to Whisper on many languages, and the failure of self-conditioned CTC adaptation, point to remaining domain-robustness and adaptation gaps that the paper itself flags.

Reading between the lines

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

  • Applying the same curriculum plus a few hundred hours of targeted read speech to another under-resourced language (e.g., Swahili or Telugu) would test whether the Japanese result is a general property of the data-mixing recipe or specific to the 30k-hour Japanese corpus.
  • Since the zero-shot language adaptation failed, an alternative is to insert the language ID into the SSL layers themselves (e.g., via adapter tokens) and re-run the FLEURS evaluation, making the paper's negative result a natural baseline.
  • The paper's comparison corpus is about 250k hours; re-running the recipe at, say, 100k hours would separate how much of the gain is pure scale versus the curriculum and data mixture, which the current ablations do not disentangle.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents Whale, a 1.87B-parameter multilingual ASR model that combines w2v-BERT SSL features, a 24-layer E-Branchformer encoder, a 6-layer Transformer decoder, and joint CTC-attention decoding. Training uses 250k hours across 144 languages, mixing public corpora, purchased datasets, and in-house Japanese recordings, with a 7-stage curriculum learning schedule and a final stage that updates previously frozen SSL parameters. Evaluation is reported on LibriSpeech, CSJ, CommonVoice, and FLEURS, with headline results of 2.4% WER on LibriSpeech test-clean and 3.4% CER on CSJ eval3, which the abstract claims outperform Whisper large-v3 and OWSM v3.1.

Significance. If the benchmark test splits are genuinely held out, the paper demonstrates a useful and credible recipe: combining a pretrained w2v-BERT representation with E-Branchformer, joint CTC/attention decoding, and a Japanese-heavy curriculum can beat large open and proprietary baselines on specific English and Japanese benchmarks. The architecture and training stages are described clearly, and the comparison tables cover a useful range of languages. However, no code or weights are released, the training/evaluation split discipline is never stated, and the abstract's 'outperforming' claim is not supported by the paper's own FLEURS and CommonVoice tables. The central empirical contribution therefore currently rests on an unverified assumption about held-out test sets.

major comments (3)
  1. [§3.2, §4] Section 3.2 lists CommonVoice, LibriSpeech, FLEURS, and CSJ among the training sources, and Section 4 reports evaluation on those same corpora, yet the manuscript never states that the official test splits were excluded from the training mixture. The CSJ result is particularly affected: CSJ is a purchased training set, and the reported 3.4% CER on eval3 versus Whisper's 16.5% cannot be interpreted as generalization unless eval1/eval2/eval3 were held out. Please state per corpus which exact split was used, confirm that those utterances were not in the training data, and ideally provide a data recipe or split-removal script; without that, the headline numbers are unverifiable.
  2. [Abstract vs. §4.4, Table 5] The abstract's claim that Whale 'achieves... outperforming Whisper large-v3 and OWSM v3.1' is contradicted by the paper's own Table 5, where Whisper large-v3 has lower WER/CER on most FLEURS languages (e.g., en 4.0 vs. 6.2, de 5.8 vs. 9.5, fr 5.3 vs. 9.7, th 30.0 vs. 37.9), and by Table 4, where Whisper wins on English CommonVoice (8.4 vs. 8.9). The claim should be scoped to the specific benchmarks and splits where it is actually true, or revised to 'comparable performance with wins on particular sets.'
  3. [§4, Tables 2–5] The manuscript does not state whether the baseline numbers for Whisper large-v3, OWSM v3.1, and OWSM CTC v3.1 were re-run under the same evaluation pipeline (including whisper-normalizer and decoding configuration) or quoted from prior publications. Because Table 2's OWSM numbers differ slightly from the cited OWSM v3.1 report, please specify the source and evaluation settings for every baseline cell so the comparisons are airtight.
minor comments (6)
  1. [Introduction] The sentence 'The total amount of our training data is 250k' should specify '250k hours' to match Section 3.2.
  2. [Index Terms; §1] There are typos: 'E-Branhformer' in the Index Terms and 'FLERUS' in the Introduction should be 'E-Branchformer' and 'FLEURS'.
  3. [§2.2.2] The language-adaptation description refers to a 'target language ID' used at inference, but no language ID embedding or input token is described in the architecture; please clarify how the LID is supplied to the model.
  4. [Tables 3–5] Only stage7 results are reported for CSJ, CommonVoice, and FLEURS; if stage6 results for these benchmarks are available, reporting them would help isolate the effect of SSL updating.
  5. [§4] Applying whisper-normalizer to both references and hypotheses is reasonable for LibriSpeech and CommonVoice, but its effect on Japanese CSJ text should be explained, since fillers, punctuation, and normalizer behavior directly affect the reported CER.
  6. [Table 1] The note 'The blanks are unknown' is ambiguous: it should state explicitly whether the blank OWSM cells mean no training data or unknown training data hours.

Circularity Check

0 steps flagged · score 2.0 of 10

No demonstrated circularity: the model is an empirical composition of externally cited components whose headline numbers are external-benchmark comparisons; the single self-citation (ref. [29]) is non-load-bearing, and the training/evaluation corpus overlap is a verification risk rather than a quoted by-construction reduction.

full rationale

The paper's claimed derivation chain is empirical, not analytic: a w2v-BERT front-end plus E-Branchformer encoder plus Transformer decoder with joint CTC-attention decoding, all attributed to external prior work ([18], [19], [20], [21]), trained with curriculum learning on 250k hours, and compared against external baselines (Whisper, OWSM) on standard benchmarks. I checked each circularity pattern. (1) No quantity is defined in terms of the quantity it is claimed to predict; there are no equations whose left-hand side collapses into its own input. (2) The only self-citation is ref. [29] (Kashiwagi/Futami/Tsunoo et al., arXiv:2406.12611), invoked in Sec. 2.2.2 as the basis of the zero-shot language-adaptation mechanism; it is not load-bearing because Sec. 4.4 tests it and reports that it 'did not have any effect on most languages,' so the headline WER/CER results do not rest on it. (3) The nearest thing to circularity is the overlap between training sources (Sec. 3.2 lists CommonVoice, LibriSpeech, FLEURS, and purchased CSJ among the training data) and the evaluation corpora (Sec. 4: LibriSpeech, CSJ, CommonVoice, FLEURS), with no sentence stating that the official test splits were excluded. This is a genuine reporting gap and a data-contamination risk: if the CSJ eval1/eval2/eval3 sets were included in the purchased CSJ used for training, the headline 3.4% CER would measure memorization rather than generalization. However, the hard rule requires exhibiting the specific reduction by quotation, and the paper nowhere states that the eval splits were included in training, so the reduction cannot be demonstrated from the text. Evidence against wholesale leakage is internal: on FLEURS, which is also listed as a training source, Whisper still beats Whale on most languages (Table 5), which is inconsistent with test-set memorization being the operative mechanism. The verdict is therefore 'no demonstrated circularity,' scored 2 rather than 0 only because the missing split-exclusion statements leave the headline CSJ comparison under-supported, and no code or weights are released to verify split discipline; this is a correctness and verifiability concern, not a definitional collapse.

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

This is an empirical systems paper, so the only free parameters are hand-set hyperparameters; there are no invented entities. The most load-bearing unstated axiom is that the benchmark test sets were not part of the training data.

free parameters (3)
  • CTC loss weight = 0.3
    The CTC branch is weighted 0.3 and the decoder 0.7 in the joint loss and in beam search decoding; this ratio is chosen by hand and no sweep is reported (Section 2.2.3).
  • Resource rank thresholds = >100h, 20-100h, <20h
    Languages are divided into High/Middle/Low resource ranks using manually chosen thresholds to design the curriculum (Section 3.2).
  • Curriculum stage schedule = Stages 1-7 with durations 1d, 1d, 1d, 1d, 3d, 2w, 3w
    The seven-stage curriculum and the duration spent in each stage are set by hand; no ablation or comparison to alternative schedules is given (Section 3.3).
assumptions (3)
  • domain assumption Evaluation splits are held out from training
    The paper lists CSJ, FLEURS, Librispeech, and CommonVoice as training sources and evaluates on the same benchmarks without stating that test subsets were excluded (Sections 3.2 and 4).
  • domain assumption whisper-normalizer is an appropriate text normalization for all evaluated languages
    All scores are computed after applying the whisper-normalizer to both references and hypotheses, without discussing its suitability for Japanese, Chinese, or other scripts (Section 4).
  • domain assumption w2v-BERT pretrained representations transfer to all 144 languages
    The model relies on a pretrained w2v-BERT SSL encoder, and the paper does not analyze whether coverage of low-resource languages is adequate (Section 2.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Whale: Large-Scale multilingual ASR model with w2v-BERT and E-Branchformer with large speech data." pith.science (2026). https://pith.science/paper/AGPWXYVB

@misc{pith2026250601439,
  author       = {Pith},
  title        = {Pith review of: Whale: Large-Scale multilingual ASR model with w2v-BERT and E-Branchformer with large speech data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AGPWXYVB}},
  note         = {Machine review of arXiv:2506.01439}
}
read the original abstract

This paper reports on the development of a large-scale speech recognition model, Whale. Similar to models such as Whisper and OWSM, Whale leverages both a large model size and a diverse, extensive dataset. Whale's architecture integrates w2v-BERT self-supervised model, an encoder-decoder backbone built on E-Branchformer, and a joint CTC-attention decoding strategy. The training corpus comprises varied speech data, of not only public corpora but also in-house data, thereby enhancing the model's robustness to different speaking styles and acoustic conditions. Through evaluations on multiple benchmarks, Whale achieved comparable performance to existing models. In particular, it achieves a word error rate of 2.4% on the Librispeech test-clean set and a character error rate of 3.4% on the CSJ eval3 set, outperforming Whisper large-v3 and OWSM v3.1.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages

  1. [1]

    Introduction Recent breakthroughs in multilingual ASR owe a great deal to the open-source community and researchers who have worked tirelessly to collect, curate, and share large-scale multilingual speech datasets. Projects such as CommonV oice, MuST-C, mT- EDx, MLS, YODAS, FLERUS and others [1–8] have enabled model developers to train and evaluate system...

    work page Pith review arXiv 2025

  2. [2]

    Next, acoustic features are extracted via SSL, w2v-BERT [18]

    Model architecture Our model assumes a single-channel input speech, and if the sampling rate differs, it is resampled to 16kHz. Next, acoustic features are extracted via SSL, w2v-BERT [18]. At this point, the frame rate is 100 frames/sec. The features are then input into an encoder based on E-Branchformer [19]. In the encoder, convolution layers first red...

    work page

  3. [3]

    Training environments The training of the Whale model was conducted on an inter- nal server infrastructure

    Training 3.1. Training environments The training of the Whale model was conducted on an inter- nal server infrastructure. Each node in our cluster is equipped with eight NVIDIA H100 GPUs, and we were able to lever- age a total of 16 nodes simultaneously. The entire training process spanned approximately six weeks. Our implementa- tion and training pipelin...

    work page

  4. [4]

    Our experiments compare Whale against state-of-the-art systems such as Whis- per [14], OWSM [15], and OWSM CTC [17]

    Evaluation The primary objective of our evaluation is to rigorously as- sess the performance and robustness of the Whale model across diverse datasets and acoustic conditions. Our experiments compare Whale against state-of-the-art systems such as Whis- per [14], OWSM [15], and OWSM CTC [17]. Before calculat- ing the scores, we applied the whisper-normaliz...

    work page

  5. [5]

    Through extensive experiments, we demonstrated that Whale achieves highly competitive performance on four benchmarks

    Conclusion In this paper, we introduced Whale, a large-scale speech recog- nition model that leverages a robust encoder–decoder archi- tecture, self-conditioned CTC, and a carefully designed train- ing strategy including curriculum learning and SSL updating. Through extensive experiments, we demonstrated that Whale achieves highly competitive performance ...

    work page

  6. [6]

    Common V oice: A Massively-Multilingual Speech Corpus,

    R. Ardila, M. Branson, K. Daviset al., “Common V oice: A Massively-Multilingual Speech Corpus,” inProceedings of the Twelfth Language Resources and Evaluation Conference, 2020, pp. 4218–4222

    work page 2020

  7. [7]

    MUST-C: a multilingual speech translation corpus,

    M. A. Di Gangi, R. Cattoni, L. Bentivogli, M. Negri, and M. Turchi, “MUST-C: a multilingual speech translation corpus,” inProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technologies, V olume 1 (Long and Short Papers). Association for Computational Linguistics, 2019, pp. 2012–2017

    work page 2019

  8. [8]

    The Multilingual TEDx Corpus for Speech Recognition and Translation,

    S. Elizabeth, W. Matthew, B. Jacobet al., “The Multilingual TEDx Corpus for Speech Recognition and Translation,” inPro- ceedings of Interspeech 2021, 2021, pp. 3655–3659

    work page 2021

Show all 39 references
  1. [9]

    MLS: A Large-Scale Multilingual Dataset for Speech Research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “MLS: A Large-Scale Multilingual Dataset for Speech Research,” 2020

    2020

  2. [10]

    YODAS: YouTube-oriented dataset for audio and speech,

    X. Li, S. Takamichi, T. Saekiet al., “YODAS: YouTube-oriented dataset for audio and speech,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  3. [11]

    FLEURS: Few-shot learn- ing evaluation of universal representations of speech,

    A. Conneau, M. Ma, S. Khanujaet al., “FLEURS: Few-shot learn- ing evaluation of universal representations of speech,” in2022 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2023, pp. 798–805

    2023

  4. [12]

    JTubeSpeech: corpus of Japanese speech collected from YouTube for speech recognition and speaker verification,

    S. Takamichi, L. K ¨urzinger, T. Saeki, S. Shiota, and S. Watan- abe, “JTubeSpeech: corpus of Japanese speech collected from YouTube for speech recognition and speaker verification,”arXiv preprint arXiv:2112.09323, 2021

    2021 arXiv

  5. [13]

    CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR V oice Cloning Toolkit,

    C. Veaux, J. Yamagishi, and K. MacDonald, “CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR V oice Cloning Toolkit,” 2017. [Online]. Available: https://doi.org/10.7488/ds/ 1994

    2017 doi

  6. [14]

    Multilingual speech recognition with a single end-to-end model,

    S. Toshniwal, T. N. Sainath, R. J. Weisset al., “Multilingual speech recognition with a single end-to-end model,” in2018 IEEE international conference on acoustics, speech and signal process- ing (ICASSP). IEEE, 2018, pp. 4904–4908

    2018

  7. [15]

    Bytes are all you need: End-to-end multilingual speech recognition and synthe- sis with bytes,

    B. Li, Y . Zhang, T. Sainath, Y . Wu, and W. Chan, “Bytes are all you need: End-to-end multilingual speech recognition and synthe- sis with bytes,” inICASSP 2019-2019 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 5621–5625

    2019

  8. [16]

    An end-to-end language-tracking speech recognizer for mixed- language speech,

    H. Seki, S. Watanabe, T. Hori, J. Le Roux, and J. R. Hershey, “An end-to-end language-tracking speech recognizer for mixed- language speech,” in2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 4919–4923

    2018

  9. [17]

    Scaling speech technology to 1,000+ languages,

    V . Pratap, A. Tjandra, B. Shiet al., “Scaling speech technology to 1,000+ languages,”Journal of Machine Learning Research, vol. 25, no. 97, pp. 1–52, 2024

    2024

  10. [18]

    Less is More: Accu- rate Speech Recognition & Translation without Web-Scale Data,

    K. C. Puvvada, P. ˙Zelasko, H. Huanget al., “Less is More: Accu- rate Speech Recognition & Translation without Web-Scale Data,” inProc. Interspeech 2024, 2024, pp. 3964–3968

    2024

  11. [19]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xuet al., “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

    2023

  12. [20]

    OWSM v3.1: Better and faster open whisper-style speech models based on e-branchformer,

    Y . Peng, J. Tian, W. Chenet al., “OWSM v3.1: Better and faster open whisper-style speech models based on e-branchformer,” arXiv preprint arXiv:2401.16658, 2024

    2024 arXiv

  13. [21]

    Reproducing whisper-style train- ing using an open-source toolkit and publicly available data,

    Y . Peng, J. Tian, B. Yanet al., “Reproducing whisper-style train- ing using an open-source toolkit and publicly available data,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  14. [22]

    OWSM-CTC: An Open Encoder-Only Speech Foundation Model for Speech Recognition, Translation, and Language Identification,

    Y . Peng, Y . Sudo, M. Shakeel, and S. Watanabe, “OWSM-CTC: An Open Encoder-Only Speech Foundation Model for Speech Recognition, Translation, and Language Identification,”arXiv preprint arXiv:2402.12654, 2024

    2024 arXiv

  15. [23]

    W2v-BERT: Combin- ing contrastive learning and masked language modeling for self- supervised speech pre-training,

    Y .-A. Chung, Y . Zhang, W. Hanet al., “W2v-BERT: Combin- ing contrastive learning and masked language modeling for self- supervised speech pre-training,” in2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 244–250

    2021

  16. [24]

    E-Branchformer: Branchformer with enhanced merging for speech recognition,

    K. Kim, F. Wu, Y . Penget al., “E-Branchformer: Branchformer with enhanced merging for speech recognition,” in2022 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2023, pp. 84–91

    2023

  17. [25]

    Joint CTC-attention based end-to-end speech recognition using multi-task learning,

    S. Kim, T. Hori, and S. Watanabe, “Joint CTC-attention based end-to-end speech recognition using multi-task learning,” in2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2017, pp. 4835–4839

    2017

  18. [26]

    Joint CTC/attention decoding for end-to-end speech recognition,

    T. Hori, S. Watanabe, and J. R. Hershey, “Joint CTC/attention decoding for end-to-end speech recognition,” inProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2017, pp. 518–529

    2017

  19. [27]

    Curricu- lum learning,

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Curricu- lum learning,” inProceedings of the 26th annual international conference on machine learning, 2009, pp. 41–48

    2009

  20. [28]

    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

  21. [29]

    Corpus of Spontaneous Japanese: Its design and evaluation,

    K. Maekawa, “Corpus of Spontaneous Japanese: Its design and evaluation,” inISCA & IEEE Workshop on Spontaneous Speech Processing and Recognition, 2003

    2003

  22. [30]

    Better Intermediates Im- prove CTC Inference,

    T. Komatsu, Y . Fujita, J. Leeet al., “Better Intermediates Im- prove CTC Inference,” inProceedings of the Annual Conference of the International Speech Communication Association, INTER- SPEECH, vol. 2022, 2022, pp. 4965–4969

    2022

  23. [31]

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

    J. Devlin, “BERT: Pre-training of deep bidirectional transformers for language understanding,”arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  24. [32]

    mHuBERT-147: A Compact Multilingual HuBERT Model,

    M. Z. Boito, V . Iyer, N. Lagos, L. Besacier, and I. Calapode- scu, “mHuBERT-147: A Compact Multilingual HuBERT Model,” arXiv preprint arXiv:2406.06371, 2024

    2024 arXiv

  25. [33]

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

    S. Chen, C. Wang, Z. Chenet al., “WavLM: Large-scale self- supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  26. [34]

    Rapid Language Adaptation for Multilingual E2E Speech Recognition Using Encoder Prompting,

    Y . Kashiwagi, H. Futami, E. Tsunoo, S. Arora, and S. Watan- abe, “Rapid Language Adaptation for Multilingual E2E Speech Recognition Using Encoder Prompting,”arXiv preprint arXiv:2406.12611, 2024

    2024 arXiv

  27. [35]

    ESPnet: End-to-End Speech Processing Toolkit,

    S. Watanabe, T. Hori, S. Karitaet al., “ESPnet: End-to-End Speech Processing Toolkit,” inProc. Interspeech, 2018, pp. 2207– 2211

    2018

  28. [36]

    VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi- supervised learning and interpretation,

    C. Wang, M. Riviere, A. Leeet al., “VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi- supervised learning and interpretation,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Internatio...

    2021

  29. [37]

    WenetSpeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,

    B. Zhang, H. Lv, P. Guoet al., “WenetSpeech: A 10000+ hours multi-domain mandarin corpus for speech recognition,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6182– 6186

    2022

  30. [38]

    The people’s speech: A large-scale diverse english speech recognition dataset for com- mercial usage,

    D. Galvez, G. Diamos, J. Ciroet al., “The people’s speech: A large-scale diverse english speech recognition dataset for com- mercial usage,”arXiv preprint arXiv:2111.09344, 2021

    2021 arXiv

  31. [39]

    The Norwegian parliamentary speech corpus,

    P. E. Solberg and P. Ortiz, “The Norwegian parliamentary speech corpus,”arXiv preprint arXiv:2201.10881, 2022

    2022 arXiv

Pith tools

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