Pith. sign in

REVIEW 4 major objections 4 minor 3 cited by

OWLS: Scaling Laws for Multilingual Speech Recognition and Translation Models

T0 review · 4 major / 4 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 Multilingual speech word error rates follow a power-law curve as model size and training compute grow, so teams can forecast final performance before a run finishes.

desk verdict OWLS is a valuable open model suite, but its headline scaling-law claim is contradicted by the paper's own appendix numbers. read the letter →

arxiv 2502.10373 v1 pith:AH47TWSE submitted 2025-02-14 cs.CL cs.AIcs.LGeess.AS

William Chen , Jinchuan Tian , Yifan Peng , Brian Yan , Chao-Han Huck Yang , Shinji Watanabe This is my paper
classification cs.CLcs.AIcs.LGeess.AS
keywords scalinglawsmultilingualASRspeechtranslationworderrorratepowerlawlow-resourcelanguagesemergentabilitiesin-contextlearning
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

This paper claims that the error rates of massively multilingual speech recognition and translation models shrink along a power-law curve as model size and training compute grow, so a practitioner can forecast final word error rate (WER) or BLEU before choosing a model size or completing a training run. The claim is built on OWLS, an open suite of 13 Whisper-style models from 0.25B to 18B parameters trained on up to 360K hours of public speech data across 150 languages. The authors report that language-specific WER-versus-model-size curves fit $\mathrm{WER}(x)=\beta x^{\alpha}$ with $R^2\approx 0.95$, and that WER during training follows a compute curve with $R^2\approx 0.82$. If these fits hold beyond the observed range, they would turn model-size selection and training-budget planning for speech models into a quantitative design step rather than trial and error. The same suite also shows that scaling improves low-resource languages and reveals abilities such as orthographic understanding, code-switching, and in-context learning that only appear in the largest models.

What carries the argument

The carrying object is the two-parameter power law $\mathrm{WER}(x)=\beta x^{\alpha}$ (the same form used for BLEU), fitted in log space for each language and task, with $x$ standing for model parameters or training FLOPs. It converts scaling experiments into a predictive formula; the paper's contribution is applying it to downstream word error rate, character error rate, and BLEU of a multilingual encoder-decoder speech model rather than to pre-training loss. The supporting machinery is the OWLS suite itself: seven model sizes with identical hyperparameters, Whisper-style task prompting, a hybrid CTC/attention Transformer, and up to 360K hours of public audio spanning 150 languages, which supplies the controlled grid of measurements the power law is fit to.

What would settle it

Train or evaluate an OWLS model at a parameter count not used in the fits—for instance 13B or 24B—or hold out one of the seven sizes, fit the law to the other six, and extrapolate; then compare the predicted WER to the measured one on FLEURS. If the errors deviate systematically, or if refitting the existing data shifts $\beta$ and $\alpha$ materially, the reliable-prediction claim fails. A cheaper check is to publish $\beta$ and $\alpha$ with confidence intervals and test whether the non-monotonic WER values visible in the appendix (e.g., Afrikaans) fall inside the fit's uncertainty.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the familiar power-law form used for language-model loss, $L(x)=L_\infty+\beta_x x^{\alpha_x}$, can be replaced in speech by a direct power law on downstream metrics, $\mathrm{WER}(x)=\beta_x x^{\alpha_x}$ (and similarly $\mathrm{BLEU}(x)=\beta_x x^{\alpha_x}$), where $x$ is model size or training compute. Fitted per language across seven model sizes (0.25B to 18B), these curves reach $R^2\approx 0.95$ on the 102-language FLEURS test set, and plotting WER against FLOPs throughout training gives $R^2\approx 0.82$, meaning later checkpoints can be projected from early ones. The paper argues this makes final ASR/ST performance reliably predictable when scaling, while also showing that data scaling alone saturates quickly unless new data brings domain or language diversity, and that extreme scaling cannot overcome tasks with an inherent shortage of training data (e.g., Japanese-to-English ST with roughly one hour of data).

Load-bearing premise

The predictive claim rests on assuming that a power law fitted to the observed model sizes, data sizes, and compute budgets continues to hold outside that observed range.

Editorial extensions

If this is right

  • A model developer can read off the expected WER for a given parameter count before spending compute, using per-language power-law fits.
  • Training runs can be shortened safely: with the compute curve, early checkpoints predict the final WER, so a run that is off-target can be stopped or adjusted.
  • Scaling model size—not just data—is the lever that lifts low-resource languages, with average WER on the 50 lowest-resource languages dropping from 59 to 45 between the 1B and 9B models.
  • Adding more data from the same distribution saturates performance; only new domains or languages bring large gains, as seen when the extra 180K YODAS hours improved Korean, Polish, and Dutch.
  • Some failures are not fixable by scale alone: tasks with only about an hour of training data, such as Japanese-to-English ST, barely improve even at 9B parameters.

Reading between the lines

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

  • A natural next check of the predictive claim is to hold out one of the seven model sizes, refit the power law on the rest, and compare the extrapolated WER to the measured value.
  • If the per-language exponents $\alpha$ vary systematically with data amount or orthographic transparency, the same fits could be turned into a data-allocation rule for new languages—a step the paper does not take.
  • The compute-scaling curve was evaluated on only three languages; whether $R^2\approx 0.82$ generalizes to low-resource languages is untested and could be checked by plotting WER-by-checkpoint for a few low-resource languages.
  • The appendix tables show some non-monotonic WER values across sizes for languages such as Afrikaans; publishing $\beta$ and $\alpha$ with confidence intervals would show whether the per-language fits are stable or noisy.
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 / 4 minor

Summary. This paper introduces OWLS, a suite of 13 Whisper-style multilingual ASR/ST models ranging from 0.25B to 18B parameters and trained on up to 360K hours of public data across 150 languages. The authors propose power-law scaling laws of the form WER(x)=β_x x^{α_x} for model size and training compute, report R²≈0.95 for language-specific model-size fits and R²≈0.82 for compute fits, and claim that final performance can be reliably predicted when scaling. The paper also studies data scaling, compares OWLS against Whisper and Canary on several benchmarks, and reports test-time behaviors including beam search, code-switching, mondegreens, contextual biasing, and in-context learning on Quechua. The appendix contains per-language WER/CER tables for all seven model sizes.

Significance. If the scaling-law claims were properly validated, this would be an important contribution: the OWLS suite would be the largest publicly released set of speech models trained on public data, and the promised checkpoints, training code, and intermediate logs would enable reproducible research on multilingual ASR/ST scaling. The resource itself, together with the exploratory results on emergent abilities and in-context learning, is valuable and likely to be widely used. However, the central predictive claim is not currently supported by the evidence in the manuscript: the fits are in-sample, no coefficients or uncertainties are reported, and the appendix tables contain many WER curves that are non-monotonic in model size and therefore incompatible with the proposed power law.

major comments (4)
  1. [§4.1, Figure 12, Tables 13–14] The claim that language-specific model-size scaling laws are highly predictive of WER/CER with R²≈0.95 is contradicted by the raw appendix tables. For example, Table 14 reports for Afrikaans the WER sequence 86.6, 102.3, 95.7, 133.7, 92.3, 152.2, 92.3 across 0.25B to 18B, and for Oromo 94.8, 89.0, 102.4, 80.5, 79.1, 77.9, 71.1; Table 13 reports for Portuguese 50.1, 29.6, 20.5, 20.5, 15.7, 14.1, 21.1 and for Vietnamese 181.2, 86.6, 83.6, 54.6, 55.5, 42.5, 47.9. A power law WER(N)=βN^α with negative α is strictly decreasing and cannot reproduce these non-monotonic sequences, so for these languages the in-sample R² cannot be near 0.95. Please report per-language fitted β, α, and R², overlay the fitted curves on the raw points for all languages, and either restrict the scaling-law claim to the subset of languages for which it actually holds or explain precisely how the aggregate R²≈0.95 is computed.
  2. [§2.1, Eq. (2), §4.1, §4.3] The predictive claim is demonstrated only in-sample. The same seven model sizes used to fit β and α are then used to show that the law 'predicts' performance, and in the compute experiments the intermediate checkpoints of the same runs are used without holding out the final checkpoint. No model size or checkpoint is held out, no extrapolation error is reported, and no fitting procedure, coefficient values, or confidence intervals are given. Please add out-of-sample validation, for example by fitting on sizes up to 4B and predicting 9B and 18B, or by fitting on early checkpoints and predicting the final checkpoint, and report the fitted parameters with uncertainty. In addition, Eq. (2) drops the irreducible-loss term L∞ that appears in Eq. (1); because a pure power law predicts WER→0 at large x, the omission should be justified or the term should be retained.
  3. [§4.3, Figure 9] The claim that for a fixed parameter size the WER of the final checkpoint can be reliably predicted from the WERs of initial checkpoints (R²≈0.82) is ambiguous about what is being fitted. If the curve for each model size includes the final checkpoint in the fit, this is again an in-sample statement rather than a prediction. The text should specify whether the fit is per model size or pooled across sizes, whether language is averaged before fitting, and whether the final checkpoint is excluded from the fit. The current evaluation also averages only English, Spanish, and Turkish, so it does not by itself support language-specific scaling-law claims.
  4. [§5.2, Table 4] The statement that 'larger models obtain consistently better perplexity scores across all model sizes' is not supported by Table 4: perplexity worsens from 372 at 9B to 429 at 18B, and the MOS decreases from 4.8 at 9B to 4.4 at 18B. Please correct this claim or discuss the non-monotonicity at the largest model size.
minor comments (4)
  1. [Headers and figures] The running header and some section headings use 'OWLs' instead of 'OWLS'; please standardize the name throughout.
  2. [Figures 4 and 11] The axis and legend labels in Figures 4 and 11 appear to contain rendering corruption, with repeated '/uni00000015/...' escape sequences; the intended text should be restored.
  3. [Tables 13–14 and Table 6] Several language names contain typos or inconsistencies: 'Portuguesse' for Portuguese, 'Tami' for Tamil, 'Igno' for Igbo, 'Ukranian' for Ukrainian, 'Pastho' for Pashto, and 'Malaysian' where 'Myanmar' appears to be intended; please correct these and check the dataset table for similar issues.
  4. [§4.1] The text reporting that average WER on the 50 lowest-resource languages decreases from 59 to 45 when scaling from 1B to 9B does not state which languages are averaged or how languages with non-monotonic WER curves are handled; please provide the averaging procedure and the per-language values.

Circularity Check

2 steps flagged · score 6.0 of 10

Model-size and compute scaling 'predictions' are in-sample power-law fits: β and α are estimated from the same seven model sizes (or same checkpoints) whose WERs are then reported as R²≈0.95/0.82, so the headline predictive claim is a goodness-of-fit statistic rather than an out-of-sample forecast.

  1. fitted input called prediction [Section 4.1, Eq. (2), Figure 3]
    ""On the other hand, language-specific model size scaling laws are highly predictive of WER/CER (R2 ≃ 0.95, Figure 3)." (Sec. 4.1); "We empirically show that this power law can also generalize to the multi-modal task of ASR (Figures 3 and 9), allowing true downstream performance to be easily predicted when x = N, B." (Sec. 2.1)."

    Equation (2) defines WER(x)=β_x x^{α_x}; Section 4.1 reports R²≈0.95 for 'language-specific model size scaling laws.' The R² is computed from the same seven model sizes (0.25B–18B) that determine β and α, with no held-out size. The fitted curve is evaluated at those same sizes and the in-sample correlation is called 'highly predictive' and 'easily predicted' downstream performance. A goodness-of-fit statistic is not an out-of-sample forecast, so the predictive claim is the fitted parameter renamed as prediction. Appendix data reinforce this: e.g., Afrikaans WERs 86.6, 102.3, 95.7, 133.7, 92.3, 152.2, 92.3 are non-monotonic, so even the in-sample power-law fit cannot be accurate for many languages.

  2. fitted input called prediction [Section 4.3, Figure 9]
    ""We find that for a fixed parameter size, the WER of the final checkpoint can be reliably predicted as a function of the training compute ( R2 ≃ 0.82). This means that one can reasonably predict the final WER of the model given the WERs of initial checkpoints.""

    The compute-scaling curve (Figure 9) is built from checkpoints of the same training runs whose final WER is being 'predicted.' The R²≈0.82 is reported on this fitting data without any held-out run or checkpoint, so the statement that 'the WER of the final checkpoint can be reliably predicted ... given the WERs of initial checkpoints' is supported only by in-sample evaluation along a curve whose parameters were estimated from those very checkpoints. This is the same fitted-input-called-prediction pattern as Section 4.1.

full rationale

The central scaling-law claims in Sections 4.1 and 4.3 are presented as predictions but are evaluated on the same data used to fit Eq. (2): the seven model sizes for the N-law and the checkpoints of the same runs for the compute-law. Reporting R² on the fitting points measures goodness-of-fit, not predictive validity, so the abstract's claim that 'final performance can be reliably predicted when scaling' is partially circular: the fitted curve is renamed a prediction. This is not a complete circularity because the OWLS suite, the training runs, the low-resource scaling observations, and the test-time/emergent-ability experiments are independent empirical contributions, and the self-citations to OWSM v3.2 and YODAS are to released datasets/prior work rather than load-bearing derivations. The non-monotonic WER sequences in Tables 13–14 are a correctness/internal-consistency concern that further undermines the claimed R²≈0.95, but the circularity finding rests on the in-sample fit being presented as prediction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central analysis adds two fitted coefficients per scaling curve (β and α) and relies on a power-law ansatz plus data-handling assumptions (uniform downsampling, FLEURS train/test separation, fixed hyperparameters across model sizes). No new physical or architectural entities are introduced.

free parameters (1)
  • Power-law coefficients β and α (per language/task) = not reported
    Fitted to WER(x)=β x^α across seven model sizes and across training checkpoints. The fitted curve is presented as the prediction; coefficients are not listed, so the 'law' cannot be reused.
assumptions (4)
  • domain assumption WER(x) = β x^α captures downstream ASR/ST scaling
    Adopted from language-model and rescoring work (Kaplan et al.; Gu et al.); validated only by R², no out-of-sample check, no coefficient reporting.
  • domain assumption Uniform downsampling of the 180K-hour corpus preserves the relevant per-language distribution
    Data-scaling runs (Section 4.2) assume the 50%, 25%, etc., splits remain representative; no analysis of distribution shift.
  • domain assumption The FLEURS test set is not contaminated by the FLEURS training split
    Table 6 lists FLEURS as part of training data; the paper does not explicitly state that the evaluation split is disjoint.
  • domain assumption Identical hyperparameters across model sizes isolate scale as the only variable
    Section 3.2 fixes hyperparameters for controllability; if larger models need different learning rates or regularization, fitted exponents are confounded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OWLS: Scaling Laws for Multilingual Speech Recognition and Translation Models." pith.science (2026). https://pith.science/paper/AH47TWSE

@misc{pith2026250210373,
  author       = {Pith},
  title        = {Pith review of: OWLS: Scaling Laws for Multilingual Speech Recognition and Translation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AH47TWSE}},
  note         = {Machine review of arXiv:2502.10373}
}
read the original abstract

Neural scaling laws offer valuable insights for designing robust sequence processing architectures. While these laws have been extensively characterized in other modalities, their behavior in speech remains comparatively underexplored. In this work, we introduce OWLS, an open-access, reproducible suite of multilingual speech recognition and translation models spanning 0.25B to 18B parameters, with the 18B version being the largest speech model, to the best of our knowledge. OWLS leverages up to 360K hours of public speech data across 150 languages, enabling a systematic investigation into how data, model, and compute scaling each influence performance in multilingual speech tasks. We use OWLS to derive neural scaling laws, showing how final performance can be reliably predicted when scaling. One of our key findings is that scaling enhances performance on low-resource languages/dialects, helping to mitigate bias and improve the accessibility of speech technologies. Finally, we show how OWLS can be used to power new research directions by discovering emergent abilities in large-scale speech models. Model checkpoints will be released on https://huggingface.co/collections/espnet/owls-scaling-laws-for-speech-recognition-and-translation-67ab7f991c194065f057ce8d for future studies.

Figures

Figures reproduced from arXiv: 2502.10373 by the authors.

Figure 1
Figure 1. Comparison of previous open models and our OWLS models (blue) by parameter count and training dataset size. Whisper (Radford et al., 2023) and Canary (Puvvada et al., 2024) are trained on undisclosed data, while OWSM (Peng et al., 2023b) and the presented OWLS use public data. (ASR) and Speech Translation (ST). The state-of-the-art (SOTA) in ASR/ST has now progressed to not only scaling in terms of model and data si… view at source ↗
Figure 2
Figure 2. The effect of scaling model size on the 102 FLEURS languages, plotted as WER (or CER) versus available training data. Although WER/CER generally decreases with more training data, the relationship is only moderately correlated, as indicated by the R2 values in the legend. Model performance is also influenced by domain alignment and orthographic transparency: for instance, more transparent languages (e.g., Spanish, I… view at source ↗
Figure 3
Figure 3. The effect of model scaling on WER/CER on FLEURS. Languages are color-coded by the amount of training data. For readability, we only show the top-20 languages (by data amount) in our training corpus. We find that model scaling is consistently predictive of downstream WER/CER across languages. Scaling curves for other languages can be found in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: WERs on multi-domain English ASR by model size. 6 standard ASR benchmarks: AMI (Carletta, 2007), Lib￾riSpeech (Panayotov et al., 2015), SPGISpeech (O’Neill et al., 2021), Tedlium (Hernandez et al., 2018), VoxPopuli (Wang et al., 2021b), and GigaSpeech (Chen et al., 202…
Figure 5
Figure 5. Figure 5: The evolution of FLEURS WER/CER for the top 20 languages by data size, as more training data is added for each language and given a fixed model capacity. Left: impact on WER/CER when scaling from 11K to 180K total hours, when all data is from the same distribution. Rig…
Figure 7
Figure 7. Figure 7: BLEU scores on X to English speech translation. hours from YODAS (Section 3.1). For these experiments, we fix the model size at 1B parameters. This leads to a total of 6 different models trained on 360K, 180K, 90K, 45K, 22.5K, and 11.25K hours of speech respectively. W…
Figure 8
Figure 8. Figure 8: BLEU scores on EN to X (top) and X to EN (bottom) ST with different dataset sizes. stream performance (R2 ≃ 0.55). The top and bottom portions of [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: shows the evolution of average WER from the 3 languages for each model size as training progresses. We find that for a fixed parameter size, the WER of the final checkpoint can be reliably predicted as a function of the training compute (R2 ≃ 0.82). This means that one…
Figure 10
Figure 10. Figure 10: Effects of model scaling on orthographic understand￾ing on Chinese (left) and Japanese (right). The quick saturation in PN-CER shows that scaling does not have a large effect on the phonetic understanding in ASR models. However, the raw CER trend shows that large-scal…
Figure 11
Figure 11. Figure 11: CER on zero-shot English-X code-switching [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: Model scaling laws for all languages in FLEURS. For almost all languages, WER/CER strongly correlated with the power law w.r.t. model parameter size. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Change in WER/CER when adding more data per language. For most languages, we observe the same trend as [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. OLMoASR: Open Models and Data for Training Robust Speech Recognition Models

    cs.SD 2025-08 conditional novelty 7.0 of 10

    An open 1M-hour English speech dataset plus Whisper-architecture models trained on it match Whisper's word error rates on short and long-form benchmarks.

  2. Benchmarking Training Paradigms, Dataset Composition, and Model Scaling for Child ASR in ESPnet

    cs.LG 2025-08 conditional novelty 5.0 of 10

    A systematic child ASR benchmark shows adult-trained SSL features are biased against child speech, flat-start training on child data helps, and zero-shot scaling plateaus near 1B parameters.

  3. Robust Assamese Speech Recognition through Controlled Fine-Tuning of Whisper Models

    cs.LG 2026-07 conditional novelty 4.0 of 10

    Fine-tuning Whisper-Small on 3,520 Assamese clips from Common Voice cuts word error rate from 201% to 44% and character error rate from 191% to 13%.

Reference graph

Works this paper leans on

83 extracted references · 60 canonical work pages · cited by 3 Pith papers

  1. [1]

    S., Anastasopoulos, A., Bojar, O., Borg, C., Carpuat, M., Cattoni, R., Cettolo, M., Chen, W., Dong, Q., Federico, M., Haddow, B., Javorsk \'y , D., Krubi \'n ski, M., Lam, T

    Ahmad, I. S., Anastasopoulos, A., Bojar, O., Borg, C., Carpuat, M., Cattoni, R., Cettolo, M., Chen, W., Dong, Q., Federico, M., Haddow, B., Javorsk \'y , D., Krubi \'n ski, M., Lam, T. K., Ma, X., Mathur, P., Matusov, E., Maurya, C., McCrae, J., Murray, K., Nakamura, S., Negri, M., Niehues, J., Niu, X., Ojha, A. K., Ortega, J., Papi, S., Pol \'a k, P., Po...

    2024

  2. [2]

    Common voice: A massively-multilingual speech corpus

    Ardila, R., Branson, M., Davis, K., Kohler, M., Meyer, J., Henretty, M., Morais, R., Saunders, L., Tyers, F., and Weber, G. Common voice: A massively-multilingual speech corpus. In LREC 2020, pp.\ 4218--4222, 2020

    2020

  3. [3]

    XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale

    Babu, A., Wang, C., Tjandra, A., Lakhotia, K., Xu, Q., Goyal, N., Singh, K., von Platen , P., Saraf, Y., Pino, J., Baevski, A., Conneau, A., and Auli, M. XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale . In Interspeech 2022, pp.\ 2278--2282, 2022. doi:10.21437/Interspeech.2022-143

    doi:10.21437/interspeech.2022-143 2022

  4. [4]

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

    Baevski, A., Zhou, Y., Mohamed, A., and Auli, M. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NeurIPS 2020, volume 33, 2020

    2020

  5. [5]

    Bang, J.-U. et al. K spon S peech: Korean spontaneous speech corpus for automatic speech recognition. Applied Sciences, 2020

    work page 2020

  6. [6]

    C., Dale, D., Dong, N., Duppenthaler, M., Duquenne, P.-A., Ellis, B., Elsahar, H., Haaheim, J., et al

    Barrault, L., Chung, Y.-A., Meglioli, M. C., Dale, D., Dong, N., Duppenthaler, M., Duquenne, P.-A., Ellis, B., Elsahar, H., Haaheim, J., et al. Seamless: Multilingual expressive and streaming speech translation. arxiv:2312.05187, 2023 a

    arXiv 2023

  7. [7]

    C., Dale, D., Dong, N., Duquenne, P.-A., Elsahar, H., Gong, H., Heffernan, K., Hoffman, J., et al

    Barrault, L., Chung, Y.-A., Meglioli, M. C., Dale, D., Dong, N., Duquenne, P.-A., Elsahar, H., Gong, H., Heffernan, K., Hoffman, J., et al. SeamlessM4T -massively multilingual & multimodal machine translation. arxiv:2308.11596, 2023 b

    arXiv 2023

  8. [8]

    aidatatang\_200zh, a free Chinese Mandarin speech corpus

    Beijing DataTang Technology Co., L. aidatatang\_200zh, a free Chinese Mandarin speech corpus

    work page

Show all 83 references
  1. [9]

    A., Purohit, S., Prashanth, U

    Biderman, S., Schoelkopf, H., Anthony, Q., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., Skowron, A., Sutawika, L., and Van Der Wal, O. Pythia: a suite for analyzing large language models across training and scaling. In Proc. ICM...

    2023

  2. [10]

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litw...

    1901

  3. [11]

    Bu, H. et al. AISHELL-1: An open-source Mandarin speech corpus and a speech recognition baseline . In O-COCOSDA, 2017

    2017

  4. [12]

    Siminchik: A speech corpus for preservation of southern Quechua

    Cardenas, R., Zevallos, R., Baquerizo, R., and Camacho, L. Siminchik: A speech corpus for preservation of southern Quechua . ISI-NLP, 2018

    2018

  5. [13]

    Unleashing the killer corpus: experiences in creating the multi-everything AMI meeting corpus

    Carletta, J. Unleashing the killer corpus: experiences in creating the multi-everything AMI meeting corpus. Springer, 2007

    2007

  6. [14]

    Cattoni, R. et al. MuST-C : A multilingual corpus for end-to-end speech translation. Computer speech & language, 66, 2021

    2021

  7. [15]

    Chen, G. et al. GigaSpeech : An evolving, multi-domain ASR corpus with 10,000 hours of transcribed audio. In Interspeech 2021, 2021

    2021

  8. [16]

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

    Chen, S., Wang, C., Chen, Z., Wu, Y., Liu, S., Chen, Z., Li, J., Kanda, N., Yoshioka, T., Xiao, X., Wu, J., Zhou, L., Ren, S., Qian, Y., Qian, Y., Wu, J., Zeng, M., Yu, X., and Wei, F. WavLM : Large-scale self-supervised pre-training for full stack speech processing. IEEE JSTS...

    2022

  9. [17]

    Joint prediction and denoising for large-scale multilingual self-supervised learning

    Chen, W., Shi, J., Yan, B., Berrebbi, D., Zhang, W., Peng, Y., Chang, X., Maiti, S., and Watanabe, S. Joint prediction and denoising for large-scale multilingual self-supervised learning. In ASRU 2023, 2023 a

    2023

  10. [18]

    Improving massively multilingual ASR with auxiliary CTC objectives

    Chen, W., Yan, B., Shi, J., Peng, Y., Maiti, S., and Watanabe, S. Improving massively multilingual ASR with auxiliary CTC objectives. In ICASSP 2023, 2023 b

    2023

  11. [19]

    Towards robust speech representation learning for thousands of languages

    Chen, W., Zhang, W., Peng, Y., Li, X., Tian, J., Shi, J., Chang, X., Maiti, S., Livescu, K., and Watanabe, S. Towards robust speech representation learning for thousands of languages. In Al-Onaizan, Y., Bansal, M., and Chen, Y.-N. (eds.), Proceedings of the 2024 Conference on ...

    2024 doi

  12. [20]

    Unsupervised Cross-Lingual Representation Learning for Speech Recognition

    Conneau, A., Baevski, A., Collobert, R., Mohamed, A., and Auli, M. Unsupervised Cross-Lingual Representation Learning for Speech Recognition . In Interspeech 2021, pp.\ 2426--2430, 2021. doi:10.21437/Interspeech.2021-329

    2021 doi

  13. [21]

    Conneau, A. et al. FLEURS : Few-shot learning evaluation of universal representations of speech. In SLT 2022, 2022

    2022

  14. [22]

    and Marxer, R

    Cuervo, S. and Marxer, R. Scaling properties of speech language models. In Al-Onaizan, Y., Bansal, M., and Chen, Y.-N. (eds.), Proc. EMNLP, pp.\ 351--361, Miami, Florida, USA, November 2024. Association for Computational Linguistics

    2024

  15. [23]

    Leveraging transparency

    Dabbish, L., Stuart, C., Tsay, J., and Herbsleb, J. Leveraging transparency. IEEE software, 30 0 (1): 0 37--43, 2012

    2012

  16. [24]

    Flashattention-2: Faster attention with better parallelism and work partitioning

    Dao, T. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations, 2024

    2024

  17. [25]

    and Elibol, O

    Droppo, J. and Elibol, O. Scaling laws for acoustic models. In Proc. Interspeech, pp.\ 2576--2580, 2021. doi:10.21437/Interspeech.2021-1644

    2021 doi

  18. [26]

    Scaling laws for multilingual neural machine translation

    Fernandes, P., Ghorbani, B., Garcia, X., Freitag, M., and Firat, O. Scaling laws for multilingual neural machine translation. In Proc. ICML, ICML'23, 2023

    2023

  19. [27]

    Scaling laws for neural machine translation

    Ghorbani, B., Firat, O., Freitag, M., Bapna, A., Krikun, M., Garcia, X., Chelba, C., and Cherry, C. Scaling laws for neural machine translation. In Proc. ICLR, 2022. URL https://openreview.net/forum?id=hR_SMu8cxCV

    2022

  20. [28]

    A., Duh, K., and Kaplan, J

    Gordon, M. A., Duh, K., and Kaplan, J. Data and parameter scaling laws for neural machine translation. In Proc. EMNLP, pp.\ 5915--5922, Online and Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics. doi:10.18653/v1/2021.emnlp-main.478. URL...

    2021 doi

  21. [29]

    Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks

    Graves, A., Fern \'a ndez, S., Gomez, F., and Schmidhuber, J. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In ICML 2006, pp.\ 369--376, 2006

    2006

  22. [30]

    Scaling laws for discriminative speech recognition rescoring models

    Gu, Y., Gurunath Shivakumar , P., Kolehmainen, J., Gandhe, A., Rastrow, A., and Bulyko, I. Scaling laws for discriminative speech recognition rescoring models. In Proc. Interspeech, pp.\ 471--475, 2023

    2023

  23. [31]

    B., Dhariwal, P., Gray, S., et al

    Henighan, T., Kaplan, J., Katz, M., Chen, M., Hesse, C., Jackson, J., Jun, H., Brown, T. B., Dhariwal, P., Gray, S., et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020

    2010 arXiv

  24. [32]

    Scaling laws for transfer

    Hernandez, D., Kaplan, J., Henighan, T., and McCandlish, S. Scaling laws for transfer. arXiv preprint arXiv:2102.01293, 2021

    2021 arXiv

  25. [33]

    TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation

    Hernandez, F., Nguyen, V., Ghannay, S., Tomashenko, N., and Esteve, Y. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In Speech and Computer: 20th International Conference, SPECOM 2018. Springer, 2018

    2018

  26. [34]

    Hestness, J., Narang, S., Ardalani, N., Diamos, G., Jun, H., Kianinejad, H., Patwary, M. M. A., Yang, Y., and Zhou, Y. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017

    2017 arXiv

  27. [35]

    The Babel Program

    IARPA . The Babel Program . URL www.iarpa.gov/index.php/research-programs/babel

  28. [36]

    B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  29. [37]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. ICLR 2015, 2015

    2015

  30. [38]

    Subword regularization: Improving neural network translation models with multiple subword candidates

    Kudo, T. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 66--75, Melbourne, Australia, July 2018

    2018

  31. [39]

    Le, D., Jain, M., Keren, G., Kim, S., Shi, Y., Mahadeokar, J., Chan, J., Shangguan, Y., Fuegen, C., Kalinli, O., Saraf, Y., and Seltzer, M. L. Contextualized streaming end-to-end speech recognition with trie-based deep biasing and shallow fusion. In Proc. Interspeech, pp.\ 177...

    2021

  32. [40]

    N., Gulati, A., Zhang, Y., Qin, J., Haghani, P., Huang, W

    Li, B., Pang, R., Sainath, T. N., Gulati, A., Zhang, Y., Qin, J., Haghani, P., Huang, W. R., Ma, M., and Bai, J. Scaling end-to-end models for large-scale multilingual asr. In Proc. ASRU, pp.\ 1011--1018, 2021

    2021

  33. [41]

    R., Black, A

    Li, X., Metze, F., Mortensen, D. R., Black, A. W., and Watanabe, S. ASR2K: Speech Recognition for Around 2000 Languages without Audio . In Interspeech 2022, 2022. doi:10.21437/Interspeech.2022-10712

    2000 doi

  34. [42]

    YODAS : Y outube-oriented dataset for audio and speech

    Li, X., Takamichi, S., Saeki, T., Chen, W., Shiota, S., and Watanabe, S. YODAS : Y outube-oriented dataset for audio and speech. In ASRU 2023, 2023

    2023

  35. [43]

    K., Lavrukhin, V., Majumdar, S., Noroozi, V., Zhang, Y., Kuchaiev, O., Balam, J., Dovzhenko, Y., Freyberg, K., Shulman, M

    O’Neill, P. K., Lavrukhin, V., Majumdar, S., Noroozi, V., Zhang, Y., Kuchaiev, O., Balam, J., Dovzhenko, Y., Freyberg, K., Shulman, M. D., Ginsburg, B., Watanabe, S., and Kucsko, G. SPGISpeech : 5,000 hours of transcribed financial audio for fully formatted end-to-end speech r...

    2021

  36. [44]

    Panayotov, V. et al. Librispeech: An ASR corpus based on public domain audio books. In ICASSP 2015, 2015

    2015

  37. [45]

    Prompting the hidden talent of web-scale speech models for zero-shot task generalization

    Peng, P., Yan, B., Watanabe, S., and Harwath, D. Prompting the hidden talent of web-scale speech models for zero-shot task generalization. In Proc. Interspeech, 2023 a

    2023

  38. [46]

    Reproducing W hisper-style training using an open-source toolkit and publicly available data

    Peng, Y., Tian, J., Yan, B., Berrebbi, D., Chang, X., Li, X., Shi, J., Arora, S., Chen, W., Sharma, R., Zhang, W., Sudo, Y., Shakeel, M., weon Jung, J., Maiti, S., and Watanabe, S. Reproducing W hisper-style training using an open-source toolkit and publicly available data. In...

    2023

  39. [47]

    Peng, Y., Tian, J., Chen, W., Arora, S., Yan, B., Sudo, Y., Shakeel, M., Choi, K., Shi, J., Chang, X., et al. OWSM v3. 1: Better and Faster Open Whisper-Style Speech Models based on E-Branchformer . arXiv preprint arXiv:2401.16658, 2024

    2024 arXiv

  40. [48]

    Post, M. et al. Improved speech-to-text translation with the fisher and callhome S panish- E nglish speech translation corpus. In IWSLT 2013, 2013

    2013

  41. [49]

    MLS : A large-scale multilingual dataset for speech research

    Pratap, V., Xu, Q., Sriram, A., Synnaeve, G., and Collobert, R. MLS : A large-scale multilingual dataset for speech research. In Interspeech 2020, pp.\ 2757--2761

    2020

  42. [50]

    Scaling speech technology to 1,000+ languages

    Pratap, V., Tjandra, A., Shi, B., Tomasello, P., Babu, A., Kundu, S., Elkahky, A., Ni, Z., Vyas, A., Fazel-Zarandi, M., et al. Scaling speech technology to 1,000+ languages. arxiv:2305.13516, 2023

    2023 arXiv

  43. [51]

    C., \.Z elasko, P., Huang, H., Hrinchuk, O., Koluguri, N

    Puvvada, K. C., \.Z elasko, P., Huang, H., Hrinchuk, O., Koluguri, N. R., Dhawan, K., Majumdar, S., Rastorgueva, E., Chen, Z., Lavrukhin, V., et al. Less is more: Accurate speech recognition & translation without web-scale data. arXiv preprint arXiv:2406.19674, 2024

    2024 arXiv

  44. [52]

    W., Xu, T., Brockman, G., Mcleavey, C., and Sutskever, I

    Radford, A., Kim, J. W., Xu, T., Brockman, G., Mcleavey, C., and Sutskever, I. Robust speech recognition via large-scale weak supervision. In ICML 2023, 2023

    2023

  45. [53]

    Zero: Memory optimizations toward training trillion parameter models

    Rajbhandari, S., Rasley, J., Ruwase, O., and He, Y. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pp.\ 1--16, 2020

    2020

  46. [54]

    Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters

    Rasley, J., Rajbhandari, S., Ruwase, O., and He, Y. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '20, pp.\ 3505–35...

    2020

  47. [55]

    S., Rosenfeld, A., Belinkov, Y., and Shavit, N

    Rosenfeld, J. S., Rosenfeld, A., Belinkov, Y., and Shavit, N. A constructive prediction of the generalization error across scales. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=ryenvpEKDr

    2020

  48. [56]

    Are emergent abilities of large language models a mirage? In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S

    Schaeffer, R., Miranda, B., and Koyejo, S. Are emergent abilities of large language models a mirage? In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.), Proc. NeurIPS, volume 36, pp.\ 55565--55581. Curran Associates, Inc., 2023

    2023

  49. [57]

    Slizhikova, A. et al. Russian Open Speech To Text (STT/ASR) Dataset , 2020. URL https://github.com/snakers4/open_stt

    2020

  50. [58]

    and Chiang, D

    Taguchi, C. and Chiang, D. Language complexity and speech recognition accuracy: Orthographic complexity hurts, phonological complexity doesn't. arXiv preprint arXiv:2406.09202, 2024

    2024 arXiv

  51. [59]

    Scaling laws vs model architectures: How does inductive bias influence scaling? In Bouamor, H., Pino, J., and Bali, K

    Tay, Y., Dehghani, M., Abnar, S., Chung, H., Fedus, W., Rao, J., Narang, S., Tran, V., Yogatama, D., and Metzler, D. Scaling laws vs model architectures: How does inductive bias influence scaling? In Bouamor, H., Pino, J., and Bali, K. (eds.), Findings of EMNLP, pp.\ 12342--12...

    2023 doi

  52. [60]

    On the effects of heterogeneous data sources on speech-to-text foundation models

    Tian, J., Peng, Y., Chen, W., Choi, K., Livescu, K., and Watanabe, S. On the effects of heterogeneous data sources on speech-to-text foundation models. In Interspeech 2024, pp.\ 3959--3963, 2024. doi:10.21437/Interspeech.2024-1938

    2024 doi

  53. [61]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need. In NeurIPS 2017, 2017

    2017

  54. [62]

    VoxForge

    VoxForge. VoxForge . URL http://www.voxforge.org/

  55. [63]

    Wang, C. et al. CoVoST 2 and Massively Multilingual Speech Translation . In Interspeech, 2021 a

    2021

  56. [64]

    Wang, C. et al. VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation . In ACL 2021, 2021 b

    2021

  57. [65]

    Can whisper perform speech-based in-context learning? In Proc

    Wang, S., Yang, C.-H., Wu, J., and Zhang, C. Can whisper perform speech-based in-context learning? In Proc. ICASSP, pp.\ 13421--13425, 2024 a

    2024

  58. [66]

    H., Wu, J., and Zhang, C

    Wang, S., Yang, C.-H. H., Wu, J., and Zhang, C. B ayesian example selection improves in-context learning for speech, text and visual modalities. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp.\ 20812--20828, Miami, Florida, USA, N...

    2024 doi

  59. [67]

    Watanabe, S., Hori, T., and Hershey, J. R. Language independent end-to-end architecture for joint language identification and speech recognition. In 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp.\ 265--271, 2017 a . doi:10.1109/ASRU.2017.8268945

    2017

  60. [68]

    R., and Hayashi, T

    Watanabe, S., Hori, T., Kim, S., Hershey, J. R., and Hayashi, T. Hybrid CTC /attention architecture for end-to-end speech recognition. IEEE Journal of Selected Topics in Signal Processing, 2017 b

    2017

  61. [69]

    ESP net: End-to-end speech processing toolkit

    Watanabe, S., Hori, T., Karita, S., Hayashi, T., Nishitoba, J., Unno, Y., Enrique Yalta Soplin , N., Heymann, J., Wiesner, M., Chen, N., Renduchintala, A., and Ochiai, T. ESP net: End-to-end speech processing toolkit. In Interspeech 2018, 2018

    2018

  62. [70]

    H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., and Fedus, W

    Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., and Fedus, W. Emergent abilities of large language models. Transactions on Machine Learning Researc...

    2022

  63. [71]

    Yamagishi, J. et al. CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit , 2019

    2019

  64. [72]

    Code-switching Evaluation Set , 2024

    Yan, B., Shimizu, S., and Watanabe, S. Code-switching Evaluation Set , 2024. URL https://github.com/brianyan918/sentence-recorder/tree/codeswitching/

    2024

  65. [73]

    H., Li, B., Zhang, Y., Chen, N., Prabhavalkar, R., Sainath, T

    Yang, C.-H. H., Li, B., Zhang, Y., Chen, N., Prabhavalkar, R., Sainath, T. N., and Strohman, T. From english to more languages: Parameter-efficient model reprogramming for cross-lingual speech recognition. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech ...

    2023

  66. [74]

    Open source MagicData-RAMC : A rich annotated mandarin conversational ( RAMC ) speech dataset

    Yang, Z., Chen, Y., Luo, L., Yang, R., Ye, L., Cheng, G., Xu, J., Jin, Y., Zhang, Q., Zhang, P., Xie, L., and Yan, Y. Open source MagicData-RAMC : A rich annotated mandarin conversational ( RAMC ) speech dataset. In Interspeech 2022, pp.\ 1736--1740, 2022

    2022

  67. [75]

    Ye, R. et al. Giga ST : A 10,000-hour pseudo speech translation corpus. In Interspeech 2023, 2023

    2023

  68. [76]

    ReazonSpeech: A Free and Massive Corpus for Japanese ASR , 2023

    Yin, Y., Mori, D., et al. ReazonSpeech: A Free and Massive Corpus for Japanese ASR , 2023

    2023

  69. [77]

    H., Kolehmainen, J., Shivakumar, P

    Yu, Y., Yang, C.-H. H., Kolehmainen, J., Shivakumar, P. G., Gu, Y., Ren, S. R. R., Luo, Q., Gourav, A., Chen, I.-F., Liu, Y.-C., et al. Low-rank adaptation of large language model rescoring for parameter-efficient speech recognition. In 2023 IEEE Automatic Speech Recognition a...

    2023

  70. [78]

    Scaling vision transformers

    Zhai, X., Kolesnikov, A., Houlsby, N., and Beyer, L. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 12104--12113, 2022

    2022

  71. [79]

    Zhang, B. et al. Wenet S peech: A 10000+ hours multi-domain mandarin corpus for speech recognition. In ICASSP 2022, 2022 a

    2022

  72. [80]

    V., et al

    Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X. V., et al. OPT : Open pre-trained transformer language models. arxiv:2205.01068, 2022 b

    2022 arXiv

  73. [81]

    Google USM : Scaling automatic speech recognition beyond 100 languages

    Zhang, Y., Han, W., Qin, J., Wang, Y., Bapna, A., Chen, Z., Chen, N., Li, B., Axelrod, V., Wang, G., et al. Google USM : Scaling automatic speech recognition beyond 100 languages. arxiv:2303.01037, 2023

    2023 arXiv

  74. [82]

    Scaling asr improves zero and few shot learning

    Zheng, W., Xiao, A., Keren, G., Le, D., Zhang, F., Fuegen, C., Kalinli, O., Saraf, Y., and Mohamed, A. Scaling asr improves zero and few shot learning. In Proc. Interspeech, pp.\ 5135--5139, 2022

    2022

  75. [83]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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