Pith. sign in

REVIEW 5 major objections 6 minor 37 references

Self-supervised learning of speech representations with Dutch archival data

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 This paper claims that a wav2vec 2.0 model continued from XLS-R on 55.7k hours of cleaned Dutch TV audio reaches 9.5% WER on the N-Best benchmark, below the 10% reported for Whisper large-v2.

desk verdict Solid empirical study of Dutch broadcast pre-training, but the headline mono-lingual advantage and SOTA WER rest on a possible test-set overlap that the paper never rules out. read the letter →

arxiv 2507.04554 v2 pith:YZGYRX6D submitted 2025-07-06 cs.SD cs.CLcs.LGeess.AS

Nik Vaessen , Roeland Ordelman , David A. van Leeuwen This is my paper · ORCID
classification cs.SDcs.CLcs.LGeess.AS
keywords self-supervisedlearningwav2vec2.0DutchspeechrecognitionarchivaltelevisionbroadcastWhisperXsegmentationmono-lingualpre-trainingdataqualitysimulationN-Bestbenchmark
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 tries to establish that a large self-supervised speech model trained only on Dutch audio can outperform multi-lingual models on Dutch recognition, and that noisy television archives can be turned into a usable pre-training corpus. The authors argue that wav2vec 2.0 pre-training is not robust to data-quality defects, with music and especially vocal music the most destabilizing, and that a Whisper/WhisperX segmentation pipeline can turn 81k hours of broadcast audio into 55.7k usable hours. On an out-of-domain Dutch benchmark (N-Best), a LARGE wav2vec 2.0 model continued from an XLS-R checkpoint reaches 9.5% WER, below the 10% reported for Whisper large-v2. The significance, if true, is that mono-lingual pre-training is a viable alternative to multi-lingual pre-training for languages with access to large raw archives, and that existing multi-lingual checkpoints can be adapted cheaply.

What carries the argument

The argument is carried by the contrastive masked-prediction objective of wav2vec 2.0, applied in two regimes: controlled data-quality simulations on LibriSpeech, where batches are deliberately contaminated with speaker overlap, MUSAN noise, or FMA music to expose which defects break pre-training, and a Whisper/WhisperX segmentation pipeline for the Dutch archive that uses transcribability and speaker diarization as a proxy for clean, single-speaker, speech-only segments. The mono-lingual comparison is carried by matched-scale pre-training on nibg-pp versus published multi-lingual checkpoints, plus continued pre-training of XLS-R on the Dutch data.

What would settle it

Pre-train a multi-lingual wav2vec 2.0 model on broadcast-domain audio from several languages using the same WhisperX segmentation and training recipe as nibg-pp, then evaluate on N-Best; if its WER matches or beats the 9.5% mono-lingual result, the claim that mono-lingual pre-training is more robust than multi-lingual pre-training is not supported.

Watch

Extended reading notes

Core claim

The core discovery is that a 55.7k-hour Dutch pre-training corpus extracted from archival television broadcasts, segmented with Whisper and WhisperX and then filtered and speaker-diarized, can support a LARGE wav2vec 2.0 model that beats published multi-lingual models on an out-of-domain Dutch evaluation. Continued pre-training of the XLS-R checkpoint on this Dutch-only data yields 9.5% WER on the N-Best evaluation, lower than the 10% WER reported for Whisper large-v2. The paper also finds, by deliberately contaminating LibriSpeech batches, that music in pre-training data is the most harmful quality defect, while a second speaker mixed in at low signal-to-noise ratio is comparatively tolerable. Finally, it claims that with equivalent data size, mono-lingual pre-training is more robust than multi-lingual pre-training when the evaluation data is out-of-domain, although multi-lingual pre-training can still match it on in-domain data.

Load-bearing premise

The mono-lingual versus multi-lingual comparison assumes the difference between nibg-pp and the XLSR/XLS-R baselines is essentially language composition, but the datasets also differ in recording domain, sampling, batch schedule, and checkpoint history, so the out-of-domain advantage could be a domain effect.

Editorial extensions

If this is right

  • Dutch ASR can reach competitive accuracy with a single-language wav2vec 2.0 model, provided enough clean mono-lingual audio is available.
  • Existing broadcast archives in under-resourced languages can be repurposed as SSL corpora using Whisper-based filtering and diarization, avoiding the need for manually curated read-speech data.
  • For out-of-domain evaluation, mono-lingual pre-training gives a robustness advantage, while multi-lingual pre-training may still win on in-domain benchmarks.
  • Continued pre-training from XLS-R with mono-lingual data is a cheap path to a high-performing Dutch model: 9.5% WER on N-Best versus 10% for Whisper large-v2.
  • Pre-training corpora for wav2vec 2.0 should filter music, especially vocal music, rather than only noise or overlapping speech, to avoid unstable training.

Reading between the lines

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

  • The paper does not test the domain confound: a multi-lingual model pre-trained on broadcast-domain audio from several languages might show the same out-of-domain robustness, which would place the mono-lingual advantage on domain rather than language.
  • Because the Whisper/WhisperX pipeline keeps only audio Whisper can transcribe, dialectal or heavily overlapping speech is likely discarded; a purpose-built speech-activity detector could preserve more useful hours, an extension the paper does not run.
  • The jump from a 5-minute to a 40-minute batch roughly halved WER for the BASE model, so part of the reported mono-lingual edge may be attributable to training recipe and throughput rather than language composition; isolating the two would require a cross-language batch-size sweep.
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

5 major / 6 minor

Summary. This paper investigates whether Dutch archival television data can be used for self-supervised pre-training of wav2vec 2.0 speech representations. It contributes three sets of experiments: (1) controlled LibriSpeech augmentations simulating speaker overlap, music, and noise to identify data-quality requirements; (2) a comparison of Whisper/WhisperX-based segmentation and cleaning pipelines applied to 2,500 hours of NIBG broadcast audio, with the best pipeline scaled to produce a 55.7k-hour Dutch dataset (nibg-pp); and (3) a comparison of mono-lingual Dutch pre-training with existing multi-lingual checkpoints, including a continuation of XLS-R. The headline result is a LARGE wav2vec 2.0 model that reaches 9.5% WER on the N-Best Dutch broadcast-news benchmark, compared with 10% reported for Whisper large-v2, and the paper concludes that mono-lingual pre-training is more robust on out-of-domain data.

Significance. Taken at face value, the paper offers a practical and useful account of how a noisy national broadcast archive can be converted into an SSL pre-training corpus, and its data-quality simulations provide concrete guidance on what types of corruption matter most. A credible Dutch ASR result at this scale would be valuable to the community if the evaluation is clean. The paper is less successful as a controlled scientific comparison of mono-lingual versus multi-lingual pre-training, because the comparisons in Table 3 are confounded by domain, recipe, and checkpoint history. The absence of any overlap check between the N-Best evaluation data and nibg-pp is the most serious limitation and must be resolved before the headline numbers can be trusted. The manuscript also does not state a code or model release plan, which limits independent verification of the reported numbers.

major comments (5)
  1. [§3.1, §3.3, §4.3] The N-Best evaluation material and the nibg-pp pre-training corpus are not explicitly disjoint. N-Best's Northern Dutch Broadcast News test section is likely drawn from the same Dutch broadcast archive (Netherlands Institute for Sound and Vision, 1972–2022) used to build nibg-pp; if any test utterance or a near-duplicate segment of the same broadcast appears in the 55.7k-hour pre-training set, the Table 3 N-Best WERs and the headline comparison against Whisper large-v2 are invalid. The manuscript must state whether N-Best audio was excluded or deduplicated from nibg-pp, and if not, quantify the overlap and re-run the affected conditions.
  2. [§4.3, §5] The claim that N-Best is 'out-of-domain for all pre-training data' is not supported. nibg-pp consists of Dutch television broadcasts from the same archive and time period from which N-Best's Northern Dutch Broadcast News material is drawn, so N-Best is at least partially in-domain for the mono-lingual nibg-pp condition. The comparison against multi-lingual models trained on read/web speech is also confounded by domain, sampling, training recipe, and checkpoint history. The conclusion in §5 that mono-lingual pre-training is more robust on out-of-domain data therefore cannot be attributed to language composition on the evidence presented.
  3. [Abstract; §4.3, Table 3] The abstract's claim that mono-lingual and multi-lingual pre-training are compared 'with equivalent amounts of data' is only weakly realized. The only multi-lingual row with roughly the same data amount as nibg-pp is LARGE XLSR-53 (56k h), and it differs in domain, batch schedule, and number of pre-training steps; the smaller mono-lingual baselines (50h–1.5k h) are not matched by any multi-lingual condition. The comparison therefore does not isolate mono-linguality from recipe and domain differences.
  4. [Tables 1–3] All WERs are reported from single training runs without error bars, seeds, or significance tests. Several numerical differences that support qualitative claims are small (e.g., Table 1, 1spk-cat versus baseline on test-clean: 12.8 versus 12.4; Table 3, LARGE nl 56k versus LARGE XLS-R on MLS-t: 13.3 versus 13.4). It is impossible to tell which of these differences are meaningful. I recommend at least three seeds for the key conditions in Tables 1 and 3, or a variance estimate for the headline N-Best result.
  5. [§4.1, §4.2, Tables 1 and 2] The paper omits runs that diverged: w-raw and w-pp-1s in Table 2, and mix-music and mix-vocal in Table 1. Because these divergences are reported only as 'DIV', the reader cannot know whether the omission reflects a genuine data-quality effect or a hyperparameter/optimization issue. The paper should report the number of diverged runs, the learning rate scan outcomes, and ideally make the non-converged curves available; otherwise the 'best preprocessing' ranking in Table 2 is selected conditional on convergence.
minor comments (6)
  1. [§1] There are several typos: 'There models' should be 'These models', 'have have similar' should be 'have similar', and 'expect for' in §3.1 should be 'except for'.
  2. [Table 2 caption] The caption reads 'We show also show average utterance length'; it should be 'We also show average utterance length'.
  3. [Table 3 caption] The caption contains 'afer' instead of 'after', and the abbreviated column headers such as 'ft. /w' and 'MLS-t / NBest' should be defined in the caption or in the text.
  4. [§4.2] The text says 'pre-training ... 40 minuntes' and 'the runtime for creating the wx-diar-3s dataset is lower (40 h) than w-pp-3s (50 h)'; these should be corrected for spelling and the runtime units made consistent.
  5. [§3.4] The WhisperX alignment/diarization stage uses a Dutch wav2vec 2.0 model (jonatasgrosman/wav2vec2-large-xlsr-53-dutch). The paper should discuss whether this biases the cleaned dataset toward audio that is easily recognized by Dutch ASR, and how that affects the generality of the preprocessing conclusions.
  6. [§4.3] The comparison with Whisper large-v2 relies on a WER of 10% reported in [8]; the paper should state whether the evaluation protocols (normalization, decoding, test subset) are identical, since the N-Best number reported here uses caseless word comparison and greedy letter decoding.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical WER results are measured on external held-out benchmarks; the self-referential data-cleaning pipeline is a validity concern, not a derivational reduction.

full rationale

This is an empirical systems paper with no first-principles derivation whose output could be equivalent to its inputs by construction. The central results in Table 3 are fine-tuned WERs measured on external evaluation sets (MLS test, CommonVoice test, and the N-Best Northern Dutch Broadcast News section), not quantities fitted to those same targets. The comparison against Whisper large-v2 uses the externally reported 10% figure from [8]; although [8] shares an author with the present paper, the present paper's 9.5% result is an independent measurement on the N-Best evaluation, so the citation is not load-bearing in a circular sense. The use of WhisperX, which internally applies a Dutch wav2vec 2.0 model for forced alignment when constructing nibg-pp, is a pre-processing recursion rather than a mathematical equivalence: the cleaned segments are still judged on external benchmarks, and the aligner is a third-party artifact, not the paper's own fitted output. The mono-lingual versus multi-lingual comparison is confounded by differences in domain, data size, and checkpoint history (notably the best LARGE model is initialized from XLS-R), but confounding is a validity and interpretation risk, not circularity. Similarly, the absence of an explicit N-Best test-set overlap check is a possible data-leakage concern, not a self-consistency circularity. No specific step can be exhibited in which a claimed prediction reduces by definition or by the paper's own equations to its inputs, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper is entirely empirical, with no theoretical derivation. The ledger therefore captures the background modeling assumptions and the hand-chosen hyperparameters that the results depend on.

free parameters (6)
  • Peak learning rate = grid-searched 1e-5 to 1e-3, best values not reported
    Used for all pre-training runs; final values are not listed in the paper.
  • Batch size = 5 min and 40 min of audio
    The 40-minute batch was chosen by scaling from the 5-minute batch using the square root scaling law.
  • Mixing fraction for data-quality simulations = 10% and 33%
    Chosen to simulate different levels of contamination; not derived from theory.
  • Signal-to-noise ratio range for speaker/audio mixing = 0 to 15 dB
    Uniformly sampled in simulations; no justification beyond plausible ranges.
  • Segment length thresholds in preprocessing = 1 s and 3 s
    Filtering thresholds for Whisper/WhisperX post-processing; chosen by hand.
  • Pre-training steps = 200k, 400k, or 500k
    Step counts differ across rows in Table 3 and were not systematically justified.
assumptions (4)
  • domain assumption Librispeech-like clean single-speaker data is the appropriate target for wav2vec 2.0 pre-training quality.
    Stated as the authors' hypothesis in Sections 1 and 3.2; the entire data-cleaning strategy depends on it.
  • domain assumption Audio segments that Whisper can successfully transcribe contain speech of sufficient quality for SSL pre-training.
    Core to the Whisper/WhisperX preprocessing in Section 3.4; accepted without independent verification.
  • domain assumption WhisperX diarization produces single-speaker segments suitable for SSL.
    The wx-diar variants in Section 3.4 assume the diarization stage yields accurate speaker boundaries.
  • domain assumption The genre and heuristic metadata filters in Section 3.3 retain speech-rich broadcasts without systematic bias.
    Filtering removed nightly news, sports, and broadcasts over three hours; the effect on final model quality is not separately evaluated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-supervised learning of speech representations with Dutch archival data." pith.science (2026). https://pith.science/paper/YZGYRX6D

@misc{pith2026250704554,
  author       = {Pith},
  title        = {Pith review of: Self-supervised learning of speech representations with Dutch archival data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YZGYRX6D}},
  note         = {Machine review of arXiv:2507.04554}
}
read the original abstract

This paper explores the use of Dutch archival television broadcast data for self-supervised learning of speech foundation models, specifically wav2vec 2.0. We first study data quality assumptions for pre-training, and show how music, noise and speaker overlap affect SSL convergence and downstream fine-tuning performance. Secondly, we explore effectively pre-processing strategies to convert the noisy broadcast dataset into a qualitative dataset for pre-training, by using Whisper and WhisperX. Thirdly, we compare mono-lingual and multi-lingual pre-training with equivalent amounts of data, and show that mono-lingual pre-training is more robust to out-of-domain data. Lastly, we achieve a state-of-the-art LARGE wav2vec 2.0 model for the Dutch language, by a continuation of pre-training a wav2vec 2.0 XLS-R model checkpoint with our 55k hour archival dataset.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

  1. [1]

    There models use self-supervised learning (SSL) to learn speech representations on unlabeled data, and evidently generalize well to downstream speech tasks

    Introduction Speech foundation models, such as wav2vec 2.0 [1], Hu- BERT [2], and WavLM [3], see wide-spread usage in various speech technology tasks [4]. There models use self-supervised learning (SSL) to learn speech representations on unlabeled data, and evidently generalize well to downstream speech tasks. One particular use-case of SSL is the paradig...

    work page

  2. [2]

    What constitutes a high quality dataset for self-supervised speech representation learning?

    work page

  3. [3]

    Can we construct such a dataset from television broadcast data?

    work page

  4. [4]

    We note that most self-supervised learning methods are solely developed on LS-like data, simply due to ease of ac- cess and availability of read speech

    Is pre-training on mono-lingual audio data better than pre- training on multi-lingual audio data? Firstly, as speculated earlier, our hypothesis on high-quality datasets is that they need to be as similar to Librispeech (LS) as possible. We note that most self-supervised learning methods are solely developed on LS-like data, simply due to ease of ac- cess...

    work page

  5. [5]

    Related work The authors of [10] compare mono-lingual pre-training with Japanese data to fine-tuning multi-lingual pre-trained mod- els, specifically, XLSR from [5]. They show that a BASE wav2vec 2.0 model, pre-trained on a dataset of 500 hours of spontaneous Japanese speech, outperforms the LARGE XLSR model, when fine-tuned (50 hours) and evaluated (3.5 ...

    work page Pith review arXiv 2025

  6. [6]

    sessions

    Methodology 3.1. Pre-training and fine-tuning In this paper we use the contrastive learning approach of wav2vec 2.0 for pre-training speech representations. We also fine-tune the network for automatic speech recognition with la- beled data using CTC loss [16]. We do not make any modifi- cation to the network architecture, making use of the BASE (12 transf...

    work page 1972

  7. [7]

    Data quality We pre-train the BASE wav2vec 2.0 variant on the 960 hours of training data from LS, but apply various augmentations as described in Section 3.2

    Experiments 4.1. Data quality We pre-train the BASE wav2vec 2.0 variant on the 960 hours of training data from LS, but apply various augmentations as described in Section 3.2. If not otherwise mentioned, we use the default configuration for pre-training and fine-tuning as de- scribed in [1]. For pre-training, we use AdamW [26] with a weight decay of 0.1, ...

    work page

  8. [8]

    We conclude that wav2vec 2.0 is not robust to the presence of (vocal) music

    Conclusion We have shown that there are (implicit) assumptions on data quality for self-supervised learning of speech representations. We conclude that wav2vec 2.0 is not robust to the presence of (vocal) music. We found that using Whisper, alongside simple heuristics, is an effective strategy to pre-process a noisy archival dataset into a qualitative pre...

    work page

Show all 37 references
  1. [9]

    Acknowledgment This work was sponsored by NWO - Domain Science for the use of supercomputer facilities (Snellius, www.surf.nl)

  2. [10]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” in Advances in Neural Information Processing Systems , vol. 33. Curran Associates, Inc., 2020, pp. 12 449–12 460

    2020

  3. [11]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhut- dinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460, 2021. [...

    2021 arXiv

  4. [12]

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

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

    2022

  5. [13]

    SUPERB: Speech Processing Universal PER- formance Benchmark,

    S. Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakho- tia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Lin, T.-H. Huang, W.-C. Tseng, K. tik Lee, D.-R. Liu, Z. Huang, S. Dong, S.-W. Li, S. Watanabe, A. Mohamed, and H. yi Lee, “SUPERB: Speech Processing Universal PER- ...

    2021

  6. [14]

    Unsupervised Cross-Lingual Representation Learning for Speech Recognition,

    A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli, “Unsupervised Cross-Lingual Representation Learning for Speech Recognition,” inInterspeech 2021. ISCA, Aug. 2021, pp. 2426–2430. [Online]. Available: https://www.isca-speech. org/archive/interspeech 2021/conneau21 ...

    2021

  7. [15]

    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,” in Proceedings of the 40th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, vol. 202. PMLR, 23–2...

    2023

  8. [16]

    Benefits of pre-trained mono-and cross-lingual speech representations for spoken language under- standing of dutch dysarthric speech,

    P. Wang and H. Van Hamme, “Benefits of pre-trained mono-and cross-lingual speech representations for spoken language under- standing of dutch dysarthric speech,”EURASIP Journal on Audio, Speech, and Music Processing, vol. 2023, no. 1, p. 15, 2023

    2023

  9. [17]

    Benchmarking and research infrastructures: Evaluating dutch automatic speech recognition,

    D. A. B ˘alan, R. J. Ordelman, K. Truong, and H. van den Heuvel, “Benchmarking and research infrastructures: Evaluating dutch automatic speech recognition,” in CLARIAH Annual Conference 2024, 2024

    2024

  10. [18]

    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,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  11. [19]

    Exploration of language dependency for japanese self-supervised speech rep- resentation models,

    T. Ashihara, T. Moriya, K. Matsuura, and T. Tanaka, “Exploration of language dependency for japanese self-supervised speech rep- resentation models,” in ICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  12. [20]

    Jnas: Japanese speech corpus for large vocabulary continuous speech recognition re- search,

    K. Itou, M. Yamamoto, K. Takeda, T. Takezawa, T. Matsuoka, T. Kobayashi, K. Shikano, and S. Itahashi, “Jnas: Japanese speech corpus for large vocabulary continuous speech recognition re- search,” Journal of the Acoustical Society of Japan (E) , vol. 20, no. 3, pp. 199–206, 1999

    1999

  13. [21]

    Don’t speak too fast: The impact of data bias on self-supervised speech models,

    Y . Meng, Y .-H. Chou, A. T. Liu, and H.-y. Lee, “Don’t speak too fast: The impact of data bias on self-supervised speech models,” in ICASSP 2022 - 2022 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) , 2022, pp. 3258– 3262

    2022

  14. [22]

    Compar- ing self-supervised pre-training and semi-supervised training for speech recognition in languages with weak language models,

    L.-M. Lam-Yee-Mui, L. O. Yang, and O. Klejch, “Compar- ing self-supervised pre-training and semi-supervised training for speech recognition in languages with weak language models,” in INTERSPEECH 2023, 2023, pp. 87–91

    2023

  15. [23]

    Com- bining multilingual resources and models to develop state-of-the- art e2e asr for swedish,

    L. Mateju, J. Nouza, P. ˇCerva, J. Zdansky, and F. Kynych, “Com- bining multilingual resources and models to develop state-of-the- art e2e asr for swedish,” inINTERSPEECH 2023, 2023, pp. 3252– 3256

    2023

  16. [24]

    A comparative analysis of bilingual and trilingual wav2vec mod- els for automatic speech recognition in multilingual oral history archives,

    J. Lehe ˇcka, J. V . Psutka, L. Smidl, P. Ircing, and J. Psutka, “A comparative analysis of bilingual and trilingual wav2vec mod- els for automatic speech recognition in multilingual oral history archives,” in Interspeech 2024, 2024, pp. 1285–1289

    2024

  17. [25]

    Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning, 2006, pp. 369–376

    2006

  18. [26]

    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,” in Inter- speech 2020, 2020, pp. 2757–2761

    2020

  19. [27]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the Twelfth Language Resources and Evaluation Conference. Marseille, France: European La...

    2020

  20. [28]

    The Spoken Dutch Corpus and its exploitation environment,

    N. H. J. Oostdijk and D. Broeder, “The Spoken Dutch Corpus and its exploitation environment,” inProceedings of the 4th Inter- national Workshop on Linguistically Interpreted Corpora (LINC- 03)., Budapest, Hungary, 2003

    2003

  21. [29]

    Results of the N-Best 2008 Dutch speech recognition evaluation,

    D. A. van Leeuwen, J. Kessens, E. Sanders, and H. van den Heuvel, “Results of the N-Best 2008 Dutch speech recognition evaluation,” in Proc. Interspeech. Brighton: ISCA, September 2009, pp. 2571–2574

    2008

  22. [30]

    Musan: A music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “Musan: A music, speech, and noise corpus,” arXiv preprint arXiv:1510.08484, 2015

    2015 arXiv

  23. [31]

    Fma: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “Fma: A dataset for music analysis,” in ISMIR 2017, 2017

    2017

  24. [32]

    Hybrid transformers for music source separation,

    S. Rouard, F. Massa, and A. D ´efossez, “Hybrid transformers for music source separation,” in ICASSP 2023 - 2023 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  25. [33]

    Whisperx: Time- accurate speech transcription of long-form audio,

    M. Bain, J. Huh, T. Han, and A. Zisserman, “Whisperx: Time- accurate speech transcription of long-form audio,” in INTER- SPEECH 2023, 2023, pp. 4489–4493

    2023

  26. [34]

    Pyannote.audio: Neural building blocks for speaker diarization,

    H. Bredin, R. Yin, J. M. Coria, G. Gelly, P. Korshunov, M. Lavechin, D. Fustes, H. Titeux, W. Bouaziz, and M.-P. Gill, “Pyannote.audio: Neural building blocks for speaker diarization,” in ICASSP 2020 - 2020 IEEE International Conference on Acous- tics, Speech and Signal Proces...

    2020

  27. [35]

    Decoupled weight de- cay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight de- cay regularization,” in International Conference on Learn- ing Representations , 2019. [Online]. Available: https: //openreview.net/forum?id=Bkg6RiCqY7

    2019

  28. [36]

    Xls-r: Self-supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Con- neau, and M. Auli, “Xls-r: Self-supervised cross-lingual speech representation learning at scale,” in Interspeech 2022, 2022, pp. 2278–2282

    2022

  29. [37]

    On the sdes and scaling rules for adaptive gradient algorithms,

    S. Malladi, K. Lyu, A. Panigrahi, and S. Arora, “On the sdes and scaling rules for adaptive gradient algorithms,” in Advances in Neural Information Processing Systems , S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Associates, Inc., 2...

    2022

Pith tools

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