Pith. sign in

REVIEW 3 major objections 6 minor 35 references

MSDA: Combining Pseudo-labeling and Self-Supervision for Unsupervised Domain Adaptation in ASR

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 Pseudo-labels after self-supervision cut Greek ASR domain-shift WER.

desk verdict Decent empirical ASR adaptation recipe with a genuinely useful cascade finding, but the Meta PL implementation is under-specified and the SOTA claim is qualified by a modified CASTLE baseline. read the letter →

arxiv 2505.24656 v2 pith:PD36CAWB submitted 2025-05-30 cs.CL cs.SDeess.AS

Dimitrios Damianos , Georgios Paraskevopoulos , Alexandros Potamianos This is my paper
classification cs.CLcs.SDeess.AS
keywords unsuperviseddomainadaptationautomaticspeechrecognitionMetaPseudoLabelspseudo-labelingself-supervisionlow-resourcelanguagesweaklysuperviseddataGreek
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 Meta Pseudo Labels teacher–student feedback mechanism, previously used for image recognition, can be brought to automatic speech recognition, and that it works best when it runs as the second stage of a cascade rather than being merged into a single objective. The first stage adapts a pre-trained Wav2Vec 2.0 model to both source and target domains with mixed contrastive self-supervision; the second stage lets the resulting teacher produce CTC pseudo-labels for unlabeled target speech while a student trains on those labels and its own source-domain CTC loss feeds back into the teacher. If this works as claimed, it gives a sample-efficient recipe for unsupervised domain adaptation that needs no transcriptions in the target domain, which matters for low-resource languages and for weakly supervised or noisy label sources. On Greek broadcast, crowd-sourced, parliamentary, and podcast data, the cascade reports lower word error rates than fine-tuning, continual pre-training, M2DS2, M2DS2-Meta PL, and the CASTLE baseline in most adaptation directions.

What carries the argument

The load-bearing mechanism is the two-stage teacher–student cascade built on XLSR-53/Wav2Vec 2.0. In Stage 1, the M2DS2 objective combines source CTC loss with contrastive self-supervised losses on source and target speech, producing a teacher already attuned to the target domain. In Stage 2, that teacher emits CTC pseudo-labels for unlabeled target audio; the student trains on those pseudo-labels, and the student's CTC loss on the source domain is added to the teacher's objective as feedback, together with a target-domain diversity loss. The design choices that carry the argument are the strict sequencing, the restriction of SpecAugment to the student in Stage 2, and the small coefficients $\gamma=\delta$ in Eq. (2) that keep source knowledge and output diversity from collapsing.

What would settle it

Compare the Stage 2 teacher update with the student's source-domain feedback loss removed, keeping the diversity loss and all other settings identical: if WER on the target test set is unchanged, the Meta PL feedback loop is not the source of the reported gains. Conversely, an implementation that truly unrolls the student's training steps to meta-optimize the teacher should beat the simple joint objective if the paper's attribution to Meta PL is correct.

Watch

Extended reading notes

Core claim

The central discovery is that the order of operations is load-bearing: first adapt with M2DS2 self-supervision on both domains, then apply Meta PL pseudo-labeling, and the combination outperforms either method alone. Stage 1 trains a teacher with $\mathcal{L}_{M2DS2} = \mathcal{L}_{CTC}(x_s,y_s)+\alpha \mathcal{L}_s(x_s)+\beta \mathcal{L}_s(x_t)$, keeping source CTC alignment while learning contrastive speech representations from both domains. Stage 2 trains a student on teacher-produced CTC pseudo-labels for the target domain while the teacher is refined with $\mathcal{L} = \mathcal{L}_{feedback}(x_s,y_s)+\gamma \mathcal{L}_{CTC}(x_s,y_s)+\delta \mathcal{L}_d(x_t)$, where $\mathcal{L}_{feedback}$ is the student's source-domain CTC loss, $\mathcal{L}_d$ is a diversity loss preventing mode collapse, and SpecAugment is applied only to the student's inputs. The paper argues that this cascade yields the reported WER reductions and that non-cascading integrations of self-supervision with Meta PL fail.

Load-bearing premise

The approach assumes that the Stage 2 objective, adding the student's source-domain CTC loss to the teacher's own loss, faithfully implements Meta Pseudo Labels' feedback mechanism; if that implementation is instead only a joint multi-task loss, the claim that Meta PL transfers to ASR is not supported.

Editorial extensions

If this is right

  • Meta Pseudo Labels, originally designed for images, becomes a viable unsupervised domain adaptation method for CTC-based ASR models.
  • A strictly ordered cascade (self-supervision before pseudo-labeling) should be preferred over joint multi-task combinations, because the paper's non-cascading variants fail.
  • Unsupervised adaptation with this recipe remains effective with as little as 10% of the target training data, making it practical when target transcription is scarce.
  • Weakly supervised target data, such as automatically transcribed podcast audio, can serve as a usable target domain, with MSDA improving over baselines in most GPC adaptation directions.

Reading between the lines

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

  • Inference: if the feedback-loss implementation is a valid Meta PL surrogate, the same two-stage scaffold should transfer to other self-supervised speech encoders beyond XLSR-53, though the paper only tests one backbone.
  • Inference: the diversity loss and the SpecAugment scheduling may account for a substantial share of the gains; an ablation with the feedback term removed but those components retained would separate the Meta PL contribution from the regularization contribution.
  • Inference: coupling MSDA with language-model beam-search pseudo-label refinement, as CASTLE does, could push WER further, since the paper compares against a greedy-decoding CASTLE variant.
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 proposes MSDA, a two-stage unsupervised domain adaptation method for ASR built on XLSR-53. Stage 1 uses the authors' M2DS2 objective (Eq. 1), combining CTC loss on the source domain with contrastive self-supervised losses on source and target speech. Stage 2 uses a teacher–student pseudo-labeling scheme whose teacher objective is given in Eq. (2) as L_feedback + gamma L_CTC + delta L_d, and is claimed to realize Meta Pseudo Labels. The method is evaluated on Greek LG/CV/HP datasets and on weakly supervised GPC categories, comparing with FT, CPT, M2DS2, FT-MP, M2DS2-MP, and CASTLE. The paper reports WER improvements across most scenarios, an ablation of gamma and delta, an experiment on non-cascading integration alternatives, and a sample-efficiency study against M2DS2 and CASTLE.

Significance. If the implementation indeed matches Meta PL's feedback mechanism, the paper would provide a useful transfer of a semi-supervised technique to low-resource ASR and demonstrate that a cascaded self-supervised plus pseudo-labeling pipeline is effective. Strengths include evaluation across six domain-pair scenarios plus weakly supervised domains, explicit hyperparameter analysis for gamma and delta, and a sample-efficiency experiment against M2DS2 and CASTLE. The main caveats are that the central attribution to Meta PL is not established by the current description of Stage 2, and the CASTLE comparison is weakened by modifying the baseline; these issues affect the interpretation of the empirical claims.

major comments (3)
  1. [§2, Eq. (2)] The description of Stage 2 does not specify an implementation of Meta Pseudo Labels. In Meta PL [11], the teacher is updated using a meta-gradient that depends on the student's training dynamics on pseudo-labeled target data and the student's loss on labeled source data; the paper instead presents L_feedback as an additive term in the teacher's loss and does not state whether the student's parameters are updated in an inner loop, whether gradients are unrolled through student updates, or how L_feedback is computed. As written, Eq. (2) could be a joint multi-task objective rather than Meta PL, and the paper's claim to have transferred Meta PL to ASR is unsupported. Please provide a complete algorithm or pseudocode for Stage 2, including student and teacher parameter updates, stop-gradient behavior, and the exact computation of L_feedback; this is load-bearing because every table and figure is interpreted through that claim.
  2. [§3.3–3.4, Table 1] The CASTLE baseline is modified by removing its language model and beam search in favor of greedy decoding from the acoustic model, as stated in Section 3.4. The reported "outperforms CASTLE" claim therefore compares against a weakened version of CASTLE rather than the published state-of-the-art method. Please either report CASTLE with its original LM-assisted beam-search pseudo-labeling, or clearly state that the comparison is against a reduced CASTLE variant and qualify the state-of-the-art claim accordingly.
  3. [§5, Table 3] The ablation claims that cascading is necessary, but the alternatives in Eqs. (3) and (4) are not faithful Meta PL implementations either, because they omit the unrolled feedback loop that defines Meta PL. Table 3 therefore demonstrates only that these particular additive objectives fail, not that non-cascading integration of Meta PL with self-supervision fails. The conclusion in Section 6 that cascading methods "show promising results" is stronger than the evidence presented.
minor comments (6)
  1. [Abstract and Section 1] The phrase "achieving state-of-the-art results, significantly outperforming state-of-the-art methods" is redundant and overreaches given the modified CASTLE baseline and the absence of significance tests; please qualify the claim.
  2. [Section 4, Tables 1 and 2] The tables report WER without confidence intervals or significance tests; several improvements over baselines are small (for example, LG→HP: 60.98 vs 61.98), so error bars or repeated-run variance would strengthen the conclusions.
  3. [Section 4, text] The wording "WER reductions of 1–4%" should be clarified as absolute percentage-point reductions or relative reductions; the current phrasing is ambiguous.
  4. [Figure 2 caption] The caption says "different setting of gamma" and "delta" but does not state which coefficient is held fixed or the exact values tested; please clarify the experimental setup and ensure the figure is legible.
  5. [Figure 3 caption] The caption does not identify the curves or the x-axis units beyond "50%, 25%, and 10% of the original dataset"; please state which method each line represents and the exact WER values.
  6. [Section 3.2] The line "Logotypografia: Logotypografia (LG)" appears to contain a formatting artifact; please use the corpus name consistently throughout.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: MSDA's central claim is grounded in controlled experiments against independent baselines; the reuse of the authors' M2DS2 method as Stage 1 is a minor self-citation, while Eq. (2)'s fidelity to Meta PL is an implementation-validity concern, not a circular reduction.

full rationale

The paper's derivation chain does not contain a circular step. Stage 1 reuses M2DS2 from the authors' prior work [19], but this is not circular: Eq. (1) is explicitly stated in the paper, and M2DS2 is independently evaluated as a baseline in the same experimental setup, so the central claim about the cascade does not reduce to the citation itself. No fitted parameter is renamed as a prediction, and no equation is defined in terms of the quantity it is claimed to predict. The main vulnerability is whether Eq. (2) faithfully realizes Meta Pseudo Labels as defined in [11], since the original method uses an unrolled meta-gradient through the student while the paper specifies only an additive L_feedback term in the teacher objective. That is a fidelity and correctness concern, not circularity, because even if Eq. (2) is a different objective, the reported WER comparisons stand on their own held-out target test sets against FT, CPT, CASTLE, and other baselines. The self-citation to [19] is a component choice, but the paper's results are externally grounded in independent evaluation, so the circularity score is low.

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

The central claim rests on the fidelity of the M2DS2 stage and the Meta PL stage; both are taken from prior work, and the boundary between them is the paper's main design choice. The free parameters (alpha, beta, gamma, delta) are manually tuned and shown to be influential. No new entities are introduced.

free parameters (4)
  • alpha = 0.01
    Mixing weight for source self-supervised loss in Eq. (1), chosen manually.
  • beta = 0.02
    Mixing weight for target self-supervised loss in Eq. (1), chosen manually.
  • gamma = 0.0001 (clean) / 0.001 (weakly-supervised)
    Teacher's CTC loss coefficient in Eq. (2); Fig. 2 shows sharp sensitivity to this value.
  • delta = 0.0001 (clean) / 0.001 (weakly-supervised)
    Diversity loss coefficient in Eq. (2); Fig. 2 shows sensitivity.
assumptions (4)
  • domain assumption Multi-domain self-supervision as in M2DS2 [19] prevents mode collapse and yields a good teacher
    Inherited from the authors' prior work [19]; the paper does not re-derive or validate this claim independently.
  • domain assumption WhisperX-generated transcripts are accurate enough to serve as labels for weakly supervised training and as reference for WER evaluation
    Section 3.2 uses GPC with WhisperX transcripts, and Table 2 reports WER against these transcripts as ground truth.
  • domain assumption Teacher-student pseudo-labeling with confidence-free greedy decoding produces useful training signal
    The paper does not use a confidence filter (contrary to CEM), yet assumes pseudo-labels from XLSR teacher are useful.
  • ad hoc to paper The simplified teacher loss in Eq. (2) approximates Meta Pseudo Labels
    No meta-gradient or unrolled optimization is described; if the original Meta PL requires such machinery, this assumption is load-bearing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MSDA: Combining Pseudo-labeling and Self-Supervision for Unsupervised Domain Adaptation in ASR." pith.science (2026). https://pith.science/paper/PD36CAWB

@misc{pith2026250524656,
  author       = {Pith},
  title        = {Pith review of: MSDA: Combining Pseudo-labeling and Self-Supervision for Unsupervised Domain Adaptation in ASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PD36CAWB}},
  note         = {Machine review of arXiv:2505.24656}
}
read the original abstract

In this work, we investigate the Meta PL unsupervised domain adaptation framework for Automatic Speech Recognition (ASR). We introduce a Multi-Stage Domain Adaptation pipeline (MSDA), a sample-efficient, two-stage adaptation approach that integrates self-supervised learning with semi-supervised techniques. MSDA is designed to enhance the robustness and generalization of ASR models, making them more adaptable to diverse conditions. It is particularly effective for low-resource languages like Greek and in weakly supervised scenarios where labeled data is scarce or noisy. Through extensive experiments, we demonstrate that Meta PL can be applied effectively to ASR tasks, achieving state-of-the-art results, significantly outperforming state-of-the-art methods, and providing more robust solutions for unsupervised domain adaptation in ASR. Our ablations highlight the necessity of utilizing a cascading approach when combining self-supervision with self-training.

Figures

Figures reproduced from arXiv: 2505.24656 by the authors.

Figure 1
Figure 1. The Multi-Stage Domain Adaptation (MSDA). At Stage 1, we use self-supervision to adapt our pre-trained model to both source (xs, ys) and target (xt) domain, using the loss objective described in Eq. (1). At Stage 2, we apply an enhanced version of Meta PL to further improve the overall adaptation. In this stage, the teacher provides pseudo-labels y ∗ t for student training, and itself is trained using the objective … view at source ↗
Figure 2
Figure 2. WER comparison between different setting of γ (blue) and δ (red), FT (yellow) and M2DS2 (green) baselines , evalu￾ated in the TC → Ed adaptation setting. adapt, such as Ed→TC [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison between M2DS2,CASTLE and MSDA for the LG → CV setting, when reducing the amount of available target samples to 50%, 25%, and 10% of the origi￾nal dataset. FT performance is indicated with the red line. Finally, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages

  1. [11]

    Domain adaptation of dnn acoustic models using knowledge distillation,

    T. Asami et al., “Domain adaptation of dnn acoustic models using knowledge distillation,” inProc. ICASSP. IEEE, 2017, pp. 5185– 5189

    work page 2017

  2. [1]

    MSDA: Combining Pseudo-labeling and Self-Supervision for Unsupervised Domain Adaptation in ASR

    Introduction Automatic Speech Recognition (ASR) has transformed human- machine interaction, driving technologies such as virtual assis- tants, automated transcription services, and voice-controlled de- vices. However, ASR systems often face challenges in main- taining accuracy when exposed to data that differ from their training environment. Factors such ...

    work page Pith review arXiv 2025

  3. [2]

    1 illustrates our proposed approach, which builds on and extends the methodologies presented in [19] and [11]

    Methodology Fig. 1 illustrates our proposed approach, which builds on and extends the methodologies presented in [19] and [11]. This approach, termed Multi-Stage Domain Adaptation (MSDA), consists of two sequential adaptation stages: an initial self- supervised stage followed by a semi-supervised stage. Stage 1: In this stage, we follow the method propose...

    work page

  4. [3]

    Pre-trained model: For our base model, we utilize XLSR-53 [23] , a state-of-the-art pre-trained speech model developed on the Wav2Vec 2.0 [20] architecture

    Experimental Setup 3.1. Pre-trained model: For our base model, we utilize XLSR-53 [23] , a state-of-the-art pre-trained speech model developed on the Wav2Vec 2.0 [20] architecture. XLSR-53 stands out for its extensive training on a diverse and extensive corpus, comprising 56,000 hours of speech data across 53 languages. This diverse training set equips XL...

    work page 2018

  5. [4]

    Results In Table 1, we compare the performance of MSDA with se- lected baselines across six adaptation scenarios involving the LG, CV and HP datasets. Meta PL consistently exhibits strong adaptation capabilities, with FT-MP and M2DS2-MP achieving notable WER reductions of 1–4% and 1–2%, respectively, com- pared to the FT and M2DS2 baselines. These results...

    work page

  6. [5]

    Teacher” col- umn contains teacher’s original WER on target domain, while the “Student

    Meta PL & Self-Supervision We explored different ways to integrate self-supervised learning with the Meta PL framework in a non-cascading manner. We experimented with combining self-supervised objectives with student feedback to improve teacher performance, as described in Eq. (3). LT = Lf eedback(xs, ys) + Ls(xs) + Ls(xt) (3) In a second approach, we app...

    work page

  7. [6]

    We found that Meta PL is an ef- fective adaptation method, providing a straightforward and eas- ily implementable solution

    Conclusions & Future Work In this work, we explored the application of the Meta PL frame- work [11] in ASR tasks and investigated its integration with self-supervision objectives. We found that Meta PL is an ef- fective adaptation method, providing a straightforward and eas- ily implementable solution. However, our proposed approach, Multi-Stage Domain Ad...

    work page

  8. [7]

    Competitiveness

    Acknowledgements This work is co-funded by the European Union’s Digital Eu- rope Programme and by the European Union through the Na- tional Strategic Reference Framework (NSRF) 2021–2027, un- der the Operational Programme “Competitiveness” Grant No 101083565 and Grant No 6002637 respectively, for the project SmartAttica-AtHeNAI DIH (the Attica region - Gr...

    work page 2021

Show all 35 references
  1. [8]

    Learning transferable features with deep adapta- tion networks,

    M. Long et al., “Learning transferable features with deep adapta- tion networks,” in Proc. ICML. PMLR, 2015, pp. 97–105

    2015

  2. [9]

    Domain-adversarial training of neural networks,

    Y . Ganin et al., “Domain-adversarial training of neural networks,” Journal of machine learning research , vol. 17, no. 59, pp. 1–35, 2016

    2016

  3. [10]

    Unsupervised domain adaptation schemes for building asr in low-resource languages,

    C. S. Anoop et al., “Unsupervised domain adaptation schemes for building asr in low-resource languages,” in Proc. ASRU. IEEE, 2021, pp. 342–349

    2021

  4. [12]

    Probability of error of some adaptive pattern- recognition machines,

    H. Scudder, “Probability of error of some adaptive pattern- recognition machines,” IEEE Transactions on Information The- ory, vol. 11, no. 3, pp. 363–371, 1965

    1965

  5. [13]

    Learning extraction patterns for subjec- tive expressions,

    E. Riloff and J. Wiebe, “Learning extraction patterns for subjec- tive expressions,” in Proc. EMNLP, 2003, pp. 105–112

    2003

  6. [14]

    Large-scale asr domain adaptation using self- and semi-supervised learning,

    D. Hwang et al., “Large-scale asr domain adaptation using self- and semi-supervised learning,” in Proc. ICASSP. IEEE, 2022, pp. 6627–6631

    2022

  7. [15]

    Pseudo label is better than human label,

    D. Hwang, K. C. Sim, Z. Huo, and T. Strohman, “Pseudo label is better than human label,” in Proc. Interspeech. ISCA, 2022, pp. 1421–1425

    2022

  8. [16]

    Improved noisy student training for automatic speech recognition,

    D. S. Park et al., “Improved noisy student training for automatic speech recognition,” in Proc. Interspeech . ISCA, 2020, pp. 2817–2821

    2020

  9. [17]

    Kaizen: Continuously improving teacher using exponential moving average for semi-supervised speech recognition,

    V . Manohar et al., “Kaizen: Continuously improving teacher using exponential moving average for semi-supervised speech recognition,” in Proc. ASRU. IEEE, 2021, pp. 518–525

    2021

  10. [18]

    Meta pseudo labels,

    H. Pham et al., “Meta pseudo labels,” in Proc. CVPR. IEEE/CVF, 2021, pp. 11 557–11 568

    2021

  11. [19]

    Progressive unsupervised domain adaptation for asr using ensemble models and multi-stage training,

    R. Ahmad, M. U. Farooq, and T. Hain, “Progressive unsupervised domain adaptation for asr using ensemble models and multi-stage training,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 11 466–11 470

    2024

  12. [20]

    knn-ctc: Enhancing asr via retrieval of ctc pseudo labels,

    J. Zhou, S. Zhao, Y . Liu, W. Zeng, Y . Chen, and Y . Qin, “knn-ctc: Enhancing asr via retrieval of ctc pseudo labels,” inICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 006–11 010

    2024

  13. [21]

    Udalm: Unsupervised domain adaptation through language modeling,

    C. Karouzos et al., “Udalm: Unsupervised domain adaptation through language modeling,” in Proc. NAACL: HLT , 2021, pp. 2579–2590

    2021

  14. [22]

    Don’t stop pretraining: Adapt language models to domains and tasks,

    S. Gururangan et al., “Don’t stop pretraining: Adapt language models to domains and tasks,” in Proc. 58th Annual Meeting of ACL, 2020, pp. 8342–8360

    2020

  15. [23]

    Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training,

    W. Hsu et al, “Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training,” in Proc. Interspeech. ISCA, 2021, pp. 721–725

    2021

  16. [24]

    Boosting cross-domain speech recognition with self-supervision,

    H. Zhu et al, “Boosting cross-domain speech recognition with self-supervision,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 471–485, 2024

    2024

  17. [25]

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

    J. D. Kenton et al., “Bert: Pre-training of deep bidirectional trans- formers for language understanding,” in Proceedings of NAACL- HLT, vol. 1, no. 2, 2019

    2019

  18. [26]

    Sample-efficient unsupervised domain adaptation of speech recognition systems: A case study for mod- ern greek,

    G. Paraskevopoulos et al., “Sample-efficient unsupervised domain adaptation of speech recognition systems: A case study for mod- ern greek,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 32, pp. 286–299, 2024

    2024

  19. [27]

    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 Proc. NeurIps, vol. 33, 2020, pp. 12 449–12 460

    2020

  20. [28]

    Towards end-to-end speech recognition with recurrent neural networks,

    A. Graves and N. Jaitly, “Towards end-to-end speech recognition with recurrent neural networks,” in Proc. ICML. PMLR, 2014, pp. 1764–1772

    2014

  21. [29]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    D. Park et al., “Specaugment: A simple data augmentation method for automatic speech recognition,” in Proc. Interspeech. ISCA, 2019, pp. 2613–2617

    2019

  22. [30]

    Unsupervised cross-lingual representation learning for speech recognition,

    A. Conneau et al., “Unsupervised cross-lingual representation learning for speech recognition,” in Proc. Interspeech. ISCA, 2021, pp. 2426–2430

    2021

  23. [31]

    Large vocabulary continuous speech recogni- tion in greek: corpus and an automatic dictation system,

    V . Digilakis et al, “Large vocabulary continuous speech recogni- tion in greek: corpus and an automatic dictation system,” in Proc. Eurospeech, 2003, pp. 1565–1568

    2003

  24. [32]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila et al, “Common voice: A massively-multilingual speech corpus,” in Proc. LREC, 2020, pp. 4218–4222

    2020

  25. [33]

    The greek podcast corpus: Competi- tive speech models for low-resourced languages with weakly su- pervised data,

    G. Paraskevopoulos et al., “The greek podcast corpus: Competi- tive speech models for low-resourced languages with weakly su- pervised data,” in Proc. Interspeech. ISCA, 2024, pp. 3969– 3973

    2024

  26. [34]

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

    M. Bain et al., “Whisperx: Time-accurate speech transcription of long-form audio,” in Proc. Interspeech. ISCA, 2023, pp. 4489– 4493

    2023

  27. [35]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” arXiv e-prints, pp. arXiv–1711, 2017

    2017

Pith tools

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