Pith. sign in

REVIEW 3 major objections 4 minor 34 references

The paper claims that extending consistency regularization to every step of an iterative non-autoregressive ASR model—the CTC base and each refinement pass—improves accuracy additively, and that the same loss enables self-training on unlabe

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 20:25 UTC pith:QYUGG377

load-bearing objection Supervised results are credible and the additivity claim holds, but the same semi-supervised configuration appears with two different dev WERs (5.5/12.6 in Tables 2/3 vs 4.1/9.5 in Table 4), so the self-training numbers are not internally reproducible as printed. the 3 major comments →

arxiv 2602.23171 v2 pith:QYUGG377 submitted 2026-02-26 eess.AS

Align-Consistency: Improving Non-autoregressive and Semi-supervised ASR with Consistency Regularization

Wanting Huang , Weiran Wang This is my paper
classification eess.AS
keywords consistency regularizationnon-autoregressive speech recognitioniterative refinementCTCsemi-supervised learningself-trainingpseudo-labelingAlign-Refine
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes Align-Consistency, a training objective that applies consistency regularization to Align-Refine, a non-autoregressive speech recognizer that iteratively refines frame-level alignments. The idea is to make every internal hypothesis—the initial CTC output and each refinement step—stable under random input perturbations, by encouraging two augmented views of the same audio to produce the same frame-level posteriors. The authors show that this stability training and the accuracy gained from iterative refinement are mutually additive: on LibriSpeech, adding both brings test WERs from 12.2/26.7 to 10.0/22.9 on 100 hours and from 4.3/9.9 to 3.3/7.4 on 960 hours. They then use the same objective for semi-supervised self-training, generating pseudo-labels on the fly with the model's own final refinement step, and report large gains—e.g., LS-100 test WER falling from 10.0/22.9 to 4.3/9.6 with unlabeled 960h, and to 3.8/9.1 with an additional 6000h. A sympathetic reader would care because the method shows a relatively simple regularization technique can make fast parallel decoding competitive with much larger semi-supervised systems.

Core claim

Align-Consistency is a learning objective that applies the symmetric KL consistency loss of CR-CTC not only to the base CTC model (step 0) but also to every refinement step of Align-Refine, enforcing agreement between frame-level posteriors from two differently perturbed views of the input. The paper finds that applying CR to both the base CTC and the refinement steps is critical: with loss weights (0.2, 0.2), dev WERs improve substantially over either alone, and the benefit of iterative refinement and consistency are mutually additive. In self-training, the fresh model's final-step greedy hypotheses are used as pseudo-labels for unlabeled utterances, and the same Align-Consistency loss is c

What carries the argument

The central object is the Align-Consistency loss LAC, which combines the Align-Refine non-autoregressive loss (a weighted sum of CTC log-likelihoods over the base and S refinement steps) with a symmetric KL divergence between frame-level posteriors of two input-perturbed views, with one side detached from gradients, applied at step 0 (CTC) and each refinement step s=1..S. Because the consistency term does not require ground-truth labels, the same objective transfers directly to unlabeled data, where it is computed with pseudo-labels from the model's own final refinement step.

Load-bearing premise

The semi-supervised results depend on the current model's final-step greedy hypotheses being accurate enough as pseudo-labels that self-training improves rather than reinforces systematic errors, because the method applies no confidence filtering or selection criterion beyond using the last refinement step.

What would settle it

Run the same self-training pipeline but threshold pseudo-labels by model confidence (e.g., average posterior or agreement between the two augmented views); if filtered labels yield no better or worse WER than unfiltered final-step labels, the paper's premise that raw final-step pseudo-labels are sufficient is undercut.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Non-autoregressive ASR need not sacrifice accuracy for speed: a parallel decoder with iterative refinement, trained with consistency at every step, becomes a strong standalone recognizer.
  • The additivity result implies consistency regularization should be applied to all internal hypotheses the model must refine, not just the base CTC module, offering a design rule for other iterative non-AR models.
  • Because the same unsupervised loss works with self-generated final-step pseudo-labels, large unlabeled corpora can be exploited without beam search or language-model fusion, preserving fast inference during pseudo-labeling.
  • The reported data efficiency—competitive with systems trained on an order of magnitude more unlabeled data—suggests consistency regularization may be a cheaper lever than scaling data.
  • Removing CR on unlabeled data hurts self-training, indicating consistency is beneficial even when pseudo-labels are noisy, not just as a supervised-training trick.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • One testable extension: pseudo-label selection could be strengthened by confidence filtering or by agreement between the two augmented views; the paper only compares final-step labels to CTC labels, so it leaves open whether filtering would push WERs lower.
  • The per-step consistency objective could plausibly transfer to other latent-alignment non-AR models, and to diffusion-based non-AR decoders whose denoising steps play the role of refinement steps—an extension the paper itself gestures toward.
  • If the gains come from stability of frame-level posteriors, one might expect CR to act as an inductive bias toward confident, peaked distributions; a separate analysis could measure whether the improvement stems from sharper posteriors or from reduced variance across augmentations.
  • Ablating which module receives the gradient when both consistency terms are active could reveal whether the observed additivity is genuinely independent or partly redundant.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes Align-Consistency, a consistency-regularization (CR) training objective for Align-Refine, a non-autoregressive CTC-based ASR model that iteratively refines frame-level hypotheses. The loss combines the standard Align-Refine objective with symmetric KL consistency between two SpecAugment views, applied at both the base CTC level and each refinement step. The authors evaluate on LibriSpeech. In the fully supervised setting, they report that CR and iterative refinement each independently reduce dev WER and that combining them yields the best results. In the semi-supervised setting, they generate online pseudo-labels from the final refinement step and apply the same Align-Consistency loss to unlabeled data, reporting substantial WER reductions on LS-100+LS-960, LS-100+(LS-960+LL-6000), and LS-960+LL-6000. They also compare against CR-CTC and published semi-supervised ASR baselines.

Significance. If the supervised results are reliable, this is a useful and clean extension of CR from pure CTC to a non-autoregressive refinement architecture, and the claim that non-AR decoding and CR are complementary is interesting. The semi-supervised protocol is simple and efficient: pseudo-labels are generated on the fly from the model's final-step greedy hypotheses, without external language models or confidence filtering. The paper is benchmark-oriented, comparing against external CR-CTC baselines and published systems, and no derivation is fitted to the target WERs, so circularity is low. The main caveats are the complete absence of variance estimates and an internal inconsistency between Tables 2 and 4 that directly affects the reported semi-supervised gains. These issues must be resolved before the quantitative claims can be accepted.

major comments (3)
  1. [Table 2 vs. Table 4 (Sec. 4.4)] The same semi-supervised configuration is reported with materially different dev WERs. For LS-100 labeled + LS-960 unlabeled, Align-Consistency pseudo-labels, (λ0, λ1) = (0.2, 0.2), and s=2 decoding, Table 2 gives clean/other 5.5/12.6, while Table 4 gives 4.1/9.5. Section 4.4 describes one protocol (110 epochs, final-step greedy labels, same loss weights), so no procedural difference is stated. Table 3 repeats 5.5/12.6, whereas Table 5's test result 4.3/9.6 for the same setup is close to Table 4, not Table 2. At least one of these tables is wrong or corresponds to a different configuration. This is a load-bearing internal inconsistency in the central quantitative claim and must be corrected or explained.
  2. [Table 1 / Sec. 4.3] All reported results are from single runs, with no variance or significance information. Several load-bearing comparisons are small: e.g., on LS-960 s=2, (0.2, 0.2) gives 2.9/6.9 versus 3.2/7.6 without CR, and CR on refinement alone gives 3.1/7.2. Without multiple seeds or confidence intervals, it is hard to establish that the apparent additive improvements are beyond training noise. Please report means and standard deviations over at least three seeds, or otherwise justify the stability of these differences.
  3. [Sec. 3.2, Sec. 4.4.1] Semi-supervised self-training uses the current model's final-step greedy hypothesis as pseudo-labels with no confidence filtering or selection. The ablation in Table 2 compares final-step labels against CTC labels, but does not compare against filtered pseudo-labels or a confidence threshold, despite related work (e.g., Refs. [18]–[21]) showing selection and agreement criteria matter. This matters because self-training can reinforce systematic errors, and the paper's claim of 'robustness of CR to noisy supervision' would be better supported by a direct noise-level analysis or comparison with filtering. At minimum, the authors should discuss why filtering is not needed or is expected to be subsumed.
minor comments (4)
  1. [Table 2] In the last row, the entry '13.35.5' is missing a separating space; it should read '13.3 5.5'.
  2. [Sec. 1] The first sentence of the introduction says 'End-to-end (E2E) ASR can be roughly divided into categories'; 'two' appears to be missing before 'categories'.
  3. [Sec. 4.2] 'a2×time reduction' should be 'a 2× time reduction'.
  4. [Sec. 4.3 / 4.4] It is not stated whether the hyperparameters (λ0, λ1) were tuned separately for each dataset and for the supervised vs. semi-supervised settings, or fixed across all experiments. Please clarify the tuning protocol.

Circularity Check

0 steps flagged

No significant circularity: the paper's central claims are empirical benchmark results on held-out LibriSpeech data, not derivations that reduce to fitted inputs or self-citations.

full rationale

The paper proposes Align-Consistency by combining Align-Refine with consistency regularization, and evaluates it on external LibriSpeech dev/test sets. The supervised and semi-supervised WERs are directly measured outcomes, not quantities constructed from the method's own definitions. The loss LAC in Eq. (7) is a combination of a standard CTC/NAR likelihood and a symmetric KL consistency term; it does not contain, by construction, the test/dev WERs it purports to predict. Hyperparameters λ0 and λ1 are tuned on dev sets and reported on dev and test sets; this is standard practice and does not make the reported results equivalent to the tuning objective. The semi-supervised procedure follows the standard self-training framework of [9], using the model's own pseudo-labels; although this is iterative and self-referential by design, the paper's contribution is the empirical demonstration that these labels improve external WER compared to CTC pseudo-labels, and the comparison is made against external baselines such as CR-CTC. The self-citations [7] and [9] share an author but are used as framing/baseline methodology, not as unverified proofs or uniqueness theorems, and the load-bearing evidence is the paper's own external evaluations. The internal inconsistency between Table 2 and Table 4 for the same semi-supervised configuration is a correctness/reproducibility concern, not a circularity step, because it does not show that any prediction is equivalent to an input by construction. Overall, the derivation chain is self-contained against external benchmarks, so no circularity is found.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard CTC math plus domain assumptions about augmentation invariance, pseudo-label quality, and teacher-forcing. Free hyperparameters λ0, λ1, α, S, and γ are hand-chosen and affect results. No new model entities are introduced.

free parameters (5)
  • λ0 = 0.2
    CR weight for the base CTC model; tuned over {0.1, 0.2, 0.3} on dev sets, 0.2 best in Table 1.
  • λ1 = 0.2
    CR weight for refinement steps; tuned jointly with λ0, 0.2 best in Table 1.
  • α = 0.3
    CTC loss weight in the non-AR loss Eq (4); fixed by hand, not tuned, affects the balance between CTC and refinement losses.
  • S = 2
    Number of refinement steps; fixed from an unreported preliminary study, said to capture most of the WER gain.
  • γ = 1
    Weight on the unsupervised loss in Sec 3.2; adopted from [9].
axioms (4)
  • standard math CTC frame-wise factorization: p(s)(y|x) = Σ_{a∈B^{-1}(y)} p(s)(a|...)
    Invoked after Eq (3); standard CTC marginalization over alignments consistent with the label.
  • domain assumption SpecAugment perturbations preserve the target transcript
    CR enforces agreement of posteriors under x̃1 and x̃2; if the augmentations changed the alignment, the consistency objective would be biased.
  • domain assumption Unlabeled LibriSpeech/LibriLight is in-domain and final-step greedy pseudo-labels are adequate supervision
    The semi-supervised loss in Sec 3.2 treats ŷ as ground truth without confidence filtering.
  • domain assumption Teacher-forcing with the previous greedy hypothesis is a valid training signal for iterative refinement
    Eq (4) maximizes likelihood of the current refinement step given the argmax of the previous step, with no exposure-bias correction.

pith-pipeline@v1.3.0-alltime-deepseek · 8654 in / 15273 out tokens · 128479 ms · 2026-08-02T20:25:59.408606+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Align-Consistency: Improving Non-autoregressive and Semi-supervised ASR with Consistency Regularization." pith.science (2026). https://pith.science/paper/QYUGG377

@misc{pith2026260223171,
  author       = {Pith},
  title        = {Pith review of: Align-Consistency: Improving Non-autoregressive and Semi-supervised ASR with Consistency Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QYUGG377}},
  note         = {Machine review of arXiv:2602.23171}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Consistency regularization (CR) improves the robustness and accuracy of Connectionist Temporal Classification (CTC) by ensuring predictions remain stable across input perturbations. In this work, we propose Align-Consistency, an extension of CR designed for Align-Refine -- a non-autoregressive (non-AR) model that performs iterative refinement of frame-level hypotheses. This method leverages the speed of parallel inference while significantly boosting recognition performance. The effectiveness of Align-Consistency is demonstrated in two settings. First, in the fully supervised setting, our results indicate that applying CR to both the base CTC model and the subsequent refinement steps is critical, and the accuracy improvements from non-AR decoding and CR are mutually additive. Second, for semi-supervised ASR, we employ fast non-AR decoding to generate online pseudo-labels on unlabeled data, which are used to further refine the supervised model and lead to substantial gains.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 5 linked inside Pith

  1. [1]

    Due to the language modeling component (the decoder) in autoregressive models, decoding is performed in a left-to-right fashion by beam search

    Introduction End-to-end (E2E) automatic speech recognition (ASR) can be roughly divided into categories: autoregressive models such as attention-based encoder-decoder [1, 2] and RNN- Transducer [3], and non-autoregressive (non-AR) models, with a prominent example being Connectionist Temporal Classifica- tion (CTC, [4]). Due to the language modeling compon...

  2. [2]

    Related Work 2.1. Non-autoregressive decoding and iterative refinement Non-autoregressive (non-AR) decoding avoids strictly left-to- right generation and enables parallel updates over multiple or all output positions. A common strategy is iterative refine- ment, where the model revises an initial hypothesis over sev- eral steps to gradually incorporate la...

    Pith/arXiv arXiv 2026

  3. [3]

    Method We propose a new learning objective, named Align- Consistency, which integrates alignment-based iterative refine- ment (Align-Refine) with consistency regularization for end-to- end ASR, and then extend its use to semi-supervised learning. 3.1. Align-Consistency: Align-Refine with CR Let ˜x1 and ˜x2 be two augmented versions of the clean input, ob-...

  4. [4]

    Datasets We evaluate Align-Consistency on LibriSpeech [23] and Libri- Light [24]

    Experiments 4.1. Datasets We evaluate Align-Consistency on LibriSpeech [23] and Libri- Light [24]. LibriSpeech is a widely used benchmark of approx- Table 1:Fully-supervised results on LibriSpeech development sets (WER↓).s= 0corresponds to the base CTC,s= 2uses 2 refinement steps on top of base CTC. Note Align-Consistency reduces to CR-CTC forα= 1.0. Data...

    2048

  5. [5]

    Our method encourages stable predictions with respect to input perturbations in every inference step, on both clean and noisy labels

    Conclusion We have presented Align-Consistency, which combines non- AR decoding with consistency regularization for end-to-end ASR. Our method encourages stable predictions with respect to input perturbations in every inference step, on both clean and noisy labels. Our experiments on both fully supervised and semi-supervised settings show that non-AR deco...

  6. [6]

    Attention-based models for speech recognition,

    J. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y . Ben- gio, “Attention-based models for speech recognition,” inNeurIPS, 2015

    2015

  7. [7]

    Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,

    W. Chan, N. Jaitly, Q. V . Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” inICASSP, 2016

    2016

  8. [8]

    Sequence transduction with recurrent neural net- works,

    A. Graves, “Sequence transduction with recurrent neural net- works,”arXiv preprint arXiv:1211.3711, 2012

    Pith/arXiv arXiv 2012

  9. [9]

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

    A. Graves, S. Fernandez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,” inICML, 2006

    2006

  10. [10]

    Mask- predict: Parallel decoding of conditional masked language mod- els,

    M. Ghazvininejad, O. Levy, Y . Liu, and L. Zettlemoyer, “Mask- predict: Parallel decoding of conditional masked language mod- els,” inEMNLP-IJCNLP, 2019

    2019

  11. [11]

    Align-refine: Non- autoregressive speech recognition via iterative realignment,

    E. A. Chi, J. Salazar, and K. Kirchhoff, “Align-refine: Non- autoregressive speech recognition via iterative realignment,” in NAACL-HLT, 2021

    2021

  12. [12]

    Streaming align-refine for non-autoregressive deliberation,

    W. Wang, K. Hu, and T. N. Sainath, “Streaming align-refine for non-autoregressive deliberation,” inInterspeech, 2022

    2022

  13. [13]

    Self-training for end-to-end speech recognition,

    J. Kahn, A. Lee, and A. Y . Hannun, “Self-training for end-to-end speech recognition,” inICASSP, 2020

    2020

  14. [14]

    Semi-supervised ASR by end- to-end self-training,

    Y . Chen, W. Wang, and C. Wang, “Semi-supervised ASR by end- to-end self-training,” inInterspeech, 2020

    2020

  15. [15]

    Cr-ctc: Consistency regularization on ctc for improved speech recognition,

    Z. Yao, W. Kang, X. Yang, F. Kuang, L. Guo, H. Zhu, Z. Jin, Z. Li, L. Lin, and D. Povey, “Cr-ctc: Consistency regularization on ctc for improved speech recognition,” inICLR, 2025

    2025

  16. [16]

    Mask ctc: Non-autoregressive end-to-end speech recognition with ctc and mask predict,

    Y . Higuchi, S. Sato, K. Nishida, S. Nagashima, and K. Ito, “Mask ctc: Non-autoregressive end-to-end speech recognition with ctc and mask predict,” inInterspeech, 2020

    2020

  17. [17]

    Non-autoregressive transformer for speech recognition,

    N. Chen, S. Watanabe, J. Villalba, P. Zelasko, and N. Dehak, “Non-autoregressive transformer for speech recognition,”IEEE Signal Processing Letters, vol. 28, pp. 121–125, 2021

    2021

  18. [18]

    Imputer: Sequence modelling via imputation and dy- namic programming,

    W. Chan, C. Saharia, J. Hsu, D. Rybach, M. Norouzi, and N. Jaitly, “Imputer: Sequence modelling via imputation and dy- namic programming,” inICML, 2020

    2020

  19. [19]

    Unsuper- vised training of acoustic models for large vocabulary continu- ous speech recognition,

    F. Wessel, R. Schl ¨uter, K. Macherey, and H. Ney, “Unsuper- vised training of acoustic models for large vocabulary continu- ous speech recognition,”IEEE Transactions on Speech and Audio Processing, vol. 13, no. 1, pp. 23–31, 2005

    2005

  20. [20]

    Lightly supervised and unsupervised acoustic model training,

    L. Lamel, J.-L. Gauvain, and G. Adda, “Lightly supervised and unsupervised acoustic model training,”Computer Speech & Lan- guage, vol. 16, no. 1, pp. 115–129, 2002

    2002

  21. [21]

    Unsupervised versus supervised training of acoustic models,

    J. Ma and R. Schwartz, “Unsupervised versus supervised training of acoustic models,” inInterspeech, 2008

    2008

  22. [22]

    Unsupervised train- ing and directed manual transcription for lvcsr,

    K. Yu, M. Gales, L. Wang, and P. Woodland, “Unsupervised train- ing and directed manual transcription for lvcsr,”Speech Commu- nication, vol. 52, no. 7–8, pp. 652–663, 2010

    2010

  23. [23]

    Semi-supervised training of deep neural networks,

    K. Vesel ´y, M. Hannemann, and L. Burget, “Semi-supervised training of deep neural networks,” inASRU, 2013

    2013

  24. [24]

    Word-level confidence mea- sures for semi-supervised training of dnn acoustic models,

    K. Vesel ´y, P. N´agl, and L. Burget, “Word-level confidence mea- sures for semi-supervised training of dnn acoustic models,” inIn- terspeech, 2017

    2017

  25. [25]

    High quality agreement-based semi-supervised training data for acous- tic modeling,

    F. d. C. Quitry, A. Oines, P. J. Moreno, and E. Weinstein, “High quality agreement-based semi-supervised training data for acous- tic modeling,” inSLT, 2016

    2016

  26. [26]

    Unsupervised domain adaptation for speech recognition via uncertainty driven self-training,

    S. Khurana, N. Moritz, T. Hori, and J. Le Roux, “Unsupervised domain adaptation for speech recognition via uncertainty driven self-training,” inICASSP, 2021

    2021

  27. [27]

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

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” inInterspeech, 2019

    2019

  28. [28]

    Lib- riSpeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- riSpeech: An ASR corpus based on public domain audio books,” inICASSP, 2015

    2015

  29. [29]

    Libri-Light: A benchmark for ASR with limited or no supervision,

    J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu, P.-E. Mazare, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegen, T. Likhomanenko, G. Synnaeve, A. Joulin, A. Mohamed, and E. Dupoux, “Libri-Light: A benchmark for ASR with limited or no supervision,” inICASSP, 2020

    2020

  30. [30]

    ESPnet: End-to-end speech processing toolkit,

    S. Watanabe, T. Hori, S. Karita, T. Hayashi, J. Nishitoba, Y . Unno, N. E. Y . Soplin, J. Heymann, M. Wiesner, N. Chen, A. Renduch- intala, and T. Ochiai, “ESPnet: End-to-end speech processing toolkit,” inInterspeech, 2018

    2018

  31. [31]

    Semi- supervised speech recognition via local prior matching,

    W.-N. Hsu, A. Lee, G. Synnaeve, and A. Hannun, “Semi- supervised speech recognition via local prior matching,”arXiv preprint arXiv:2002.10336, 2020

    Pith/arXiv arXiv 2002

  32. [32]

    Iterative pseudo-labeling for speech recognition,

    Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert, “Iterative pseudo-labeling for speech recognition,” arXiv preprint arXiv:2005.09267, 2020

    Pith/arXiv arXiv 2005

  33. [33]

    Improved noisy student training for automatic speech recognition,

    D. S. Park, Y . Zhang, Y . Jia, W. Han, C.-C. Chiu, B. Li, Y . Wu, and Q. V . Le, “Improved noisy student training for automatic speech recognition,”arXiv preprint arXiv:2005.09629, 2020

    Pith/arXiv arXiv 2005

  34. [34]

    Audio-conditioned diffusion llms for asr and deliberation pro- cessing,

    M. Wang, Z. Liu, Z. Jin, G. Sun, C. Zhang, and P. C. Woodland, “Audio-conditioned diffusion llms for asr and deliberation pro- cessing,”arXiv preprint arXiv:2509.16622, 2025

    arXiv 2025