Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Pitch Accent Detection improves Pretrained Automatic Speech Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Training a wav2vec2 model jointly on ASR and pitch accent detection cuts LibriSpeech test-clean word error rate from 6.0 to 4.3.

desk verdict Useful result with a real confound: the WER gains are plausible but not yet attributable to pitch accent information because there's no non-prosody auxiliary-task control. read the letter →

arxiv 2508.04814 v1 pith:VJTT3J7B submitted 2025-08-06 cs.CL cs.SDeess.AS

David Sasu , Natalie Schluter This is my paper
classification cs.CLcs.SDeess.AS
keywords speechrecognitionprosodypitchaccentdetectionwav2vec2multi-tasklearningself-traininglimited-resourceASRLibri
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 argues that prosody is information that a pretrained speech model should not lose, and that a small amount of pitch accent supervision can put it back. The authors train a wav2vec2-based model on one hour of labelled speech, Libri-light train-1h, to do two things at once: transcribe words and mark which words carry pitch accent. On LibriSpeech test-clean, the joint model lowers word error rate from 6.0 to 4.3 without a language model and from 5.0 to 3.0 with a 4-gram language model. The same prosody head also improves pitch accent F1 over the previous system at every tolerance, from 0.81 to 0.88 at 100 ms, and self-training on the one-hour set raises it to 0.90.

What carries the argument

The machinery is the joint objective $\mathcal{L}_j = \mathcal{L}_{\mathrm{asr}} + \mathcal{L}_{\mathrm{pad}}$, where a shared wav2vec2 encoder produces the representations for both a Connectionist Temporal Classification head (ASR) and a pitch accent head made of a BiLSTM, a linear layer, layer normalization, and a per-frame binary classifier. The joint loss pulls gradients from both tasks into the shared encoder during the unfrozen half of fine-tuning, which is the mechanism by which prosody is re-learned rather than discarded. The semi-supervised extension adds a voting step: three models trained on complementary subsets of BURNC label every frame of Libri-light train-1h, the majority lab

What would settle it

Train the identical joint model on LS1 with the pitch accent labels shuffled across time frames (or replaced by random labels with the same proportion of positives); if the WER drop from 6.0 to 4.3 persists, the effect is not caused by pitch accent information. A complementary check is to manually annotate a random sample of LS1 and compare it with the voting ensemble's labels; near-chance agreement would show the auxiliary task is teaching noise.

Watch

Extended reading notes

Core claim

The paper's central claim is that pitch accent detection and ASR can be jointly optimised on top of a self-supervised speech encoder, and that the prosody objective is not merely compatible with ASR but actively improves it. In the limited-resource setting, the joint model reduces WER on LibriSpeech test-clean from 6.0 to 4.3 without a language model, from 5.0 to 3.0 with a 4-gram LM, and from 5.0 to 4.0 with a Transformer LM; the corresponding BURNC test numbers are 23.0 to 20.0, 17.59 to 13.25, and 18.0 to 14.9. Character error rate stays roughly constant, so the gains are not a general accuracy boost. For the prosody task in isolation, the paper shows that replacing explicit fundamental f

Load-bearing premise

The load-bearing premise is that the automatically projected pitch accent labels on Libri-light train-1h are good enough to teach the shared encoder something real about prominence; the paper never directly measures that label quality, so the WER gain could in principle come from the extra parameters and regularisation of the added head.

Editorial extensions

If this is right

  • A one-hour labelled ASR corpus can be prosodically annotated by a voting ensemble with no new human annotation, and that relabelled hour yields better ASR fine-tuning than the same hour without pitch accent labels.
  • The WER improvement appears with and without language models, so the benefit lives in the acoustic representation rather than in the text prior.
  • Pitch accent detection no longer requires hand-engineered fundamental frequency features once wav2vec2 representations are available, which simplifies the prosody module.
  • The label transfer from broadcast news (BURNC) to audiobooks (LibriSpeech) works well enough to improve ASR, suggesting the method tolerates at least some domain shift in prosody annotation.

Reading between the lines

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

  • The paper's stopping rule optimises for BURNC test, not for label correctness on LibriSpeech; a hand-annotated sample of LS1 would let someone measure pseudo-label F1 directly and separate annotation noise from genuine domain transfer.
  • Because the relative gain is largest when no language model is used, low-resource languages without a text LM are the natural next testbed for this method.
  • The pitch accent head uses only hard binary labels; extending the same joint-loss scheme to boundary tones, nuclear accents, or soft prominence scores would show which prosodic tier carries the ASR benefit.
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 / 5 minor

Summary. This paper proposes a multi-task model that jointly performs ASR (wav2vec2 + CTC) and pitch accent detection (BiLSTM + linear + layer norm) on the same wav2vec2 representations. The authors first present a simplified prosody detector that outperforms wav2TOBI on the Boston University Radio News Corpus (BURNC) without using F0 features. They then use self-training with three-model voting to pseudo-label pitch accents for Libri-light train-1h (LS1), and fine-tune the joint ASR-prosody model on LS1. Compared with ASR-only fine-tuning, the joint model reduces WER on LibriSpeech test-clean from 6.0 to 4.3 (no LM), 5.0 to 3.0 (4-gram LM), and 5.0 to 4.0 (Transformer LM), with similar gains on BURNC. The paper concludes that explicitly modeling pitch accents improves low-resource ASR.

Significance. If the causal claim is established, this is a practically useful result: it shows that prosodic supervision can be injected into self-supervised speech representations and improve low-resource ASR, while also advancing pitch accent detection with a simpler architecture. Strengths include the release of code, use of standard benchmarks (BURNC, LibriSpeech, Libri-light), and consistent WER reductions across three language-model conditions. The reported improvement over wav2TOBI in pitch accent detection is also noteworthy. However, the central attribution of the ASR gains to pitch accent information is currently confounded by the addition of an auxiliary branch and loss, and by the use of unvalidated pseudo-labels.

major comments (3)
  1. [§4, Table 2] The central causal claim—that pitch accent information improves ASR—is not identified by the experimental design. The joint ASR-prosody model differs from the ASR-only baseline in two conflated ways: it adds an auxiliary branch (BiLSTM + linear + layer norm) and a second loss L_pad, and it uses pseudo-labels from a BURNC-trained model. Even with perfect labels, the WER reduction could come from multi-task regularization rather than from pitch accent content. A necessary control is a non-prosody auxiliary task using the same architecture and the same amount of auxiliary supervision (e.g., random labels, frame-level speaker or phone classification), or an ablation with shuffled/random pitch accent labels. Without such a control, Table 2 does not support the claim that pitch accent detection, rather than auxiliary supervision in general, drives the gains.
  2. [§3.3 and Table 1] The self-training procedure selects the number of iterations by monitoring performance on the BURNC test set: 'checked over BURNC test set for performance gains' and the process halts after four iterations. Consequently, the 'Ours Semi-Sup' row in Table 1 is a test-set-selected result and cannot be reported as an unbiased state-of-the-art F1. Moreover, the manuscript never evaluates the quality of the pseudo pitch accent labels on LibriSpeech/Libri-light audiobook audio. If these labels are noisy or domain-mismatched, the auxiliary loss may be training on incorrect targets, and any WER improvement cannot be attributed to pitch accent information. Please report pseudo-label accuracy on held-out audiobook samples (or another prosody dataset) and use a held-out development set for stopping.
  3. [§4, Table 2] All WER/CER numbers appear to come from single training runs; no error bars, confidence intervals, or significance tests are reported. In a 1-hour low-resource fine-tuning setting, run-to-run variance can be substantial. The improvements (e.g., 6.0→4.3 without LM, 5.0→3.0 with 4-gram LM) are the central quantitative evidence, so the lack of uncertainty quantification makes it difficult to assess whether the effect is reliable. Please provide multiple seeds (at least for the no-LM condition) or significance tests to confirm that the observed differences are stable.
minor comments (5)
  1. [Abstract and §1] 'complimentary' should be 'complementary' in the abstract.
  2. [§8, Reference [16]] Reference [16] is incomplete: it lacks the paper title and page numbers, reading only 'in Proceedings...'.
  3. [§3.2 and §3.3] The heading 'Results for prosody detection' appears twice (once at the end of §3.2 and again at the start of §3.3). One should be reworded to avoid duplication.
  4. [§3.1] 'LibriV ox' contains a word-break artifact; it should read 'LibriVox'.
  5. [§4] The joint loss is written as Lj = Lasr + Lpad with no weighting coefficient. Please state whether the two losses are simply summed or whether any weighting/balancing is used, and report the learning rate, batch size, and other fine-tuning hyperparameters explicitly rather than only linking to defaults.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity in the main ASR claim; one secondary prosody result is selected on its own test set via the self-training stopping rule.

  1. fitted input called prediction [Section 3.3 (Semi-supervised Prosodic Event Detection) and Table 1]
    "The prosody labelled version of LS1 was then added to the full BURNC train set, and then checked over BURNC test set for performance gains. If there were gains, we repeated the process now with the prosody labelled LS1 as part of the partitioning step, replacing the labels of LS1 at each iteration. Otherwise the process halts. Our training process halted after 4 iterations."

    The self-training loop's stopping rule is evaluated on the BURNC test set, and Table 1 reports the Ours Semi-Sup F1 on that same BURNC test set. The iteration count (and the choice among the 'number of different self-training techniques' mentioned in footnote 3) is therefore fitted to the evaluation set; the reported Semi-Sup F1 is a selected result, not an independent prediction. This is a real but localized test-set-fitting circularity. It does not affect the central ASR claim: the joint-model WERs in Table 2 are measured on LibriSpeech test-clean, disjoint from the BURNC set used for stopping, so the headline 6.0->4.3 comparison remains an externally benchmarked empirical result.

full rationale

The main derivation chain is empirical and self-contained: wav2vec2 representations are fine-tuned on LS1 with and without an auxiliary pitch accent loss, and WER is evaluated on LibriSpeech test-clean, an independent benchmark. No equation in the paper reduces the ASR objective to the prosody loss, and no fitted parameter is renamed as the LibriSpeech WER. The pseudo-labelling pipeline uses BURNC-trained models to label LS1; this is an auxiliary target, not the ASR transcript, so it does not make the ASR result definitionally circular. The one circular-adjacent issue is the BURNC test-based stopping rule for self-training, which makes the Ours Semi-Sup row of Table 1 a test-set-selected estimate rather than independent evidence; however, it is not load-bearing for the main WER claim. Two non-circular weaknesses remain: (i) the joint model differs from the ASR-only baseline by an added auxiliary branch and loss, so without a non-prosody auxiliary-task control the improvement cannot be causally attributed to pitch accent information, and (ii) the pseudo-label accuracy on LibriSpeech/audiobook audio is not reported. Both are correctness/identification concerns, not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on cross-corpus transfer of prosody labels and on the auxiliary objective being informative. The only hand-tuned controller that directly affects reported numbers is the self-training iteration count, selected on the BURNC test set. No new physical entities are introduced.

free parameters (2)
  • self-training iterations = 4
    Stopping rule uses gains on BURNC test set (Section 3.3), and the same test set is reported in Tables 1 and 2.
  • joint loss weight = 1.0 for each loss
    The joint loss is a simple sum L_asr + L_pad with no weighting or sensitivity analysis (Section 3.2).
assumptions (4)
  • domain assumption BURNC frame-level binary pitch accent labels are reliable supervision for prosody detection.
    Used as ground truth throughout (Section 3.1); no inter-annotator agreement or label quality check is reported.
  • domain assumption wav2vec2 representations retain enough pitch accent information to make F0 features unnecessary.
    The streamlined prosody head removes the F0 input used by wav2TOBI (Section 3.2, Table 1); if false, the gains could come from hidden acoustic shortcuts.
  • domain assumption Prosody models trained on BURNC broadcast news transfer to LibriSpeech audiobook audio.
    Section 3.3 uses BURNC-trained models to pseudo-label Libri-light train-1h; no direct evaluation of the pseudo-label quality on LibriSpeech is given.
  • domain assumption The auxiliary pitch accent objective, not the extra layers or regularization, causes the ASR improvement.
    No ablation with a non-prosodic auxiliary task or with shuffled labels is provided; this is the key confound for the causal claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pitch Accent Detection improves Pretrained Automatic Speech Recognition." pith.science (2026). https://pith.science/paper/VJTT3J7B

@misc{pith2026250804814,
  author       = {Pith},
  title        = {Pith review of: Pitch Accent Detection improves Pretrained Automatic Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VJTT3J7B}},
  note         = {Machine review of arXiv:2508.04814}
}
read the original abstract

We show the performance of Automatic Speech Recognition (ASR) systems that use semi-supervised speech representations can be boosted by a complimentary pitch accent detection module, by introducing a joint ASR and pitch accent detection model. The pitch accent detection component of our model achieves a significant improvement on the state-of-the-art for the task, closing the gap in F1-score by 41%. Additionally, the ASR performance in joint training decreases WER by 28.3% on LibriSpeech, under limited resource fine-tuning. With these results, we show the importance of extending pretrained speech models to retain or re-learn important prosodic cues such as pitch accent.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Afrispeech Semantics: Evaluating Audio Semantic Reasoning in Spoken Language Models Across Domains and Accents

    cs.CL 2026-05 unverdicted novelty 4.0 of 10

    Audio language models are benchmarked on five semantic and paralinguistic reasoning tasks to reveal limitations in handling spoken audio evidence, accent variation, and domain shifts.

Reference graph

Works this paper leans on

56 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [1]

    important

    Introduction Models based on self-supervised speech representations have in recent years claimed state-of-the-art performance in ASR [1, 2]. Moreover, they have been permitted to bypass both a heavy speech-science informed featurisation component, as well a language dependent acoustic dictionary resource writing com- ponent. In doing so, such models have ...

    work page

  2. [2]

    We streamline and significantly boost the performance of the current state-of-the-art model for pitch accent detection

    work page

  3. [3]

    We present a multi-task model for integrating pitch accent detection into the ASR task, which improves the performance for ASR in limited resource settings

    work page

  4. [4]

    We then automatically annotate the pitch accents of a small dataset using self-training, and then apply it in our proposed joint model, achieving even further ASR performance boosts.2

    work page

  5. [5]

    Prosody Detection There is a long line of research on automatic prosody detec- tion (for example, [ 8, 9, 10, 11, 12, 13])

    Related work 2.1. Prosody Detection There is a long line of research on automatic prosody detec- tion (for example, [ 8, 9, 10, 11, 12, 13]). With the advent of pretrained speech models, and in particular, wav2vec [ 14] and wav2vec2 [ 1], a new line of systems that builds on self- supervised speech representations has achieved the state-of-the- art in det...

    arXiv 2025

  6. [6]

    Datasets Our research uses the Boston University Radio News Corpus (BURNC) [7], Librispeech [45] and Libri-light [46] corpora

    Modelling prosody and ASR 3.1. Datasets Our research uses the Boston University Radio News Corpus (BURNC) [7], Librispeech [45] and Libri-light [46] corpora. The BURNC dataset is a broadcast news-style read speech corpus which contains 11 hours of speech, sourced from 7 differ- ent speakers (3 female and 4 male). It consists of audio snippets with their t...

    work page

  7. [7]

    Our mod- els are all trained for 30,000 steps on three v100 GPUs, using default parameters, and take approximately 8 hours to train

    Experimental setup and results We use the base-960h wav2vec2 pretrained model4. Our mod- els are all trained for 30,000 steps on three v100 GPUs, using default parameters, and take approximately 8 hours to train. 5 Results for word and character error rates (respectively WER and CER) are given in Table 2. All models were fine-tuned for ASR (resp. ASR and ...

    work page

  8. [8]

    CHIEF’S", we do not remove the apostrophe ( ’ ) during pre-processing to form the word

    Error analysis and discussion We have shown above that pitch accent detection is useful for improving the performance of pretrained speech models in ASR tasks within limited resource scenarios. However, even though we improve upon the WER in most of the experiments that we perform with our proposed joint model, we notice that experi- ments that involve th...

    work page

Show all 56 references
  1. [9]

    Conclusion In this paper we have presented an approach for leveraging prosodic information to improve the performance of a pretrained speech model in a limited resource scenario. The results from our experiments demonstrate that re-focusing self-supervised speech models on sup...

  2. [10]

    More work is required to investigate the real impact when fine-tuning with larger ASR datasets

    Limitations All experiments were carried out under the limited resource setting, with little fine-tuning data, due to the requirement of our method to use prosodic labels. More work is required to investigate the real impact when fine-tuning with larger ASR datasets. Also, for...

  3. [11]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  4. [12]

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

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

    2021

  5. [13]

    The zero resource speech challenge 2017,

    E. Dunbar, X. N. Cao, J. Benjumea, J. Karadayi, M. Bernard, L. Be- sacier, X. Anguera, and E. Dupoux, “The zero resource speech challenge 2017,” in2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2017, pp. 323–330

    2017

  6. [14]

    The zero resource speech challenge 2020: Discovering discrete subword and word units,

    E. Dunbar, J. Karadayi, M. Bernard, X. Cao, R. Algayres, L. Ondel, L. Besacier, S. Sakti, and E. Dupoux, “The zero resource speech challenge 2020: Discovering discrete subword and word units,” in NeurIPS Workshop on Self-Supervised Learning for Speech and Audio Processing, 2020

    2020

  7. [15]

    The zero resource speech benchmark 2021: Metrics and baselines for unsupervised spoken language modeling,

    T. A. Nguyen, M. de Seyssel, P. Rozé, M. Rivière, E. Kharitonov, A. Baevski, E. Dunbar, and E. Dupoux, “The zero resource speech benchmark 2021: Metrics and baselines for unsupervised spoken language modeling,” inProceedings of Interspeech 2021, 2021, pp. 1574–1578

    2021

  8. [16]

    ProsAu- dit, a prosodic benchmark for self-supervised speech models,

    M. de Seyssel, M. Lavechin, H. Titeux, A. Thomas, G. Virlet, A. S. Revilla, G. Wisniewski, B. Ludusan, and E. Dupoux, “ProsAu- dit, a prosodic benchmark for self-supervised speech models,” in Proceedings of INTERSPEECH 2023, 2023, pp. 2963–2967

    2023

  9. [17]

    The boston university radio news corpus,

    M. Ostendorf, P. J. Price, and S. Shattuck-Hufnagel, “The boston university radio news corpus,”Linguistic Data Consortium, pp. 1–19, 1995

    1995

  10. [18]

    Using neural networks to locate pitch accents

    P. A. Taylor, “Using neural networks to locate pitch accents.” 1995

    1995

  11. [19]

    Modeling phrasing and prominence using deep recurrent learning,

    A. Rosenberg, R. Fernandez, and B. Ramabhadran, “Modeling phrasing and prominence using deep recurrent learning,” inSix- teenth Annual Conference of the International Speech Communi- cation Association, 2015

    2015

  12. [20]

    Automatic classification of lexical stress in english and arabic languages using deep learning

    M. A. Shahin, J. Epps, and B. Ahmed, “Automatic classification of lexical stress in english and arabic languages using deep learning.” inInterspeech, 2016, pp. 175–179

    2016

  13. [21]

    Automatic lexi- cal stress and pitch accent detection for l2 english speech using multi-distribution deep neural networks,

    K. Li, S. Mao, X. Li, Z. Wu, and H. Meng, “Automatic lexi- cal stress and pitch accent detection for l2 english speech using multi-distribution deep neural networks,”Speech Communication, vol. 96, pp. 28–36, 2018

    2018

  14. [22]

    Acoustic and temporal representations in convolutional neural network models of prosodic events,

    S. Stehwien, A. Schweitzer, and N. T. Vu, “Acoustic and temporal representations in convolutional neural network models of prosodic events,”Speech Communication, vol. 125, pp. 128–141, 2020

    2020

  15. [23]

    Cnn encoding of acoustic param- eters for prominence detection,

    K. Sabu, M. Vaidya, and P. Rao, “Cnn encoding of acoustic param- eters for prominence detection,”arXiv preprint arXiv:2104.05488, 2021

    2021 arXiv

  16. [24]

    wav2vec: Unsupervised pre-training for speech recognition,

    S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training for speech recognition,” in20th Annual Conference of the International Speech Communication Association, Interspeech 2019, Graz, Austria, September 15-19, 2019, G. Kubin and Z. Kacic, Eds....

    2019 doi

  17. [25]

    Detection of prosodic bound- aries in speech using wav2vec 2.0,

    M. Kunešová and M. ˇRezáˇcková, “Detection of prosodic bound- aries in speech using wav2vec 2.0,” inInternational Conference on Text, Speech, and Dialogue. Springer, 2022, pp. 377–388

    2022

  18. [26]

    Zhai and M

    W. Zhai and M. Hasegawa-Johnson, inProceedings of the Annual Conference of the International Speech Communication Associa- tion, INTERSPEECH, 2023, pp. 2748–2752, publisher Copyright: © 2023 International Speech Communication Association. All rights reserved.; 24th Internation...

    2023

  19. [27]

    On the use of prosody in automatic dialogue understanding,

    E. Nöth, A. Batliner, V . Warnke, J. Haas, M. Boros, J. Buckow, R. Huber, F. Gallwitz, M. Nutt, and H. Niemann, “On the use of prosody in automatic dialogue understanding,”Speech Communi- cation, vol. 36, no. 1-2, pp. 45–62, 2002

    2002

  20. [28]

    Prosody modeling for automatic speech recognition and understanding,

    E. Shriberg and A. Stolcke, “Prosody modeling for automatic speech recognition and understanding,” inMathematical Founda- tions of Speech and Language Processing. Springer, 2004, pp. 105–114

    2004

  21. [29]

    Can prosody aid the automatic classification of dialog acts in con- versational speech?

    E. Shriberg, A. Stolcke, D. Jurafsky, N. Coccaro, M. Meteer, R. Bates, P. Taylor, K. Ries, R. Martin, and C. Van Ess-Dykema, “Can prosody aid the automatic classification of dialog acts in con- versational speech?”Language and speech, vol. 41, no. 3-4, pp. 443–492, 1998

    1998

  22. [30]

    Improving end-to-end slu performance with prosodic attention and distillation,

    S. Rajaa, “Improving end-to-end slu performance with prosodic attention and distillation,”arXiv preprint arXiv:2305.08067, 2023

    2023 arXiv

  23. [31]

    A neural prosody encoder for end-to-end dialogue act classification,

    K. Wei, D. Knox, M. Radfar, T. Tran, M. Müller, G. P. Strimel, N. Susanj, A. Mouchtaris, and M. Omologo, “A neural prosody encoder for end-to-end dialogue act classification,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)....

    2022

  24. [32]

    Parsing speech: a neural approach to integrat- ing lexical and acoustic-prosodic information,

    T. Tran, S. Toshniwal, M. Bansal, K. Gimpel, K. Livescu, and M. Ostendorf, “Parsing speech: a neural approach to integrat- ing lexical and acoustic-prosodic information,”arXiv preprint arXiv:1704.07287, 2017

    2017 arXiv

  25. [33]

    Sentence-internal prosody does not help parsing the way punctuation does,

    M. Gregory, M. Johnson, and E. Charniak, “Sentence-internal prosody does not help parsing the way punctuation does,” inPro- ceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, 2004...

    2004

  26. [34]

    Effective use of prosody in parsing conversational speech,

    J. G. Kahn, M. Lease, E. Charniak, M. Johnson, and M. Ostendorf, “Effective use of prosody in parsing conversational speech,” in Proceedings of Human Language Technology Conference and Con- ference on Empirical Methods in Natural Language Processing, 2005, pp. 233–240

    2005

  27. [35]

    Exploiting prosody for pcfgs with latent annotations

    M. Dreyer and I. Shafran, “Exploiting prosody for pcfgs with latent annotations.” inINTERSPEECH. Citeseer, 2007, pp. 450–453

    2007

  28. [36]

    Joint reranking of parsing and word recognition with automatic segmentation,

    J. G. Kahn and M. Ostendorf, “Joint reranking of parsing and word recognition with automatic segmentation,”Computer Speech & Language, vol. 26, no. 1, pp. 1–19, 2012

    2012

  29. [37]

    The use of prosody in syntactic disambiguation,

    P. J. Price, M. Ostendorf, S. Shattuck-Hufnagel, and C. Fong, “The use of prosody in syntactic disambiguation,”the Journal of the Acoustical Society of America, vol. 90, no. 6, pp. 2956–2970, 1991

    1991

  30. [38]

    The parsing of prosody,

    M. E. Beckman, “The parsing of prosody,”Language and cognitive processes, vol. 11, no. 1-2, pp. 17–68, 1996

    1996

  31. [39]

    Towards using prosody in speech recognition/understanding systems: Dif- ferences between read and spontaneous speech,

    K. E. Silverman, E. Blaauw, J. Spitz, and J. F. Pitrelli, “Towards using prosody in speech recognition/understanding systems: Dif- ferences between read and spontaneous speech,” inSpeech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26...

    1992

  32. [40]

    Prosody models for conversational speech recognition,

    M. Ostendorf, I. ShafranÞ, and R. Bates, “Prosody models for conversational speech recognition,” 2003

    2003

  33. [41]

    Use of prosodic features for speech recognition,

    K. Hirose and N. Minematsu, “Use of prosodic features for speech recognition,” inEighth International Conference on Spoken Lan- guage Processing, 2004

    2004

  34. [42]

    Prosodic and other cues to speech recognition failures,

    J. Hirschberg, D. Litman, and M. Swerts, “Prosodic and other cues to speech recognition failures,”Speech communication, vol. 43, no. 1-2, pp. 155–175, 2004

    2004

  35. [43]

    Simulta- neous recognition of words and prosody in the boston university radio speech corpus,

    M. Hasegawa-Johnson, K. Chen, J. Cole, S. Borys, S.-S. Kim, A. Cohen, T. Zhang, J.-Y . Choi, H. Kim, T. Yoonet al., “Simulta- neous recognition of words and prosody in the boston university radio speech corpus,”Speech Communication, vol. 46, no. 3-4, pp. 418–439, 2005

    2005

  36. [44]

    Improved speech recogni- tion using acoustic and lexical correlates of pitch accent in a n-best rescoring framework,

    S. Ananthakrishnan and S. Narayanan, “Improved speech recogni- tion using acoustic and lexical correlates of pitch accent in a n-best rescoring framework,” in2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP’07, vol. 4. IEEE, 2007, pp. IV–873

    2007

  37. [45]

    Using prosody to improve automatic speech recognition,

    K. Vicsi and G. Szaszák, “Using prosody to improve automatic speech recognition,”Speech Communication, vol. 52, no. 5, pp. 413–426, 2010

    2010

  38. [46]

    A new prosody-assisted mandarin asr system,

    S.-H. Chen, J.-H. Yang, C.-Y . Chiang, M.-C. Liu, and Y .-R. Wang, “A new prosody-assisted mandarin asr system,”Ieee transactions on audio, speech, and language processing, vol. 20, no. 6, pp. 1669–1684, 2012

    2012

  39. [47]

    Data aug- mentation using prosody and false starts to recognize non-native children’s speech,

    H. Kathania, M. Singh, T. Grósz, and M. Kurimo, “Data aug- mentation using prosody and false starts to recognize non-native children’s speech,”arXiv preprint arXiv:2008.12914, 2020

    2008 arXiv

  40. [48]

    Prosodic feature-based discriminatively trained low resource speech recognition system,

    T. Hasija, V . Kadyan, K. Guleria, A. Alharbi, H. Alyami, and N. Goyal, “Prosodic feature-based discriminatively trained low resource speech recognition system,”Sustainability, vol. 14, no. 2, p. 614, 2022

    2022

  41. [49]

    Explicit tone transcription improves asr per- formance in extremely low-resource languages: A case study in bribri,

    R. Coto-Solano, “Explicit tone transcription improves asr per- formance in extremely low-resource languages: A case study in bribri,” inProceedings of the first workshop on natural language processing for Indigenous languages of the Americas, 2021, pp. 173–184

    2021

  42. [50]

    Effectiveness of self- supervised pre-training for speech recognition,

    A. Baevski, M. Auli, and A. Mohamed, “Effectiveness of self- supervised pre-training for speech recognition,”arXiv preprint arXiv:1911.03912, 2019

    1911 arXiv

  43. [51]

    Decoar 2.0: Deep contextualized acous- tic representations with vector quantization,

    S. Ling and Y . Liu, “Decoar 2.0: Deep contextualized acous- tic representations with vector quantization,”arXiv preprint arXiv:2012.06659, 2020

    2012 arXiv

  44. [52]

    slimipl: Language-model-free iterative pseudo-labeling,

    T. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert, “slimipl: Language-model-free iterative pseudo-labeling,”arXiv preprint arXiv:2010.11524, 2020

    2010 arXiv

  45. [53]

    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

    2005 arXiv

  46. [54]

    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

    2005 arXiv

  47. [55]

    Librispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: an asr corpus based on public domain audio books,” in2015 IEEE international conference on acoustics, speech and signal process- ing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  48. [56]

    Libri- light: A benchmark for asr with limited or no supervision,

    J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu, P.-E. Mazaré, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegenet al., “Libri- light: A benchmark for asr with limited or no supervision,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal...

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.