pith. sign in

arxiv: 2606.28857 · v1 · pith:72AJG6LPnew · submitted 2026-06-27 · 💻 cs.SD · cs.CL

wav2VOT: Automatic estimation of voice onset time, closure duration, and burst realisation with wav2vec2

James Tanner , Morgan Sonderegger , Jane Stuart-Smith , Tyler Kendall , Jeff Mielke This is my paper

Pith reviewed 2026-06-30 08:53 UTC · model grok-4.3

classification 💻 cs.SD cs.CL
keywords wav2vec2voice onset timeautomatic annotationphonetic analysisstop consonantsfine-tuningburst realisationspeech timing
0
0 comments X

The pith

wav2VOT applies wav2vec2 to estimate voice onset time, closure duration, and burst realisation at accuracy levels matching current tools on new data.

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

The paper introduces wav2VOT as a system that takes raw audio and uses wav2vec2 to mark the timing of voice onset, closure, and burst events in stop consonants. Tests on datasets the model has not seen before show performance comparable to existing automatic methods. Fine-tuning the model on target data raises accuracy further. The estimates hold steady across voiced and voiceless stops and across different places of articulation. The work positions large pretrained speech models as ready components for phonetic annotation tasks.

Core claim

wav2VOT demonstrates that wav2vec2 can be used to automatically estimate voice onset time, closure duration, and burst realisation, performing comparably to current approaches on unseen datasets and achieving high accuracy with fine-tuning, with high fidelity across stop voicing and place of articulation.

What carries the argument

wav2vec2 representations, used directly or after fine-tuning, to predict the acoustic timing and burst events that define stop consonants.

If this is right

  • Phonetic annotation pipelines can incorporate wav2VOT to label large corpora with less manual review.
  • Fine-tuning on a small amount of target data improves timing estimates for specific research datasets.
  • Performance remains consistent across voicing categories and places of articulation without extra adjustments.
  • Large speech models can replace or supplement rule-based or smaller trained systems for stop-consonant measurements.

Where Pith is reading between the lines

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

  • The same model backbone could be adapted to measure other timing events such as vowel formant transitions.
  • Large-scale phonetic studies that were previously limited by annotation cost become feasible.
  • Domain mismatch between training and test recordings remains the main practical limit on out-of-the-box use.

Load-bearing premise

wav2vec2 representations already encode the timing and burst details of stop consonants well enough to transfer across different datasets and speakers.

What would settle it

A new test set from a clearly different domain, such as a different language or recording condition, where wav2VOT accuracy drops well below that of existing tools.

Figures

Figures reproduced from arXiv: 2606.28857 by James Tanner, Jane Stuart-Smith, Jeff Mielke, Morgan Sonderegger, Tyler Kendall.

Figure 1
Figure 1. Figure 1: Spectograms of three stop tokens from the Corpus of Spontaneous Japanese-Core (CSJ-C), overlaid with Closure + VOT (left), VOT (middle), and lenition (right) predicted from wav2VOT model trained in Section 3.1. convolutional layers which reduce the dimensionality of the speech signal to a set of features based on the product of the strides of all convolutional layers. In the standard wav2vec2 configuration… view at source ↗
Figure 2
Figure 2. Figure 2: Performance of CSJ-C-trained wav2VOT model for the CSJ-C training set, reported as a percentage of the test cor￾pora that fall within a set of fixed tolerances (e.g., <2ms refers to a proportion of predictions within 2ms of the manual anno￾tations). ommends that users train a classifier model on a subset of the target data, such that the parameters reflect the target distribu￾tion: in order to determine wh… view at source ↗
Figure 3
Figure 3. Figure 3: Performance of wav2VOT VOT (top) and closure duration (bottom left) predictions, reported as a percentage of the test corpora that fall within a set of fixed tolerances. Lenition predictions (bottom right) reported as predictive accuracy (with horizontal jitter). Purple squares lines indicate the CSJ-C-trained model (i.e. no finetuning), with circles corresponding to models finetuned with different amounts… view at source ↗
Figure 4
Figure 4. Figure 4: Model-estimated VOT (left) and closure dura￾tion (right) for manually-annotated stops (purple circles) and wav2VOT predictions (green triangles), separated by phoneme label. Points indicate posterior median, with lines indicating the 95% Highest Posterior Density (HPD) interval [40]. methods differ in the size of the estimated voicing contrast, the size of this difference is negligible, corresponding to a … view at source ↗
read the original abstract

While automatic tools for speech annotation are now commonplace within phonetic research pipelines, many tasks require substantial manual correction or training sets to perform accurately. Simultaneously, large speech models such as wav2vec2 have been shown to perform well at speech classification tasks, raising the question of how these models may be applied to phonetic annotation tasks. We introduce wav2VOT: a tool for the automatic estimation of voice onset time, closure duration, and burst realisation using wav2vec2. We demonstrate that wav2VOT performs comparably with current approaches on unseen datasets, and can estimate with high accuracy with fine-tuning. Analysis of wav2VOT predictions demonstrate high fidelity across stop voicing and place of articulation. These results demonstrate that large speech models are capable of producing accurate annotations, and further motivate exploration of large speech models as tools in phonetic research pipelines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 1 minor

Summary. The paper introduces wav2VOT, a tool based on wav2vec2 for automatic estimation of voice onset time (VOT), closure duration, and burst realisation for stop consonants. The central claim is that wav2VOT performs comparably to existing approaches on unseen datasets, achieves high accuracy after fine-tuning, and produces predictions with high fidelity across stop voicing and place of articulation.

Significance. If the quantitative results and generalization claims hold, the work would be significant for phonetic research pipelines by demonstrating that large pre-trained speech models can handle fine-grained timing and realisation tasks with reduced need for manual correction or large training sets. The explicit testing on unseen datasets and analysis across phonetic categories provide a concrete basis for further exploration of such models in annotation tasks.

minor comments (1)
  1. The abstract would be strengthened by including at least one key quantitative result (e.g., correlation or error metric) to support the performance claims, even though the full manuscript presumably contains these details in the results section.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of our work and for recommending acceptance. We are pleased that the significance of applying wav2vec2 to fine-grained phonetic timing and realisation tasks was recognised, particularly the evaluation on unseen data and across phonetic categories.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes an empirical ML application of pretrained wav2vec2 models (with optional fine-tuning) to estimate VOT, closure duration, and burst features, evaluated on unseen datasets. No derivation chain, equations, or first-principles results are present. Performance claims rest on direct comparison to existing tools and human annotations rather than any self-referential fitting or imported uniqueness theorem. The central results are externally falsifiable via replication on held-out speech data and do not reduce to the inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated premise that the wav2vec2 encoder already encodes the relevant phonetic timing information.

pith-pipeline@v0.9.1-grok · 5689 in / 998 out tokens · 31219 ms · 2026-06-30T08:53:04.635832+00:00 · methodology

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

49 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    Many analyses now utilise a pipeline of (semi-)automatic utterance- and word-level tran- scription and automatic time-alignment of phonemic labels to the speech signal (e.g

    Introduction Automatic approaches to the segmentation and annotation of speech data has become increasingly popular within phonetic and speech science research. Many analyses now utilise a pipeline of (semi-)automatic utterance- and word-level tran- scription and automatic time-alignment of phonemic labels to the speech signal (e.g. [1, 2, 3]), with furth...

  2. [2]

    wav2VOT: Automatic estimation of voice onset time, closure duration, and burst realisation with wav2vec2

    Model The architecture of wav2vec2 consists of a Feature Encoder (FE) block and Transformer Encoder (TE) block. Feature en- coding in wav2vec2 is performed via a set of 1-dimensional 1Source code and models available at https://github.com/james-tanner/wav2VOT. arXiv:2606.28857v1 [cs.SD] 27 Jun 2026 0.000 0.015 0.030 0.045 0.060 0.075 0.090 0.105 0.120 Tim...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    3.1), and 2) across a variety of unseen datasets, along with changes in performance when fine-tuned on varying amounts of data (Sec

    Experiment 1 The goals of Experiment 1 are to demonstrate that wav2VOT is capable of learning and capturing the fine-grained tempo- ral properties of stops in 1) a large set of training data with a matched test set (Sec. 3.1), and 2) across a variety of unseen datasets, along with changes in performance when fine-tuned on varying amounts of data (Sec. 3.2...

    1903

  4. [4]

    Experiment 2 Following previous work evaluating the performance of auto- matic VOT annotations [10, 12], we also evaluate the predictive performance of wav2VOT by comparing predictions directly with manual annotations in a hypothetical phonetic study con- text. Specifically, we take a regression modelling approach to estimate the degree of similarity betw...

  5. [5]

    Discussion The goal of this study has been to further explore the applica- tion of large speech models to the task of phonetic annotation. In particular, we sought to explore the extent to which wav2vec2 can accurately estimate the acoustic properties of stops, and introduce wav2VOT: an application of wav2vec2 for the auto- matic estimation of VOT, closur...

  6. [6]

    Computational resources were provided by the Digital Research Alliance of Canada

    Acknowledgements The authors thank the SPADE Data Guardians, Rachel Macdon- ald, Michael McAuliffe, and Vanna Willerton. Computational resources were provided by the Digital Research Alliance of Canada. This research was supported by a T-AP Digging into Data award in the form of the following grants: ESRC Grant #ES/R003963/1, NSERC/CRSNG Grants #RGPDD-501...

    2023

  7. [7]

    Generative AI Use Disclosure The authors declare that no Generative AI tools were used in the research, design, analysis, writing, or editing of this work

  8. [8]

    Articulation rate in American English in a corpus of YouTube videos,

    S. Coats, “Articulation rate in American English in a corpus of YouTube videos,”Language and Speech, vol. 63, pp. 799–831, 2020

    2020

  9. [9]

    Towards “English

    J. Tanner, M. Sonderegger, J. Stuart-Smith, and J. Fruehwald, “Towards “English” phonetics: variability in the pre-consonantal voicing effect across English dialects and speakers,”Frontiers in Artificial Intelligence, vol. 3, 2020

    2020

  10. [10]

    Sys- tematic co-variation of monophthongs across speakers of New Zealand English,

    J. Brand, J. Hay, L. Clark, K. Watson, and M. S ´oskuthy, “Sys- tematic co-variation of monophthongs across speakers of New Zealand English,”Journal of Phonetics, vol. 88, 2021

    2021

  11. [11]

    V oice Onset Time (VOT) at 50: Theoretical and practical issues in measuring voicing distinc- tions,

    A. S. Abramson and D. H. Whalen, “V oice Onset Time (VOT) at 50: Theoretical and practical issues in measuring voicing distinc- tions,”Journal of Phonetics, vol. 63, pp. 75–86, 2017

    2017

  12. [12]

    A cross-language study of voicing in initial stops: Acoustical measurements,

    L. Lisker and A. S. Abramson, “A cross-language study of voicing in initial stops: Acoustical measurements,”Word, vol. 20, no. 3, pp. 384–422, 1964

    1964

  13. [13]

    Rapid versus rabid: A catalogue of acoustic features that may cue the distinction,

    L. Lisker, “Rapid versus rabid: A catalogue of acoustic features that may cue the distinction,”Journal of the Acoustical Society of America, vol. 62, pp. S77–S78, 1977

    1977

  14. [14]

    Relation between voice-onset time and vowel duration,

    R. F. Port and R. Rotunno, “Relation between voice-onset time and vowel duration,”Journal of the Acoustical Society of America, vol. 66, pp. 654–662, 1979

    1979

  15. [15]

    Variation and universals in VOT: evi- dence from 18 languages,

    T. Cho and P. Ladefoged, “Variation and universals in VOT: evi- dence from 18 languages,”Journal of Phonetics, vol. 27, pp. 207– 229, 1999

    1999

  16. [16]

    Laryngeal phonetics, phonology, assimilation and final neutralization,

    J. Salmons, “Laryngeal phonetics, phonology, assimilation and final neutralization,” inCambridge Handbook of Germanic Lin- guistics, R. Page and M. T. Putnam, Eds. Cambridge: Cambridge University Press, 2019, pp. 119–142

    2019

  17. [17]

    Automatic measurement of voice onset time using discriminative structured predictions,

    M. Sonderegger and J. Keshet, “Automatic measurement of voice onset time using discriminative structured predictions,”Journal of the Acoustical Society of America, vol. 132, pp. 3965–3979, 2012

    2012

  18. [18]

    Structure in talker-specific phonetic realization: Covariation of stop consonant VOT in American En- glish,

    E. Chodroff and C. Wilson, “Structure in talker-specific phonetic realization: Covariation of stop consonant VOT in American En- glish,”Journal of Phonetics, vol. 61, pp. 30–47, 2017

    2017

  19. [19]

    Assessing auto- matic VOT annotation using unimpaired and impaired speech,

    E. Buz, A. Buchwald, T. Fuchs, and J. Keshet, “Assessing auto- matic VOT annotation using unimpaired and impaired speech,” International Journal of Speech-Language Pathology, vol. 20, pp. 624–634, 2018

    2018

  20. [20]

    Structured heterogeneity in Scottish stops over the twentieth century,

    M. Sonderegger, J. Stuart-Smith, T. Knowles, R. MacDonald, and T. Rathcke, “Structured heterogeneity in Scottish stops over the twentieth century,”Language, vol. 96, pp. 94–125, 2020

    2020

  21. [21]

    Automatic estimation of voice onset time for word-initial stops by applying random forest to onset detec- tion,

    H.-C. W. Chi-Yueh Lin, “Automatic estimation of voice onset time for word-initial stops by applying random forest to onset detec- tion,”Journal of thr Acoustical Society of America, vol. 130, p. 514–525, 2011

    2011

  22. [22]

    Estimation of voice-onset time in continuous speech using temporal measures,

    A. P. Prathosh, A. G. Ramakrishnan, and T. V . Ananthapadman- abha, “Estimation of voice-onset time in continuous speech using temporal measures,”Journal of the Acoustical Society of America, vol. 136, p. EL122–EL128, 2014

    2014

  23. [23]

    Sequence segmentation using joint rnn and structured prediction models,

    Y . Adi, J. Keshet, E. Cibelli, and M. A. Goldrick, “Sequence segmentation using joint rnn and structured prediction models,” 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2422–2426, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:885991

    2017

  24. [24]

    Dr. vot: Measuring posi- tive and negative voice onset time in the wild,

    Y . Shrem, M. Goldrick, and J. Keshet, “Dr. vot: Measuring posi- tive and negative voice onset time in the wild,”Proc. Interspeech 2019, pp. 629–633, 2019

    2019

  25. [25]

    Puggaard-Rode, “getVOT,” 2024, version 0.2.0

    R. Puggaard-Rode, “getVOT,” 2024, version 0.2.0. [Online]. Available: https://github.com/rpuggaardrode/getVOT

    2024

  26. [26]

    Segmental and prosodic effects on intervocalic voiced stop reduction in connected speech,

    D. Bouavichith and L. Davidson, “Segmental and prosodic effects on intervocalic voiced stop reduction in connected speech,”Pho- netica, vol. 70, pp. 182–206, 2013

    2013

  27. [27]

    The causal structure of lenition: a case for the causal precedence of durational shortening,

    U. Cohen Priva and E. Gleason, “The causal structure of lenition: a case for the causal precedence of durational shortening,”Lan- guage, vol. 96, pp. 413–448, 2020

    2020

  28. [28]

    wav2vec: Unsupervised pre-training for speech recognition,

    S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training for speech recognition,” CoRR, vol. abs/1904.05862, 2019. [Online]. Available: http: //arxiv.org/abs/1904.05862

    work page arXiv 1904

  29. [29]

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

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,”CoRR, vol. abs/2006.11477, 2020. [Online]. Available: https://arxiv.org/abs/2006.11477

    work page arXiv 2006

  30. [30]

    Automatic classification of stop realisation with wav2vec2.0,

    J. Tanner, M. Sonderegger, J. Stuart-Smith, J. Mielke, and T. Kendall, “Automatic classification of stop realisation with wav2vec2.0,” inInterspeech 2025, 2025, pp. 2270–2274

    2025

  31. [31]

    Using wav2vec 2.0 for phonetic classi- fication tasks: methodological aspects,

    L. Kim and C. Gendrot, “Using wav2vec 2.0 for phonetic classi- fication tasks: methodological aspects,” inProceedings of Inter- speech 2024, 2024, pp. 1530–1534

    2024

  32. [32]

    Comparing human and ma- chine’s use of coarticulatory vowel nasalization for linguistic clas- sification,

    G. Zellou, L. Kim, and C. Gendrot, “Comparing human and ma- chine’s use of coarticulatory vowel nasalization for linguistic clas- sification,”Journal of the Acoustical Society of America, vol. 156, pp. 489–502, 2024

    2024

  33. [33]

    Speech recognition in adverse con- ditions by humans and machines,

    C. Patman and E. Chodroff, “Speech recognition in adverse con- ditions by humans and machines,”JASA Express Letters, vol. 4, 2024

    2024

  34. [34]

    AI-assisted analysis of phonological variation in English,

    V . Partridge, J. Pater, P. Bhangla, A. Nirheche, and B. Prickett, “AI-assisted analysis of phonological variation in English,” inSpecial session on Deep Phonology, Annual Meeting on Phonology, University of California Berkeley, 2025. [Online]. Available: https://github.com/ginic/wav2ipa

    2025

  35. [35]

    Phone-to-audio alignment without text: A semi-supervised approach,

    J. Zhu, C. Zhang, and D. Jurgens, “Phone-to-audio alignment without text: A semi-supervised approach,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 8167–8171

    2022

  36. [36]

    Tradition or innovation: A comparison of modern ASR methods for forced alignment,

    R. Rousso, E. Cohen, J. Keshet, and E. Chodroff, “Tradition or innovation: A comparison of modern ASR methods for forced alignment,” inInterspeech 2024, 2024, pp. 1525–1529

    2024

  37. [37]

    Spontaneous speech corpus of Japanese,

    K. Maekawa, H. Koiso, S. Furui, and H. Isahara, “Spontaneous speech corpus of Japanese,” inProceedings of the Second In- ternational Conference of Language Resources and Evaluation (LREC), vol. 2, 2000, pp. 946–952

    2000

  38. [38]

    Design and development of an RDB version of the Corpus of Spontaneous Japanese,

    H. Koiso, Y . Den, K. Nishikawa, and K. Maekawa, “Design and development of an RDB version of the Corpus of Spontaneous Japanese,” inProceedings of the Ninth International Conference on Language Resources and Evaluation, 2014, pp. 1471–1476

    2014

  39. [39]

    Shimizu,A cross-language study of the voicing contrasts of stop consonants in Asian languages

    K. Shimizu,A cross-language study of the voicing contrasts of stop consonants in Asian languages. Tokyo: Seibido, 1996

    1996

  40. [40]

    The intermediate degree of VOT in Japanese initial stops,

    T. J. Riney, N. Takagi, K. Ota, and Y . Uchida, “The intermediate degree of VOT in Japanese initial stops,”Journal of Phonetics, vol. 35, pp. 439–443, 2007

    2007

  41. [41]

    DARPA TIMIT acoustic phonetic continuous speech corpus CD-ROM,

    J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren, “DARPA TIMIT acoustic phonetic continuous speech corpus CD-ROM,” 1993

    1993

  42. [42]

    Changing sounds in a changing city: An acoustic phonetic investigation of real-time change over a century of Glaswegian,

    J. Stuart-Smith, B. Jose, T. Rathcke, R. MacDonald, and E. Law- son, “Changing sounds in a changing city: An acoustic phonetic investigation of real-time change over a century of Glaswegian,” inLanguage and a Sense of Place: Studies in Language and Re- gion, C. Montgomery and E. Moore, Eds. Cambridge: Cam- bridge University Press, 2017, pp. 38–65

    2017

  43. [43]

    Managing data for integrated speech corpus anal- ysis in SPeech Across Dialects of English (SPADE),

    M. Sonderegger, J. Stuart-Smith, M. McAuliffe, R. Macdonald, and T. Kendall, “Managing data for integrated speech corpus anal- ysis in SPeech Across Dialects of English (SPADE),” inOpen Handbook of Linguistic Data Management. Cambridge: MIT Press, 2022

    2022

  44. [44]

    SWITCH- BOARD: telephone speech corpus for research and development,

    J. J. Godfrey, E. C. Holliman, and J. McDaniel, “SWITCH- BOARD: telephone speech corpus for research and development,” inProceedings of the 1992 IEEE international conference on Acoustics, speech and signal processing - Volume 1, 1992, pp. 517–520

    1992

  45. [45]

    The medium-term dy- namics of accents on reality television,

    M. Sonderegger, M. Bane, and P. Graff, “The medium-term dy- namics of accents on reality television,”Language, vol. 93, pp. 598–640, 2017

    2017

  46. [46]

    Advanced Bayesian multilevel modeling with the R package brms,

    P.-C. B ¨urkner, “Advanced Bayesian multilevel modeling with the R package brms,”The R Journal, vol. 10, no. 1, pp. 395–411, 2018

    2018

  47. [47]

    R. V . Lenth,emmeans: Estimated Marginal Means, aka Least-Squares Means, 2024, r package version 1.10.0. [Online]. Available: https://CRAN.R-project.org/package=emmeans

    2024

  48. [48]

    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 Selected Topics in Signal Processing, vol. 16, no. 6, p. 1505–1518, Oct. 2022. [Online]....

    work page doi:10.1109/jstsp.2022.3188113 2022

  49. [49]

    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,” 2022. [Online]. Available: https: //arxiv.org/abs/2212.04356

    work page internal anchor Pith review Pith/arXiv arXiv 2022