Pith. sign in

REVIEW 3 major objections 4 minor 43 references

Word stress in self-supervised speech models: A cross-linguistic comparison

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Word stress is encoded in Wav2vec 2.0 XLS-R embeddings for five languages, with representations separated by fixed-versus-variable stress type.

desk verdict Nice cross-linguistic stress-probing study with an unaddressed position confound in the fixed-stress languages; the variable-stress results and the typological clustering claim are conditional until a position-only baseline is run. read the letter →

arxiv 2507.04738 v1 pith:RUJ2LUJI submitted 2025-07-07 cs.CL cs.AI

Martijn Bentum , Louis ten Bosch , Tomas O. Lentz This is my paper
classification cs.CLcs.AI
keywords wordstressself-supervisedspeechmodelsdiagnosticprobingwav2vec2.0cross-linguisticprosodyvariablevsfixedrepresentations
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

The paper sets out to show that a self-supervised speech model, Wav2vec 2.0 XLS-R, retains information about word stress when processing connected speech, not just isolated words. Diagnostic classifiers trained on the model's layer embeddings distinguish stressed from unstressed vowels with high accuracy in Dutch, English, German, Polish, and Hungarian, with the strongest signal at transformer layer 17. The authors also argue that these stress representations are language-specific: classifiers generalize worse across languages, and the similarity structure of classifier performance groups the variable-stress languages (Dutch, English, German) apart from the fixed-stress languages (Polish, Hungarian). If the claim holds, prosodic information is available in self-supervised speech embeddings and could support downstream tasks such as word segmentation and lexical access without dedicated stress annotations.

What carries the argument

The machinery is diagnostic probing. Vowel frames are extracted from several model layers (codevector, CNN, and transformer layers 5, 11, 17, and 23), mean-pooled, and fed to multilayer-perceptron classifiers that predict a syllable's stress label; performance is summarized by the Matthews correlation coefficient. Stress labels come from lexical lookup for the variable-stress languages and from positional rules for the fixed-stress languages. Language similarity is read off from the matrix of cross-lingual classifier performances using LDA and agglomerative hierarchical clustering, which is what yields the variable-versus-fixed grouping.

What would settle it

A concrete control: train the same MLP probes on the same layer-17 embeddings but with the label changed from stress to syllable position (first versus second syllable) for Polish and Hungarian. If the position-only classifier matches or exceeds the stress classifier's MCC, the fixed-stress results are explained by position or word-boundary encoding rather than by stress.

Watch

Extended reading notes

Core claim

The central claim is that word stress is represented in Wav2vec 2.0 XLS-R embeddings for all five languages, and that the representational geometry mirrors the typological distinction between variable and fixed stress. The authors show that simple multilayer-perceptron probes can classify stressed versus unstressed syllables from mean-pooled vowel embeddings at well above chance in all five languages, strongest at transformer layer 17, while standard acoustic correlates (duration, intensity, pitch, spectral tilt, formant position) are much weaker predictors, especially for Polish and Hungarian. When the same classifiers are applied across languages, performance drops, showing language-specificity. Hierarchical clustering and LDA on the cross-lingual performance vectors place the three variable-stress languages on one branch and the two fixed-stress languages on another, which the authors interpret as evidence that the model maintains language-specific stress representations organized by stress system type rather than a single cross-lingual stress code.

Load-bearing premise

In the fixed-stress languages the stress label is assigned by rule to the first syllable, so the high accuracy and the fixed-versus-variable split could be driven by syllable position or word-boundary cues instead of stress itself.

Editorial extensions

If this is right

  • Word stress can be recovered from self-supervised speech embeddings in read-aloud connected speech, so stress information survives sentence-level processing in Wav2vec 2.0 XLS-R.
  • Stress representations are not collapsed across languages; classifiers trained on one language lose accuracy on others, indicating language-specific encoding at every layer.
  • The model's stress geometry separates variable-stress languages from fixed-stress languages, matching a typological distinction that is otherwise visible in human stress deafness.
  • Conventional acoustic features are weak predictors of stress, particularly in Polish and Hungarian, so the high probe accuracy must come from contextual or representational structure beyond raw vowel acoustics.
  • The strongest stress signal at layer 17 and the reduced language-specificity at deeper layers suggest stress becomes more abstract at higher layers, consistent with the earlier isolated-word finding the paper extends.

Reading between the lines

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

  • A natural control the paper leaves open is a position-only classifier (first versus second syllable) for Polish and Hungarian, because their stress labels are identical to syllable position; if it matches the stress classifier's accuracy, the fixed-stress results would be explained by position or word-boundary encoding rather than stress.
  • The fixed-versus-variable split might be driven by how predictable the stress system is rather than its acoustic realization; testing a fixed-stress language with non-initial stress would show whether the fixed group clusters by predictability.
  • The strong probe accuracy at layer 17 suggests a weak-supervision route to stress annotation for languages that lack prosodic corpora, an application the paper does not claim.
  • Per-word or per-lemma analysis would show whether a few lexical items drive the stress-probe accuracy, testing whether the effect is truly a category-level property of stress.
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 / 4 minor

Summary. The paper investigates whether word stress is encoded in the self-supervised speech model Wav2vec 2.0 XLS-R, using diagnostic classifiers on vowel-level representations extracted from connected speech in five languages: Dutch, English, German (variable stress) and Polish, Hungarian (fixed stress). Stress labels come from CELEX for the variable-stress languages and from positional rules for Polish and Hungarian. The authors report high classification accuracy at transformer layer 17 across all languages, language-specific classifier performance, and hierarchical clustering that separates variable-stress from fixed-stress languages. They conclude that word stress is represented in the model and that those representations are language-specific, with a greater difference between variable- and fixed-stress language groups.

Significance. If the central claim holds, the paper would contribute to interpretability of self-supervised speech models by showing that abstract prosodic structure is recoverable from internal embeddings in connected speech, and that its organization mirrors a typological distinction. The study is well positioned within the probing literature and has several strengths: labels are drawn from an external lexical database and explicit rules, evaluation uses 20-fold cross-validation on held-out data, and the comparison across five languages is broader than most prior stress-probing work. The authors also make their rule-based labeling code available. However, the fixed-stress results are currently confounded with syllable position, and the variable-stress results are not controlled for vowel identity; these issues directly affect the paper's main claims.

major comments (3)
  1. [2.1 / 3.1 / 3.3] The fixed-stress labels are perfectly collinear with syllable position. For Polish and Hungarian, Section 2.1 assigns primary stress to the first syllable of every bisyllabic word, so the binary stress label is definitionally identical to 'first vowel versus second vowel.' Any representation in W2V layer 17 that encodes word boundaries, word onset, vowel position, or position-dependent phonotactics will therefore produce a high-MCC stress classifier, and the Section 3.1 claim that stress is represented in all five languages is not uniquely supported for these two languages. The Section 3.3 variable-versus-fixed clustering inherits the same problem, since the first split in Figure 3 depends on the fixed languages having high classification accuracy. The authors acknowledge this confound in Section 4, but they do not run the control they identify as necessary. I request a position-only baseline: train the same classifier pipeline (layer 17, MLP, 20-fold CV) to classify first versus second syllable for each language and report MCC; for the variable-stress languages, where SFS is 75--84%, this baseline also quantifies how much of the reported stress accuracy is generic position rather than stress. A stricter control would restrict evaluation to words or vowels that appear in both stress positions, where feasible. Without these controls, the fixed-stress accuracy and the variable/fixed clustering should not be interpreted as evidence about stress specifically.
  2. [2.3 / 3.1] There is no control for vowel identity in the variable-stress languages. In Dutch, English, and German, unstressed syllables are frequently reduced (e.g., to schwa) while stressed syllables are not, and the features are extracted exclusively from the vowel segment. A classifier could therefore distinguish stressed from unstressed syllables by recognizing vowel categories rather than prominence. This directly affects the claim that stress is represented in the model for all three variable-stress languages. I request a control that conditions on vowel identity, for example by including the vowel label as an additional feature, by testing only on vowel tokens that appear in both stressed and unstressed positions, or by evaluating on stress minimal pairs such as Dutch /'kaːnɔn/ versus /kaː'nɔn/. Without such a control, the variable-language results overstate the stress-specific content of the embeddings.
  3. [3.1 / 4] The acoustic-feature comparison for fixed-stress languages is similarly exposed to the position confound. Since Polish and Hungarian stress is always on the first syllable, acoustic features such as duration and intensity are also correlated with syllable position, and the paper's finding that acoustic cues are weaker in fixed-stress languages (Section 3.1, Figure 1) is not cleanly interpretable without a position baseline. A concrete control would be to train the same acoustic classifiers to predict first versus second syllable for each language and compare those MCC values with the stress-classification MCCs. This would show whether the apparent cross-language difference in acoustic cue reliability is an artifact of different label--position correlations. This issue is secondary to the main W2V claim, but it feeds into the Section 4 discussion and should be addressed.
minor comments (4)
  1. [2.2] The intensity formula is hard to read: '10 log10(x2/4 ∗ 10−10)' lacks parentheses and does not define x or the reference value; please clarify the formula and the units.
  2. [2.3] The MLP classifier is not specified beyond 'multilayer Perceptron as implemented in Scikit-learn.' For reproducibility, please report the architecture, hidden-layer sizes, activation, regularization, and number of training epochs, or state that default settings were used.
  3. [3.3] The text refers to 'The LDA results (section 2.3)' but LDA on performance vectors is described in Section 2.4; please correct the cross-reference.
  4. [References] There are a few typographical issues in the reference list: 'Needlemann-Wunch' should be 'Needleman-Wunsch', and 'V on Platen' should be 'Von Platen'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: stress labels come from external CELEX lookups and explicit positional rules; classifiers are evaluated on held-out folds, and the acknowledged word-boundary confound is a validity threat rather than a circular derivation.

full rationale

The paper's central derivation is not circular. Ground-truth stress labels are obtained from an external lexical database (CELEX) for Dutch, English, and German, and from explicit phonological rules for Polish and Hungarian (Section 2.1). Classifier performance is measured with 20-fold cross-validation on held-out data (Section 2.3), so the reported MCC values are not produced by fitting the target labels into the model. The language-specificity and variable-vs-fixed comparisons are computed from these held-out classifier performance vectors (Sections 2.4, 3.2, 3.3), not from the labels themselves. The only self-citation to the authors' prior work ([9], used for context and as a comparison of stress representations in isolated English words) is not load-bearing: the present results for five languages in connected speech are evaluated independently. The main validity concern is the perfect collinearity of fixed-stress labels with syllable position in Polish and Hungarian, which the authors themselves flag in Section 4 ('it could serve as a highly reliable predictor of word stress in fixed stress languages and would be a confounding factor'). That is a potential confound or ambiguity about what the classifier learns, not a circular reduction of the paper's outputs to its inputs; the labels still originate outside the model and the evaluation is held-out. No step in the derivation chain reduces by construction to its own input.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No free parameters are fitted to data in a derivation sense. The central claim rests on external lexical labels, rule-based fixed-stress labels, forced alignment, and hand-chosen analysis thresholds. The most consequential assumption is the positional rule for Polish and Hungarian, which ties stress labels to syllable position.

free parameters (2)
  • MLP classifier hyperparameters = not reported
    Architecture, hidden units, regularization, and training schedule for the stress MLP classifiers are unspecified (Section 2.3); classification performance could depend on these choices.
  • Vowel frame overlap threshold = 50%
    Frames with at least 50% overlap with a vowel were retained and mean-pooled (Section 2.2); this threshold is chosen by hand and affects the embeddings fed to classifiers.
assumptions (5)
  • domain assumption CELEX stress labels are accurate for the Common Voice recordings of Dutch, English, and German.
    Section 2.1: lexical stress status is taken from CELEX and aligned to forced-aligned transcriptions; alignment or lexicon errors would propagate to labels.
  • domain assumption All Polish and Hungarian bisyllabic words carry primary stress on the first syllable.
    Section 2.1: rule-based labeling; exceptions and stress shift are ignored, and the rule makes stress equivalent to syllable position.
  • domain assumption MAUS forced alignment provides accurate vowel and syllable boundaries.
    Section 2.1: boundaries drive which frames are pooled into vowel embeddings and all acoustic duration measures.
  • domain assumption Mean-pooled frames with 50% vowel overlap represent the syllable's stress-relevant content in W2V.
    Section 2.2: no check that pooling preserves stress cues or that the overlap criterion is sufficient.
  • domain assumption MCC values are comparable across languages despite different class imbalances and corpus sizes.
    Sections 2.3-3: performance comparisons pool MCC across languages with different stress distributions (75% to 100% first-syllable stress); MCC is class-imbalance robust, but language differences in corpus size and speaker diversity are not controlled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Word stress in self-supervised speech models: A cross-linguistic comparison." pith.science (2026). https://pith.science/paper/RUJ2LUJI

@misc{pith2026250704738,
  author       = {Pith},
  title        = {Pith review of: Word stress in self-supervised speech models: A cross-linguistic comparison},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RUJ2LUJI}},
  note         = {Machine review of arXiv:2507.04738}
}
read the original abstract

In this paper we study word stress representations learned by self-supervised speech models (S3M), specifically the Wav2vec 2.0 model. We investigate the S3M representations of word stress for five different languages: Three languages with variable or lexical stress (Dutch, English and German) and two languages with fixed or demarcative stress (Hungarian and Polish). We train diagnostic stress classifiers on S3M embeddings and show that they can distinguish between stressed and unstressed syllables in read-aloud short sentences with high accuracy. We also tested language-specificity effects of S3M word stress. The results indicate that the word stress representations are language-specific, with a greater difference between the set of variable versus the set of fixed stressed languages.

Figures

Figures reproduced from arXiv: 2507.04738 by the authors.

Figure 1
Figure 1. The performance in terms of Matthew’s correlation coefficient (MCC) of classifiers trained on acoustic correlates of stress and various layers of the W2V model for Dutch, German, English, Polish and Hungarian. Results are shown for duration (dur), intensity (int), pitch (pit), formants (for), spectral-tilt (st), combined-features (cf), codevectors (cv), cnn and transformer layers 5 - 23. Error bars contain the 99% C… view at source ↗
Figure 4
Figure 4. The first six panels show the acoustic feature results, [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 3
Figure 3. Resulting clusters from applying Agglomerative Hierarchical Clustering on the classifier performance vectors based on the best performing acoustic correlate (left) and best performing transformer layer (right). 0 10 2.5 0.0 2.5 5.0 7.5 duration Dutch English German Polish Hungarian 0 5 2 0 2 intensity 10 0 10 5.0 2.5 0.0 2.5 5.0 pitch 0 10 2 0 2 formant 5 0 5 4 2 0 2 4 spectral-tilt 2.5 0.0 2.5 4 2 0 2 combined 25 0… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: LDA-scatterplots based on language specific and cross-lingual performance of stress classifiers for acoustic cor￾relates, codevectors, CNN output and transformer layers. The x- and y-axes represent the first two linear discriminants respec￾tively. Each dot represents 1…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages

  1. [1]

    Word stress in self-supervised speech models: A cross-linguistic comparison

    Introduction Self-supervised speech models (S3Ms) learn to represent spo- ken language and can be effectively fine-tuned for downstream tasks such as automatic speech recognition [1], speaker identi- fication [2], and emotion recognition [3]. However, their end- to-end nature makes their inner workings difficult to interpret. One approach to improve inter...

    work page Pith review arXiv 2025

  2. [2]

    This model is pre-trained on 500,000 hours of speech recordings across 128 languages, including the five languages featuring in our study

    Methods We used the multilingual Wav2vec 2.0 XLS-R model 1[26], henceforth denoted W2V . This model is pre-trained on 500,000 hours of speech recordings across 128 languages, including the five languages featuring in our study. 2.1. Materials We used language materials from the Common V oice corpus [27], which consists of recordings of short read-aloud se...

    work page

  3. [3]

    Word stress representations in connected speech The results for the language-specific stress classification test are summarized in Figure 1

    Results 3.1. Word stress representations in connected speech The results for the language-specific stress classification test are summarized in Figure 1. The error bars indicate the 99% inter- val of the mean stress classification performance. For each of the five languages (Dutch, English, German, Polish and Hun- garian), we found evidence for stress rep...

    work page

  4. [4]

    Our study fo- cused on three languages with variable (i.e

    Discussion We examined word stress representations in a self-supervised speech model (S3M) across five languages. Our study fo- cused on three languages with variable (i.e. lexical) word stress (Dutch, English, and German) and two with fixed (i.e. demarca- tive) word stress (Polish and Hungarian). Using short, read- aloud sentences, we applied the model a...

    work page

  5. [5]

    Zuidema (Univ

    Acknowledgements All authors participate in the Dutch NWO/NW A project In- Deep (https://www.nwo.nl/en/projects/nwa129219399), led by J. Zuidema (Univ. of Amsterdam)

    work page

  6. [6]

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

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

    2020

  7. [7]

    Large-scale self-supervised speech representation learning for automatic speaker verification,

    Z. Chen, S. Chen, Y . Wu, Y . Qian, C. Wang, S. Liu, Y . Qian, and M. Zeng, “Large-scale self-supervised speech representation learning for automatic speaker verification,” in ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6147–6151

    work page 2022

  8. [8]

    Evaluating self-supervised speech representations for speech emotion recognition,

    B. T. Atmaja and A. Sasou, “Evaluating self-supervised speech representations for speech emotion recognition,” IEEE Access , vol. 10, pp. 124 396–124 407, 2022

    work page 2022

Show all 43 references
  1. [9]

    Understanding intermediate layers using linear classifier probes,

    G. Alain and Y . Bengio, “Understanding intermediate layers using linear classifier probes,” arXiv preprint arXiv:1610.01644, 2018

    2018 arXiv

  2. [10]

    What you can cram into a single $&!#* vector: Probing sentence embeddings for linguistic properties,

    A. Conneau, G. Kruszewski, G. Lample, L. Barrault, and M. Ba- roni, “What you can cram into a single $&!#* vector: Probing sentence embeddings for linguistic properties,” in Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018, pp. 2126–2136

    2018

  3. [11]

    Domain- informed probing of wav2vec 2.0 embeddings for phonetic fea- tures,

    P. C. English, J. Kelleher, and J. Carson-Berndsen, “Domain- informed probing of wav2vec 2.0 embeddings for phonetic fea- tures,” in Proceedings of the 19th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphol- ogy, 2022, pp. 83–91

    2022

  4. [12]

    Self-supervised speech representations are more phonetic than semantic,

    K. Choi, A. Pasad, T. Nakamura, S. Fukayama, K. Livescu, and S. Watanabe, “Self-supervised speech representations are more phonetic than semantic,” in Interspeech 2024, 2024, pp. 4578– 4582

    2024

  5. [13]

    Wave to syn- tax: Probing spoken language models for syntax,

    G. Shen, A. Alishahi, A. Bisazza, and G. Chrupała, “Wave to syn- tax: Probing spoken language models for syntax,” in Interspeech 2023, 2023, pp. 1259–1263

    2023

  6. [14]

    The processing of stress in end-to-end automatic speech recognition models,

    M. Bentum, L. ten Bosch, and T. Lentz, “The processing of stress in end-to-end automatic speech recognition models,” in Proc. In- terspeech 2024, 2024, pp. 2350–2354

    2024

  7. [15]

    Gussenhoven, The phonology of tone and intonation

    C. Gussenhoven, The phonology of tone and intonation . Cam- bridge University Press, 2004

    2004

  8. [16]

    Rounds, Hungarian: An essential grammar

    C. Rounds, Hungarian: An essential grammar. Routledge, 2009

    2009

  9. [17]

    Booij, The phonology of Dutch

    G. Booij, The phonology of Dutch . Oxford University Press, 1999

    1999

  10. [18]

    The role of strong syllables in segmen- tation for lexical access

    A. Cutler and D. Norris, “The role of strong syllables in segmen- tation for lexical access.” Journal of Experimental Psychology: Human perception and performance, vol. 14, no. 1, p. 113, 1988

    1988

  11. [19]

    Bossong, B

    G. Bossong, B. Comrie, Y . Matras, and H. v. d. Hulst, Word prosodic systems in the languages of Europe. Mouton de Gruyter, 1999

    1999

  12. [20]

    Acoustic correlates and perceptual cues of word and sentence stress,

    V . J. van Heuven, “Acoustic correlates and perceptual cues of word and sentence stress,” in The Study of Word Stress and Ac- cent: Theories, Methods and Data , R. Goedemans, J. Heinz, and H. van der Hulst, Eds. Cambridge University Press, 2018, pp. 15–59

    2018

  13. [21]

    Duration and intensity as physical correlates of lin- guistic stress,

    D. B. Fry, “Duration and intensity as physical correlates of lin- guistic stress,” Journal of the Acoustical Society of America , vol. 27, no. 4, pp. 765–768, 1955

    1955

  14. [22]

    Spectral balance as an acoustic correlate of linguistic stress,

    A. M. Sluijter and V . J. Van Heuven, “Spectral balance as an acoustic correlate of linguistic stress,” Journal of the Acoustical society of America, vol. 100, no. 4, pp. 2471–2485, 1996

    1996

  15. [23]

    Acoustic correlates of word stress: A cross-linguistic survey,

    M. Gordon and T. Roettger, “Acoustic correlates of word stress: A cross-linguistic survey,” Linguistics Vanguard, vol. 3, no. 1, p. 20170007, 2017

    2017

  16. [24]

    Articulation rate and the duration of syllables and stress groups in connected speech,

    T. H. Crystal and A. S. House, “Articulation rate and the duration of syllables and stress groups in connected speech,”Journal of the Acoustical Society of America, vol. 88, no. 1, pp. 101–112, 1990

    1990

  17. [25]

    Coarticulation and connected speech processes,

    E. Farnetani and D. Recasens, “Coarticulation and connected speech processes,” Handbook of phonetic sciences , vol. 371, p. 404, 1997

    1997

  18. [26]

    Lexical stress in Polish: evidence from focus and phrase-position differentiated production data,

    Z. Malisz and M. Zygis, “Lexical stress in Polish: evidence from focus and phrase-position differentiated production data,” in Pro- ceedings of the 9th International Conference on Speech Prosody, Poznan, Poland, 2018, pp. 13–16

    2018

  19. [27]

    Perception of pre- dictable stress: A cross-linguistic investigation,

    S. Peperkamp, I. Vendelin, and E. Dupoux, “Perception of pre- dictable stress: A cross-linguistic investigation,” Journal of Pho- netics, vol. 38, no. 3, pp. 422–430, 2010

    2010

  20. [28]

    Wiese, The phonology of German

    R. Wiese, The phonology of German. Oxford University Press, 2000

    2000

  21. [29]

    McMahon, An introduction to English phonology

    A. McMahon, An introduction to English phonology. Edinburgh: University Press, 2020

    2020

  22. [30]

    A grid theory of stress in Polish,

    J. Rubach and G. E. Booij, “A grid theory of stress in Polish,” Lingua, vol. 66, no. 4, pp. 281–320, 1985

    1985

  23. [31]

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

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. V on Platen, Y . Saraf, J. Pino et al. , “Xls-r: Self- supervised cross-lingual speech representation learning at scale,” arXiv preprint arXiv:2111.09296, 2021

    2021 arXiv

  24. [32]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,”arXiv preprint arXiv:1912.06670, 2019

    1912 arXiv

  25. [33]

    Multilingual processing of speech via web services,

    T. Kisler, U. Reichel, and F. Schiel, “Multilingual processing of speech via web services,”Computer Speech & Language, vol. 45, pp. 326–347, 2017

    2017

  26. [34]

    The CELEX lexical database (cd-rom),

    R. H. Baayen, R. Piepenbrock, and L. Gulikers, “The CELEX lexical database (cd-rom),” 1996

    1996

  27. [35]

    A general method applicable to the search for similarities in the amino acid sequence of two proteins,

    S. B. Needleman and C. D. Wunsch, “A general method applicable to the search for similarities in the amino acid sequence of two proteins,” Journal of Molecular Biology, vol. 48, no. 3, pp. 443– 453, 1970

    1970

  28. [36]

    librosa: Audio and music signal analysis in python,

    B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Batten- berg, and O. Nieto, “librosa: Audio and music signal analysis in python,” in Proceedings of the 14th python in science conference, vol. 8, 2015, pp. 18–25

    2015

  29. [37]

    Comparison of the predicted and observed sec- ondary structure of T4 phage lysozyme,

    B. W. Matthews, “Comparison of the predicted and observed sec- ondary structure of T4 phage lysozyme,” Biochimica et Biophys- ica Acta (BBA)-Protein Structure , vol. 405, no. 2, pp. 442–451, 1975

    1975

  30. [38]

    Assessing the accuracy of prediction algorithms for classifica- tion: an overview,

    P. Baldi, S. Brunak, Y . Chauvin, C. A. Andersen, and H. Nielsen, “Assessing the accuracy of prediction algorithms for classifica- tion: an overview,” Bioinformatics, vol. 16, no. 5, pp. 412–424, 2000

    2000

  31. [39]

    The advantages of the Matthews cor- relation coefficient (MCC) over F1 score and accuracy in binary classification evaluation,

    D. Chicco and G. Jurman, “The advantages of the Matthews cor- relation coefficient (MCC) over F1 score and accuracy in binary classification evaluation,”BMC genomics, vol. 21, no. 1, pp. 1–13, 2020

    2020

  32. [40]

    Layer-wise analysis of a self-supervised speech representation model,

    A. Pasad, J.-C. Chou, and K. Livescu, “Layer-wise analysis of a self-supervised speech representation model,” in2021 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 914–921

    2021

  33. [41]

    Wav2vec behind the scenes: How end2end models learn phonet- ics,

    T. tom Dieck, P. A. P´erez-Toro, T. Arias, E. Noeth, and P. Klumpp, “Wav2vec behind the scenes: How end2end models learn phonet- ics,” in Interspeech 2022, 2022, pp. 5130–5134

    2022

  34. [42]

    An information-theoretic analysis of self-supervised discrete repre- sentations of speech,

    B. M. Abdullah, M. M. Shaik, B. M ¨obius, and D. Klakow, “An information-theoretic analysis of self-supervised discrete repre- sentations of speech,” in Interspeech 2023, 2023, pp. 2883–2887

    2023

  35. [43]

    Phonemic competition in end-to-end ASR models,

    L. ten Bosch, M. Bentum, and L. Boves, “Phonemic competition in end-to-end ASR models,” in Interspeech 2023, 2023, pp. 586– 590

    2023

Pith tools

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