The reviewed record of science sign in
Pith

arxiv: 2605.25596 · v1 · pith:NG6RA3JL · submitted 2026-05-25 · cs.CL

Multilingual Phonological Feature Recognition with Self-Supervised Speech Models

Abner Hernandez , Tom\'as Arias-Vergara , Daiqi Liu , Andreas Maier , Paula Andrea P\'erez-Toro This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-29 21:32 UTCgrok-4.3pith:NG6RA3JLrecord.jsonopen to challenge →

classification cs.CL
keywords phonological featuresself-supervised speech modelsmultilingual recognitionframe-level predictionmanner-conditioned gatingphoneme baseline comparisonunseen language evaluation
0
0 comments X

The pith

PhonoQ-2.0 directly predicts a 22-dimensional phonological feature vector from self-supervised speech models and exceeds a phoneme baseline by eight points on average.

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

The paper introduces PhonoQ-2.0 to recognize phonological features such as manner, place, vowel quality, and voicing directly at each frame of audio rather than deriving them after phoneme recognition. It builds the recognizer on self-supervised speech models and adds a gating step that conditions later features on the predicted manner to keep outputs phonologically valid. The method records macro-F1 scores of 91.3 percent in-domain and 88.9 percent out-of-domain while beating the baseline by roughly nine points in both settings. Gains also appear when the system is tested on languages absent from training data. A reader would care because the results indicate that phonological features can serve as a more language-general representation of speech than phoneme sequences.

Core claim

PhonoQ-2.0 predicts phonological features directly rather than deriving them from phoneme sequences. The model achieves an average macro-F1 of 91.3 percent in-domain and 88.9 percent out-of-domain, with gains of 8.8 and 8.6 points over a strong CTC phoneme baseline. In unseen-language tests the improvement reaches 6.7 points on average.

What carries the argument

The manner-conditioned gating mechanism that activates only valid combinations of features based on the predicted manner class.

If this is right

  • Direct frame-level prediction of the 22-dimensional feature vector yields consistent accuracy gains over phoneme-based derivation.
  • The gains hold for both familiar languages and languages never seen during training.
  • Self-supervised speech models supply enough signal to support feature prediction without an intermediate phoneme step.
  • The gating step maintains phonological coherence while preserving coverage of valid feature groups.

Where Pith is reading between the lines

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

  • If the gains persist, the method could reduce dependence on language-specific phoneme inventories in downstream speech tools.
  • Similar conditioning on one feature group could be tested for other structured speech attributes such as prosody.
  • Measuring prediction consistency on a wider set of language families would show how far the language-general property extends.

Load-bearing premise

The manner-conditioned gating produces phonologically coherent predictions without adding systematic bias or missing valid feature combinations.

What would settle it

An evaluation on new languages where the full model with gating shows macro-F1 no higher than the phoneme baseline or produces invalid feature combinations at a higher rate than a version without gating.

Figures

Figures reproduced from arXiv: 2605.25596 by Abner Hernandez, Andreas Maier, Daiqi Liu, Paula Andrea P\'erez-Toro, Tom\'as Arias-Vergara.

Figure 1
Figure 1. Figure 1: Pipeline comparison and PhonoQ-2.0 architecture. Left: CTC-Phoneme predicts phonemes, which are mapped to phonological classes. PhonoQ-2.0 predicts them directly from speech. Right: a fine-tuned XLSR encoder feeds a 2-layer Conformer with structured classification heads (manner, vowel, place, voice) and conditional decoding. 2.3. Phonetic Baseline with Class Mapping We fine-tune the same pretrained XLSR-ft… view at source ↗
Figure 2
Figure 2. Figure 2: Per-feature segment-level F1 (%) on the in-domain CV-test set for CTC-Phoneme and PhonoQ-2.0. 0.4 0.5 0.6 0.7 0.8 0.9 0.5 0.6 0.7 0.8 0.9 Precision 0.50 0.60 0.70 0.75 0.80 0.85 0.90 0.95 stop nasal rhotic fricative affricate approximant lateral vowel Manner 0.80 0.85 0.90 0.95 0.800 0.825 0.850 0.875 0.900 0.925 0.950 0.80 0.85 0.90 0.95 high mid low central front back Vowel 0.65 0.70 0.75 0.80 0.85 0.90 … view at source ↗
Figure 3
Figure 3. Figure 3: Precision and recall per phonological feature on the in-domain CV-test. Arrows connect CTC-Phoneme to PhonoQ￾2.0. tion transfers more robustly than post-hoc feature mapping un￾der cross-lingual shift ( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
read the original abstract

Phonological features provide a language-general and linguistically grounded representation of speech. We present PhonoQ-2.0, a multilingual frame-level phonological feature recognizer built on self-supervised speech models. The system directly predicts a structured 22-dimensional feature vector per frame encoding manner, vowel quality, place, and voicing, instead of deriving features from phoneme outputs. To ensure phonologically coherent predictions, we introduce a manner-conditioned gating mechanism that activates valid feature groups. Evaluated across multiple languages and corpora, PhonoQ-2.0 achieves an average macro-F1 of 91.3% in-domain and 88.9% out-of-domain. Compared to a strong CTC phoneme baseline, it delivers consistent gains of +8.8 F1 in-domain and +8.6 out-of-domain on average. In unseen-language evaluation, PhonoQ-2.0 improves macro-F1 from 66.9% to 73.6% (+6.7 on average), with gains of up to +10.8 points.

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

2 major / 2 minor

Summary. The paper introduces PhonoQ-2.0, a multilingual frame-level phonological feature recognizer built on self-supervised speech models. It directly predicts a structured 22-dimensional feature vector per frame (manner, vowel quality, place, voicing) rather than deriving features from phoneme outputs, and introduces a manner-conditioned gating mechanism to enforce phonological coherence. The system is evaluated on multiple languages and corpora, reporting average macro-F1 of 91.3% in-domain and 88.9% out-of-domain, with gains of +8.8 and +8.6 F1 over a CTC phoneme baseline; unseen-language evaluation shows improvement from 66.9% to 73.6% (+6.7 on average).

Significance. If the performance claims are supported by complete evaluation protocols and ablations, the work would offer a linguistically grounded alternative to phoneme-based pipelines for multilingual speech processing, with potential benefits for cross-lingual generalization and feature-level downstream tasks. The direct prediction of structured features and the gating idea are conceptually appealing, though the manuscript provides no machine-checked proofs, parameter-free derivations, or reproducible code artifacts to strengthen the contribution.

major comments (2)
  1. [Abstract / Evaluation] Abstract and evaluation sections: the reported macro-F1 gains (+8.8 in-domain, +6.7 unseen-language) are presented without any description of data splits, training protocols for the CTC baseline, error bars, statistical significance tests, or exact definitions of in-domain/out-of-domain/unseen-language partitions. These omissions make it impossible to assess whether the central performance claims are robust.
  2. [Model Architecture] Model description (manner-conditioned gating): the gating mechanism is introduced explicitly 'to ensure phonologically coherent predictions,' yet the results consist only of aggregate macro-F1 scores. No ablation removing the gate, no count of invalid feature combinations (e.g., impossible manner-place or voicing pairs), and no separate coherence metric are supplied, so it is unclear whether the reported gains depend on this component or would be obtained by the self-supervised backbone plus a direct 22-dimensional head alone.
minor comments (2)
  1. [Abstract] The abstract states 'consistent gains' but does not clarify whether the +8.8 / +8.6 figures are macro-averaged across all languages or weighted; a table breaking down per-language deltas would improve clarity.
  2. [Model Architecture] Notation for the 22-dimensional feature vector and the gating function should be formalized with an equation rather than prose description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript to improve clarity on evaluation protocols and to include additional analyses of the gating mechanism.

read point-by-point responses

  1. Referee: [Abstract / Evaluation] Abstract and evaluation sections: the reported macro-F1 gains (+8.8 in-domain, +6.7 unseen-language) are presented without any description of data splits, training protocols for the CTC baseline, error bars, statistical significance tests, or exact definitions of in-domain/out-of-domain/unseen-language partitions. These omissions make it impossible to assess whether the central performance claims are robust.

    Authors: We agree that the abstract and high-level evaluation overview lack sufficient detail on these elements. The full experimental section describes the corpora, language partitions (in-domain: same languages/corpora as training; out-of-domain: different corpora but overlapping languages; unseen: completely held-out languages), and baseline training (CTC phoneme model trained on the same multilingual data with identical self-supervised backbone). In revision we will add a concise protocol summary to the abstract and evaluation sections, report standard deviations across runs, and include paired statistical significance tests (e.g., McNemar or bootstrap) for the reported F1 differences. revision: yes

  2. Referee: [Model Architecture] Model description (manner-conditioned gating): the gating mechanism is introduced explicitly 'to ensure phonologically coherent predictions,' yet the results consist only of aggregate macro-F1 scores. No ablation removing the gate, no count of invalid feature combinations (e.g., impossible manner-place or voicing pairs), and no separate coherence metric are supplied, so it is unclear whether the reported gains depend on this component or would be obtained by the self-supervised backbone plus a direct 22-dimensional head alone.

    Authors: The gating mechanism conditions feature prediction on manner to block invalid combinations at inference. While aggregate F1 is the primary metric, we acknowledge the absence of an explicit ablation and coherence analysis. In the revision we will add (1) an ablation comparing the gated model against an ungated 22-dimensional head on the same backbone, (2) a coherence metric reporting the rate of phonologically invalid feature tuples (e.g., nasal + fricative manner conflicts), and (3) per-feature-group F1 to isolate the gate's contribution. These additions will clarify whether the gains are attributable to the gating component. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical metrics on held-out data

full rationale

The paper reports macro-F1 scores on explicitly held-out in-domain, out-of-domain, and unseen-language splits, with comparisons to a CTC phoneme baseline. These are direct empirical measurements on independent test data rather than quantities derived from the model definition or fitted parameters by construction. The manner-conditioned gating is presented as an architectural choice whose contribution is reflected in aggregate performance numbers; no equations, self-citations, or uniqueness claims reduce the reported gains (+8.8 in-domain, +6.7 unseen) to the inputs themselves. The derivation chain consists of standard supervised training and evaluation and remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The approach rests on the assumption that self-supervised speech models already encode phonological information and that the introduced gating mechanism adds value without side effects; no free parameters beyond standard model training are specified in the abstract.

axioms (1)
  • domain assumption Self-supervised speech models encode sufficient phonological information to support direct feature prediction
    The system is built directly on these models without additional justification in the abstract.
invented entities (1)
  • manner-conditioned gating mechanism no independent evidence
    purpose: Activates valid feature groups to ensure phonologically coherent predictions
    New component introduced in the paper; no independent evidence of its necessity or effect is provided in the abstract.

pith-pipeline@v0.9.1-grok · 5728 in / 1448 out tokens · 34723 ms · 2026-06-29T21:32:58.300898+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

35 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Most downstream systems therefore focus on phoneme prediction or end-to-end word modeling

    Introduction Self-supervised learning (SSL)-based speech models have made phoneme recognition highly accurate across many lan- guages [1]. Most downstream systems therefore focus on phoneme prediction or end-to-end word modeling. In such pipelines, phonological information is derived indirectly: mod- els predict phoneme sequences, and phonological feature...

  2. [2]

    Method 2.1. Phonological Feature Representation We compare PhonoQ-2.0, a structured multi-head phonological recognizer that predicts features directly, andCTC-Phoneme, a phoneme recognizer that derives phonological classes via post-hoc mapping. Both systems share the same pretrained acoustic backbone. PhonoQ-2.0 represents each frame as a 22- arXiv:2605.2...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Datasets We train and evaluate on four languages spanning three fam- ilies: Germanic (English, German), Romance (Spanish), and Slavic (Czech)

    Experimental Setup 3.1. Datasets We train and evaluate on four languages spanning three fam- ilies: Germanic (English, German), Romance (Spanish), and Slavic (Czech). ForSpanish, training data combine 14.5h of CommonPhone (CP) [18] and a Latin American Spanish corpus comprising publicly available speech from Argentina, Chile, Colombia, Peru, Puerto Rico, ...

  4. [4]

    Mixed-precision (fp16) training is applied, and model selec- tion is based on development phone error rate (PER). 3.3. Evaluation Protocol PhonoQ-2.0 is trained to predict structured phonological fea- ture vectors at the frame level using MFA-aligned TextGrids converted to 50 fps representations. Although supervision is frame-aligned, all results reported...

  5. [5]

    All evaluations are conducted at the phone-segment level within the shared phonological feature space

    Results We evaluate PhonoQ-2.0 and CTC-Phoneme under three in- creasingly challenging conditions:(1) In-domain(CP and CV test splits),(2) Out-of-domain(FLEURS, V oxPopuli), and(3) Zero-shot cross-lingual(French, Italian, Russian from CP, un- seen during training). All evaluations are conducted at the phone-segment level within the shared phonological feat...

  6. [6]

    2) show that improvements ex- tend across the entire phonological inventory rather than be- ing concentrated in a single category

    Per-Feature Analysis Segment-level F1 scores (Fig. 2) show that improvements ex- tend across the entire phonological inventory rather than be- ing concentrated in a single category. Gains are observed for all major manner classes, across vowel height and backness di- mensions, and for each place contrast, while even voicing, al- ready strong in the baseli...

  7. [7]

    Discussion & Conclusions We presented PhonoQ-2.0, a multilingual frame-level phono- logical feature recognizer built on a pretrained XLSR wav2vec 2.0 encoder. Importantly, CTC-Phoneme is a strong phonetic baseline, achieving an average 6.80% PER yet its post-hoc mapped phonological features fall behind PhonoQ- 2.0 across all languages and conditions. This...

  8. [8]

    Generative AI Use Disclosure Generative artificial intelligence tools were used solely to assist with language editing and clarity of presentation. All research ideas, methodology, experiments, and interpretations were con- ceived and carried out by the authors, who take full responsibil- ity for the originality, validity, and integrity of the work

  9. [9]

    Unsupervised Cross-Lingual Representation Learning for Speech Recognition,

    A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli, “Unsupervised Cross-Lingual Representation Learning for Speech Recognition,” inInterspeech 2021, 2021, pp. 2426– 2430

    2021

  10. [10]

    Panphon: A resource for mapping ipa segments to articulatory feature vectors,

    D. R. Mortensen, P. Littell, A. Bharadwaj, K. Goyal, C. Dyer, and L. Levin, “Panphon: A resource for mapping ipa segments to articulatory feature vectors,” inProceedings of COLING 2016, the 26th international conference on computational linguistics: Technical papers, 2016, pp. 3475–3484

    2016

  11. [11]

    Leveraging allophony in self-supervised speech models for atyp- ical pronunciation assessment,

    K. Choi, E. Yeo, K. Chang, S. Watanabe, and D. R. Mortensen, “Leveraging allophony in self-supervised speech models for atyp- ical pronunciation assessment,” inProceedings of the 2025 Con- ference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2025, pp. 2613–2628

    2025

  12. [12]

    Self-supervised models of speech infer universal articulatory kinematics,

    C. J. Cho, A. Mohamed, A. W. Black, and G. K. Anumanchipalli, “Self-supervised models of speech infer universal articulatory kinematics,” inICASSP 2024-2024 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12 061–12 065

    2024

  13. [13]

    The internal organization of speech sounds,

    G. N. Clements and E. V . Hume, “The internal organization of speech sounds,” inThe handbook of phonological theory/ed. by John A. Goldsmith. Blackwell, 1995, pp. 245–306

    1995

  14. [14]

    Phonet: A Tool Based on Gated Recurrent Neural Net- works to Extract Phonological Posteriors from Speech,

    J. V ´asquez-Correa, P. Klumpp, J. R. Orozco-Arroyave, and E. N¨oth, “Phonet: A Tool Based on Gated Recurrent Neural Net- works to Extract Phonological Posteriors from Speech,” inInter- speech 2019, 2019, pp. 549–553

    2019

  15. [15]

    Phonological Feature Detection for US English using the Phonet Library,

    H. V . Tadavarthy, A. Jones, and M. E. L. Renwick, “Phonological Feature Detection for US English using the Phonet Library,” in Interspeech 2024, 2024, pp. 1515–1519

    2024

  16. [16]

    Weakly supervised phonological features for pathological speech analysis,

    J. Thienpondt, G. Vanderreydt, A. Hammami, and K. Demuynck, “Weakly supervised phonological features for pathological speech analysis,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  17. [17]

    Leveraging ipa and artic- ulatory features as effective inductive biases for multilingual asr training,

    J. Lee, M. Mimura, and T. Kawahara, “Leveraging ipa and artic- ulatory features as effective inductive biases for multilingual asr training,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  18. [18]

    Improving cross-lingual phonetic representation of low-resource languages through language sim- ilarity analysis,

    M. Kim, K. Jang, and H. Kim, “Improving cross-lingual phonetic representation of low-resource languages through language sim- ilarity analysis,” inICASSP 2025-2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  19. [19]

    Phonological features in discriminative classification of dysarthric speech,

    F. Rudzicz, “Phonological features in discriminative classification of dysarthric speech,” in2009 IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2009, pp. 4605– 4608

    2009

  20. [20]

    Measuring phonological precision in children with cleft lip and palate

    T. Arias-Vergara, E. Londo ˜no-Mora, P. A. P´erez-Toro, M. Schus- ter, E. N ¨oth, J. R. Orozco-Arroyave, and A. Maier, “Measuring phonological precision in children with cleft lip and palate.” in INTERSPEECH, 2023, pp. 4638–4642

    2023

  21. [21]

    Con- trastive learning approach for assessment of phonological preci- sion in patients with tongue cancer using mri data,

    T. Arias-Vergara, P. A. P ´erez-Toro, X. Liu, F. Xing, M. Stone, J. Zhuo, J. L. Prince, M. Schuster, E. N ¨oth, J. Wooet al., “Con- trastive learning approach for assessment of phonological preci- sion in patients with tongue cancer using mri data,” inInterspeech, vol. 2024, 2024, p. 927

    2024

  22. [22]

    Audio–vision contrastive learning for phonological class recognition,

    D. Liu, T. Arias-Vergara, J. Hutter, A. Maier, and P. A. P ´erez- Toro, “Audio–vision contrastive learning for phonological class recognition,” inInternational Conference on Text, Speech, and Dialogue. Springer, 2025, pp. 60–71

    2025

  23. [23]

    Arias-Vergara,Analysis of Pathological Speech Signals

    T. Arias-Vergara,Analysis of Pathological Speech Signals. Ger- many: Logos Verlag Berlin, 2022

    2022

  24. [24]

    Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,” inInterspeech 2017, 2017, pp. 498–502

    2017

  25. [25]

    Simple and effective zero-shot cross-lingual phoneme recognition,

    Q. Xu, A. Baevski, and M. Auli, “Simple and effective zero-shot cross-lingual phoneme recognition,”arXiv preprint arXiv:2109.11680, 2021

    work page arXiv 2021

  26. [26]

    Common phone: A multilingual dataset for robust acoustic modelling,

    P. Klumpp, T. Arias, P. A. P ´erez-Toro, E. Noeth, and J. Orozco- Arroyave, “Common phone: A multilingual dataset for robust acoustic modelling,” inProceedings of the Thirteenth Language Resources and Evaluation Conference, 2022, pp. 763–768

    2022

  27. [27]

    Crowdsourcing latin american spanish for low-resource text-to- speech,

    A. Guevara-Rukoz, I. Demirsahin, F. He, S.-H. C. Chu, S. Sarin, K. Pipatsrisawat, A. Gutkin, A. Butryna, and O. Kjartansson, “Crowdsourcing latin american spanish for low-resource text-to- speech,” inProceedings of the Twelfth Language Resources and Evaluation Conference, 2020, pp. 6504–6513

    2020

  28. [28]

    Darpa timit acoustic-phonetic continous speech corpus cd-rom. nist speech disc 1-1.1,

    J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, and D. S. Pallett, “Darpa timit acoustic-phonetic continous speech corpus cd-rom. nist speech disc 1-1.1,”NASA STI/Recon technical report n, vol. 93, p. 27403, 1993

    1993

  29. [29]

    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,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  30. [30]

    Carina–a corpus of aligned german read speech including annotations,

    H. Kath, S. Stone, S. Rapp, and P. Birkholz, “Carina–a corpus of aligned german read speech including annotations,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6157–6161

    2022

  31. [31]

    Com- mon voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Hen- retty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Com- mon voice: A massively-multilingual speech corpus,” inProceed- ings of the twelfth language resources and evaluation conference, 2020, pp. 4218–4222

    2020

  32. [32]

    ParlaSpeech-HR - a freely available ASR dataset for Croatian bootstrapped from the ParlaMint corpus,

    N. Ljube ˇsi´c, D. Kor ˇzinek, P. Rupnik, and I.-P. Jazbec, “ParlaSpeech-HR - a freely available ASR dataset for Croatian bootstrapped from the ParlaMint corpus,” inProceedings of the Workshop ParlaCLARIN III within the 13th Language Resources and Evaluation Conference, D. Fi ˇser, M. Eskevich, J. Lenardi ˇc, and F. de Jong, Eds. Marseille, France: Europe...

    2022

  33. [33]

    Par- czech 3.0: A large czech speech corpus with rich metadata,

    M. Kopp, V . Stankov, J. O. Kr˚uza, P. Straˇn´ak, and O. Bojar, “Par- czech 3.0: A large czech speech corpus with rich metadata,” in Text, Speech, and Dialogue, K. Ekˇstein, F. P´artl, and M. Konop´ık, Eds. Cham: Springer International Publishing, 2021, pp. 293– 304

    2021

  34. [34]

    Fleurs: Few-shot learning evaluation of universal representations of speech,

    A. Conneau, M. Ma, S. Khanuja, Y . Zhang, V . Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna, “Fleurs: Few-shot learning evaluation of universal representations of speech,”arXiv preprint arXiv:2205.12446, 2022

    work page arXiv 2022

  35. [35]

    V oxPopuli: A large-scale multilingual speech corpus for representation learn- ing, semi-supervised learning and interpretation,

    C. Wang, M. Riviere, A. Lee, A. Wu, C. Talnikar, D. Haz- iza, M. Williamson, J. Pino, and E. Dupoux, “V oxPopuli: A large-scale multilingual speech corpus for representation learn- ing, semi-supervised learning and interpretation,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Co...

    2021