Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

A recipe of multilingual SSL, Self-Conditioned CTC, and large G2P data yields state-of-the-art phone recognition on both multilingual and accented-English speech.

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

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

T0 review · grok-4.5

2026-07-14 20:01 UTC pith:CQNSW7TO

load-bearing objection Clean empirical recipe that actually isolates SSL, SelfCTC, and data scale for multilingual phone recognition and ships a usable SOTA model with open code. the 2 major comments →

arxiv 2603.29042 v2 pith:CQNSW7TO submitted 2026-03-30 cs.CL cs.LGcs.SDeess.AS

An Empirical Recipe for Universal Phone Recognition

Shikhar Bharadwaj , Chin-Jou Li , Kwanghee Choi , Eunjung Yeo , William Chen , Shinji Watanabe , David R. Mortensen This is my paper
classification cs.CL cs.LGcs.SDeess.AS
keywords phone recognitionmultilingual speechself-supervised learningSelf-Conditioned CTCarticulatory featuresaccented speechcross-lingual transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Phone recognition turns raw speech into sequences of speech sounds and is a foundation for multilingual and low-resource speech technology, yet systems that work well on English rarely transfer and multilingual systems have left performance on the table. This paper shows that the combination of a large multilingual self-supervised speech encoder, Self-Conditioned CTC, and the full scale of a phonemic training corpus produces a single model that is state-of-the-art on both multilingual evaluation suites and accented English. Controlled ablations isolate the contribution of each piece of the recipe—pretraining, loss, and data scale—while evaluations across more than 100 languages quantify where the gains appear and where errors remain. A sympathetic reader cares because the work supplies an open, reproducible recipe rather than another black-box model, and because the analyses map which language families, accents, and articulatory features still resist accurate recognition.

Core claim

PhoneticXEUS, obtained by fine-tuning the XEUS multilingual self-supervised encoder with Self-Conditioned CTC on the full IPAPack++ phonemic corpus, reaches 17.7 percent average phone-feature error rate on the multilingual PRiSM suites and 10.6 percent on the accented-English suites, outperforming prior English-centric and multilingual phone-recognition systems under a single evaluation scheme. Ablations show that Self-Conditioned CTC improves multilingual generalization relative to vanilla and intermediate CTC, that the pretrained encoder supplies large gains over training from scratch, and that scaling non-English training data improves multilingual performance without harming English.

What carries the argument

Self-Conditioned CTC: intermediate-layer phone posteriors are projected and added back into the encoder hidden states so that deeper layers can refine predictions using soft phonetic context from earlier layers; this objective, applied on top of a massively multilingual SSL encoder and large G2P-labeled data, is the training recipe that carries the reported gains.

Load-bearing premise

The paper assumes that large-scale G2P dictionary labels, even though they are coarse, noisy, and biased toward canonical pronunciations, still provide accurate enough supervision for learning robust acoustic-to-phone mappings.

What would settle it

Retrain the identical recipe on a held-out subset of the same languages that have carefully hand-aligned, speaker- and accent-matched phonemic transcriptions instead of G2P labels, then re-evaluate on the same PRiSM suites; if the measured gains over the from-scratch baseline shrink or reverse, the central claim that the recipe itself is what produces robust universal phone recognition is falsified.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper presents PhoneticXEUS, a phone-recognition model obtained by fine-tuning the multilingual SSL encoder XEUS with Self-Conditioned CTC on the large G2P-derived IPAPack++ corpus. Under the external PRiSM evaluation protocol it reports state-of-the-art average PFER of 17.7 % on multilingual suites and 10.6 % on accented-English suites (Table 1). Controlled ablations isolate the contributions of CTC-loss variants (Table 2), SSL backbones versus training from scratch (Table 3), and multilingual data scale (Figure 2). Additional analyses quantify cross-lingual transfer across language families, error rates on articulatory features, and robustness to English accents.

Significance. If the reported ranking holds, the work supplies a concrete, reproducible training recipe that simultaneously advances multilingual and accented-English phone recognition—two regimes previously treated as trade-offs. The open release of data and code, the use of a fixed external human-annotated benchmark, and the systematic separation of loss, initialization and data-scale effects make the result immediately usable by the community and strengthen the empirical foundation for future PR systems. The articulatory-feature and language-family breakdowns further identify concrete failure modes that can guide subsequent modeling.

major comments (2)
  1. Sections 2.3 and 3.3 acknowledge that G2P labels are coarse, noisy and biased toward canonical pronunciations, yet no quantitative estimate of label error (e.g., phone-level agreement with a human-annotated subset of IPAPack++) is supplied. Because every ablation and the final SOTA claim rest on this supervision, a short diagnostic of label noise would strengthen that the measured gains of SelfCTC and SSL are not inflated by systematic misalignment between labels and acoustics.
  2. Tables 2–3 and Figure 2 report single-run point estimates with no error bars, confidence intervals or multi-seed variance. The absolute differences that decide the recipe (SelfCTC 17.7 vs. InterCTC 18.5 vs. Vanilla 18.8; XEUS vs. E-Branchformer) are modest; without a measure of run-to-run variability it is difficult to judge whether the ranking is stable.
minor comments (4)
  1. Figure 1 caption and Table 1 both claim SOTA, but the figure y-axes begin at non-zero values and omit several baselines listed in the table; a single consistent legend would improve readability.
  2. Notation for the Self-Conditioned residual (Eq. 4) re-uses the same symbol set as the intermediate-layer posteriors; a brief clarifying sentence would help readers distinguish the projection matrices.
  3. The rank-correlation analysis (Section 3.1) reports ρ = −0.25, p = 0.096 as “marginal significance”; stating the exact sample size (number of languages) would make the claim easier to interpret.
  4. A few typographical inconsistencies appear (e.g., “PhoneticXEUS” vs. “PhoneticXeus” in figure labels; “POWSM-CTC” vs. “POWSM”).

Circularity Check

0 steps flagged

No circularity: empirical SOTA and ablations measured on held-out human-annotated PRiSM suites; no prediction reduces to a fitted input or self-definition.

full rationale

The paper's load-bearing claims are comparative empirical results (Table 1 SOTA PFER of 17.7% multilingual / 10.6% accented-English; Tables 2–3 and Figure 2 ablations of SelfCTC vs. InterCTC/vanilla, XEUS vs. MMS/E-Branchformer, and data scale). Training uses G2P-derived IPAPack++ labels; evaluation uses the external PRiSM human-annotated suites (PR-tmt/arc/saa, PR-drc/vox/tsm) that are not used to define the metric or the model. No equation equates a claimed prediction to a fitted constant by construction; no uniqueness theorem or ansatz is imported via self-citation to force the result. Self-citations (PRiSM [21], POWSM [20], ZIPA [19], XEUS [31]) supply open models/benchmarks that are independently usable and do not close a definitional loop. The acknowledged G2P label noise (Sections 2.3, 3.3) is a correctness/robustness concern, not circularity. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

Empirical ML paper; the load-bearing content is experimental comparison rather than formal derivation. Free parameters are the usual training hyperparameters; axioms are standard CTC assumptions and the validity of the G2P-derived training labels; the only invented entity is the named model itself.

free parameters (3)
  • lambda (auxiliary CTC weight) = tuned on validation
    Balances main and intermediate CTC losses in InterCTC/SelfCTC; tuned on validation set (Eq. 3).
  • lambda (CTC-Attention mix)
    Balances CTC and CE terms in joint training (Eq. 6); chosen in [0,1].
  • subset sizes for data-scale ablation = 150k-600k utts
    150k / 300k / 600k multilingual utterances while holding English fixed; discrete design choices that affect the scaling curve in Figure 2.
axioms (4)
  • domain assumption CTC marginalization over alignments is a valid training objective for unsegmented phone sequences
    Standard ASR assumption used throughout Section 2.1; not re-proved.
  • domain assumption G2P-generated IPA labels in IPAPack++ are sufficiently accurate supervision for learning acoustic-to-phone maps
    Core training-data premise (Sections 2–3); paper itself notes noise and canonical bias.
  • domain assumption PRiSM human-annotated sets constitute a fair, unified evaluation across 100+ languages and accents
    All SOTA claims rest on this benchmark (Table 1, Section 2).
  • domain assumption Phonological language vectors (URIEL/lang2vec) measure cross-lingual similarity relevant to phone recognition transfer
    Used for the rank-correlation analysis in Section 3.1.
invented entities (1)
  • PhoneticXEUS independent evidence
    purpose: Name for the concrete model obtained by the recommended recipe (XEUS + SelfCTC + full IPAPack++)
    The model is the deliverable; independent evidence is the open weights/code and the public PRiSM numbers.

pith-pipeline@v1.1.0-grok45 · 15003 in / 2767 out tokens · 27829 ms · 2026-07-14T20:01:04.320559+00:00 · methodology

0 comments
read the original abstract

Phone recognition (PR) is a key enabler of multilingual and low-resource speech processing tasks, yet robust performance remains elusive. Highly performant English-focused models do not generalize across languages, while multilingual models underutilize pretrained representations. It also remains unclear how data scale, architecture, and training objective contribute to multilingual PR. We present PhoneticXEUS -- trained on large-scale multilingual data and achieving state-of-the-art performance on both multilingual (17.7% PFER) and accented English speech (10.6% PFER). Through controlled ablations with evaluations across 100+ languages under a unified scheme, we empirically establish our training recipe and quantify the impact of SSL representations, data scale, and loss objectives. In addition, we analyze error patterns across language families, accented speech, and articulatory features. All data and code are released openly at https://github.com/changelinglab/PhoneticXeus

Figures

Figures reproduced from arXiv: 2603.29042 by Chin-Jou Li, David R. Mortensen, Eunjung Yeo, Kwanghee Choi, Shikhar Bharadwaj, Shinji Watanabe, William Chen.

Figure 1
Figure 1. Figure 1: PhoneticXEUS achieves SOTA performance on both accented English and multilingual speech. Details in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Increased language diversity in fine-tuning data ben￾efits PR performance on multilingual datasets (subsection 2.3). formance of MMS deteriorates (S4 vs. C4). However, XEUS achieves the best performance (C3) showing 2.0% improvement on English and 5.4% improvement on multilingual evaluations over the baseline trained from scratch (C5). Hence, we adopt XEUS as the backbone in our recipe. We further explore … view at source ↗
Figure 3
Figure 3. Figure 3: Performance of PhoneticXEUS across language families in VoxAngeles. SSL improves cross-lingual transfer (subsection 3.1). Back Round High Low Tense Voicing Spread Glottis Constrained Glottis Continuant Lateral Nasal Strident Delayed Release Coronal Labial Anterior Distributed E-BranchFormer PhoneticXeus (Ours) 10 20 30 40 Phonation Vowel Manner of Articulation Place of Articulation [PITH_FULL_IMAGE:figure… view at source ↗
Figure 4
Figure 4. Figure 4: Performance across articulatory features (↓). Differ￾ent features show different relative gains with SSL initialization. (subsection 3.2). SSL representations allow the model to better leverage phono￾logical proximity to languages in paired training data for mak￾ing predictions for unseen languages. To better understand the limitations of our model, we next examine errors qualitatively in the lowest-perfor… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. BranchShine: Compact Raw-Audio-to-IPA Transcription with a RoPE E-Branchformer Encoder

    cs.LG 2026-06 unverdicted novelty 4.0

    A 33M-parameter raw-audio CTC model with 19-block RoPE E-Branchformer achieves 9.19% whitespace-insensitive IPA CER on a 16,660-utterance 41-language test set, outperforming a 575M-parameter PhoneticXEUS baseline at 9...

Reference graph

Works this paper leans on

57 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    PR is also heav- ily employed in atypical speech assessment [7, 8], computer- assisted language learning [9–11] and linguistic fieldwork [12– 14]

    Introduction Phone Recognition (PR) enables important multilingual speech processing technologies, especially for zero text-resource lan- guages [1–3] via cross-lingual transfer [4–6]. PR is also heav- ily employed in atypical speech assessment [7, 8], computer- assisted language learning [9–11] and linguistic fieldwork [12– 14]. For English PR, several s...

  2. [2]

    Conse- quently, it remains unclear to what extent data scale, model ar- chitecture, and training objective contribute to multilingual PR performance

    uses Zipformer [33] with CR-CTC [34] for training ef- ficiency, and POWSM [20] uses CTC-Attention joint training in an autoregressive setup [35] for multi-task learning. Conse- quently, it remains unclear to what extent data scale, model ar- chitecture, and training objective contribute to multilingual PR performance. To cover these gaps, we provide contr...

    Pith/arXiv arXiv 2026

  3. [3]

    We evaluate the model with PRiSM [21], a benchmark for PR systems, and report the results in Table 1

    Experiments We organize our experiments around three research questions: •RQ1: Which CTC training objective best supports cross- lingual generalization in PR? •RQ2: Do SSL representations pretrained on massively mul- tilingual speech improve PR over training from scratch? •RQ3: How does the scale of multilingual training-data affect English versus multili...

  4. [4]

    2) MMS [45] (S3, S4), pretrained on 1k languages

    (S1,S2) as baselines without pretraining. 2) MMS [45] (S3, S4), pretrained on 1k languages. 3) XEUS [31] (S5), an E- Branchformer speech encoder trained with HuBERT-style [47] masked prediction on 4k languages. In Table 3, MMS and XEUS both consistently show im- provements over similar sized baselines (S1, S2). We also ex- periment with MMS-1B on full IPA...

  5. [5]

    We analyze the role of SSL in cross-lingual transfer via comparison with E-Branchformer (C5) which differs only in initialization to our recipe (C3)

    Analyses Cross-lingual transfer is critical for PR in under-resourced lan- guages. We analyze the role of SSL in cross-lingual transfer via comparison with E-Branchformer (C5) which differs only in initialization to our recipe (C3). Also, we aim to provide a deeper understanding of the model behavior and its limitations. 3.1. Cross-lingual Performance and...

  6. [6]

    Conclusion We present PhoneticXEUS, a SOTA PR system along with sys- tematic ablation study that disentangles the effects of training data, initialization, and training objective. Analyses reveal the role of SSL in cross-lingual transfer, and variance in perfor- mance across articulatory attributes, along with the need to de- velop better quality evaluation sets

  7. [7]

    All outputs were reviewed and validated by the authors

    Generative AI Use Disclosure Generative AI tools were used to improve the clarity and gram- mar of the manuscript and to assist with portions of the code. All outputs were reviewed and validated by the authors

  8. [8]

    The Zero Resource Speech Challenge 2019: TTS Without T,

    E. Dunbaret al., “The Zero Resource Speech Challenge 2019: TTS Without T,” inProc. Interspeech, 2019, pp. 1088–1092

    2019

  9. [9]

    The Zero Resource Speech Challenge 2021: Spoken Language Mod- elling,

    E. Dunbar, M. Bernard, N. Hamilakis, T. A. Nguyen, M. d. Seyssel, P. Roz´e, M. Rivi`ere, E. Kharitonov, and E. Dupoux, “The Zero Resource Speech Challenge 2021: Spoken Language Mod- elling,” inProc. Interspeech, 2021, pp. 1574–1578

    2021

  10. [10]

    Self-supervised lan- guage learning from raw audio: Lessons from the zero resource speech challenge,

    E. Dunbar, N. Hamilakis, and E. Dupoux, “Self-supervised lan- guage learning from raw audio: Lessons from the zero resource speech challenge,”IEEE Journal of Selected Topics in Signal Pro- cessing, vol. 16, no. 6, pp. 1211–1226, 2022

    2022

  11. [11]

    That Sounds Familiar: An Analysis of Phonetic Representations Transfer Across Lan- guages,

    P. ˙Zelasko, L. Moro-Vel ´azquez, M. Hasegawa-Johnson, O. Scharenborg, and N. Dehak, “That Sounds Familiar: An Analysis of Phonetic Representations Transfer Across Lan- guages,” inProc. Interspeech, 2020, pp. 3705–3709

    2020

  12. [12]

    Multilingual and crosslin- gual speech recognition using phonological-vector based phone embeddings,

    C. Zhu, K. An, H. Zheng, and Z. Ou, “Multilingual and crosslin- gual speech recognition using phonological-vector based phone embeddings,” inProc. ASRU, 2021, pp. 1034–1041

    2021

  13. [13]

    Differentiable Allophone Graphs for Language-Universal Speech Recognition,

    B. Yan, S. Dalmia, D. R. Mortensen, F. Metze, and S. Watan- abe, “Differentiable Allophone Graphs for Language-Universal Speech Recognition,” inProc. Interspeech, 2021, pp. 2471–2475

    2021

  14. [14]

    Learner training for phonetic transcription of typical and/or disordered speech: A scoping review,

    A. Lee and N. Bessell, “Learner training for phonetic transcription of typical and/or disordered speech: A scoping review,”Interna- tional Journal of Language & Communication Disorders, vol. 59, no. 6, pp. 2926–2945, 2024

    2024

  15. [15]

    L. D. Shriberg, R. D. Kent, T. McAllister, J. L. Preston, and M. L. Speights,Clinical phonetics. Plural Publishing, 2025

    2025

  16. [16]

    Advances in phone-based modeling for automatic accent classification,

    P. Angkititrakul and J. H. Hansen, “Advances in phone-based modeling for automatic accent classification,”IEEE transactions on audio, speech, and language processing, vol. 14, no. 2, pp. 634–646, 2006

    2006

  17. [17]

    EduSpeak®: A speech recognition and pronunciation scoring toolkit for computer-aided language learning applications,

    H. Franco, H. Bratt, R. Rossier, V . R. Gadde, E. Shriberg, V . Abrash, and K. Precoda, “EduSpeak®: A speech recognition and pronunciation scoring toolkit for computer-aided language learning applications,”Language Testing, vol. 27, pp. 401 – 418, 2010. [Online]. Available: https://api.semanticscholar.org/ CorpusID:143273296

    2010

  18. [18]

    Automatic Pronuncia- tion Assessment-A Review,

    Y . El Kheir, A. Ali, and S. A. Chowdhury, “Automatic Pronuncia- tion Assessment-A Review,” inProc. EMNLP, 2023

    2023

  19. [19]

    S. L. Chelliah and W. J. De Reuse,Handbook of descriptive lin- guistic fieldwork. Springer Science & Business Media, 2010

    2010

  20. [20]

    The UCLA Phonetics Lab Archive,

    P. Ladefogedet al., “The UCLA Phonetics Lab Archive,” 2009. [Online]. Available: https://archive.phonetics.ucla.edu/

    2009

  21. [21]

    Tusom2021: A Phonetically Transcribed Speech Dataset from an Endangered Language for Universal Phone Recognition Experiments,

    D. R. Mortensen, J. Picone, X. Li, and K. Siminyu, “Tusom2021: A Phonetically Transcribed Speech Dataset from an Endangered Language for Universal Phone Recognition Experiments,” in Proc. Interspeech, 2021, pp. 3660–3664

    2021

  22. [22]

    Automatic recognition of phonetic patterns in speech,

    H. Dudley and S. Balashek, “Automatic recognition of phonetic patterns in speech,”The Journal of the Acoustical Society of Amer- ica, vol. 30, no. 8, pp. 721–732, 1958

    1958

  23. [23]

    Acoustic phonetic approach for speech recognition: A review,

    P. Bhagath and P. K. Das, “Acoustic phonetic approach for speech recognition: A review,”Language, vol. 77, p. 93, 2004

    2004

  24. [24]

    HuPER: A Human-Inspired Framework for Phonetic Perception,

    C. Guo, J. Lian, Y . Liu, B. Huang, S. Narayanan, C. J. Cho, and G. Anumanchipalli, “HuPER: A Human-Inspired Framework for Phonetic Perception,”arXiv preprint arXiv:2602.01634, 2026

    arXiv 2026

  25. [25]

    XLSR Inclusive English Speech-to-IPA,

    K. Labs, “XLSR Inclusive English Speech-to-IPA,” 2025. [Online]. Available: https://huggingface.co/collections/KoelLabs/ xlsr-inclusive-english-speech-to-ipa

    2025

  26. [26]

    ZIPA: A family of efficient models for multilingual phone recognition,

    J. Zhu, F. Samir, E. Chodroff, and D. R. Mortensen, “ZIPA: A family of efficient models for multilingual phone recognition,” in Proc. ACL, 2025, pp. 19 568–19 585

    2025

  27. [27]

    POWSM: A Phonetic Open Whisper-Style Speech Foundation Model,

    C.-J. Li, K. Chang, S. Bharadwaj, E. Yeo, K. Choi, J. Zhu, D. Mortensen, and S. Watanabe, “POWSM: A Phonetic Open Whisper-Style Speech Foundation Model,”arXiv preprint 2510.24992, 2025

    arXiv 2025

  28. [28]

    PRiSM: Benchmarking Phone Realization in Speech Models,

    S. Bharadwajet al., “PRiSM: Benchmarking Phone Realization in Speech Models,”arXiv preprint 2601.14046, 2026

    Pith/arXiv arXiv 2026

  29. [29]

    Gemini 2.5: Pushing the frontier with ad- vanced reasoning, multimodality, long context, and next genera- tion agentic capabilities,

    G. Comaniciet al., “Gemini 2.5: Pushing the frontier with ad- vanced reasoning, multimodality, long context, and next genera- tion agentic capabilities,”arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Qwen3-Omni Technical Report,

    J. Xuet al., “Qwen3-Omni Technical Report,”arXiv preprint 2509.17765, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    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,” inInterspeech, 2022, pp. 2113–2117

    2022

  32. [32]

    Universal Au- tomatic Phonetic Transcription into the International Phonetic Al- phabet,

    C. Taguchi, Y . Sakai, P. Haghani, and D. Chiang, “Universal Au- tomatic Phonetic Transcription into the International Phonetic Al- phabet,” inInterspeech, 2023, pp. 2548–2552

    2023

  33. [33]

    Towards zero-shot learning for automatic phonemic transcrip- tion,

    X. Li, S. Dalmia, D. Mortensen, J. Li, A. Black, and F. Metze, “Towards zero-shot learning for automatic phonemic transcrip- tion,” inProceedings of the AAAI Conference on Artificial Intelli- gence, vol. 34, no. 05, 2020, pp. 8261–8268

    2020

  34. [34]

    Universal phone recognition with a multilingual al- lophone system,

    X. Liet al., “Universal phone recognition with a multilingual al- lophone system,” inProc. ICASSP, 2020, pp. 8249–8253

    2020

  35. [35]

    Allophant: Cross- lingual Phoneme Recognition with Articulatory Attributes,

    K. Glocker, A. Herygers, and M. Georges, “Allophant: Cross- lingual Phoneme Recognition with Articulatory Attributes,” inIn- terspeech, 2023, pp. 2258–2262

    2023

  36. [36]

    The taste of IPA: Towards open-vocabulary keyword spotting and forced alignment in any language,

    J. Zhu, C. Yang, F. Samir, and J. Islam, “The taste of IPA: Towards open-vocabulary keyword spotting and forced alignment in any language,” inProc. NAACL, Jun. 2024, pp. 750–772

    2024

  37. [37]

    Multilingual Dysarthric Speech Assessment Using Universal Phone Recogni- tion and Language-Specific Phonemic Contrast Modeling,

    E. Yeo, J. M. Liss, V . Berisha, and D. R. Mortensen, “Multilingual Dysarthric Speech Assessment Using Universal Phone Recogni- tion and Language-Specific Phonemic Contrast Modeling,”arXiv preprint arXiv:2601.21205, 2026

    arXiv 2026

  38. [38]

    Towards robust speech representation learning for thousands of languages,

    W. Chen, W. Zhang, Y . Peng, X. Li, J. Tian, J. Shi, X. Chang, S. Maiti, K. Livescu, and S. Watanabe, “Towards robust speech representation learning for thousands of languages,” inProc. EMNLP, 2024, pp. 10 205–10 224

    2024

  39. [39]

    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,” inProc. ASRU, 2021, pp. 914–921

    2021

  40. [40]

    Zipformer: A faster and better encoder for automatic speech recognition,

    Z. Yao, L. Guo, X. Yang, W. Kang, F. Kuang, Y . Yang, Z. Jin, L. Lin, and D. Povey, “Zipformer: A faster and better encoder for automatic speech recognition,”Proc. ICLR, 2024

    2024

  41. [41]

    CR-CTC: Consistency regularization on CTC for improved speech recognition,

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

    2025

  42. [42]

    Joint CTC-attention based end-to-end speech recognition using multi-task learning,

    S. Kim, T. Hori, and S. Watanabe, “Joint CTC-attention based end-to-end speech recognition using multi-task learning,” inProc. ICASSP, 2017, pp. 4835–4839

    2017

  43. [43]

    PanPhon: A Resource for Mapping IPA Segments to Articulatory Feature Vectors,

    D. R. Mortensen, P. Littell, A. Bharadwaj, K. Goyal, C. Dyer, and L. S. Levin, “PanPhon: A Resource for Mapping IPA Segments to Articulatory Feature Vectors,” inProc. COLING, 2016, pp. 3475– 3484

    2016

  44. [44]

    Relaxing the Conditional Indepen- dence Assumption of CTC-Based ASR by Conditioning on Inter- mediate Predictions,

    J. Nozaki and T. Komatsu, “Relaxing the Conditional Indepen- dence Assumption of CTC-Based ASR by Conditioning on Inter- mediate Predictions,” inProc. Interspeech, 2021, pp. 3735–3739

    2021

  45. [45]

    Connectionist temporal classification: labelling un- segmented sequence data with recurrent neural networks,

    A. Graves, “Connectionist temporal classification: labelling un- segmented sequence data with recurrent neural networks,” in Proc. ICML, 2006, pp. 369–376

    2006

  46. [46]

    Hierarchical Multitask Learning With CTC,

    R. Sanabria and F. Metze, “Hierarchical Multitask Learning With CTC,” inProc. SLT, 2018, pp. 485–490

    2018

  47. [47]

    Intermediate loss regularization for ctc- based speech recognition,

    J. Lee and S. Watanabe, “Intermediate loss regularization for ctc- based speech recognition,” inProc. ICASSP, 2021, pp. 6224– 6228

    2021

  48. [48]

    Deja-vu: Double feature presen- tation and iterated loss in deep transformer networks,

    A. Tjandra, C. Liu, F. Zhang, X. Zhang, Y . Wang, G. Synnaeve, S. Nakamura, and G. Zweig, “Deja-vu: Double feature presen- tation and iterated loss in deep transformer networks,” inProc. ICASSP, 2020, pp. 6899–6903

    2020

  49. [49]

    Im- proving massively multilingual asr with auxiliary ctc objectives,

    W. Chen, B. Yan, J. Shi, Y . Peng, S. Maiti, and S. Watanabe, “Im- proving massively multilingual asr with auxiliary ctc objectives,” inProc. ICASSP, 2023, pp. 1–5

    2023

  50. [50]

    Hierarchi- cal conditional end-to-end asr with ctc and multi-granular sub- word units,

    Y . Higuchi, K. Karube, T. Ogawa, and T. Kobayashi, “Hierarchi- cal conditional end-to-end asr with ctc and multi-granular sub- word units,” inProc. ICASSP, 2022, pp. 7797–7801

    2022

  51. [51]

    E-branchformer: Branchformer with enhanced merging for speech recognition,

    K. Kim, F. Wu, Y . Peng, J. Pan, P. Sridhar, K. J. Han, and S. Watanabe, “E-branchformer: Branchformer with enhanced merging for speech recognition,” inProc. SLT, 2023, pp. 84–91

    2023

  52. [52]

    Scaling Speech Technology to 1,000+ Lan- guages,

    V . Pratapet al., “Scaling Speech Technology to 1,000+ Lan- guages,”arXiv, 2023

    2023

  53. [53]

    [b]=[d]-[t]+[p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic,

    K. Choi, E. Yeo, C. J. Choet al., “[b]=[d]-[t]+[p]: Self-supervised Speech Models Discover Phonological Vector Arithmetic,”arXiv preprint arXiv:2602.18899, 2026

    Pith/arXiv arXiv 2026

  54. [54]

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

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  55. [55]

    Phonetic Seg- mentation of the UCLA Phonetics Lab Archive,

    E. Chodroff, B. Pa ˇzon, A. Baker, and S. Moran, “Phonetic Seg- mentation of the UCLA Phonetics Lab Archive,” inProc. LREC- COLING, 2024, pp. 12 724–12 733

    2024

  56. [56]

    URIEL and lang2vec: Representing languages as typological, geographical, and phylogenetic vectors,

    P. Littell, D. R. Mortensen, K. Lin, K. Kairis, C. Turner, and L. Levin, “URIEL and lang2vec: Representing languages as typological, geographical, and phylogenetic vectors,” in Proc. EACL, Apr. 2017, pp. 8–14. [Online]. Available: https: //aclanthology.org/E17-2002/

    2017

  57. [57]

    Speech Accent Archive,

    S. Weinberger, “Speech Accent Archive,” 2015, retrieved from https://accent.gmu.edu

    2015