pith. sign in

arxiv: 2606.19910 · v2 · pith:KERCT6SSnew · submitted 2026-06-18 · 💻 cs.CL · cs.SD· eess.AS

Light-weight Pronunciation Assessment via Discrete Speech Token Surprisal

Syeda Faiza Ahmed Sara , Shammur Absar Chowdhury This is my paper

Pith reviewed 2026-06-26 17:35 UTC · model grok-4.3

classification 💻 cs.CL cs.SDeess.AS
keywords pronunciation assessmentsurprisaldiscrete speech tokenstoken language modelDTW alignmentself-supervised speechnative resourcesL2 speech evaluation
0
0 comments X

The pith

Surprisal from a native-trained token language model on discretized speech, fused with DTW alignment features, raises pronunciation assessment correlation from 0.60 to 0.66 on SpeechOcean762 using only native resources.

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

The paper establishes a lightweight pronunciation assessment framework that trains exclusively on native speech. Learner utterances are discretized via an SSL encoder and K-means codebook; a token language model then computes surprisal to flag phonotactic deviations. A transcript-guided Text2DUnit-DTW module supplies additional alignment features that are regressed together with surprisal scores. This combination lifts Pearson correlation from a 0.60 baseline to 0.66 on SpeechOcean762 and yields consistent gains on L2-ARCTIC while approaching supervised methods. A sympathetic reader would care because the approach removes the need for costly labeled learner-error corpora.

Core claim

The central claim is that surprisal computed by a token language model trained on native discretized sequences, when combined with transcript-guided Text2DUnit-DTW alignment features and fused by simple regression, produces pronunciation scores whose correlation with human judgments improves from 0.60 to 0.66 on SpeechOcean762 and shows consistent gains on L2-ARCTIC, all while using only native speech or light calibration.

What carries the argument

Token surprisal from a language model trained on native K-means discretized sequences (higher surprisal marks phonotactic deviation), fused via regression with transcript-guided Text2DUnit-DTW alignment features.

If this is right

  • The framework operates unsupervised or with only a small set of scored utterances for calibration.
  • Transcript guidance is required to reach the 0.66 PCC; without it performance stays at the 0.60 level.
  • Consistent cross-dataset gains appear on both SpeechOcean762 and L2-ARCTIC.
  • Results approach those of supervised baselines that require labeled non-native data.

Where Pith is reading between the lines

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

  • The same surprisal signal could be tested on other L2 dimensions such as fluency or prosody without new labeled data.
  • If the discretization step proves robust across languages, the method could support pronunciation assessment in low-resource target languages.
  • Replacing K-means with a learned codebook might tighten the mapping from acoustic tokens to phonotactic expectations.

Load-bearing premise

Phonotactic deviations in learner speech are reliably captured by surprisal in the K-means discretized token space, and the alignment module plus small calibration set generalizes without overfitting to test conditions.

What would settle it

A new L2 dataset with unseen accents or error types where adding the surprisal-plus-alignment features fails to raise PCC above the 0.60 baseline or drops performance.

Figures

Figures reproduced from arXiv: 2606.19910 by Shammur Absar Chowdhury, Syeda Faiza Ahmed Sara.

Figure 1
Figure 1. Figure 1: Training overview. Training uses only standard na￾tive speech (ASR) resources and requires no learner data, man￾ual annotation, or forced alignment. out reference text, and can be lightly supervised when a small set of scored learner utterances is available, making it suitable for zero-resource settings. First, we discretize speech using a frozen self-supervised encoder and a K-means codebook trained on na… view at source ↗
Figure 2
Figure 2. Figure 2: Inference overview. At inference, we compute audio-only surprisal features and optional transcript-guided DTW alignment features. A simple regressor can be trained with a small set of annotated learner samples, but the features also act as direct pronunci￾ation quality indicators. reference transcripts into the same discrete unit space. At inference ( [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
read the original abstract

Training automated pronunciation assessment often relies on labeled learner errors or non-native corpora that are costly to collect. We propose a lightweight framework trained only on native speech resources, operating unsupervised or lightly calibrated with a small set of scored utterances. At inference, learner speech is discretized with an SSL encoder and a K-means codebook. A token language model trained on native sequences computes surprisal where higher surprisal indicates phonotactic deviation. We add a transcript-guided Text2DUnit--DTW module that predicts native token sequences from reference text and aligns them to acoustic tokens to derive error-sensitive features. Surprisal and alignment features are fused via simple regression. On SpeechOcean762, PCC improves from 0.60 to 0.66 with transcript guidance, near supervised baselines. Cross-dataset evaluation on L2-ARCTIC shows consistent gains.

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

1 major / 1 minor

Summary. The paper proposes a lightweight pronunciation assessment framework trained only on native speech resources, operating unsupervised or with light calibration on a small set of scored utterances. Learner speech is discretized via an SSL encoder and K-means codebook; a token language model trained on native sequences computes surprisal to flag phonotactic deviations. A transcript-guided Text2DUnit-DTW alignment module derives error-sensitive features from reference text, and surprisal plus alignment features are fused via simple regression. On SpeechOcean762 the method improves PCC from 0.60 to 0.66 with transcript guidance, nearing supervised baselines; cross-dataset evaluation on L2-ARCTIC shows consistent gains.

Significance. If the empirical results hold, the work is significant for offering a practical alternative to supervised pronunciation assessment that largely avoids costly labeled non-native corpora. The unsupervised pipeline trained solely on native data, the use of discrete-token surprisal, and the reported cross-dataset consistency are clear strengths. The approach is falsifiable via the stated PCC metrics and could reduce data-collection barriers in the field.

major comments (1)
  1. [Abstract] Abstract: the central claim rests on the reported PCC improvement from 0.60 to 0.66 and cross-dataset consistency, yet the abstract supplies neither error bars, statistical significance tests, nor explicit baseline descriptions; without these the magnitude and reliability of the gains cannot be verified from the provided text.
minor comments (1)
  1. [Abstract] Abstract: the Text2DUnit-DTW module is referenced without a brief definition or citation, which reduces immediate clarity for readers outside the immediate sub-area.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive evaluation and recommendation for minor revision. We address the single major comment on the abstract below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim rests on the reported PCC improvement from 0.60 to 0.66 and cross-dataset consistency, yet the abstract supplies neither error bars, statistical significance tests, nor explicit baseline descriptions; without these the magnitude and reliability of the gains cannot be verified from the provided text.

    Authors: We agree that the abstract would be strengthened by including error bars, a note on statistical significance, and more explicit baseline references. The main text reports these details (including confidence intervals and comparisons to supervised systems), but space constraints in the original abstract omitted them. In revision we will expand the abstract to incorporate error bars on the PCC values, state the significance of the 0.06 gain, and name the supervised baselines referenced. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents an unsupervised pipeline that trains a token LM and alignment module exclusively on native speech resources, then applies surprisal and DTW-derived features to L2 utterances via simple regression. No equations reduce the final PCC to a fitted parameter by construction, no self-citation chain supports the central claim, and the discretization plus regression steps operate on external native data without importing uniqueness theorems or ansatzes from prior author work. The reported gains (0.60 to 0.66) are empirical outcomes of the described method rather than definitional identities.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; full text would be needed to audit any K-means codebook size, LM training details, or regression coefficients.

pith-pipeline@v0.9.1-grok · 5676 in / 1042 out tokens · 29052 ms · 2026-06-26T17:35:31.384634+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

58 extracted references · 18 canonical work pages · 1 internal anchor

  1. [1]

    The most widely used approach computes Goodness of Pronunciation (GoP) scores from ASR acoustic models [1], which requires forced alignment and reference transcriptions

    Introduction Pronunciation assessment is central to computer-assisted lan- guage learning, yet building reliable automatic systems remains difficult when labeled data are scarce. The most widely used approach computes Goodness of Pronunciation (GoP) scores from ASR acoustic models [1], which requires forced alignment and reference transcriptions. Regressi...

  2. [2]

    Light-weight Pronunciation Assessment via Discrete Speech Token Surprisal

    Method Figures 1 and 2 illustrate the proposed pronunciation assess- ment framework. During training (Figure 1), we use only native speech to learn a discrete unit vocabulary and a native phono- tactic prior through two modules: Audio2DUnit and the Token- level Language Model (TLM). We also train Text2DUnit to map arXiv:2606.19910v2 [cs.CL] 23 Jun 2026 Fi...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Data Training.The Audio2DUnit codebook and Token Lan- guage Model (TLM) are trained on LibriSpeech 960 h of native English read speech [11]

    Experiments 3.1. Data Training.The Audio2DUnit codebook and Token Lan- guage Model (TLM) are trained on LibriSpeech 960 h of native English read speech [11]. The Text2DUnit model is trained on (transcript, collapsed-token) pairs extracted from the LibriSpeech 960-hour setup and tokens from the same Audio2DUnit K-means codebook, with a 90/10 random train/v...

    2048

  4. [4]

    First, anunsupervisedset- ting in which the proposed features are used directly as pro- nunciation quality indicators without any supervised regression model (see Table 2)

    Results and Discussion Evaluation Settings: For the study, we evaluatethree infer- ence settingson SpeechOcean762. First, anunsupervisedset- ting in which the proposed features are used directly as pro- nunciation quality indicators without any supervised regression model (see Table 2). Second, anaudio-onlysetting (light- supervised), where pronunciation ...

  5. [5]

    These methods typically require phoneme inventories, forced alignment, and labeled learner data

    Related Work Automated pronunciation assessment has traditionally relied on GoP from forced-aligned ASR models [1], with later work extending this through end-to-end mispronunciation detection, alignment-aware training, phonetic/acoustic cue modeling, and Transformer-based regressors such as GOPT [18, 19, 20, 2]. These methods typically require phoneme in...

  6. [6]

    The approach combines discrete-token surprisal from a native token language model with optional transcript-guided Text2DUnit– DTW alignment in the same discrete space

    Conclusion We presented a lightweight pronunciation assessment frame- work that learns from standard native speech resources and is designed for settings with little or no labeled learner data. The approach combines discrete-token surprisal from a native token language model with optional transcript-guided Text2DUnit– DTW alignment in the same discrete sp...

  7. [7]

    These tools were not used to generate experimental results, design the methodology, analyze data, or produce data or tables

    Use of Generative AI Generative AI tools were used during the preparation of this manuscript to assist with language editing, grammar correction, and improving clarity of the written text. These tools were not used to generate experimental results, design the methodology, analyze data, or produce data or tables. All technical content, experimental design,...

  8. [8]

    The findings achieved herein are solely the responsibility of the authors

    Acknowledgement The work is supported by HBKU flagship research grant (HBKU-INT-VPR-FRG-03-09). The findings achieved herein are solely the responsibility of the authors

  9. [9]

    Phone-level pronunciation scoring and assessment for interactive language learning,

    S. M. Witt and S. J. Young, “Phone-level pronunciation scoring and assessment for interactive language learning,” Speech communication, vol. 30, no. 2-3, pp. 95–108,

  10. [10]

    Available: https://www.semanticscholar.org/ paper/317c3e997171ec9a1d15928ef8c996b5f4b8d7ea

    [Online]. Available: https://www.semanticscholar.org/ paper/317c3e997171ec9a1d15928ef8c996b5f4b8d7ea

  11. [11]

    Transformer-based multi-aspect multi-granularity non-native english speaker pronunciation assessment,

    Y . Gong, Z. Chen, I.-H. Chu, P. Chang, and J. Glass, “Transformer-based multi-aspect multi-granularity non-native english speaker pronunciation assessment,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, May 2022, p. 7262–7266. [Online]. Available: http://dx.doi.org/10.1109/ICASSP43922. 2022.9746743

    work page doi:10.1109/icassp43922 2022

  12. [14]

    Children’s speech recog- nition through discrete token enhancement,

    V . N. Sukhadia and S. A. Chowdhury, “Children’s speech recog- nition through discrete token enhancement,” inProc. Interspeech 2024, 2024, pp. 5143–5147

    2024

  13. [15]

    On generative spoken language modeling from raw audio,

    K. Lakhotia, E. Kharitonov, W.-N. Hsu, Y . Adi, A. Polyak, B. Bolte, T.-A. Nguyen, J. Copet, A. Baevski, A. Mohamed et al., “On generative spoken language modeling from raw audio,”Transactions of the Association for Computational Linguistics, vol. 9, pp. 1336–1354, 2021. [Online]. Available: https://arxiv.org/abs/2102.01192v2

    work page arXiv 2021

  14. [16]

    Exploring self-supervised audio models for generalized anomalous sound detection,

    B. Han, A. Jiang, X. Zheng, W.-Q. Zhang, J. Liu, P. Fan, and Y . Qian, “Exploring self-supervised audio models for generalized anomalous sound detection,”IEEE Transactions on Audio, Speech and Language Processing, 2025. [Online]. Available: https://arxiv.org/abs/2508.12230v1

    work page arXiv 2025

  15. [17]

    A probabilistic earley parser as a psycholin- guistic model,

    J. Hale, “A probabilistic earley parser as a psycholin- guistic model,” inSecond meeting of the north american chapter of the association for computational linguistics,

  16. [18]

    Available: https://www.semanticscholar.org/ paper/a959ae0f08228d5d04eb46d5dc1ec3f4c9ee91d7

    [Online]. Available: https://www.semanticscholar.org/ paper/a959ae0f08228d5d04eb46d5dc1ec3f4c9ee91d7

  17. [19]

    Expectation-based syntactic comprehension,

    R. Levy, “Expectation-based syntactic comprehension,” Cognition, vol. 106, no. 3, pp. 1126–1177, 2008. [Online]. Available: https://www.semanticscholar.org/paper/ 2a92864d3f44e1b34da0fa17e13d5752d089eddf

    2008

  18. [20]

    Computing and visualizing dynamic time warping alignments in r: the dtw package,

    T. Giorgino, “Computing and visualizing dynamic time warping alignments in r: the dtw package,”Jour- nal of statistical Software, vol. 31, pp. 1–24, 2009. [Online]. Available: https://www.semanticscholar.org/paper/ 51bcfef92aef96df37b25bd1ed7c096142107007

    2009

  19. [21]

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

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: an asr corpus based on public domain audio books,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210. [Online]. Available: https://www.semanticscholar. org/paper/34038d9424ce602d7ac917a4e582d977725d4393

    2015

  20. [22]

    speechocean762: An open-source non-native english speech corpus for pronunciation assessment,

    J. Zhang, Z. Zhang, Y . Wang, Z. Yan, Q. Song, Y . Huang, K. Li, D. Povey, and Y . Wang, “speechocean762: An open-source non-native english speech corpus for pronunciation assessment,” arXiv preprint arXiv:2104.01378, 2021. [Online]. Available: https://arxiv.org/abs/2104.01378v2

    work page arXiv 2021

  21. [23]

    L2-ARCTIC: A Non-native English Speech Corpus,

    G. Zhao, S. Sonsaat, A. Silpachai, I. Lucic, E. Chukharev- Hudilainen, J. Levis, and R. Gutierrez-Osuna, “L2-ARCTIC: A Non-native English Speech Corpus,” inInterspeech 2018, pp. 2783–2787. [Online]. Available: https://www.isca-archive.org/ interspeech 2018/zhao18b interspeech.html

    2018

  22. [24]

    Ca- nine: Pre-training an efficient tokenization-free encoder for language representation,

    J. H. Clark, D. Garrette, I. Turc, and J. Wieting, “Ca- nine: Pre-training an efficient tokenization-free encoder for language representation,”Transactions of the Associa- tion for Computational Linguistics, vol. 10, pp. 73–91,

  23. [25]

    Available: https://www.semanticscholar.org/ paper/969287b8a96e242793b11f0dbb99ec341228106f

    [Online]. Available: https://www.semanticscholar.org/ paper/969287b8a96e242793b11f0dbb99ec341228106f

  24. [26]

    Zero-Shot Automatic Pronunciation Assessment,

    H. Liu, M. Shi, and Y . Wang, “Zero-Shot Automatic Pronunciation Assessment,” inInterspeech 2023, 2023, pp. 1009–

    2023

  25. [27]

    Available: https://arxiv.org/abs/2305.19563v1

    [Online]. Available: https://arxiv.org/abs/2305.19563v1

    work page arXiv

  26. [28]

    MultiPA: A Multi-task Speech Pronunciation Assessment Model for Open Response Scenarios,

    Y .-W. Chen, Z. Yu, and J. Hirschberg, “MultiPA: A Multi-task Speech Pronunciation Assessment Model for Open Response Scenarios,” inInterspeech 2024, 2024, pp. 297–301. [On- line]. Available: https://www.isca-archive.org/interspeech 2024/ chen24c interspeech.pdf

    2024

  27. [29]

    Towards efficient and multifaceted computer-assisted pronunciation training leveraging hierarchical selective state space model and decoupled cross-entropy loss,

    F.-A. Chao and B. Chen, “Towards efficient and multifaceted computer-assisted pronunciation training leveraging hierarchical selective state space model and decoupled cross-entropy loss,” inProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (V olume 1: Lon...

    2025

  28. [30]

    Cnn-rnn-ctc based end- to-end mispronunciation detection and diagnosis,

    W.-K. Leung, X. Liu, and H. Meng, “Cnn-rnn-ctc based end- to-end mispronunciation detection and diagnosis,” inICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 8132–

    2019

  29. [31]

    Available: https://www.semanticscholar.org/ paper/cea19ebccad6c798292e3a0f42ac247e85b1719c

    [Online]. Available: https://www.semanticscholar.org/ paper/cea19ebccad6c798292e3a0f42ac247e85b1719c

  30. [32]

    Phoneme mispronunciation detection by jointly learning to align,

    B. Lin and L. Wang, “Phoneme mispronunciation detection by jointly learning to align,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6822–6826. [Online]. Available: https://ieeexplore.ieee.org/document/9746727

    work page arXiv 2022

  31. [33]

    Peppanet: Effective mispronunciation detection and diagnosis leveraging phonetic, phonological, and acoustic cues,

    B.-C. Yan, H.-W. Wang, and B. Chen, “Peppanet: Effective mispronunciation detection and diagnosis leveraging phonetic, phonological, and acoustic cues,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 1045–1051. [Online]. Available: https://ieeexplore.ieee.org/document/10022472

    work page arXiv 2023

  32. [34]

    Automatic pronunciation assessment using self-supervised speech representation learning,

    E. Kim, J. Jeon, H. Seo, and H. Kim, “Automatic pronunciation assessment using self-supervised speech representation learning,”

  33. [35]

    Available: https://www.semanticscholar.org/ paper/5bb169370c6458654031c659f3c1f2837066b0f9

    [Online]. Available: https://www.semanticscholar.org/ paper/5bb169370c6458654031c659f3c1f2837066b0f9

  34. [36]

    Exploiting information from native data for non-native automatic pronunciation assessment,

    B. Lin and L. Wang, “Exploiting information from native data for non-native automatic pronunciation assessment,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 708–714. [Online]. Available: https://ieeexplore.ieee.org/ document/10022486

    work page arXiv 2023

  35. [37]

    Improving mispronunciation detection with wav2vec2-based momentum pseudo-labeling for accentedness and intelligibility assessment,

    M. Yang, K. Hirschi, S. D. Looney, O. Kang, and J. H. Hansen, “Improving mispronunciation detection with wav2vec2-based momentum pseudo-labeling for accentedness and intelligibility assessment,”arXiv preprint arXiv:2203.15937, 2022. [Online]. Available: https://par.nsf.gov/servlets/purl/10358184

    work page arXiv 2022

  36. [38]

    Multi-lingual pronunciation assessment with unified phoneme set and language-specific embeddings,

    B. Lin and L. Wang, “Multi-lingual pronunciation assessment with unified phoneme set and language-specific embeddings,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5. [Online]. Available: https://ieeexplore.ieee.org/document/ 10095673

    2023

  37. [39]

    Multi-view multi- task representation learning for mispronunciation detection,

    Y . EL Kheir, S. Chowdhury, and A. Ali, “Multi-view multi- task representation learning for mispronunciation detection,” inProc. SLaTE 2023, 2023, pp. 86–90. [Online]. Available: https://arxiv.org/abs/2306.01845v2

    work page arXiv 2023

  38. [40]

    Improving non-native word-level pronunciation scoring with phone-level mixup data augmentation and multi-source information,

    K. Fu, S. Gao, K. Wang, W. Li, X. Tian, and Z. Ma, “Improving non-native word-level pronunciation scoring with phone-level mixup data augmentation and multi-source information,”ArXiv, vol. abs/2203.01826, 2022. [Online]. Available: https://api. semanticscholar.org/CorpusID:247223012

    work page arXiv 2022

  39. [41]

    A full text-dependent end to end mispronunciation detection and diagnosis with easy data augmentation techniques,

    K. Fu, J. Lin, D. Ke, Y . Xie, J. Zhang, and B. Lin, “A full text-dependent end to end mispronunciation detection and diagnosis with easy data augmentation techniques,”arXiv preprint arXiv:2104.08428, 2021. [Online]. Available: https: //arxiv.org/abs/2104.08428v1

    work page arXiv 2021

  40. [42]

    L2-gen: A neural phoneme paraphrasing approach to l2 speech synthesis for mispronunciation diagnosis,

    D. Zhang, A. Ganesan, S. Campbell, and D. Korzekwa, “L2-gen: A neural phoneme paraphrasing approach to l2 speech synthesis for mispronunciation diagnosis,”Interspeech,

  41. [43]

    Available: https://www.semanticscholar.org/ paper/487f9e74fff5e29b8aa37ba551edab71cfcbd256

    [Online]. Available: https://www.semanticscholar.org/ paper/487f9e74fff5e29b8aa37ba551edab71cfcbd256

  42. [44]

    Detection of lexical stress errors in non-native (l2) english with data augmentation and attention,

    D. Korzekwa, R. Barra-Chicote, S. Zaporowski, G. Beringer, J. Lorenzo-Trueba, A. Serafinowicz, J. Droppo, T. Drugman, and B. Kostek, “Detection of lexical stress errors in non-native (l2) english with data augmentation and attention,” inInterspeech,

  43. [45]

    Available: https://www.semanticscholar.org/ paper/b721c93c3fb97c0cff2ce95d2558195ad02934ea

    [Online]. Available: https://www.semanticscholar.org/ paper/b721c93c3fb97c0cff2ce95d2558195ad02934ea

  44. [46]

    Speechblender: Speech augmentation framework for mispronunciation data generation,

    Y . EL Kheir, S. Chowdhury, A. Ali, H. Mubarak, and S. Afzal, “Speechblender: Speech augmentation framework for mispronunciation data generation,” inProc. 9th Workshop on Speech and Language Technology in Education (SLaTE), 2023, pp. 26–30. [Online]. Available: https://www.isca-archive.org/ slate 2023/elkheir23 slate.pdf

    2023

  45. [47]

    L1-aware multilingual mispronunciation detection framework,

    Y . Kheir, S. A. Chowdhury, and A. Ali, “L1-aware multilingual mispronunciation detection framework,” inICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12 752– 12 756. [Online]. Available: https://www.semanticscholar.org/ paper/a5064b6f851eaae6ff63a6179bc8070f2f2a517d

    2024

  46. [48]

    Automatic pro- nunciation assessment - a review,

    Y . Kheir, A. Ali, and S. Chowdhury, “Automatic pro- nunciation assessment - a review,” inFindings of the Association for Computational Linguistics: EMNLP 2023, H. Bouamor, J. Pino, and K. Bali, Eds., Dec. 2023. [Online]. Available: https://www.semanticscholar.org/paper/ 7e52ffa0ecdaa1b8eb6dc0757de532894bfd24b1

    2023

  47. [49]

    A comparison-based approach to mispronunciation detection,

    A. Lee and J. Glass, “A comparison-based approach to mispronunciation detection,” in2012 IEEE Spoken Lan- guage Technology Workshop (SLT). IEEE, 2012, pp. 382–

    2012

  48. [50]

    Available: https://www.semanticscholar.org/paper/ 621af45b51cf6a369a8946d9f3f006b0138b43d7

    [Online]. Available: https://www.semanticscholar.org/paper/ 621af45b51cf6a369a8946d9f3f006b0138b43d7

  49. [51]

    Relative dynamic time warp- ing comparison for pronunciation errors,

    C. Richter and J. Gudhnason, “Relative dynamic time warp- ing comparison for pronunciation errors,” inICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–

    2023

  50. [52]

    Available: https://www.semanticscholar.org/paper/ c17f97083f096cec719fe0d9106380b9e116cf8e

    [Online]. Available: https://www.semanticscholar.org/paper/ c17f97083f096cec719fe0d9106380b9e116cf8e

  51. [53]

    Zero-shot text- to-speech as golden speech generator: A systematic framework and its applicability in automatic pronunciation assessment,

    T.-H. Lo, M.-T. Tsai, Y .-T. Sung, and B. Chen, “Zero-shot text- to-speech as golden speech generator: A systematic framework and its applicability in automatic pronunciation assessment,” arXiv preprint arXiv:2409.07151, 2024. [Online]. Available: https://arxiv.org/abs/2409.07151v2

    work page arXiv 2024

  52. [54]

    Towards real-time mispronun- ciation detection in kids’ speech,

    P. Plantinga and E. Fosler-Lussier, “Towards real-time mispronun- ciation detection in kids’ speech,” in2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2019, pp. 690–696. [Online]. Available: https://www.semanticscholar. org/paper/b48db573051b3e5dd28673cde3c145be7760363f

    2019

  53. [56]

    [Online]. Available: https://arxiv.org/abs/2507.12217v1 Supplementary: Light-weight Pronunciation Assessment via Discrete Speech Token Surprisal Syeda Faiza Ahmed Sara ID 1, Shammur Absar Chowdhury ID 1,∗∗ 1 Qatar Computing Research Institute, Doha, Qatar syeda.faiza.ahmed@gmail.com, shchowdhury@hbku.edu.qa

    work page arXiv

  54. [57]

    All three studies are un- supervised, with no regression and no learner labels

    Tokenizer Selection Ablations This supplementary reports the full ablations used to select the discrete tokenizer in the main paper. All three studies are un- supervised, with no regression and no learner labels. For each configuration we compute the mean token surprisal on the Spee- chOcean762 test set and measure its correlation with the hu- man accurac...

    2048

  55. [58]

    HuBERT: Self-Supervised Speech Representa- tion Learning by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021. [Online]. Available: https://arxiv.org/abs/ 2106.07447v1

    work page arXiv 2021

  56. [59]

    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. Xiaoet al., “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022. [Online]. Available: http://dx.doi.org/10.1109/JSTSP.2022.3188113

    work page doi:10.1109/jstsp.2022.3188113 2022

  57. [60]

    Xls-r: Self-supervised cross-lingual speech representation learning at scale.arXiv preprint arXiv:2111.09296, 2021

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. V on Platen, Y . Saraf, J. Pinoet al., “Xls-r: Self- supervised cross-lingual speech representation learning at scale,” arXiv preprint arXiv:2111.09296, 2021. [Online]. Available: https: //www.isca-archive.org/interspeech 2022/babu22 interspeech.pdf

    work page arXiv 2021

  58. [61]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020. [On- line]. Available: https://proceedings.neurips.cc/paper/2020/file/ 92d1e1eb1cd6f9fba3227870bb6d7f07-Paper.pdf

    2020