Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Enhancing GOP in CTC-Based Mispronunciation Detection with Phonological Knowledge

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Phonologically restricted substitutions cut alignment-free GOP computation by about 90% while keeping accuracy close to unrestricted search.

desk verdict A useful but unevenly evidenced idea: restricting CTC substitution sets with a phoneme confusion map, where the headline efficiency win is an arithmetic example rather than a measured runtime saving. read the letter →

arxiv 2506.02080 v2 pith:ZKLF5U5M submitted 2025-06-02 eess.AS cs.AI

Aditya Kamlesh Parikh , Cristian Tejedor-Garcia , Catia Cucchiarini , Helmer Strik This is my paper
classification eess.AScs.AI
keywords goodnessofpronunciationGOPphonemerecognitionComputer-AssistedTrainingmispronunciationdetectionConnectionistTemporalClassificationconfusionmapalignment-free
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

Mispronunciation detection in language-learning apps needs to score pronunciation in real time, but the standard goodness-of-pronunciation (GOP) measure depends on forced alignments that are unreliable on non-native speech. Alignment-free GOP avoids that problem by computing sequence-level connectionist-temporal-classification (CTC) likelihoods directly, yet it scales poorly: with 39 phonemes and a 10-phoneme utterance it must consider 390 substitutions and deletions. The paper proposes a substitution-aware variant that restricts replacements to a handcrafted Phoneme Confusion Map built from articulatory similarity, common learner errors, and phonological rules, cutting the count to about 40 (a 90% computation reduction). In restricted (RPS) and unrestricted (UPS) configurations on two L2 English datasets—Dutch children and Mandarin speakers—the alignment-free methods outperform the forced-alignment baseline on most metrics, and the paper concludes that substitution-aware GOP gives the efficiency-accuracy balance needed for instantaneous feedback.

What carries the argument

The central object is the Phoneme Confusion Map, a handcrafted mapping from each phoneme to a small set of allowable substitutions derived from three criteria: phonetic proximity (shared place or manner of articulation), common L2 learner errors, and phonological rules such as allophonic variants and vowel mergers. The map does the work of shrinking the substitution search space from the whole inventory ($V-1$ alternatives per phoneme) to a few phonologically plausible candidates. It appears in two instantiations: PA-AF GOP integrates substitution and deletion directly into a modified CTC forward pass through position-specific perturbations and state-dependent token masking, while PP-AF GOP generates perturbed label sequences outside the CTC computation and uses the minimum CTC-loss difference as the GOP score. Because every detected mispronunciation must pass through the map, the map simultaneously carries the claimed computation reduction and the accuracy ceiling of the method.

What would settle it

On SpeechOcean762, restrict the evaluation to annotated errors whose actual phoneme is not among the map's allowed substitutions for the canonical phoneme and compare RPS recall on that subset with UPS recall; if RPS recall is near zero, the map's coverage is the bottleneck. As a second check, replace MPC's artificially injected errors (generated from the same phonological rules that built the map) with independently collected learner mispronunciations and see whether the restricted-substitution advantage over forced alignment persists.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that an alignment-free CTC-based GOP can be made computationally practical for real-time Computer-Assisted Pronunciation Training by constraining, for each canonical phoneme, the set of phonemes that are allowed to substitute it. The constraint is the Phoneme Confusion Map, built from phonetic proximity, common L2 learner errors, and phonological rules, and it is applied in two ways: the Phoneme-Adaptive variant (PA-AF GOP), which masks non-confusable transitions inside the CTC forward pass, and the Phoneme-Perturbed variant (PP-AF GOP), which externally perturbs the label sequence at each position and scores by the minimum CTC-loss difference between original and perturbed sequences. With three confusable alternatives per phoneme the calculation count drops from $n(V-1)+n = 390$ to roughly 40, about 90 percent fewer, for the paper's worked example. In the experiments, both alignment-free variants outperform the forced-alignment baseline on most metrics on MPC and SpeechOcean762, restricted substitution gives the highest precision (PP-AF RPS, 0.509 on MPC), and unrestricted substitution gives the highest recall and MCC (PP-AF UPS, 0.595 on MPC; MCC 0.395, high-confidence PCC 0.502, and MSE 0.104 on SpeechOcean762). The paper's stated conclusion is that substitution-aware methods 'provide a balance between efficiency and accuracy' suitable for applications requiring instantaneous feedback.

Load-bearing premise

The load-bearing premise is that the handcrafted Phoneme Confusion Map covers the substitutions learners actually make; if a real mispronunciation is absent from the map, the restricted system cannot flag it because neither variant ever considers that alternative.

Editorial extensions

If this is right

  • Per-phoneme computation in alignment-free GOP drops from $O(nV)$ to $O(nk)$, where $k$ is the size of the confusion set, making the method fast enough for real-time feedback in CAPT.
  • Restricted substitution trades some recall for higher precision, so fewer correctly pronounced phonemes are flagged as errors in read-aloud exercises.
  • Because the approach is model-agnostic and the map is a linguistic resource, the same machinery transfers to other CTC-based phoneme recognizers and, with a different map, to other learner populations and languages.
  • Even restricted alignment-free GOP aligns better with human pronunciation ratings than forced alignment on non-native speech, as seen in the higher PCC and lower MSE on SpeechOcean762.
  • Unrestricted substitution remains the accuracy upper bound, and the paper's contribution is to show how much of that bound survives at roughly one-tenth of the computation.

Reading between the lines

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

  • Beyond the paper: because MPC's artificial errors were generated from the same phonological rules used to build the confusion map, the restricted-substitution results on MPC likely overstate real-world coverage; an independent corpus of naturally occurring learner errors would be a fairer test.
  • Beyond the paper: the extra recall of unrestricted substitution on SpeechOcean762 can be decomposed; if most of it comes from substitutions outside the map, the 90% computation saving is paid for with blindness to exactly the errors learners make.
  • Beyond the paper: the same restriction idea could model insertion errors in spontaneous speech, which the paper deliberately excludes; conversational CAPT, where insertions are common, would need an insertion confusion set to keep the efficiency gain.
  • Beyond the paper: a hybrid deployment could run restricted substitution for instant feedback during practice and defer unrestricted (or richer contextual) scoring to offline reports, following the efficiency-accuracy trade-off the paper maps out.
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

4 major / 5 minor

Summary. The paper proposes substitution-aware alignment-free CTC-based GOP methods that restrict phoneme substitutions using a handcrafted Phoneme Confusion Map. Two variants are introduced: PA-AF GOP, which integrates position-specific substitutions and deletions into a modified CTC forward pass, and PP-AF GOP, which evaluates externally perturbed label sequences. Both are tested under restricted (RPS) and unrestricted (UPS) substitutions on the MPC and SpeechOcean762 datasets. The central claim is that Substitution-Aware Alignment-Free GOP methods balance efficiency and accuracy for instantaneous CAPT feedback.

Significance. If the efficiency claim were substantiated, the paper would offer a useful practical contribution to CAPT: restricting CTC substitutions to phonologically plausible alternatives is an intuitive way to reduce the search space of alignment-free GOP. The authors also provide positive evidence that alignment-free methods can outperform a forced-alignment baseline, and the use of two datasets with different learner populations is a strength. However, the efficiency half of the balance claim rests on an arithmetic example rather than on runtime measurements, and the accuracy evaluation is affected by test-set threshold selection and, on MPC, by artificial errors generated from the same phonological rules that define the substitution map. These issues need to be addressed before the central claim can be accepted. The paper is clearly written and the code is made available, which supports reproducibility.

major comments (4)
  1. [§2.2–§2.3, §5] The '90% reduction in computation' claim is not supported by the evidence presented. The calculation in §2.2 counts the number of candidate perturbed sequences (n(V−1)+n = 390 for V=39, n=10), but this is not the number of operations performed by the implemented methods. PP-AF GOP indeed evaluates each perturbed sequence as a separate CTC forward pass, but PA-AF GOP computes the denominator in a single modified CTC forward pass with position-specific perturbations and state-dependent token masking; its cost scales with sequence length, label count, and per-state emission complexity, not with the number of separately evaluated sequences. Moreover, the arithmetic uses V=39 while the actual acoustic model has 387 phonetic labels (§3.2). No wall-clock time, FLOPs, or any runtime measurement is reported. Since the paper explicitly motivates the method by efficiency and the accuracy tables show RPS is consistently worse than UPS on both datasets (e.g., MCC 0.489 vs 0.595 on MPC; 0.327 vs 0.395 on SpeechOcean762), the claimed 'balance between efficiency and accuracy' in §5 is not established. Please provide runtime measurements for both PA-AF and PP-AF GOP under RPS and UPS, or a complexity analysis that accounts for the actual implementation and the 387-label inventory.
  2. [§3.3] The threshold is optimized on the test set: 'we optimized the threshold by selecting the GOP percentile that maximized MCC' and then report MCC, F1, precision, and recall at that threshold. This leaks test-set labels into model selection and yields optimistically biased results. The reported numbers in Tables 1 and 2 therefore cannot be taken at face value as estimates of how the methods would perform in practice. Please use a development set for threshold selection (or nested cross-validation) and report test-set metrics obtained with the selected threshold.
  3. [§3.1.1, §2.3.1] The MPC evaluation uses artificial errors that are generated from the same set of phonological rules and learner-error patterns used to build the Phoneme Confusion Map. Replacing /D/ with /d/, /T/ with /s/, /æ/ with /e/, and simplifying diphthongs are exactly the kinds of substitutions the map is designed to capture. The RPS results on MPC are therefore partly self-confirming: the artificial test errors are drawn from the map's own substitution sets. This does not independently validate the map's coverage. The SpeechOcean762 natural annotations provide an independent check, but the artificial-error design should be acknowledged as a limitation and, where possible, a held-out set of naturally occurring learner mispronunciations should be used for evaluation.
  4. [§2.3.1, §4] The handcrafted Phoneme Confusion Map is the key component of the proposed method, but it is not validated against independent learner error data. The paper itself notes that if the correct mispronunciation is not in the predefined confusing phoneme pairs, the CTC function cannot align it correctly. This coverage risk is real: for the MPC dataset (Dutch children) and SpeechOcean762 (Mandarin-speaking adults), the map is expected to generalize across very different L1 backgrounds, yet no evidence is given that the map actually covers the substitutions that occur in these populations. Please provide either an analysis of the map's coverage on an independent corpus with annotated mispronunciations or a comparison against a data-driven confusion set estimated from the acoustic model.
minor comments (5)
  1. [§2.2] The phrase 'if each phoneme is associated with at least three confusable phoneme pairs, the total number of calculations is reduced from 390 to 40' assumes each phoneme has exactly three alternatives; for phonemes with fewer than three, the reduction is smaller, and the actual inventory size of the acoustic model (387 labels) is not the V=39 used in the example.
  2. [§3.2] Section 3.2 refers to 'the classical approach, as outlined in Equation 2,' but Equation 2 is the DNN-based GOP; the classical HMM-GMM GOP is defined in Equation 1. Please correct the cross-reference.
  3. [Tables 1 and 2] The header of Table 1 contains a typo ('setups' for 'setups'), and the acronyms PA-AF and PP-AF are not expanded in the table captions.
  4. [§4] The paper does not report confidence intervals or significance tests for any of the metrics except PCC; given the moderate dataset sizes, it would be helpful to know whether differences such as PP-AF RPS vs UPS on SpeechOcean762 (MCC 0.327 vs 0.395) are statistically meaningful.
  5. [References] Reference [25] lists the publication year as 2011 in the entry while the text says "Interspeech 2021"; please correct the inconsistency.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: MPC artificial errors are generated from the same phonological rules that define the Phoneme Confusion Map, making RPS coverage on MPC partly self-referential; natural SpeechOcean762 annotations and the separate efficiency count keep the central claim partly independent.

  1. self definitional [Section 2.3.1 (Substitution Mapping Construction) and Section 3.1.1 (My Pronunciation Coach Dataset)]
    "A handcrafted Phoneme Confusion Map is developed based on three main criteria: ... (2) common L2 learner errors, based on empirical observations of non-native speech patterns, reflecting frequent pronunciation mistakes (e.g., dental fricative substitutions: / T/ → [/ D/, /f/]); and (3) phonological rules, prioritizing allophonic variants ... and vowel mergers. ... As MPC lacks annotated mispronunciations, we introduced artificial errors by modifying phoneme sequences."

    The substitution map and the artificial MPC labels are constructed from the same phonological knowledge: the map is explicitly based on common learner errors and phonological rules, while the injected mispronunciations are dental-fricative substitutions, vowel mergers, and diphthong simplifications of exactly that type. Hence RPS coverage on MPC is partly guaranteed by construction: the test set does not provide an independent check of whether the map covers real learner errors; it checks whether the map contains the same rules used to fabricate the errors. This is a self-definitional element in one of the two evaluations, not a formal equivalence of the derived GOP formulas.

full rationale

The paper's central derivation (alignment-free GOP as a CTC loss ratio, and the substitution-aware restricted variants) is mathematically self-contained and not equivalent to its inputs; no equation is fitted to the quantity it predicts, and the self-citations to earlier learner-error studies are empirical references rather than load-bearing uniqueness claims. The main circularity concern is the MPC evaluation: the artificial mispronunciations introduced there are generated from the same phonological rules and common learner-error patterns that define the handcrafted Phoneme Confusion Map, so the RPS performance on MPC is partly testing the map against its own construction criteria. The 90% computation-reduction claim is a definitional count of candidate substitution sequences, not a circular derivation, though the paper does not measure whether that count translates into runtime for PA-AF GOP's single masked CTC pass. SpeechOcean762 provides independent human-annotated evidence, and on both datasets the unrestricted UPS baseline generally matches or beats RPS, which shows the comparison is not simply rigged in favor of the restricted map. Overall, there is one partial self-definitional element in the evaluation, but the central method and its main efficiency argument are not circular, so a moderate score of 4 is appropriate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central method rests on a handcrafted confusion map (a non-numeric design choice), on the validity of the underlying alignment-free GOP framework, on the reliability of the chosen wav2vec2 acoustic model for L2 child and Mandarin-accented speech, and on the representativeness of artificially injected errors for real learner mistakes. The reported evaluation metrics further depend on a threshold chosen to maximize MCC on the test set and on a polynomial regression whose fitting protocol is not specified.

free parameters (3)
  • GOP decision threshold (MCC-maximizing percentile) = Not reported; selected per dataset to maximize MCC
    Section 3.3: threshold chosen as the GOP percentile that maximizes MCC on the same dataset, so reported metrics are conditional on test-set fitting.
  • Second-order polynomial regression coefficients for GOP-to-human-score mapping = Not reported
    Section 3.3: PCC and MSE for SpeechOcean762 are computed after fitting a second-order polynomial to relate GOP scores to human ratings; the train/test split protocol is not specified.
  • Phoneme Confusion Map = Handcrafted, no numeric values
    Section 2.3.1: handcrafted substitution sets based on phonetic proximity, common learner errors, and phonological rules; this mapping is a design choice not fitted to data, and the full map is not reproduced in the paper.
assumptions (4)
  • domain assumption The CTC alignment-free GOP framework of Cao et al. [19] is a valid basis for phoneme-level pronunciation scoring.
    Section 2.2 adapts Eq. 3 from [19] without re-derivation.
  • domain assumption The wav2vec2-xlsr-53-espeak-cv-ft acoustic model produces reliable phoneme posteriors for L2 child and Mandarin-accented speech.
    Section 3.2 uses this openly available model as the acoustic backbone for all alignment-free GOP variants.
  • ad hoc to paper Artificial errors inserted into MPC sequences resemble naturally occurring learner mispronunciations.
    Section 3.1.1: MPC lacks mispronunciation annotations, so errors are simulated by substituting a small set of phonemes; these substitutions align with the phonological rules used to build the confusion map.
  • domain assumption Forced alignment from a Kaldi HMM-GMM trained on LibriSpeech 100h provides a fair baseline.
    Section 3.2: the FA baseline is the classical GOP; the choice of LibriSpeech-trained model may not be optimal for child L2 speech.
invented entities (1)
  • Handcrafted Phoneme Confusion Map
    purpose: Constraint set limiting allowed substitution targets in RPS alignment-free GOP
    The paper provides no independent validation of the map's coverage or accuracy; on MPC the artificial errors are generated from similar phonological rules, so map adequacy is not tested against naturally occurring error distributions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing GOP in CTC-Based Mispronunciation Detection with Phonological Knowledge." pith.science (2026). https://pith.science/paper/ZKLF5U5M

@misc{pith2026250602080,
  author       = {Pith},
  title        = {Pith review of: Enhancing GOP in CTC-Based Mispronunciation Detection with Phonological Knowledge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKLF5U5M}},
  note         = {Machine review of arXiv:2506.02080}
}
read the original abstract

Computer-Assisted Pronunciation Training (CAPT) systems employ automatic measures of pronunciation quality, such as the goodness of pronunciation (GOP) metric. GOP relies on forced alignments, which are prone to labeling and segmentation errors due to acoustic variability. While alignment-free methods address these challenges, they are computationally expensive and scale poorly with phoneme sequence length and inventory size. To enhance efficiency, we introduce a substitution-aware alignment-free GOP that restricts phoneme substitutions based on phoneme clusters and common learner errors. We evaluated our GOP on two L2 English speech datasets, one with child speech, My Pronunciation Coach (MPC), and SpeechOcean762, which includes child and adult speech. We compared RPS (restricted phoneme substitutions) and UPS (unrestricted phoneme substitutions) setups within alignment-free methods, which outperformed the baseline. We discuss our results and outline avenues for future research.

Figures

Figures reproduced from arXiv: 2506.02080 by the authors.

Figure 1
Figure 1. An illustrative example of the transition of the phonemes in the substitution mapping construction. 3. Experimental Procedure 3.1. Datasets 3.1.1. My Pronunciation Coach Dataset To answer our RQ, we conducted experiments with two datasets of L2 English speech. The first one, the MPC speech database [27], is particularly challenging as it contains L2 speech of chil￾dren (124 in total) learning English in Dutch second… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    Introduction Language is a fundamental skill that shapes communication, cognitive development, and cultural integration [1] and learning languages other than the native one is essential in our globalized society. Traditional classroom settings often make it difficult for teachers to provide the degree of individualized attention that is required for high-...

    work page

  2. [2]

    Enhancing GOP in CTC-Based Mispronunciation Detection with Phonological Knowledge

    Methodology 2.1. GOP Definitions First, we follow the definition of GOP by Witt and Young [11]. They compute GOP using the sequence of feature vec- tors OT 1 = {o1, . . . ,oT } of length T and the corresponding arXiv:2506.02080v2 [eess.AS] 8 Jul 2025 canonical phoneme transcription Lcano = {l1, . . . , l|Lcano|}. For a given phoneme li ∈ Lcano, the origin...

    work page Pith review arXiv 2025

  3. [3]

    Datasets 3.1.1

    Experimental Procedure 3.1. Datasets 3.1.1. My Pronunciation Coach Dataset To answer our RQ, we conducted experiments with two datasets of L2 English speech. The first one, the MPC speech database [27], is particularly challenging as it contains L2 speech of chil- dren (124 in total) learning English in Dutch secondary schools. Child speech presents speci...

    work page

  4. [4]

    Results Table 1 shows the experimental results on the MPC dataset. We report results for both forced alignment (FA) as the base- line, and two alignment-free approaches (PA-AF GOP, PP-AF GOP), with the latter evaluated under both RPS and UPS se- tups. Both PA-AF GOP and PP-AF GOP outperform FA GOP scores in most metrics, except recall, where FA achieves t...

    work page

  5. [5]

    However, it is important to note that these ap- proaches may also lead to a decline in performance compared to unrestricted phoneme substitutions

    Discussion and Conclusion The results obtained in this work answered our RQ, demonstrat- ing that phoneme clustering and learner-specific error modeling can reduce computational costs in an alignment-free CTC-based MDD system. However, it is important to note that these ap- proaches may also lead to a decline in performance compared to unrestricted phonem...

    work page

  6. [6]

    Acknowledgements This publication is part of the project Responsible AI for V oice Diagnostics (RAIVD) with file number NGF.1607.22.013 of the research programme NGF AiNed Fellowship Grants which is financed by the Dutch Research Council (NWO)

  7. [7]

    Language, culture, and cog- nition in cross-cultural communication

    L. Nardon, R. Steers, and C. Stone, “Language, culture, and cog- nition in cross-cultural communication.” in Proceedings of the new frontiers in management and organizational cognition con- ference. National University of Maynooth, 2012

    work page 2012

  8. [8]

    Teachers’ perspectives on the use of differentiated instruction in inclusive classrooms: Implica- tion for teacher education

    C. N. Onyishi and M. M. Sefotho, “Teachers’ perspectives on the use of differentiated instruction in inclusive classrooms: Implica- tion for teacher education.” International Journal of Higher Edu- cation, vol. 9, no. 6, pp. 136–150, 2020

    work page 2020

Show all 42 references
  1. [9]

    Feedback in computer assisted pronunciation training: When technology meets peda- gogy,

    A. Neri, C. Cucchiarini, and H. Strik, “Feedback in computer assisted pronunciation training: When technology meets peda- gogy,” in 10th International CALL Conference on CALL profes- sionals and the future of CALL research . Antwerpen: Univer- siteit Antwerpen, 2002, pp. 179–188

    2002

  2. [10]

    Oral proficiency training in dutch l2: The contribution of asr-based corrective feedback,

    C. Cucchiarini, A. Neri, and H. Strik, “Oral proficiency training in dutch l2: The contribution of asr-based corrective feedback,” Speech Communication, vol. 51, no. 10, pp. 853–863, 2009

    2009

  3. [11]

    ASR corrective feedback on pronunciation: Does it really work?

    A. Neri, C. Cucchiarini, and H. Strik, “ASR corrective feedback on pronunciation: Does it really work?” in Interspeech 2006 . ISCA, 2006, pp. 1372–Wed3A3O.2

    2006

  4. [12]

    Computer-assisted pronunciation training: A systematic review,

    M. Amrate and P.-h. Tsai, “Computer-assisted pronunciation training: A systematic review,” ReCALL, vol. 37, no. 1, p. 22–42, 2025

    2025

  5. [13]

    Teaching pronunciation with computer assisted pronunciation instruction in a technological university

    S.-C. Liu and P.-Y . Hung, “Teaching pronunciation with computer assisted pronunciation instruction in a technological university.” Universal Journal of Educational Research , vol. 4, no. 9, pp. 1939–1943, 2016

    1939

  6. [14]

    Corrective feedback accuracy and pronunciation improvement: Feedback that is ‘good enough’,

    A. Silpachai, R. Neiriz, M. Novotny, R. Gutierrez-Osuna, J. Levis, and E. Chukharev-Hudilainen, “Corrective feedback accuracy and pronunciation improvement: Feedback that is ‘good enough’,” Language Learning & Technology, vol. 28, pp. 1–16, 2024

    2024

  7. [15]

    Automatic pronun- ciation error detection in non-native speech: The case of vowel er- rors in dutch,

    J. v. Doremalen, C. Cucchiarini, and H. Strik, “Automatic pronun- ciation error detection in non-native speech: The case of vowel er- rors in dutch,” The Journal of the Acoustical Society of America , vol. 134, no. 2, pp. 1336–1347, 2013

    2013

  8. [16]

    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 com- munication, vol. 30, no. 2-3, pp. 95–108, 2000

    2000

  9. [17]

    Use of speech recognition in computer-assisted lan- guage learning

    S. M. Witt, “Use of speech recognition in computer-assisted lan- guage learning.” Ph.D. dissertation, University of Cambridge, 2000

    2000

  10. [18]

    Using non-native error patterns to improve pronunciation verification,

    J. van Doremalen, C. Cucchiarini, and H. Strik, “Using non-native error patterns to improve pronunciation verification,” in Inter- speech 2010. ISCA, 2010, pp. 590–593

    2010

  11. [19]

    Lattice-based gop in automatic pronunciation evaluation,

    Y . Song, W. Liang, and R. Liu, “Lattice-based gop in automatic pronunciation evaluation,” in2010 The 2nd International Confer- ence on Computer and Automation Engineering (ICCAE), vol. 3. IEEE, 2010, pp. 598–602

    2010

  12. [20]

    Context-aware goodness of pronuncia- tion for computer-assisted pronunciation training,

    J. Shi, N. Huo, and Q. Jin, “Context-aware goodness of pronuncia- tion for computer-assisted pronunciation training,” in Interspeech

  13. [21]

    CTC Variations Through New WFST Topologies,

    A. Laptev, S. Majumdar, and B. Ginsburg, “CTC Variations Through New WFST Topologies,” inInterspeech 2022, 2022, pp. 1041–1045

    2022

  14. [22]

    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

  15. [23]

    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

  16. [24]

    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

    2022

  17. [25]

    Self-supervised models for phoneme recognition: Applications in children’s speech for reading learning,

    L. B. Medin, T. Pellegrini, and L. Gelin, “Self-supervised models for phoneme recognition: Applications in children’s speech for reading learning,” in Interspeech 2024. ISCA, 2024, pp. 5168– 5172

    2024

  18. [26]

    A framework for phoneme-level pronunciation assessment using ctc,

    X. Cao, Z. Fan, T. Svendsen, and G. Salvi, “A framework for phoneme-level pronunciation assessment using ctc,” in Inter- speech 2024. ISCA, 2024, pp. 302–306

    2024

  19. [27]

    The Impact of Forced-Alignment Errors on Automatic Pronunciation Evaluation,

    V . C. Mathad, T. J. Mahr, N. Scherer, K. Chapman, K. C. Hus- tad, J. Liss, and V . Berisha, “The Impact of Forced-Alignment Errors on Automatic Pronunciation Evaluation,” in Interspeech, 2021, pp. 1922–1926

    2021

  20. [28]

    Phonological difficulties encountered by italian learners of english: An error analysis,

    A. Wheelock, “Phonological difficulties encountered by italian learners of english: An error analysis,” Hawaii Pacific University TESOL Working Paper Series, vol. 14, pp. 41–61, 2016

    2016

  21. [29]

    Clustering approach in speech phoneme recognition based on statistical analysis,

    G. K. Tak and V . Bhargava, “Clustering approach in speech phoneme recognition based on statistical analysis,” in RTNSA, CNSA 2010, Chennai, India, July 23-25, 2010. Proceedings 3 . Springer, 2010, pp. 483–489

    2010

  22. [30]

    Multilingual acoustic modeling method based on phoneme clustering,

    M. Meng, J. Liang, and B. Xu, “Multilingual acoustic modeling method based on phoneme clustering,” Pattern Recognition and Artificial Intelligence, 2009

    2009

  23. [31]

    Hierarchical phoneme classification for improved speech recognition,

    D. Oh, J.-S. Park, J.-H. Kim, and G.-J. Jang, “Hierarchical phoneme classification for improved speech recognition,”Applied Sciences, vol. 11, no. 1, p. 428, 2021

    2021

  24. [32]

    Error selection for asr-based english pronunciation training in’my pro- nunciation coach’,

    C. Cucchiarini, H. v. d. Heuvel, E. Sanders, and H. Strik, “Error selection for asr-based english pronunciation training in’my pro- nunciation coach’,” in Interspeech 2021. Florence, Italy: sn, 2011

    2021

  25. [33]

    Pronunciation errors made by Dutch secondary school students in English,

    V . Kruitbosch, “Pronunciation errors made by Dutch secondary school students in English,” Master’s thesis, Radboud University, 2020

    2020

  26. [34]

    My pronunciation coach: Improving english pronunciation with an automatic coach that listens,

    C. Cucchiarini, W. Nejjari, and H. Strik, “My pronunciation coach: Improving english pronunciation with an automatic coach that listens,” Language Learning in Higher Education , vol. 1, no. 2, pp. 365–376, 2012

    2012

  27. [35]

    A hi- erarchical context-aware modeling approach for multi-aspect and multi-granular pronunciation assessment,

    F.-A. Chao, T.-H. Lo, T.-I. Wu, Y .-T. Sung, and B. Chen, “A hi- erarchical context-aware modeling approach for multi-aspect and multi-granular pronunciation assessment,” in Interspeech 2023 . ISCA, 2023, pp. 974–978

    2023

  28. [36]

    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,” inIn- terspeech 2021. ISCA, 2021, pp. 3710–3714

    2021

  29. [37]

    The kaldi speech recognition toolkit,

    D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek, Y . Qian, P. Schwarzet al., “The kaldi speech recognition toolkit,” inIEEE 2011 workshop on automatic speech recognition and understanding. IEEE Signal Processing Society, 2011

    2011

  30. [38]

    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,” in ICASSP, 2015, pp. 5206–5210

    2015

  31. [39]

    Phonemizer: Text to phones transcrip- tion for multiple languages in python,

    M. Bernard and H. Titeux, “Phonemizer: Text to phones transcrip- tion for multiple languages in python,” Journal of Open Source Software, vol. 6, no. 68, p. 3958, 2021

    2021

  32. [40]

    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,” in Interspeech 2022, 2022, pp. 2113–2117

    2022

  33. [41]

    Common 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, “Common voice: A massively-multilingual speech corpus,” in LREC 2020. Marseille, France: ELRA, May 2020, pp. 4218–4222

    2020

  34. [2020]

    3057–3061

    ISCA, 2020, pp. 3057–3061

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.