Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Scaling and Prompting for Improved End-to-End Spoken Grammatical Error Correction

T0 review · 3 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 The paper claims that end-to-end Whisper models, trained on pseudo-labelled learner audio and prompted with fluent transcriptions, can outperform a cascaded ASR-plus-GEC pipeline for spoken grammatical error correction.

desk verdict Solid empirical study of E2E spoken GEC, but the headline scaling claim rests on a small.en WER gap that the paper's own significance markers don't back. read the letter →

arxiv 2505.21137 v1 pith:MFSL6STC submitted 2025-05-27 cs.CL cs.SDeess.AS

Mengjie Qian , Rao Ma , Stefano Bannò , Kate M. Knill , Mark J.F. Gales This is my paper
classification cs.CLcs.SDeess.AS
keywords spokengrammaticalerrorcorrectionend-to-endspeechmodelWhisperfine-tuningpseudo-labellingpromptingwithfluenttranscriptionslearnerfeedbackgenerationsecondlanguageassessmentdisfluencyremoval
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that an end-to-end spoken grammatical error correction system can match or beat the traditional three-module cascade of speech recognition, disfluency removal, and text correction. The first lever is scale: a fully automatic pseudo-labelling pipeline turns roughly 2,500 hours of unlabelled learner audio into training material, about thirty times the 77 hours of manually annotated data. The second lever is prompting: feeding the model a fluent transcription alongside the audio gives it cleaner context, improving both correction quality and learner feedback. On the Linguaskill test set the best small model reaches 12.72% WER versus 13.24% for the cascade, and the best large model with prompts reaches 11.04%.

What carries the argument

The central mechanism is a four-step pseudo-labelling loop that converts unlabelled audio into GEC training pairs: Whisper small.en ASR with timestamp and truecasing information, punctuation-based segmentation, Whisperflt for fluent transcripts, and BART GEC for phrase-level corrected text. This loop produces roughly 2,500 hours of training audio from the unlabelled Linguaskill data. The second mechanism is prompt conditioning: Whispergec is fine-tuned with audio as input, GEC transcriptions as references, and Whisperflt transcriptions as prompts, with SpecAugment applied during Whisperflt decoding on the training set to align the fluency of training prompts with dev-time prompts.

What would settle it

Regenerate the pseudo-labels with a large-v2 teacher instead of the small.en teacher and retrain Whispergec large-v2 on them; if WER does not improve over the no-pseudo-label large-v2 model, the paper's teacher-size explanation is falsified. In parallel, a paired significance test on the small.en S&I dev comparison (16.84% vs 16.91%) would settle whether that cross-corpus gain is real.

Watch

Extended reading notes

Core claim

The central claim is that Whispergec, an end-to-end Whisper model fine-tuned to output grammatically corrected transcriptions directly from audio, can outperform the cascaded Whisperflt-plus-BART-GEC baseline when trained with pseudo-labelled data and prompted with fluent transcriptions. The pseudo-labelling pipeline generates disfluent ASR transcriptions with Whisper small.en, segments audio by punctuation, produces fluent transcriptions with Whisperflt, and applies a BART text GEC system to yield phrase-level corrected transcriptions. Training on these pseudo-labels first and then fine-tuning on labelled data lowers small.en WER from 13.48% to 12.72% on Linguaskill test and from 17.76% to 16.84% on the S&I dev set. Prompting with fluent transcriptions adds further gains, and for large-v2 the best prompted model reaches 11.04% WER on Linguaskill and 13.08% on S&I, with feedback F0.5 scores of 41.08 and 42.62, closely matching or exceeding the cascade.

Load-bearing premise

The whole result rests on the assumption that the automatic pseudo-labelling pipeline, built from a small ASR model, a fluency filter, and a text GEC model, produces labels good enough to teach a larger end-to-end model, and that the gains seen on one test set carry over to other learner populations.

Editorial extensions

If this is right

  • On the Linguaskill test set, the best small.en model reaches 12.72% WER, beating the 13.24% cascaded baseline, and the best large-v2 model with prompts reaches 11.04%.
  • Training on pseudo-labelled audio alone gives 14.16% WER, within about 5% relative of training on manually labelled data at 13.48%, showing unlabelled learner audio can substitute for scarce annotation.
  • Prompting with fluent transcriptions improves feedback generation as well as correction, with large-v2 Whispergec+text-flt reaching F0.5 of 41.08 on Linguaskill and 42.62 on S&I, matching or exceeding the cascade and shrinking the feedback gap reported in prior work.
  • Larger models benefit more from labelled data but less from pseudo-labels; the paper attributes this to the small.en teacher used in pseudo-labelling, though it does not directly test that explanation.

Reading between the lines

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

  • An iterative relabelling loop, using the best prompted E2E model to regenerate pseudo-labels and retrain, is a natural extension that might restore pseudo-label gains for large-v2; the paper stops at a fixed small.en teacher.
  • Because the fluent transcription prompt is already available at inference time from Whisperflt, a two-pass decoding setup could apply prompting at test time rather than only during training; the paper does not report this variant.
  • The small.en advantage on the S&I dev set is only 0.07% WER with no significance test reported, so the cross-corpus generalisation claim rests mainly on the large-v2 numbers rather than on the small.en comparison.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper investigates end-to-end spoken grammatical error correction (SGEC) using Whisper-based models. It proposes a pseudo-labelling pipeline that expands the training set from 77 hours to about 2500 hours of unlabelled Linguaskill audio, and it also explores prompting the Whisper decoder with fluent transcriptions generated by a Whisperflt model. Experiments on the Linguaskill test set and the Speak & Improve (S&I) dev set compare Whispergec models of two sizes (small.en and large-v2) against a cascaded Whisperflt+GEC baseline. The results show that for small.en, pseudo-labelled data followed by labelled fine-tuning improves WER over the labelled-only model (12.72 vs 13.48 on LNG; 16.84 vs 17.76 on S&I) and slightly surpasses the cascaded baseline; for large-v2, pseudo-labelling does not help, while prompting with fluent transcriptions yields small improvements. Feedback performance, measured with M2/ERRANT F0.5, improves with prompting and model size, closing much of the gap to the cascaded system.

Significance. If the results are robust, this is a useful contribution to SGEC: it shows a practical way to leverage large amounts of unlabelled speech via pseudo-labelling, demonstrates that prompting a Whisper-based GEC model with fluent transcriptions helps both correction and feedback generation, and provides a clear comparison across model sizes. The paper is honest in reporting that pseudo-labelling does not help the large-v2 model, and it evaluates on held-out human-annotated test sets (Linguaskill test, S&I dev), which mitigates circularity concerns. The main reservation is statistical: the headline small.en improvements over the cascaded baseline are not marked as significant in the paper's own significance convention, and the S&I gains are very small in absolute terms.

major comments (3)
  1. [§4.1, Table 3] The central scaling claim rests on small.en WER differences that Table 3 does not mark as statistically significant. The footnote to Table 3 defines † as an improvement over the cascaded system significant at p<0.001, and only the large-v2 rows carry †; no small.en row is marked. Nevertheless, §4.1 states that the small.en model fine-tuned on LNGlbl after pseudo-labelling 'outperforms the cascaded system by 4.0% relatively' (12.72 vs 13.24 on LNG), while the same row is only 0.07% absolute better than the cascaded system on S&I (16.84 vs 16.91). Because the large-v2 pseudo-labelling result is negative (Table 2: 11.10 vs 11.10 on LNG; 13.93 vs 13.21 on S&I), the positive evidence for the scaling contribution is exactly this unannotated small.en comparison. Please report significance tests or confidence intervals for all pairwise comparisons, including the small.en deltas and the large-v2 negative result, and adjust the abstract, §4.1, and §5 claims accordingly.
  2. [§4.1, §4.2] The explanation offered for the large-v2 pseudo-labelling failure is post hoc and untested. The text attributes the lack of gain to the fact that 'the pseudo-GEC transcriptions are generated using the small.en model,' citing size mismatch and lower transcription quality, but no experiment replaces the teacher (for example, with a large-v2 Whisperflt or a stronger GEC system) and no measure of pseudo-label quality is reported. As written, this is a conjecture that cannot be distinguished from other explanations, such as overfitting to pseudo-label noise or differences in the fine-tuning schedule. Please either remove the explanation or support it with an ablation.
  3. [§4.1, title] The paper uses 'scaling' for a single comparison between 77 hours (LNGlbl) and approximately 2500 hours (LNGlbl + pseudo-labelled LNGunl). Without a learning curve or an experiment that varies the amount of pseudo-labelled data, the observed gain cannot be attributed to data scale rather than to the domain or noise characteristics of the pseudo-labelled set, or to the two-stage fine-tuning schedule. A learning curve using subsets of LNGunl would directly support the scaling claim made in the title and abstract.
minor comments (5)
  1. [§5, Table 3] The conclusion states that prompting the large-v2 Whispergec model with fluent transcriptions 'achieves the best SGEC performance, with a WER of 11.08%,' but Table 3 reports the best large-v2 WER as 11.04% for Whispergec+text-flt-SA; please correct the number or clarify which model is referenced.
  2. [Table 3] Table 3 uses the undefined abbreviations 'Whsflt' and 'Whsgec,' which are inconsistent with the 'Whisperflt' and 'Whispergec' notation used in the text; please harmonize the notation.
  3. [§1] There is a stray parenthesis in the first paragraph: 'Incomplete sentences commonly found in spoken language).'
  4. [§3.3, Table 3 caption] The paper does not state which statistical test was used to obtain the p<0.001 markers in Table 3 (for example, paired bootstrap or McNemar); please specify the test and whether it is paired.
  5. [Throughout] Spelling is inconsistent: 'pseudo-labeled' and 'pseudo-labelled' both appear, as do 'LNG unl' and 'LNGunl'; please use one form throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all central claims are evaluated on held-out, human-annotated test sets (Linguaskill test, Speak & Improve dev), and no equation reduces a target to its own inputs.

full rationale

This paper presents no significant circularity. The central claims—that pseudo-labelled data (LNGunl, ~2500 h from a Whisper small.en + Whisperflt + BART GEC pipeline) plus fine-tuning on human-labelled data (LNGlbl) improves end-to-end SGEC, and that prompting Whispergec with Whisperflt fluent transcriptions improves SGEC and feedback—are all evaluated on held-out, human-annotated test sets (Linguaskill test and the open-source Speak & Improve dev corpus). No equation defines a target in terms of its predictor: the pseudo-labelling pipeline produces labels for unlabelled audio, and the trained model is scored against manual references on sentences the teacher never labelled. This is a standard teacher/student setup, and observing the student beat its teacher on held-out data (12.72 vs 13.24 on LNG; 16.84 vs 16.91 on S&I for small.en) is an empirical finding, not a construction-level identity. Although the authors lean on their own prior work [10] for the Whisperflt component and the GEC setup, this is not load-bearing circularity: the paper re-fine-tunes, replicates, and re-evaluates these components on the external test sets within this paper (Table 2; 'the replicated models show results consistent with [10]'), so the citation is backed by independent, reproducible measurement in the same paper. No fitted parameter is renamed as a prediction; no uniqueness theorem is imported; no ansatz is smuggled in via citation. The skeptic's concern—that Table 3's daggar marker (p<0.001) appears only on large-v2 rows while the small.en gains driving the scaling narrative are unmarked—is a statistical robustness/reporting issue, explicitly out of scope for a circularity analysis. Overall, the evaluation protocol is self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims rest on standard domain assumptions about Whisper, pseudo-labelling quality, and evaluation metrics. The only hand-fitted choices are the SpecAugment settings and the teacher model selection, neither of which is theoretically motivated. No new entities are introduced.

free parameters (2)
  • SpecAugment configuration = F=22, T=50, W=5
    Chosen to align WER between the training set and dev set when generating fluent prompts; the paper does not justify these values theoretically and they affect the prompted model's performance.
  • Pseudo-labelling teacher model = Whisper small.en + BART GEC
    The teacher size is a design choice; the paper attributes the large-v2 failure to this choice but does not empirically test an alternative teacher, making it a hand-picked configuration.
assumptions (4)
  • domain assumption Whisper fine-tuning on labelled SGEC data yields a valid end-to-end GEC model
    Adopted from [10]; the paper fine-tunes Whisper on GEC references and assumes the model learns to correct grammar from speech without evaluating the intermediate representation.
  • domain assumption The cascaded pseudo-labelling pipeline produces labels of sufficient quality to train an E2E model
    Section 2.3; the paper assumes that Whisper ASR + Whisperflt + BART GEC can generate correct GEC transcriptions for unlabelled audio, without measuring label accuracy or comparing to human annotations.
  • domain assumption WER is an adequate metric for SGEC quality
    Section 3.3; the authors adopt WER as the primary metric citing prior work, but do not report TER or human evaluation, and WER penalizes all word differences equally, which may not align with correction quality.
  • domain assumption M2 with ERRANT F0.5 measures feedback quality
    Section 3.3; feedback is evaluated via edit-based metrics, which may not capture the pedagogical usefulness or clarity of the feedback text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scaling and Prompting for Improved End-to-End Spoken Grammatical Error Correction." pith.science (2026). https://pith.science/paper/MFSL6STC

@misc{pith2026250521137,
  author       = {Pith},
  title        = {Pith review of: Scaling and Prompting for Improved End-to-End Spoken Grammatical Error Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFSL6STC}},
  note         = {Machine review of arXiv:2505.21137}
}
read the original abstract

Spoken Grammatical Error Correction (SGEC) and Feedback (SGECF) are crucial for second language learners, teachers and test takers. Traditional SGEC systems rely on a cascaded pipeline consisting of an ASR, a module for disfluency detection (DD) and removal and one for GEC. With the rise of end-to-end (E2E) speech foundation models, we investigate their effectiveness in SGEC and feedback generation. This work introduces a pseudo-labelling process to address the challenge of limited labelled data, expanding the training data size from 77 hours to approximately 2500 hours, leading to improved performance. Additionally, we prompt an E2E Whisper-based SGEC model with fluent transcriptions, showing a slight improvement in SGEC performance, with more significant gains in feedback generation. Finally, we assess the impact of increasing model size, revealing that while pseudo-labelled data does not yield performance gain for a larger Whisper model, training with prompts proves beneficial.

Figures

Figures reproduced from arXiv: 2505.21137 by the authors.

Figure 1
Figure 1. Illustration of the E2E SGEC and cascaded systems. speech, producing grammatically correct transcriptions. This modular approach combines speech recognition data, disfluency detection data, and text-based GEC data, all of which are more readily available than annotated spoken GEC data, helping to address the challenge of limited annotated spoken GEC data. Previous work [10] introduced an end-to-end DD model us￾ing W… view at source ↗
Figure 2
Figure 2. The E2E SGEC system prompting with additional ASR transcriptions. To generate grammatically correct transcriptions, the model can benefit from additional contextual information. Our approach builds on this idea by prompting the model with fluent tran￾scriptions, which have disfluencies removed. This extra guid￾ance helps the model better understand the structure of the spo￾ken language. Specifically, we fine-tune a … view at source ↗
Figure 3
Figure 3. shows performance breakdown by grade lev￾els. Whispergec+text-flt outperforms the cascaded system for the LNGlbl test set at levels B1, B2 and C. However, feed￾back remains more challenging for the E2E model, with Whispergec+text-flt only outperforming the cascaded system at level C [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 32 canonical work pages

  1. [1]

    Introduction Spoken Grammatical Error Correction (SGEC) has emerged as a critical task in the field of computer-assisted language learn- ing (CALL), providing learners with essential feedback on their spoken language use. Unlike traditional text-based Grammati- cal Error Correction (GEC), which focuses on written content, SGEC must handle the complexities...

    work page

  2. [2]

    Scaling and Prompting for Improved End-to-End Spoken Grammatical Error Correction

    Method 2.1. Cascaded System A traditional cascaded spoken GEC system consists of three dis- tinct modules: ASR, DD and GEC (Figure 1). First, the ASR module transcribes the speech into text. Then, the DD mod- ule identifies and removes disfluencies, such as interruptions, repetitions and hesitations, from the text transcription. Finally, the GEC module co...

    work page Pith review arXiv 2025

  3. [3]

    Experimental Setup 3.1. Datasets This paper uses Linguaskill [24] labelled and unlabelled train- ing sets to build systems, Linguaskill dev set to select hyper- parameters, and Linguaskill test set and Speak & Improve Cor- pus [25] dev set for system evaluation. Linguaskill: The data used in our study are obtained from can- didate responses to the Speakin...

    work page 2025

  4. [4]

    Scaling Training Data Using Pseudo-labelling Previous work [10] demonstrated promising results with the E2E Whisper gec, achieving a WER of 13.49% on the LNG lbl test set

    Experiments 4.1. Scaling Training Data Using Pseudo-labelling Previous work [10] demonstrated promising results with the E2E Whisper gec, achieving a WER of 13.49% on the LNG lbl test set. However, a performance gap remains compared to the cascaded system, which achieves a lower WER of 12.96% [10]. In this experiment, we replicate their models and investi...

    work page

  5. [5]

    Increasing the model size also improves the capability of an E2E model for both SGEC and feedback tasks

    Conclusions In this work, we explore the use of pseudo-labelled GEC data to scale up the training size of an E2E SGEC model, demon- strating its effectiveness in enhancing the model’s performance. Increasing the model size also improves the capability of an E2E model for both SGEC and feedback tasks. Additionally, incorporating extra information through m...

    work page

  6. [6]

    Spoken Language ‘Grammati- cal Error Correction’,

    Y . Lu, M. J. Gales, and Y . Wang, “Spoken Language ‘Grammati- cal Error Correction’,” inInterspeech 2020, 2020, pp. 3840–3844

    work page 2020

  7. [7]

    Grammatical Error Correction: A Survey of the State of the Art,

    C. Bryant, Z. Yuan, M. R. Qorib, H. Cao, H. T. Ng, and T. Briscoe, “Grammatical Error Correction: A Survey of the State of the Art,” Computational Linguistics, vol. 49, no. 3, pp. 643–701, 09

    work page

  8. [8]

    Grammatical Error Cor- rection for L2 Speech Using Publicly Available Data,

    S. Bann `o, M. Rais, and M. Matassoni, “Grammatical Error Cor- rection for L2 Speech Using Publicly Available Data,” in 9th Workshop on Speech and Language Technology in Education (SLaTE), 2023, pp. 136–140

    work page 2023

Show all 40 references
  1. [9]

    The CoNLL-2014 shared task on grammatical er- ror correction,

    H. T. Ng, S. M. Wu, T. Briscoe, C. Hadiwinoto, R. H. Susanto, and C. Bryant, “The CoNLL-2014 shared task on grammatical er- ror correction,” in Proceedings of the 18th conference on compu- tational natural language learning: shared task, 2014, pp. 1–14

    2014

  2. [10]

    The BEA- 2019 shared task on grammatical error correction,

    C. Bryant, M. Felice, Ø. E. Andersen, and T. Briscoe, “The BEA- 2019 shared task on grammatical error correction,” inProceedings of the fourteenth workshop on innovative use of NLP for building educational applications, 2019, pp. 52–75

    2019

  3. [11]

    Shared task on Multilingual Grammatical Error Cor- rection 2025,

    “Shared task on Multilingual Grammatical Error Cor- rection 2025,” https://www.aclweb.org/portal/content/ shared-task-multilingual-grammatical-error-correction-2025, 2024

    2025

  4. [12]

    Au- tomatic error detection in the Japanese learners’ English spoken data,

    E. Izumi, K. Uchimoto, T. Saiga, T. Supnithi, and H. Isahara, “Au- tomatic error detection in the Japanese learners’ English spoken data,” in The Companion Volume to the Proceedings of 41st An- nual Meeting of the Association for Computational Linguistics , 2003, pp. 145–148

    2003

  5. [13]

    Enhancing grammatical error correction systems with explanations,

    Y . Fei, L. Cui, S. Yang, W. Lam, Z. Lan, and S. Shi, “Enhancing grammatical error correction systems with explanations,” in Pro- ceedings of the 61st Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers). Association for Computational Lingu...

    2023

  6. [14]

    On assessing and developing spoken ’grammatical error correction’ systems,

    Y . Lu, S. Bann `o, and M. Gales, “On assessing and developing spoken ’grammatical error correction’ systems,” in Proceedings of the 17th Workshop on Innovative Use of NLP for Building Ed- ucational Applications (BEA 2022), 2022, pp. 51–60

    2022

  7. [15]

    GEE! grammar error explanation with large language models,

    Y . Song, K. Krishna, R. Bhatt, K. Gimpel, and M. Iyyer, “GEE! grammar error explanation with large language models,” in Findings of the Association for Computational Linguistics: NAACL 2024. Mexico City, Mexico: Association for Computa- tional Linguistics, Jun. 2024, pp. 754–...

    2024

  8. [16]

    Back to grammar: Using gram- matical error correction to automatically assess l2 speaking profi- ciency,

    S. Bann `o and M. Matassoni, “Back to grammar: Using gram- matical error correction to automatically assess l2 speaking profi- ciency,”Speech Communication, vol. 157, p. 103025, 2024

    2024

  9. [17]

    Towards End-to-End Spoken Grammatical Error Correction,

    S. Bann `o, R. Ma, M. Qian, K. M. Knill, and M. J. Gales, “Towards End-to-End Spoken Grammatical Error Correction,” in IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2024, pp. 10 791–10 795

    2024

  10. [18]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

    2023

  11. [19]

    Grammati- cal error correction based on learner comprehension model in oral conversation,

    K. Lee, S. Ryu, P. H. Seo, S. Kim, and G. G. Lee, “Grammati- cal error correction based on learner comprehension model in oral conversation,” in2014 IEEE Spoken Language Technology Work- shop (SLT). IEEE, 2014, pp. 283–287

    2014

  12. [20]

    Cross-Lingual Transfer Learning for Speech Translation,

    R. Ma, M. Qian, Y . Fathullah, S. Tang, M. Gales, and K. Knill, “Cross-Lingual Transfer Learning for Speech Translation,” inAn- nual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL), 2025

    2025

  13. [21]

    Controlled generation with prompt insertion for natural language explanations in grammatical error correction,

    M. Kaneko and N. Okazaki, “Controlled generation with prompt insertion for natural language explanations in grammatical error correction,” in Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 202...

    2024

  14. [22]

    Investigating the Emer- gent Audio Classification Ability of ASR Foundation Models,

    R. Ma, A. Liusie, M. Gales, and K. Knill, “Investigating the Emer- gent Audio Classification Ability of ASR Foundation Models,” in Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technologies (...

    2024

  15. [23]

    Speak & Improve Challenge 2025: Tasks and Baseline Systems,

    M. Qian, K. Knill, S. Banno, S. Tang, P. Karanasou, M. J. Gales, and D. Nicholls, “Speak & Improve Challenge 2025: Tasks and Baseline Systems,” arXiv preprint arXiv:2412.11985, 2024

    2025 arXiv

  16. [24]

    Speak & Improve Corpus 2025: an L2 English Speech Cor- pus for Language Assessment and Feedback,

    K. Knill, D. Nicholls, M. J. Gales, M. Qian, and P. Stroinski, “Speak & Improve Corpus 2025: an L2 English Speech Cor- pus for Language Assessment and Feedback,” arXiv preprint arXiv:2412.11986, 2024

    2025 arXiv

  17. [25]

    Learn and Don’t Forget: Adding a New Language to ASR Foundation Models,

    M. Qian, S. Tang, R. Ma, K. M. Knill, and M. J. Gales, “Learn and Don’t Forget: Adding a New Language to ASR Foundation Models,” in Interspeech 2024, 2024, pp. 2544–2548

    2024

  18. [26]

    Fine- tuning Whisper on Low-Resource Languages for Real-World Ap- plications,

    V . Timmel, C. Paonessa, R. Kakooee, M. V ogel et al. , “Fine- tuning Whisper on Low-Resource Languages for Real-World Ap- plications,” arXiv preprint arXiv:2412.15726, 2024

    2024 arXiv

  19. [27]

    Annotation of L2 English Speech for Developing and Evalu- ating End-to-End Spoken Grammatical Error Correction,

    K. M. Knill, D. Nicholls, M. Gales, P. Stroinski, and A. Watkin- son, “Annotation of L2 English Speech for Developing and Evalu- ating End-to-End Spoken Grammatical Error Correction,” inProc. 9th Workshop on Speech and Language Technology in Education (SLaTE), 2023, pp. 146–150

    2023

  20. [28]

    Task Arithmetic for Language Ex- pansion in Speech Translation,

    Y .-F. Cheng, H. Futami, Y . Kashiwagi, E. Tsunoo, W. S. Teo, S. Arora, and S. Watanabe, “Task Arithmetic for Language Ex- pansion in Speech Translation,”arXiv preprint arXiv:2409.11274, 2024

    2024 arXiv

  21. [29]

    BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Gener- ation, Translation, and Comprehension,

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer, “BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Gener- ation, Translation, and Comprehension,” in Proceedings of the 58th Annual Meeting of the Associa...

    2020

  22. [30]

    ( facebook/bart-base). The model is trained on the EF- CAMDAT and BEA-2019 data for 19 epochs with a maximum sequence length of 256, a batch size of 16, a gradient accu- mulation step of 4, and a learning rate of 2e-6. It is then further fine-tuned on the Linguaskill data for ...

    2019

  23. [31]

    Improving Domain Generalization in Speech Emotion Recognition with Whisper,

    E. Goron, L. Asai, E. Rut, and M. Dinov, “Improving Domain Generalization in Speech Emotion Recognition with Whisper,” in 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 631–11 635

    2024

  24. [32]

    Ludlow, Official Quick Guide to Linguaskill

    K. Ludlow, Official Quick Guide to Linguaskill . Cambridge: Cambridge University Press & Assessment, 2020

    2020

  25. [33]

    The Speak & Improve Corpus 2025: an L2 English Speech Corpus for Language Assessment and Feedback,

    K. Knill, D. Nicholls, M. J. Gales, M. Qian, and P. Stroinski, “The Speak & Improve Corpus 2025: an L2 English Speech Corpus for Language Assessment and Feedback,” 2025. [Online]. Available: https://doi.org/10.17863/CAM.114333

    2025 doi

  26. [34]

    Cambridge: Cambridge University Press, 2001

    Council of Europe, Common European Framework of Reference for Languages: Learning, Teaching, Assessment . Cambridge: Cambridge University Press, 2001. [Online]. Available: https: //rm.coe.int/1680459f97

    2001

  27. [35]

    Speak & Improve: L2 English Speaking Practice Tool,

    D. Nicholls, K. M. Knill, M. J. F. Gales, A. Ragni, and P. Rick- etts, “Speak & Improve: L2 English Speaking Practice Tool,” in Interspeech 2023, 2023, pp. 3669–3670

    2023

  28. [37]

    Transformers: State-of-the-art natural language processing,

    T. Wolf et al., “Transformers: State-of-the-art natural language processing,” inProc. Conf. on Empirical Methods in Natural Lan- guage Processing: System Demonstrations, 2020, pp. 38–45

    2020

  29. [38]

    Better evaluation for grammatical error correction,

    D. Dahlmeier and H. T. Ng, “Better evaluation for grammatical error correction,” in Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2012, pp. 568–572

    2012

  30. [39]

    Automatic annotation and evaluation of error types for grammatical error correction,

    C. Bryant, M. Felice, and T. Briscoe, “Automatic annotation and evaluation of error types for grammatical error correction,” inPro- ceedings of the 55th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers). Association for Computational Lin...

    2017

  31. [40]

    SpecAugment: A Simple Data Augmenta- tion Method for Automatic Speech Recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “SpecAugment: A Simple Data Augmenta- tion Method for Automatic Speech Recognition,” in Interspeech 2019, 2019, pp. 2613–2617

    2019

  32. [2023]

    Available: https://doi.org/10.1162/coli a 00478

    [Online]. Available: https://doi.org/10.1162/coli a 00478

    doi

Pith tools

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