Pith. sign in

REVIEW 3 major objections 5 minor 54 references

End-to-End Spoken Grammatical Error Correction

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Spoken grammatical error correction can be done end-to-end with a single Whisper-based model that beats cascaded ASR-plus-text systems in both transcription accuracy and feedback precision.

desk verdict Serious E2E SGEC paper with real WER gains; feedback claims are plausible but need fixed thresholds and an error-category audit of reference alignment. read the letter →

arxiv 2506.18532 v1 pith:QHLN4MLF submitted 2025-06-23 cs.CL cs.LGcs.SDeess.AS

Mengjie Qian , Rao Ma , Stefano Bannò , Mark J.F. Gales , Kate M. Knill This is my paper
classification cs.CLcs.LGcs.SDeess.AS
keywords spokengrammaticalerrorcorrectionend-to-endspeechprocessingWhisperpseudo-labelingprompt-basedtrainingreferencealignmentconfidenceestimationlearnerfeedback
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

This paper argues that spoken grammatical error correction (SGEC) need not be split into a cascaded pipeline of speech recognition, disfluency removal, and text correction. It claims that a single end-to-end model built on the Whisper speech foundation model can match or beat that pipeline for both transcription and learner feedback, provided it is trained and evaluated with four supporting techniques: large-scale pseudo-labeled data, prompt conditioning on fluent transcriptions, references aligned to strip ASR-induced edits, and confidence-based filtering of low-confidence corrections. The best configuration reaches an F0.5 of 44.73 on the Linguaskill test set and 46.94 on the Speak & Improve dev set, against 40.51 and 43.10 for the partial-cascaded baseline. If correct, this simplifies SGEC systems and makes automated grammar feedback for language learners more practical.

What carries the argument

The central object is the Whisper speech foundation model fine-tuned into two variants: Whisper flt, which removes disfluencies, and Whisper gec, which produces grammatically corrected text directly from audio. The argument is carried by three mechanisms: (1) prompt-based training, where the fluent transcription from Whisper flt is fed as a prompt to Whisper gec so the model focuses on grammar rather than disfluency; (2) a reference modification algorithm that aligns the GEC reference (gec-ref) against the FLT hypothesis (flt-hyp) and the FLT reference (flt-ref), and rewrites the training target to remove edits caused by ASR errors, with a mild and a strong variant; and (3) edit confidence estimation, which maps token-level confidence from both FLT and GEC models onto each MaxMatch edit (using averages or minima) and filters out edits below a threshold, preferring higher precision in learner-facing feedback.

What would settle it

Build a test set of learner utterances in which the Whisper flt hypothesis itself contains a genuine grammar error that is also a plausible ASR confusion, for instance 'yesterday I goed' where the learner said 'yesterday I go' and the ASR transcribed 'goed'. If a model trained with Strong aligned references fails to correct 'goed' to 'went' while a model trained on original references does correct it, the assumption that flt-hyp is a safe source for feedback is falsified.

Watch

Extended reading notes

Core claim

The central claim is that an end-to-end Whisper-based SGEC system, trained to map learner audio directly to grammatically corrected text, can outperform cascaded and partial-cascaded baselines built from the same foundation model, in both transcription accuracy (WER) and feedback quality (F0.5). This is achieved by scaling training data with an automated pseudo-labeling pipeline that grows the corpus from 77 to over 2,500 hours; conditioning the model on a fluent transcription prompt; modifying the GEC training references so that edits caused by ASR noise are removed while genuine grammatical corrections are kept; and filtering low-confidence edits during evaluation using combined FLT and GEC token confidence. With Whisper large-v2, the end-to-end model reduces WER on the Linguaskill test set from 12.11 (cascaded) or 11.81 (partial-cascaded) to 11.10 without prompting, and with prompting, aligned references, and edit-average filtering it reaches F0.5 of 44.73 on LNG and 46.94 on S&I Dev, exceeding the partial-cascaded baseline of 40.51 and 43.10.

Load-bearing premise

The method assumes that every difference between the fluent transcription produced by the Whisper flt model and the human fluent reference is an ASR error, so deleting those differences from the training target cannot suppress a real grammar correction.

Editorial extensions

If this is right

  • End-to-end SGEC with large-v2 Whisper can replace the three-stage pipeline of ASR, disfluency detection, and GEC with a single model, removing error propagation between modules.
  • Pseudo-labeled data is most valuable for small-capacity models; large models trained on 77 hours of manual labels already match or slightly beat the full pseudo-labeling pipeline.
  • Prompting with fluent transcriptions improves feedback precision by over three F0.5 points without hurting WER, making it a low-cost addition to E2E training.
  • Strong reference alignment plus edit-average confidence filtering yields the best feedback F0.5, at a small WER cost, so systems can trade transcription accuracy for feedback precision depending on the application.

Reading between the lines

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

  • We infer that the reference alignment and confidence filtering techniques are not specific to grammar; the same pattern of aligning training targets to a noisy intermediate hypothesis and pruning low-confidence edits could apply to other spoken language understanding tasks, such as disfluency removal or named entity correction in learner speech.
  • We infer that because edit confidence filtering is applied at evaluation time, it could be repurposed as a deployment-time gate that withholds feedback from learners when the model is uncertain, which would directly affect learner trust; this is a testable extension the paper does not run.
  • We infer that the assumption that flt-hyp is a reliable source for feedback could be checked per learner group: if Whisper flt makes more recognition errors for certain L1 backgrounds, feedback precision under Strong alignment would drop for those groups; the paper does not report this breakdown.
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. This manuscript examines end-to-end spoken grammatical error correction (SGEC) built on Whisper, comparing cascaded, partial-cascaded, and E2E systems. It proposes four techniques: a pseudo-labeling pipeline that scales training data from 77 to over 2500 hours, prompt-based conditioning on fluent transcriptions, a reference alignment algorithm that rewrites GEC training targets to match the fluent hypothesis, and edit confidence estimation/filtering. Experiments on in-house Linguaskill and public Speak & Improve (S&I) corpora report that the best E2E system (Whisper gec+pt with strong reference alignment and edit-avg filtering) achieves F0.5 of 44.73 on LNG test and 46.94 on S&I Dev, outperforming the partial-cascaded baseline (40.51 and 43.10), and reduces WER on LNG from 11.81 (partial-cascaded) to 11.10.

Significance. If the findings hold, the paper is a solid contribution: it demonstrates that E2E SGEC can surpass modular pipelines in both transcription and feedback quality, provides a large pseudo-labeled resource, and introduces a reusable confidence-filtering framework. The WER results are internally consistent and show clear gains for the large-v2 E2E model. The public S&I Eval results are a useful benchmark. However, the feedback-quality claims are weakened by two methodological issues: the edit confidence threshold is optimized on the test set, and the reference alignment couples the training target to the same ASR hypothesis used at test time without an error-category audit. These issues affect the headline F0.5 improvements and must be addressed before the feedback results can be fully trusted.

major comments (3)
  1. [Section V-E, Figure 4] The edit confidence threshold (0.67 for LNG) is chosen by maximizing F0.5 on the LNG test set, and the resulting F0.5 is then reported for that same set. This is a direct form of evaluation-based fitting that inflates the reported precision and F0.5. The authors should select thresholds on a held-out development set (e.g., LNG Dev or S&I Dev) and apply them to the test set, or report threshold-independent summaries (e.g., area under the P/R curve). The same concern applies to the S&I Dev results if thresholds were tuned there.
  2. [Section III-C, Algorithm 1] The reference modification assumes that any discrepancy between the fluent hypothesis (flt-hyp) and the manual fluent reference (flt-ref) is an ASR error, and it rewrites the GEC training reference to agree with flt-hyp whenever edit types and boundaries match. Because the same flt-hyp is used at test time to derive the predicted edit set, the model is trained to suppress exactly those corrections whose absence would be scored as high-precision edits under the modified target. Table V shows the expected precision-up/recall-down pattern (Strong ref P 52.20/R 26.45 vs Orig P 43.92/R 32.63 on LNG), but the paper does not verify that the dropped edits are actually ASR artifacts. A concrete error-category audit (e.g., ERRANT types of removed edits, with examples) or a human evaluation is needed to rule out that genuine learner errors that coincide with ASR confusions are being suppressed.
  3. [Table V] The F0.5 differences between configurations are reported without significance tests or confidence intervals. Given that the threshold selection issue already affects these numbers, it is important to know whether the gains from reference alignment and confidence filtering are statistically reliable. Please provide bootstrap confidence intervals or significance tests for the feedback metrics.
minor comments (5)
  1. [Section V-A] The text reports a '6.9% relative WER reduction' on LNG for large-v2 E2E versus partial-cascaded, but the numbers in Table II (11.10 vs 11.81) give 6.0%; please correct the arithmetic or clarify the baseline.
  2. [Throughout] Model names are frequently written without spaces (e.g., 'Whispergec', 'Whisperflt', 'Whsgec'), which makes the text harder to read; please normalize the notation.
  3. [Section V-F, Table VI] In the Strong + edit avg row, the WER column appears to be left blank; please either report the value or explain that filtering does not change WER.
  4. [Section III-D] The effect of the fixed confidence value (1.0) for missing/unnecessary edits is described qualitatively; please include the actual values swept in the grid or in an appendix.
  5. [Section IV-A] The description of the Switchboard Reannotated Dataset is longer than needed for the paper's contribution; consider compressing it or moving the details to a footnote.

Circularity Check

1 steps flagged · score 6.0 of 10

Confidence-filtering threshold is selected on the LNG test set, so the headline F0.5 of 44.73 is a fitted maximum rather than an independent prediction; the remaining contributions are not circular.

  1. fitted input called prediction [Section V-E (Feedback with Confidence), Figure 4 and Table V; cf. Section IV-A evaluation protocol]
    "Figure 4 plots Precision, Recall, and F 0.5 against different confidence thresholds on LNG. ... F 0.5, which prioritizes Precision, peaks at a threshold of 0.67. ... The best-performing configuration on LNG is Whisper gec+pt with the Strong GEC reference and edit avg, achieving a F 0.5 of 44.73. ... The Linguaskill Dev set is used for hyperparameter tuning, while the Linguaskill Test set and the S&I Dev set serve as primary evaluation sets."

    The threshold 0.67 is explicitly selected by maximizing F0.5 on the LNG test set, and the same LNG test set is then used to report the headline F0.5 of 44.73 with edit-avg filtering. The reported value is therefore the argmax of the F0.5-versus-threshold curve on the evaluation set, not an out-of-sample prediction. Any rejection threshold tuned on the test labels will produce a non-negative F0.5 gain by construction, so the paper's statement that confidence filtering 'improves F 0.5' on LNG reduces to this test-set fit. The S&I Dev result with the same threshold is less affected, but the central LNG number is not an independent evaluation of the filtering method.

full rationale

Most of the paper's derivation chain is self-contained. The pseudo-labeling pipeline, prompt conditioning, and reference alignment are not circular: each introduces external data or manual references, and the evaluation reference edits are computed from manual flt-ref and gec-ref, not from the model's own flt-hyp. The self-citations [11] and [12] are prior building blocks, not load-bearing uniqueness claims. The one reduction-by-construction is in Section V-E: the confidence-filtering threshold is selected by maximizing F0.5 on the LNG test set, and the same test set is then used to report the headline F0.5 with edit-avg filtering. This makes the LNG filtering gain a fitted maximum rather than an independent prediction. The reference-alignment coupling flagged in the skeptic analysis is a real validity concern but not circularity, because the training targets and test references are not definitionally identical. The score reflects the partial circularity of the confidence-filtering result, while acknowledging that the other contributions retain independent empirical content.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The only invented entities are algorithmic components (reference alignment, confidence filtering), not physical entities. The free parameters are the confidence threshold (tuned on test), fixed confidence defaults, and the mild/strong mode choice. The load-bearing axioms are the evaluation paradigm and the reliability of the text GEC model used to create pseudo-labels.

free parameters (3)
  • edit confidence threshold (LNG test) = 0.67
    Selected by maximizing F0.5 on the LNG test set, as stated in Section V-E and Figure 4. This is a free parameter fit to the evaluation data, inflating the reported F0.5.
  • fixed confidence value for missing/unnecessary edits = 1.0
    Assigned to FLT side for missing edits and GEC side for unnecessary edits (Section III-D). The authors say they tried other values and found peak F0.5 unaffected, but it remains a hand-set parameter.
  • mod-mild vs mod-strong mode = chosen per experiment
    The choice of reference alignment mode is a design parameter, not fitted to data, but it changes training targets and thus results; treated as a free choice in the paper.
assumptions (3)
  • domain assumption M2/ERRANT edit-based evaluation is a valid measure of feedback quality
    The paper adopts MaxMatch (M2) and ERRANT to compute Precision/Recall/F0.5 for feedback edits (Section IV-C). This is the standard in written GEC but may not fully capture the pedagogical value of spoken feedback.
  • domain assumption The text-based GEC model (BART trained on EFCAMDAT and BEA-2019, fine-tuned on Linguaskill) produces correct GEC annotations for fluent transcriptions in pseudo-labeling
    Section III-A Step 4 uses this model to generate all pseudo-labels. No evaluation of the pseudo-label quality is provided; systematic errors in this model would directly train the E2E model to mimic them.
  • standard math Alright, let me compute the approximate average of 26.61, 27.71, and 26.29
    No, this is from the table; I am not adding this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End-to-End Spoken Grammatical Error Correction." pith.science (2026). https://pith.science/paper/QHLN4MLF

@misc{pith2026250618532,
  author       = {Pith},
  title        = {Pith review of: End-to-End Spoken Grammatical Error Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QHLN4MLF}},
  note         = {Machine review of arXiv:2506.18532}
}
read the original abstract

Grammatical Error Correction (GEC) and feedback play a vital role in supporting second language (L2) learners, educators, and examiners. While written GEC is well-established, spoken GEC (SGEC), aiming to provide feedback based on learners' speech, poses additional challenges due to disfluencies, transcription errors, and the lack of structured input. SGEC systems typically follow a cascaded pipeline consisting of Automatic Speech Recognition (ASR), disfluency detection, and GEC, making them vulnerable to error propagation across modules. This work examines an End-to-End (E2E) framework for SGEC and feedback generation, highlighting challenges and possible solutions when developing these systems. Cascaded, partial-cascaded and E2E architectures are compared, all built on the Whisper foundation model. A challenge for E2E systems is the scarcity of GEC labeled spoken data. To address this, an automatic pseudo-labeling framework is examined, increasing the training data from 77 to over 2500 hours. To improve the accuracy of the SGEC system, additional contextual information, exploiting the ASR output, is investigated. Candidate feedback of their mistakes is an essential step to improving performance. In E2E systems the SGEC output must be compared with an estimate of the fluent transcription to obtain the feedback. To improve the precision of this feedback, a novel reference alignment process is proposed that aims to remove hypothesised edits that results from fluent transcription errors. Finally, these approaches are combined with an edit confidence estimation approach, to exclude low-confidence edits. Experiments on the in-house Linguaskill (LNG) corpora and the publicly available Speak & Improve (S&I) corpus show that the proposed approaches significantly boost E2E SGEC performance.

Figures

Figures reproduced from arXiv: 2506.18532 by the authors.

Figure 1
Figure 1. Illustration of the cascaded, partial-cascaded , and E2E SGEC systems. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The E2E SGEC system prompting with additional ASR transcriptions. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Precision, Recall, and F0.5 scores at different edit confidence thresholds for the Whispergec+pt model trained with Strong GEC references. TABLE VI SGEC AND FEEDBACK PERFORMANCE ON S&I EVAL SET. † DENOTES STATISTICALLY SIGNIFICANT IMPROVEMENT OVER THE PARTIAL-CASCADED SYSTEM (p < 0.001). Model Train S&I Name Ref WER P R F0.5 Whsflt + GEC 14.37 48.97 27.25 42.24 Whsgec+pt Mild 13.44† 47.76 29.96 42.69 Strong 13.72 53… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 31 canonical work pages

  1. [1]

    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 2023. [Online]. Available: https://doi.org/10.1162/coli a 00478

    doi:10.1162/coli 2023

  2. [2]

    Automatic error detection in the Japanese learners’ English spoken data,

    E. Izumi, K. Uchimoto, T. Saiga, T. Supnithi, and H. Isahara, “Automatic error detection in the Japanese learners’ English spoken data,” inThe Companion Volume to the Proceedings of 41st Annual Meeting of the Association for Computational Linguistics, 2003, pp. 145–148. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    work page 2003

  3. [3]

    The CoNLL-2014 Shared Task on Grammatical Error Correction,

    H. T. Ng, S. M. Wu, T. Briscoe, C. Hadiwinoto, R. H. Susanto, and C. Bryant, “The CoNLL-2014 Shared Task on Grammatical Error Correction,” inProceedings of the 18th conference on computational natural language learning: shared task, 2014, pp. 1–14. [Online]. Available: https://aclanthology.org/W14-1701/

    work page 2014

  4. [4]

    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. [Online]. Available: https://aclanthology.org/W19-4406/

    work page 2019

  5. [5]

    Shared task on Multilingual Grammatical Error Correction 2025,

    E. V olodina, A. Masciolini, A. Caines, O. D. Clercq, J. Kruijsbergen, M. Kurfali, R. M. S´anchez, and R. ¨Ostling, “Shared task on Multilingual Grammatical Error Correction 2025,” https://www.aclweb.org/portal/ content/shared-task-multilingual-grammatical-error-correction-2025, 2025

    work page 2025

  6. [6]

    Spoken Language ‘Grammatical Error Correction’,

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

    work page 2020

  7. [7]

    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,” inProceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Appli- cations (BEA 2022), 2022, pp. 51–60

    work page 2022

  8. [8]

    Grammatical Error Correction for L2 Speech Using Publicly Available Data,

    S. Bann `o, M. Rais, and M. Matassoni, “Grammatical Error Correction for L2 Speech Using Publicly Available Data,” in9th Workshop on Speech and Language Technology in Education (SLaTE), 2023, pp. 136– 140

    work page 2023

Show all 54 references
  1. [9]

    Back to grammar: Using grammatical error correction to automatically assess L2 speaking proficiency,

    S. Bann `o and M. Matassoni, “Back to grammar: Using grammatical error correction to automatically assess L2 speaking proficiency,”Speech Communication, vol. 157, p. 103025, 2024

    2024

  2. [10]

    Robust speech recognition via large-scale weak supervi- sion,

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

    2023

  3. [11]

    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,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 791–10 795

    2024

  4. [12]

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

    M. Qian, R. Ma, S. Bann `o, K. M. Knill, and M. J. Gales, “Scaling and Prompting for Improved End-to-End Spoken Grammatical Error Correction,” inInterspeech 2025, 2025

    2025

  5. [13]

    Speak & Improve Challenge 2025: Tasks and Baseline Systems,

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

    2025 arXiv

  6. [14]

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

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

    2025 arXiv

  7. [15]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Rayet al., “Training language models to follow instructions with human feedback,”Advances in neural information processing systems, vol. 35, pp. 27 730–27 744, 2022

    2022

  8. [16]

    GPT-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  9. [17]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient foundation language models,” 2023. [Online]. Available: https: //arxiv.org/abs/2...

    2023 arXiv

  10. [18]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stoneet al., “Llama 2: Open foundation and fine-tuned chat models,” 2023. [Online]. Available: https: //arxiv.org/abs/2307.09288

    2023 arXiv

  11. [19]

    The llama 3 herd of models,

    A. Grattafiori, A. Dubey, A. Jauhriet al., “The llama 3 herd of models,” 2024. [Online]. Available: https://arxiv.org/abs/2407.21783

    2024 arXiv

  12. [20]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,”Journal of machine learning research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  13. [21]

    Gemma: Open models based on gemini research and technology,

    G. Team, T. Mesnard, C. Hardin, R. Dadashi, S. Bhupatiraju, S. Pathak, L. Sifre, M. Rivi `ere, M. S. Kale, J. Loveet al., “Gemma: Open models based on gemini research and technology,”arXiv preprint arXiv:2403.08295, 2024

    2024 arXiv

  14. [22]

    Gemma 2: Improving open language models at a practical size,

    G. Team, M. Riviere, S. Pathak, P. G. Sessa, C. Hardin, S. Bhupatiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ram ´eet al., “Gemma 2: Improving open language models at a practical size,”arXiv preprint arXiv:2408.00118, 2024

    2024 arXiv

  15. [23]

    Gemma 3 technical report,

    G. Team, A. Kamath, J. Ferret, S. Pathak, N. Vieillard, R. Merhej, S. Perrin, T. Matejovicova, A. Ram ´e, M. Rivi `ereet al., “Gemma 3 technical report,”arXiv preprint arXiv:2503.19786, 2025

    2025 arXiv

  16. [24]

    Mistral 7b,

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. de las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M.-A. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, and W. E. Sayed, “Mistral 7b,” 2023. [Online]. Available: ht...

    2023 arXiv

  17. [25]

    Qwen technical report,

    J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huanget al., “Qwen technical report,”arXiv preprint arXiv:2309.16609, 2023

    2023 arXiv

  18. [26]

    Qwen2.5-1m technical report,

    A. Yang, B. Yu, C. Li, D. Liu, F. Huang, H. Huang, J. Jiang, J. Tu, J. Zhang, J. Zhou, J. Lin, K. Dang, K. Yang, L. Yu, M. Li, M. Sun, Q. Zhu, R. Men, T. He, W. Xu, W. Yin, W. Yu, X. Qiu, X. Ren, X. Yang, Y . Li, Z. Xu, and Z. Zhang, “Qwen2.5-1m technical report,”

  19. [27]

    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,” inProceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguist...

    2023

  20. [28]

    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,” inFindings of the Association for Computational Linguistics: NAACL 2024. Mexico City, Mexico: Association for Computational Linguistics, Jun. 2024, pp. 754–781...

    2024

  21. [29]

    Audiopalm: A large language model that can speak and listen,

    P. K. Rubenstein, C. Asawaroengchai, D. D. Nguyenet al., “Audiopalm: A large language model that can speak and listen,” 2023. [Online]. Available: https://arxiv.org/abs/2306.12925

    2023 arXiv

  22. [30]

    Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,

    Y . Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models,”arXiv preprint arXiv:2311.07919, 2023

    2023 arXiv

  23. [31]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Linet al., “Qwen2-audio technical report,”arXiv preprint arXiv:2407.10759, 2024

    2024 arXiv

  24. [32]

    SALMONN: Towards generic hearing abilities for large language models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. MA, and C. Zhang, “SALMONN: Towards generic hearing abilities for large language models,” inThe Twelfth International Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum?id=14rn7HpKVk

    2024

  25. [33]

    SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities,

    D. Zhang, S. Li, X. Zhang, J. Zhan, P. Wang, Y . Zhou, and X. Qiu, “SpeechGPT: Empowering Large Language Models with Intrinsic Cross-Modal Conversational Abilities,” inFindings of the Association for Computational Linguistics: EMNLP 2023. Singapore: Association for Computation...

    2023

  26. [34]

    Grammatical error correction based on learner comprehension model in oral conversation,

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

    2014

  27. [35]

    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. F. Gales, “Learn and Don’t Forget: Adding a New Language to ASR Foundation Models,” inInterspeech 2024, 2024, pp. 2544–2548. [Online]. Available: https://doi.org/10.21437/Interspeech.2024-1045

    2024 doi

  28. [36]

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

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

    2024 arXiv

  29. [37]

    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,” inAnnual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL), 2025, pp. 33–43. [Online]. Available: ...

    2025

  30. [38]

    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: Human Language Technologies (Vo...

    2024

  31. [39]

    Improving Domain Gener- alization in Speech Emotion Recognition with Whisper,

    E. Goron, L. Asai, E. Rut, and M. Dinov, “Improving Domain Gener- alization in Speech Emotion Recognition with Whisper,” in2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 631–11 635

    2024

  32. [40]

    Ludlow,Official Quick Guide to Linguaskill

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

    2020

  33. [41]

    Better evaluation for grammatical error correction,

    D. Dahlmeier and H. T. Ng, “Better evaluation for grammatical error correction,” inProceedings of the 2012 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, 2012, pp. 568–572. JOURNAL OF LATEX CLASS FILES,...

    2012

  34. [42]

    Automatic linguis- tic annotation of large scale L2 databases: The EF-Cambridge Open Language Database (EFCAMDAT),

    J. Geertzen, T. Alexopoulou, and A. Korhonen, “Automatic linguis- tic annotation of large scale L2 databases: The EF-Cambridge Open Language Database (EFCAMDAT),” inProc. 31st Second Language Research Forum, 2013, pp. 240–254

    2013

  35. [43]

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

    K. Knill, D. Nicholls, M. J. F. 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

  36. [44]

    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

  37. [45]

    Speak & Improve: L2 English Speaking Practice Tool,

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

    2023

  38. [46]

    SWITCHBOARD: Telephone speech corpus for research and development,

    J. J. Godfrey, E. C. Holliman, and J. McDaniel, “SWITCHBOARD: Telephone speech corpus for research and development,” inAcoustics, speech, and signal processing, ieee international conference on, vol. 1. IEEE Computer Society, 1992, pp. 517–520

    1992

  39. [47]

    Switchboard-1 Release 2 LDC97S62,

    J. Godfrey and E. Holliman, “Switchboard-1 Release 2 LDC97S62,” Linguistic Data Consortium, vol. 34, 1993

    1993

  40. [48]

    Resegmentation of SWITCHBOARD,

    N. Deshmukh, A. Ganapathiraju, A. Gleeson, J. Hamaker, and J. Picone, “Resegmentation of SWITCHBOARD,” inICSLP. Syndey, 1998

    1998

  41. [49]

    Treebank-3 Tech Rep,

    M. P. Marcus, B. Santorini, M. A. Marcinkiewicz, and A. Taylor, “Treebank-3 Tech Rep,”Linguistic Data Consortium, 1999

    1999

  42. [50]

    Edit Detection and Parsing for Transcribed Speech,

    E. Charniak and M. Johnson, “Edit Detection and Parsing for Transcribed Speech,” inSecond Meeting of the North American Chapter of the Association for Computational Linguistics, 2001. [Online]. Available: https://aclanthology.org/N01-1016

    2001

  43. [51]

    BART: Denoising Sequence-to- Sequence Pre-training for Natural Language Generation, 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 Generation, Translation, and Comprehension,” inProceedings of the 58th Annual Meeting of the Associati...

    2020

  44. [52]

    Transformers: State-of-the-art natural language pro- cessing,

    T. Wolfet al., “Transformers: State-of-the-art natural language pro- cessing,” inProc. Conf. on Empirical Methods in Natural Language Processing: System Demonstrations, 2020, pp. 38–45

    2020

  45. [53]

    Automatic annotation and eval- uation of error types for grammatical error correction,

    C. Bryant, M. Felice, and T. Briscoe, “Automatic annotation and eval- uation of error types for grammatical error correction,” inProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Lingu...

    2017

  46. [2025]

    Available: https://arxiv.org/abs/2501.15383

    [Online]. Available: https://arxiv.org/abs/2501.15383

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.