CoChat

Error-preserving Automatic Speech Recognition of Young English Learners’ Language

Open access
Michot, Janick, Hürlimann, Manuela, Deriu, Jan, Sauer, Luzia, Mlynchyk, Katsiaryna, Cieliebak, Mark · 2024 · Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

The DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index; none of the 15 sources its claims cite is checked for support yet.

Abstract

Janick Michot, Manuela Hürlimann, Jan Deriu, Luzia Sauer, Katsiaryna Mlynchyk, Mark Cieliebak. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Overview

CoChat Summary

TL;DR

The paper builds an error-preserving ASR system for young English learners by collecting 85 hours of spontaneous speech from Swiss grades 4–6 and fine-tuning a Wav2Vec-XLSR-300M model (ChaLL-300M). It introduces WEPR to measure error preservation and shows fine-tuning on learner data improves both WER and error preservation over pre-trained ASR systems.

Key findings

  • The collected dataset contains 85 hours, 45,004 utterances, 327 distinct speakers, 485,770 tokens, and 14,396 error-annotated tokens.
  • Among pre-trained systems, Whisper-Large achieved the best overall WER and chrF, while XLSR-1B models fine-tuned on CommonVoice 6.1 achieved the best CER and WEPR; CTC models without a language model better preserved speaker errors.
  • Fine-tuned ChaLL-300M achieved the best WEPR, improving over the comparable XLSR-53 adult-trained model by 8 points in WER and 12 points in WEPR.
  • On the 20 most frequent speaker errors, ChaLL-300M preserved 82.7% versus Whisper-Large’s 67.3% and XLSR-1B’s 72.3%; it mistakenly corrected 15% versus 26% and 22.9%, respectively.

Methods

  • Collected 85 hours of English audio from learners in Switzerland grades 4–6 using playful speaking activities, then transcribed verbatim with error annotations (`@!` for errors, `@g` for German words).
  • Split data into five folds of about 16 hours each, with each class/speaker in only one fold, and fine-tuned Wav2Vec-XLSR-300M via 5-fold cross-validation for 4000 steps with learning rate 3e-5, total batch size 1260, and 8-bit AdamW.
  • Evaluated with WEPR plus WER, CER, and chrF, comparing seven pre-trained ASR systems (four CTC-based, three encoder-decoder) against the fine-tuned model.

Limitations

  • Stated: the dataset is limited to Swiss school children in grades 4–6 learning English in German-speaking Switzerland, so transferability must be confirmed with other populations.
  • Stated: transcription and error annotation were outsourced, risking erroneous data; the model was limited to 300M parameters due to compute cost; no hyperparameter tuning or joint error prediction was performed; and data cannot be freely shared because it contains minors’ speech.
  • Evident from the text: WEPR is computed only over manually annotated error tokens, so it does not measure preservation of unannotated errors, and the corpus was not annotated for CEFR levels.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.18653/v1/2024.acl-long.348

Cites

31 of 31 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
CC-BY · From the PDF
Figure 2 Distribution of utterance lengths.
Figure 2 Distribution of utterance lengths.CC-BY · From the PDF
Figure 3 Duration and grade distribution of the data folds.
Figure 3 Duration and grade distribution of the data folds.CC-BY · From the PDF

Verification

Claims the paper makes

0 Valid · 6 Not valid · 11 Not checked · 52 Own results and interpretations

21 more claims

Create a free account to see all 69 claims and to verify the 9 references still open.

See all claims — free account
  • Probably not a substantial claim. 26
  • It may state two separate claims as one. 27

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 31 references

Cited by

See all 5

Metadata

Authors
Michot, J., Hürlimann, M., Deriu, J., Sauer, L., Mlynchyk, K., & Cieliebak, M.
Affiliations
  • ZHAW Zurich University of Applied Sciences (CH)
  • Zurich University of Teacher Education (CH)
Venue
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Published
2024
Pages
6444-6454
ISSN
0736-587X (ISSN-L)
OpenAlex
W4402671667
Cited by
5
References
28
Citations by year
2026: 4 · 2025: 1
FWCI
0.41
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
Proceedings of the conference - Association for Computational Linguistics. Meeting (publishedVersion, cc-by)
Funders
Innosuisse - Schweizerische Agentur für Innovationsförderung
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Phonetics and Phonology Research
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
Keywords
speech recognition, young learners, English as a second language, speech recognition errors, child language acquisition, low-resource speech recognition

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-20, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-20 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Error-preserving Automatic Speech Recognition of Young English Learners’ Language. https://app.cochat.ai/papers/10.18653/v1/2024.acl-long.348
The paper
Michot, J., Hürlimann, M., Deriu, J., Sauer, L., Mlynchyk, K., & Cieliebak, M. (2024). Error-preserving Automatic Speech Recognition of Young English Learners’ Language. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). https://doi.org/10.18653/v1/2024.acl-long.348

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.