Pith. sign in

REVIEW 4 major objections 4 minor 34 references

Improving Child Speech Recognition and Reading Mistake Detection by Using Prompts

T0 review · 4 major / 4 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 Prompting Whisper with error-laced text and an LLM with alignment hypotheses cuts Dutch child read-speech word-error rate from 9.4% to 5.1% and lifts reading-mistake detection F1 from 0.39 to 0.73.

desk verdict A genuinely new prompting combination that likely helps, but the headline numbers need a check on the hallucination fallback and missing statistics. read the letter →

arxiv 2506.11079 v1 pith:4VDMWME6 submitted 2025-06-04 eess.AS cs.AIcs.CLcs.SD

Lingyun Gao , Cristian Tejedor-Garcia , Catia Cucchiarini , Helmer Strik This is my paper
classification eess.AScs.AIcs.CLcs.SD
keywords childspeechrecognitionreadingmistakedetectionmiscueWhisperpromptingLLMalignmenthypothesesDutchreadpromptengineering
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 is trying to establish that off-the-shelf speech and language models, steered entirely by text prompts, can do a substantial part of what a teacher does when listening to a child read aloud: transcribe the words and spot the reading mistakes. The central claim is that prompting Whisper with an irrelevant passage full of deliberately generated mistakes, and then prompting a large language model with alignment-based hypotheses that combine Whisper's output, a wav2vec2-CTC transcript, and the original text, cuts Dutch child read-speech word error rate from 9.4% to 5.1% and raises reading-mistake detection F1 from 0.39 to 0.73, which the authors present as state of the art. The authors care because reading-aloud assessment is labor-intensive and error-prone, and because standard ASR models — built for adult, fluent, correctly-spoken speech — tend to repair children's mistakes into the intended words instead of writing the mistakes down, which makes automated detection of reading errors fail. If the claim holds, an automated reading tutor needs no task-specific training data, only prompting recipes.

What carries the argument

The load-bearing object is the prompt itself: an irrelevant text passage into which a rule-based algorithm has inserted mistakes at a 300% rate — several generated mistakes per word — which Whisper treats as a transcription-style guide rather than as lexical content. Because the prompt is not the text being read, Whisper cannot use it as a spelling reference; instead it adopts the prompt's error-tolerant style and transcribes the child's actual self-repetitions, short-word insertions, and other mistakes. A second mechanism is the heuristic hallucination filter: any prompted hypothesis that is more than 20% longer or more than 5% shorter than the read text is treated as a hallucination and replaced by the unprompted Whisper transcription. The third mechanism is the alignment-hypothesis refinement: the LLM receives forced alignments of (read text, Whisper hypothesis) and (read text, wav2vec2-CTC hypothesis) and must emit a forced alignment for (read text, refined hypothesis), a format that turns the LLM into a fusion and correction layer; the paper shows this format is decisive, because the same LLMs with plain-hypothesis input and output degrade performance.

What would settle it

Rerun the best pipeline on the same 2.05-hour test set with the hallucination filter removed and with its thresholds swept (for example, 10% longer/2% shorter and 30% longer/10% shorter), then recompute WER and F1; if the reported 5.1% WER and 0.73 F1 degrade toward the unprompted baseline (9.4% WER, 0.39 F1) as the filter changes, the gains are produced by which hypotheses the filter keeps rather than by the prompts. Reporting the share of test utterances replaced by the fallback in each condition would settle the question more cheaply.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that a prompted ASR pipeline can transcribe reading mistakes faithfully instead of silently correcting them. On Dutch child read speech, unprompted Whisper attains 9.4% WER, but its transcriptions are fluent in the wrong way: it detects reading mistakes at F1 0.39, weighted toward deletions (F1 0.71) while largely missing insertions (F1 0.27) and substitutions (F1 0.48). The authors show that prompting Whisper Large v2 with four sentences of irrelevant text containing 300% generated mistakes (about three random mistakes per word) lowers WER to 8.1% and raises detection F1 to 0.61, and that feeding both this hypothesis and a wav2vec2-CTC hypothesis — in the forced-alignment form (read text, hypothesis) — to an instruction-tuned LLM that must also produce an aligned output yields 5.1% WER and 0.73 F1, with balanced precision (0.76) and recall (0.70). The crucial conditioning detail is the input-output format: asking the LLM to rewrite a bare hypothesis makes WER worse (9.1% to 9.8%), whereas the alignment format improves it, which the authors take as evidence that the LLM works as an alignment-level fusion layer that uses the reading text and Dutch phoneme-grapheme knowledge to repair words while keeping the mistakes the CTC model flagged.

Load-bearing premise

The load-bearing premise is the hallucination filter that decides which prompted transcriptions are trustworthy: any hypothesis more than 20% longer or more than 5% shorter than the text being read is thrown away and replaced with the unprompted Whisper output, and this rule was calibrated on a 55-second, single-speaker validation clip with no reported sensitivity analysis.

Editorial extensions

If this is right

  • At the reported F1 of 0.73 (precision 0.76, recall 0.70), a prompted pipeline would catch 70% of a child's actual reading mistakes with about three of every four flagged mistakes being real, where the unprompted baseline caught only 31% of mistakes.
  • Because the entire improvement comes from text prompts and API-callable models, the pipeline can be applied to any written-Dutch reading material without retraining or fine-tuning ASR models on child speech.
  • The LLM step generalizes across similarly sized instruction-tuned models (GPT-4o-mini, Llama 3.3-70B-Instruct, DeepSeek-R1-Distill-70B), and the size of the WER gain tracks each model's Dutch-language proficiency, giving practitioners a cheap benchmark-based way to select the LLM.
  • Because the alignment format, not the LLM identity, is what makes refinement work in the paper's data, the method is presented as generalizable across instruction-tuned LLMs, and its failure mode—plain-hypothesis prompting hurts WER—is the same for all three models tested.

Reading between the lines

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

  • If the style-transfer account is correct, saturated-error prompting should work as a general 'stop normalizing' switch for Whisper on other phenomena it erases—fillers, false starts, self-corrections, dialect spellings—which could be tested in any language on a few minutes of labeled read-aloud or spontaneous speech.
  • The results suggest a reusable division of labor the paper does not generalize: the CTC decoder is the high-recall detector of mispronunciations, Whisper is the high-precision transcriber, and the LLM is a cheap fusion layer over their alignments; the same reconciliation could combine any phone-level error detector with any attention-based decoder.
  • Because the hallucination filter silently replaces flagged hypotheses with unprompted Whisper output, the headline numbers are a mixture of prompted and unprompted transcriptions; reporting the replacement rate per condition and sweeping the two thresholds would reveal how much of the gain is attributable to prompting alone.
  • The correlation the paper notes between LLM Dutch proficiency and WER gain yields a testable prediction: applying the same prompts in a lower-resource language should shrink the LLM step's contribution, so practitioners could decide between better ASR and better LLM support on the basis of a language-model benchmark rather than an ASR experiment.
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 / 4 minor

Summary. This paper proposes two prompt-based pipelines for Dutch child read speech recognition and reading mistake detection. The first prompts Whisper with text containing generated reading mistakes before transcription; the second prompts an instruction-tuned LLM, optionally with forced-alignment hypotheses, to refine Whisper output. On the Jasmin-CGN corpus, the authors report that prompting Whisper with irrelevant text containing 300% generated mistakes reduces WER from 9.4% to 8.1%, and that the LLM alignment-hypothesis setting further reduces WER to 5.1%, while reading-mistake detection F1 increases from 0.39 to 0.73. The paper also analyzes frequent ASR substitution and deletion errors and claims state-of-the-art performance for Dutch child read speech.

Significance. If the reported results are sound, the work is practically relevant for automated reading assessment, a low-resource and educationally important setting. The paper's strengths include a detailed, transferable prompting methodology, evaluation over multiple Whisper and LLM models, an explicit error taxonomy, and a public code/prompt repository, which support reproducibility. The main empirical claims, however, rest on a comparison whose interpretation is complicated by an asymmetric hallucination-fallback rule and by the absence of statistical and comparative evidence for the state-of-the-art assertion.

major comments (4)
  1. [Section 2.3.1, Tables 1 and 4] The hallucination check is applied only to prompted Whisper hypotheses: outputs more than 20% longer or 5% shorter than the read text are replaced with standard unprompted Whisper output, while the "no prompt" baseline receives no equivalent rescue. Because the fallback can only replace a prompt-induced outlier with the baseline output, it imposes a one-sided cap on the prompted pipeline's errors, so the reported WER gain from 9.4% to 8.1% (and the F1 improvements in Table 4) may partly reflect the rescue mechanism rather than the prompt itself. The paper reports neither the fraction of test hypotheses that trigger the fallback nor the prompted-pipeline performance with the fallback disabled, so the central claim that prompting Whisper improves child read speech recognition is not fully supported by the reported comparison.
  2. [Section 2.3.1 and Section 2.1] The hallucination thresholds (20% and 5%) are heuristic and are said to be confirmed only on a 55-second, single-speaker, five-sentence validation set drawn from the third story, whereas the evaluation test set is 2.05 hours from the first story. No sensitivity analysis is given for these thresholds, and the 300% mistake-density choice in the prompt is likewise not justified by a systematic search. If the length filter removes a biased subset of prompted outputs, the magnitude of the reported prompting gains could be overstated, so the paper should quantify fallback rates and test at least a few threshold settings.
  3. [Abstract, Section 2.3, and Section 4] The abstract states that the best system achieved state-of-the-art recognition performance in Dutch child read speech, but no comparison table with prior published WER results on Dutch child read speech is provided. The only support cited is the authors' own prior work [7], which is not an independent benchmark. The state-of-the-art claim is therefore not substantiated by the evidence in this manuscript and should either be removed or backed by a direct comparison to the best previously reported numbers on the same or comparable data.
  4. [Tables 1, 2, and 4] The paper uses "significantly" in the abstract and conclusion to describe improvements, but no significance tests, confidence intervals, or error bars are reported for any of the WER or F1 differences. With 14,251 reading attempts in the test set, paired or per-utterance significance tests are feasible and would strengthen the claim that the differences are not due to chance. Without such statistics, the reader cannot assess whether the differences between, for example, 9.4% and 8.1% in Table 1 or 8.1% and 5.1% in Table 2 are reliable.
minor comments (4)
  1. [Table 3] The table header contains inconsistent formatting ("LLama 70 b" and "DeepSeek-R1-distill 70b") and the value "9,4" uses a comma as decimal separator, differing from the rest of the paper.
  2. [Table 4 caption] The caption says the table is "based on whisper with irrelevant text + 300% mistakes," but the first two rows report "Whisper without prompt" and "+read text 10% mistakes," which are not based on that setting; the caption should be reworded to avoid ambiguity.
  3. [Section 2.3.1] The sentence "Since Whisper has a limited token capacity for nearly reached the maximum token size" is grammatically incomplete and obscures the intended reason for choosing the 300% mistake setting; this should be rewritten.
  4. [Discussion] The statement "Our results suggested broad applicability across LLMs" overstates the evidence, since only three LLMs of similar size were tested in a single pipeline configuration; the claim should be softened or qualified.

Circularity Check

1 steps flagged · score 2.0 of 10

No construction-level circularity; the prompt-vs-baseline comparison is internally controlled, and only the SOTA framing rests on a self-citation.

  1. self citation load bearing [Section 1 (Introduction)]
    "We use Whisper models as our speech recognition module as they support text prompt transfer styles of transcriptions and appear to be the state-of-the-art (SOTA) models for Dutch child speech recognition [7]."

    The abstract's headline claim of state-of-the-art recognition performance, and the choice of Whisper as the base ASR module, are justified by [7], a prior Interspeech 2024 paper by the same authors. The present paper does not benchmark against independent published Dutch child-read-speech systems, so the SOTA label is inherited from the authors' own prior work rather than established by an external comparison in this paper. This self-citation carries the framing claim, although the core prompt-versus-no-prompt WER and F1 comparisons are measured on the same held-out test set against independent annotations and are not constructed from the prompt inputs.

full rationale

The paper is an empirical evaluation rather than a derivation. Reported WER and F1 scores are computed from ASR/LLM outputs on a held-out 2.05-hour test set (14,251 reading attempts) using SCTK force alignment and ground-truth mistake annotations; no equation defines the reported metrics in terms of the prompt text, and the read text is a legitimate prior for read-speech evaluation because the ground truth includes the children's actual misreadings (615 incorrect words or fragments). The hallucination fallback in Section 2.3.1 is a methodological confound: it can replace out-of-bound prompted hypotheses with the unprompted baseline while the baseline gets no equivalent rescue, and the paper does not report the fallback rate or a no-fallback ablation, so part of the reported prompting gain may reflect this asymmetric rescue rather than the prompt itself. However, this is a statistical/validity concern, not circularity in the sense of the target being equivalent to the input by construction. The only self-citation that plays a load-bearing role is the SOTA status inherited from [7]; the central prompt-versus-baseline comparisons are internally controlled and self-contained. Hence a low score of 2.

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

The ledger has two hand-set thresholds and four domain assumptions, but no invented entities. The paper's gain is empirical, not theoretical, and the main burden is whether the prompt and threshold choices generalize.

free parameters (2)
  • Hallucination length thresholds = 20% longer / 5% shorter than read text
    Hand-set in Section 2.3.1, confirmed on a 55-second validation set; hypotheses outside this band are replaced by unprompted Whisper output.
  • Prompt synthetic mistake density = 10%, 100%, 300% mistake-injected prompts
    Experimental prompt configurations chosen by hand; the 300% irrelevant prompt is selected as optimal without a full sensitivity sweep.
assumptions (4)
  • domain assumption Whisper text prompts steer transcription style but do not act as a spelling reference when the prompt is irrelevant to the read text.
    Section 2.3.1 builds the 300% irrelevant prompt design on this assumption.
  • ad hoc to paper A length-based hallucination threshold reliably identifies Whisper hallucinations in prompted output.
    Section 2.3.1; threshold set heuristically and checked on a 55-second validation set, with no sensitivity analysis.
  • domain assumption LLM refinement using read text, common mistake examples, and CTC hypotheses improves rather than corrupts the transcript.
    Section 2.3.2; this is the premise of the prompting LLM pipeline.
  • domain assumption The Jasmin-CGN orthographic and mistake annotations are sufficiently reliable ground truth under the loose alignment criteria of [22].
    Sections 2.1 and 2.2; all WER and F1 measures inherit this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Child Speech Recognition and Reading Mistake Detection by Using Prompts." pith.science (2026). https://pith.science/paper/4VDMWME6

@misc{pith2026250611079,
  author       = {Pith},
  title        = {Pith review of: Improving Child Speech Recognition and Reading Mistake Detection by Using Prompts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VDMWME6}},
  note         = {Machine review of arXiv:2506.11079}
}
read the original abstract

Automatic reading aloud evaluation can provide valuable support to teachers by enabling more efficient scoring of reading exercises. However, research on reading evaluation systems and applications remains limited. We present a novel multimodal approach that leverages audio and knowledge from text resources. In particular, we explored the potential of using Whisper and instruction-tuned large language models (LLMs) with prompts to improve transcriptions for child speech recognition, as well as their effectiveness in downstream reading mistake detection. Our results demonstrate the effectiveness of prompting Whisper and prompting LLM, compared to the baseline Whisper model without prompting. The best performing system achieved state-of-the-art recognition performance in Dutch child read speech, with a word error rate (WER) of 5.1%, improving the baseline WER of 9.4%. Furthermore, it significantly improved reading mistake detection, increasing the F1 score from 0.39 to 0.73.

Figures

Figures reproduced from arXiv: 2506.11079 by the authors.

Figure 1
Figure 1. Overviews of prompting Whisper pipeline (a) on the left side and prompting LLM pipeline (b) on the right side. 2.3.1. Prompting Whisper Pipeline We form a three-step reading mistake detection pipeline for prompting Whisper as shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 33 canonical work pages

  1. [7]

    Read - ing miscue detection in primary school through automatic speech recognition,

    L. Gao, C. Tejedor -Garcia, H. Strik, and C. Cucchiarini, “Read - ing miscue detection in primary school through automatic speech recognition,” in Interspeech 2024, 2024, pp. 5153–5157

    work page 2024

  2. [1]

    Automating this process could improve efficiency, which is important as declining literacy rates highlight the need for innovative solu - tions [2]

    Introduction Reading aloud is regularly assessed in primary education as a measure of reading skills, but this form of assessment is labor - intensive [1] and error-prone, as teachers must manually evaluate read speech, identify mistakes, and assign scores. Automating this process could improve efficiency, which is important as declining literacy rates hi...

    work page

  3. [2]

    het is zondag

    Method 2.1. Data This study uses read speech data from the Jasmin-CGN Corpus [20], featuring recordings of 71 Dutch primary school pupils (ages 6–13, 35 female, 36 male) reading three stories aloud at their mastery levels. The dataset includes orthographic and phonemic annotations, as well as reading mistake annotations for the first story [21]. The data ...

    work page 2025

  4. [3]

    No Prompt

    Results We experimented with a single A100 GPU, WhisperX [27], huggingface pipeline and OpenAI API, and DeepSeek API. The code and prompts used in this project can be found online6. 3.1. Prompting Whisper Table 1 shows the child speech recognition performance of the prompting Whisper method under various prompt configurations for Whisper Large v2 and Larg...

    work page 2025

  5. [4]

    This setting demonstrates balanced improvements across all metric and mistake types, with insertion, substitution, and deletion F1 scores of 0.76, 0.75, and 0.71, respectively

    with the LLM using hypothesis alignment, which achieves an F1 score of 0.73 for all mistakes. This setting demonstrates balanced improvements across all metric and mistake types, with insertion, substitution, and deletion F1 scores of 0.76, 0.75, and 0.71, respectively. These findings suggest that combining Whis- per with irrelevant prompts for transcript...

    work page

  6. [5]

    Acknowledgments This publication is part of the project Responsible AI for Voice 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)

    work page

  7. [6]

    read text with 100% mistakes

    Discussion and Conclusion In this work, we combined Whisper with prompts and LLMs with prompts to enhance child speech recognition and reading mistake detection, particularly in settings where resources are limited. By addressing two RQs, we evaluated the effectiveness of these methods in improving child speech recognition and reading mistake detection. T...

    work page

  8. [8]

    Wetenschappelijke verantwoording LVS-toetsen DMT,

    A. van Til, F. Kamphuis, J. Keuning, M. Gijsel, J. Vloedgraven, and A. de Wijs, “Wetenschappelijke verantwoording LVS-toetsen DMT,” 2018

    work page 2018

Show all 34 references
  1. [9]

    PIRLS-2021: Trends in leesprestaties, leesattitude en leesgedrag van tienjarigen uit Nederland

    N. Swart, J. Gubbels, M. in ‘t Zandt, M. Wolbers, and E. Segers, “PIRLS-2021: Trends in leesprestaties, leesattitude en leesgedrag van tienjarigen uit Nederland.” Nijmegen, 2023

    2021

  2. [10]

    A review of recent ad - vances on deep learning methods for audio-visual speech recogni- tion,

    D. Ivanko, D. Ryumin, and A. Karpov, “A review of recent ad - vances on deep learning methods for audio-visual speech recogni- tion,” Mathematics, vol. 11, no. 12, p. 2665, 2023

    2023

  3. [11]

    Scaling speech technology to 1,000+ languages,

    V. Pratap, A. Tjandra, B. Shi, P. Tomasello, A. Babu, S. Kundu, A. Elkahky, Z. Ni, A. Vyas, M. Fazel -Zarandi et al. , “Scaling speech technology to 1,000+ languages,” Journal of Machine Learning Research, vol. 25, no. 97, pp. 1–52, 2024

    2024

  4. [12]

    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

  5. [13]

    Benchmarking chil - dren’s asr with supervised and self-supervised speech foundation models,

    R. Fan, N. Balaji Shankar, and A. Alwan, “Benchmarking chil - dren’s asr with supervised and self-supervised speech foundation models,” in Interspeech 2024, 2024, pp. 5173–5177

    2024

  6. [14]

    Computer -guided oral reading versus independent practice: Comparison of sustained silent reading to an automated reading tutor that listens,

    J. Mostow, J. Nelson -Taylor, and J. E. Beck, “Computer -guided oral reading versus independent practice: Comparison of sustained silent reading to an automated reading tutor that listens,” Journal of Educational Computing Research, vol. 49, no. 2, pp. 249–276, 2013

    2013

  7. [15]

    Measuring word correctness in young initial readers: Compar- ing assessments from teachers, phoneticians, and asr models,

    W. Harmsen, F. Hubers, R. van Hout, C. Cucchiarini, and H. Strik, “Measuring word correctness in young initial readers: Compar- ing assessments from teachers, phoneticians, and asr models,” in 9th Workshop on Speech and Language Technology in Education (SLaTE), 2023, pp. 11–15

    2023

  8. [16]

    [8] examine its application to reading word lists

    highlight its effectiveness in transcribing fillers in child read- aloud speech, while Harmsen et al. [8] examine its application to reading word lists. Although Whisper often corrects reading mis- takes rather than directly transcribing them, its confidence scores aid in dete...

  9. [17]

    Improving child speech recognition with augmented child -like speech,

    Y. Zhang, Z. Yue, T. Patel, and O. Scharenborg, “Improving child speech recognition with augmented child -like speech,” in Inter- speech 2024, 2024, pp. 5183–5187

    2024

  10. [18]

    End-to-end neural systems for automatic children speech recognition: An empirical study,

    P. G. Shivakumar and S. Narayanan, “End-to-end neural systems for automatic children speech recognition: An empirical study,” Computer Speech & Language, vol. 72, p. 101289, 2022

    2022

  11. [19]

    A dataset and two-pass system for reading miscue detection,

    R. Gothi, R. Kumar, M. Pereira, N. Nayak, and P. Rao, “A dataset and two-pass system for reading miscue detection,” in Interspeech 2024, 2024, pp. 4014–4018

    2024

  12. [20]

    Automatic Assessment of Oral Reading Accuracy for Reading Diagnostics,

    B. Molenaar, C. Tejedor -Garcia, C. Cucchiarini, and H. Strik, “Automatic Assessment of Oral Reading Accuracy for Reading Diagnostics,” in Proc. Interspeech 2023, 2023, pp. 5232–5236

    2023

  13. [21]

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

    A. Baevski, Y. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in neural information processing systems , vol. 33, pp. 12 449–12 460, 2020

    2020

  14. [22]

    The Science of Reading Pro - gresses: Communicating Advances Beyond the Simple View of Reading,

    N. K. Duke and K. B. Cartwright, “The Science of Reading Pro - gresses: Communicating Advances Beyond the Simple View of Reading,” Reading Research Quarterly, vol. 56, no. S1, pp. S25– S44, 2021

    2021

  15. [23]

    Computergestützte bestimmung des sprechflusses bei vorschulkindern,

    V. Kany and J. Trouvain, “Computergestützte bestimmung des sprechflusses bei vorschulkindern,” in Elektronische Sprachsig- nalverarbeitung 2024, Tagungsband der 35. Konferenz, Regens- burg, 6.-8. März 2024. TUDpress, 2024, pp. 62–69

    2024

  16. [24]

    Do Prompts Really Prompt? Exploring the Prompt Understanding Capability of Whis- per,

    C.-K. Yang, K. -P. Huang, and H. -Y. Lee, “Do Prompts Really Prompt? Exploring the Prompt Understanding Capability of Whis- per,” in 2024 IEEE Spoken Language Technology Workshop (SLT), 2024, pp. 1–8

    2024

  17. [25]

    Automatic speech recognition tuned for child speech in the classroom,

    R. Southwell, W. Ward, V. A. Trinh, C. Clevenger, C. Clevenger, E. Watts, J. Reitman, S. D’Mello, and J. Whitehill, “Automatic speech recognition tuned for child speech in the classroom,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Process...

    2024

  18. [26]

    Progres: Prompted generative rescoring on asr n-best,

    A. D. Tur, A. Moumen, and M. Ravanelli, “Progres: Prompted generative rescoring on asr n-best,” in 2024 IEEE Spoken Language Technology Workshop (SLT), 2024, pp. 600–607

    2024

  19. [27]

    Recording speech of children, non -natives and elderly people for HLT applications: the JASMIN-CGN corpus

    C. Cucchiarini, J. Driesen, H. Van hamme, and E. Sanders, “Recording speech of children, non -natives and elderly people for HLT applications: the JASMIN-CGN corpus.” in Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08) . Marrakec...

    2008

  20. [28]

    Analyzing read aloud speech by primary school pupils: Insights for research and development,

    S. Limonard, C. Cucchiarini, R. van Hout, and H. Strik, “Analyzing read aloud speech by primary school pupils: Insights for research and development,” in Interspeech 2020, 2020, pp. 3710–3714

    2020

  21. [29]

    Automatic detection of word-level reading errors in non-native english speech based on asr output,

    Y. Qin, Y. Qian, A. Loukina, P. Lange, A. Misra, K. Evanini, and T. Lee, “Automatic detection of word-level reading errors in non-native english speech based on asr output,” in 2021 12th International Symposium on Chinese Spoken Language Processing (ISCSLP). IEEE, 2021, pp. 1–5

    2021

  22. [30]

    Developing an automatic assessment tool for children’s oral reading,

    L. Cleuren, J. Duchateau, A. Sips, P. Ghesquière, and H. Van Hamme, “Developing an automatic assessment tool for children’s oral reading,” in Interspeech 2006. ISCA, Sep. 2006, pp. paper 1113–Tue1WeS.1–0

    2006

  23. [31]

    Chil- dren’s oral reading corpus (chorec): Description and assessment of annotator agreement

    L. Cleuren, J. Duchateau, P. Ghesquiere, and H. V. hamme, “Chil- dren’s oral reading corpus (chorec): Description and assessment of annotator agreement.” in LREC, 2008

    2008

  24. [32]

    Quantifying language variation acoustically with few resources,

    M. Bartelds and M. Wieling, “Quantifying language variation acoustically with few resources,” in Proceedings of the 2022 Con- ference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2022, pp. 3735–3741

    2022

  25. [33]

    En- coder vs Decoder: Comparative Analysis of Encoder and Decoder Language Models on Multilingual NLU Tasks,

    D. Saattrup Nielsen, K. Enevoldsen, and P. Schneider-Kamp, “En- coder vs Decoder: Comparative Analysis of Encoder and Decoder Language Models on Multilingual NLU Tasks,” in Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference ...

    2025

  26. [34]

    WhisperX: Time - Accurate Speech Transcription of Long -Form Audio,

    M. Bain, J. Huh, T. Han, and A. Zisserman, “WhisperX: Time - Accurate Speech Transcription of Long -Form Audio,” in Proc. Interspeech 2023, 2023, pp. 4489–4493

    2023

Pith tools

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