Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Prompting Whisper for Improved Verbatim Transcription and End-to-end Miscue Detection

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 the target reading text improves verbatim transcription over fine-tuning alone, and adding special miscue tokens makes end-to-end detection of reading errors feasible.

desk verdict The central prompting claim holds up across the table; the real weakness is that every metric rests on unvalidated ground-truth labels. read the letter →

arxiv 2505.23627 v1 pith:BCZGVNF2 submitted 2025-05-29 cs.LG

Griffin Dietz Smith , Dianna Yee , Jennifer King Chen , Leah Findlater This is my paper
classification cs.LG
keywords speechrecognitionverbatimtranscriptionmiscuedetectionWhisperpromptconditioningchildren'satypicalreadingerrors
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 claims that a speech recognition model called Whisper transcribes read-aloud speech more accurately when it is given the text the speaker was supposed to read as part of the prompt, compared with fine-tuning the model on such speech without that context. The authors prepend the tokenized reading text to Whisper's start-of-sequence token, train the model to predict only the verbatim transcript, and report lower speaker-level word-error rates on children's read-aloud and adult atypical-speech corpora. They also add three special vocabulary tokens -- , , and -- so the same model can output reading-error annotations directly, making end-to-end miscue detection feasible without a separate post-hoc comparison step. If correct, these results would let educators and clinicians automate reading-error annotation with a single pass, instead of the current slow manual labeling or fragile post-hoc transcript matching.

What carries the argument

The load-bearing mechanism is prompt conditioning within Whisper's encoder-decoder structure: the tokenized reading text is prepended to the start-of-sequence token, and the training loss is computed only over the verbatim transcript and miscue tokens, so the model learns to use the reading text as context without being trained to reproduce it. The second mechanism is tokenizer augmentation, where <OMIT>, <SUBSTITUTE>, and <INSERT> are added to the vocabulary so the decoder can interleave miscue annotations with words. The evaluation pipeline aligns predicted and ground-truth transcriptions after removing miscue tokens, then computes speaker-level word-error rate and per-type F1, which is what carries the quantitative claims.

What would settle it

Take a random sample of several hundred utterances from each dataset and have independent annotators re-transcribe them with the same verbatim guidelines; if token-level agreement falls well below near-perfect, the ground-truth labels are too noisy to support the reported numbers. Alternatively, generate a read-aloud test set from text-to-speech with known injected miscues, run the tuned-prompt and tuned-unprompted recipes on it, and check whether the prompted model still achieves the lower word-error rate.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that conditioning Whisper on the target reading text is not a minor auxiliary cue but the primary driver of verbatim transcription improvement. Prepending the tokenized reading text to the start-of-sequence token and fine-tuning with loss restricted to the transcript yields lower speaker-level word-error rates across Whisper sizes (tiny, small, medium) on both case-study datasets than fine-tuning without the prompt; on the children's corpus, tuned prompted models reach word-error rates near 4-5 percent with medium and small models, roughly half the error of tuned unprompted models. Adding the miscue tokens preserves most of that transcription gain and lets the decoder predict <OMIT>, <SUBSTITUTE>, and <INSERT> tokens directly, with F1 values that are competitive with post-hoc calculation on smaller models, while post-hoc calculation from the improved transcripts remains the most accurate miscue signal overall.

Load-bearing premise

The results stand or fall with the accuracy of the human transcriptions used as ground truth: if those labels do not capture exactly what was said, including every omission, substitution, and insertion, then the reported word-error rates and miscue F1 scores no longer measure verbatim reading performance.

Editorial extensions

If this is right

  • Reading-annotation systems should condition on the target text: fine-tuning with the reading-text prompt is the strongest single improvement reported, beating fine-tuning without the prompt on both case-study datasets.
  • A single end-to-end model can emit both the verbatim transcript and the miscue labels, removing the need for a separate post-hoc alignment module in applications where real-time annotation matters.
  • Smaller speech models gain disproportionately from the miscue-prediction task, so deployable reading tutors do not necessarily need the largest available ASR models.
  • Transfer to a held-out children's corpus is positive but slightly weaker than same-distribution performance, so models trained this way still face a distribution-shift penalty when readers or texts change.
  • For the most accurate miscue labels, post-hoc calculation from prompted transcripts remains the safer choice, while E2E prediction is competitive and improving for smaller models.

Reading between the lines

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

  • If the prompting benefit transfers beyond read-aloud speech, the same recipe could improve ASR for other constrained tasks, such as dictating into forms or transcribing read medical instructions, where the expected text is known in advance.
  • The miscue-token vocabulary is explicitly extensible: adding tokens for pauses, fillers, or repetition subtypes would let the same architecture address stuttering assessment or fluency scoring without redesigning the model.
  • A direct test of the label assumption would be to re-annotate a subset of the two datasets with multiple transcribers and compare agreement; low agreement would mean the reported WER and F1 numbers need recalibration.
  • An ablation that removes the miscue tokens while keeping the reading-text prompt would isolate whether the observed benefit for smaller models comes from the extra prediction task itself or from the prompt, a distinction the paper leaves implicit.
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. The paper proposes two modifications to Whisper for reading-error annotation: (1) prepending the target reading text as a prompt during fine-tuning, and (2) augmenting the tokenizer with miscue-event tokens (<OMIT>, <SUBSTITUTE>, <INSERT>) to enable end-to-end miscue detection. The authors evaluate on children's read-aloud speech (Xc), adult atypical speech (Xa), and an out-of-distribution set (XCMU), reporting speaker-level WER and miscue F1 across three random splits. They claim that prompting with the reading text improves verbatim transcription over fine-tuning without prompting, and that end-to-end miscue detection is feasible, with post-hoc calculation still more accurate. I verified the Table 1 WER values: the example cited as a possible counterexample (Xc, medium.en) actually supports the paper's claim, since tuned-unprompted WER is 8.9±3.6 and tuned-prompted WER is 4.0±1.0.

Significance. If the results hold, the paper makes a useful practical contribution: conditioning an ASR model on the target reading text via prompting is a simple and effective alternative to task-specific language models, and extending the output vocabulary with miscue tokens demonstrates a plausible route to end-to-end miscue detection. The experiments cover two relevant case studies and an out-of-distribution generalization test, with three-trial averages and standard deviations reported. The central comparison in Table 1 is consistently in favor of tuned-prompted models across all datasets and model sizes, and the tables report speaker-level metrics with a defined alignment procedure. The main weakness is that the ground-truth labels—especially the re-annotated CMU Kids corpus — are asserted to be verbatim without any supporting annotation-quality evidence, which is a load-bearing precondition for every reported number.

major comments (4)
  1. [§3.2, §3.3, Tables 1 and 2] The validity of all reported WER and miscue F1 scores depends on the accuracy of ground-truth transcriptions that are stated to be verbatim, but the manuscript provides no evidence that the labels are reliable. For Xc and Xa, the paper simply states that miscues are transcribed verbatim; for XCMU, it states that the dataset was re-annotated for verbatim capture of miscues, yet gives no annotation protocol, no inter-annotator agreement, no adjudication procedure, and no label-error analysis. This is not a cosmetic omission: if the labels systematically miss or mislabel substitutions, omissions, or insertions, the relative ordering of tuned-prompt versus tuned-unprompt WER could change, and the miscue F1 numbers become uninterpretable. I ask the authors to either report annotation-quality statistics (e.g., inter-annotator agreement, adjudication rates, a sample audit) or explicitly assess the error rate in the re-annotated labels and its potential impact on the conclusions.
  2. [§3.3, Figure 3] The miscue-detection evaluation procedure is under-specified. The text states that predicted and ground-truth miscue event tokens are aligned and that a 'no tag' is inserted where needed, but it does not define the alignment algorithm, how insertion versus omission is aligned to specific words, or how ties are broken when multiple miscue types are adjacent. Without a precise, reproducible definition of this alignment and the resulting F1 computation, the numbers in Table 2 cannot be independently verified or compared to future work. Please provide a formal description of the alignment and scoring procedure.
  3. [§2.1, §2.2, §3.1] Training details are insufficient for reproducibility. The manuscript does not state whether fine-tuning updates all parameters or uses adapters/LoRA, the learning rate, number of epochs, batch size, optimizer, or how the 30-second and 448-token limits are enforced (truncation versus splitting). It also does not specify the exact output sequence format for the E2E model: are miscue tokens inserted immediately after each affected word, before the word, or in place of the word? The loss-masking description is clear in principle (loss is computed only on transcript and miscue tokens, not the prompt), but the sequence-format choices materially affect both training and evaluation. Please include a more complete specification.
  4. [Abstract, §1, §4] The abstract and conclusion claim improvement 'compared to current state-of-the-art,' but the experiments include no direct state-of-the-art baseline. The comparison to the prior E2E work [17] on XCMU is explicitly non-direct because of taxonomy differences, and no post-hoc reading-text-constrained language-model approach (e.g., [12], [13]) is benchmarked on the same data. I recommend either adding a comparable baseline or tempering the 'state-of-the-art' claim to 'over the included baselines'.
minor comments (4)
  1. [Table 2] In the Xc, medium.en, <INSERT> row, the parenthesized recall value appears as '64.9' rather than '0.649'; this looks like a formatting typo.
  2. [Tables 1 and 2] The column headers for the END-TO-END conditions are visually ambiguous ('END-TO-END [ours] CALC[ours] PRED[ours]'); the caption should make explicit that these are two separate columns, one for post-hoc-computed miscues and one for directly predicted miscues.
  3. [§4.1] Several central comparisons have overlapping standard deviations across only three trials (e.g., Xc medium: tuned unprompted 8.9±3.6 versus tuned prompted 4.0±1.0). The paper would be strengthened by reporting paired significance tests or at least explicitly noting the overlap when drawing conclusions from individual rows.
  4. [§3.2] The description of the Speech Accessibility Project data says 'read speech utterances' and 'miscues which may arise either naturally or due to dysfluencies,' but it is not stated whether the provided transcriptions already followed verbatim conventions or whether the authors re-annotated them; please clarify this in the data description.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central WER and miscue-F1 comparisons are held-out empirical measurements against independently defined labels, with no fitted quantity renamed as a prediction.

full rationale

The paper's derivation chain is empirical rather than definitional. The target reading text is provided as a decoder prompt, and the model is trained with loss over verbatim transcriptions and miscue tokens; the claim that prompting helps is supported by Table 1 comparing tuned-prompted, tuned-unprompted, untuned, and naive-copy baselines on held-out speaker-disjoint splits. The token-level miscue labels are the prediction targets, not fitting parameters, and the F1 computation aligns predicted tokens to ground-truth tokens, so the E2E numbers do not reduce to the training objective by construction. The re-annotated CMU Kids corpus is an external held-out set used only for evaluation, and the comparison against the prior LSTM-based approach [17] is an external benchmark. The only author self-citation is [9], used as background for stuttering-related ASR; it is not load-bearing and no uniqueness claim is imported from it. Concerns about unvalidated ground-truth label quality, annotation consistency, or the absence of adversarial hyperparameter variation are threats to measurement validity and statistical robustness, but they are not circularity: the reported quantities are not equivalent to their inputs by definition.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The central contribution rests on three types of assumptions: accurate labels in both proprietary datasets, a consistent re-annotation of CMU Kids, and standard fine-tuning practice. The only hand-chosen design parameters that materially affect results are the model size and the three-token miscue taxonomy. No new physical entities or mathematical axioms are introduced; the invented tokens are evaluated directly against held-out labels.

free parameters (2)
  • Whisper model size = tiny.en, small.en, medium.en
    Three model sizes are tested, and conclusions depend on model size, e.g., E2E benefit is strongest for tiny.en. The choice of which size to report as the headline result is not specified, and the medium.en results sometimes contradict the prompting-vs-fine-tuning narrative.
  • Miscue taxonomy = <OMIT>, <SUBSTITUTE>, <INSERT>
    The set of three miscue types is chosen to be computable with string comparison operations, and the paper explicitly excludes pauses and filler words. The taxonomy choice affects both training and evaluation, and is a human design decision rather than derived from the data.
assumptions (3)
  • domain assumption Ground-truth transcriptions in Xc and Xa are accurate verbatim captures of what was said.
    All reported WER and miscue F1 numbers are computed against these labels. Section 3.2 states transcriptions capture miscues verbatim, but no label-quality or agreement analysis is provided.
  • domain assumption The re-annotation of the CMU Kids Corpus is consistent with the Xc annotation guidelines.
    Section 3.2 says the dataset 'was re-annotated for a verbatim capture of miscues,' but no annotation protocol, inter-annotator agreement, or validation is reported. XCMU is used to claim generalization.
  • domain assumption Fine-tuning Whisper with LoRA or full fine-tuning preserves the pretrained model's general speech capability.
    The paper does not specify whether fine-tuning was full or parameter-efficient, and assumes that the fine-tuned model remains a valid ASR system. This is standard practice but unstated in detail.
invented entities (1)
  • Miscue event tokens <OMIT>, <SUBSTITUTE>, <INSERT> independent evidence
    purpose: Augment the tokenizer vocabulary so the model can directly output reading-error annotations alongside the verbatim transcript.
    The tokens are a new interface for the model, but their meaning is defined by the dataset labels and the evaluation protocol: they are directly evaluated against held-out ground-truth miscue annotations, so their usefulness is independently measurable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prompting Whisper for Improved Verbatim Transcription and End-to-end Miscue Detection." pith.science (2026). https://pith.science/paper/BCZGVNF2

@misc{pith2026250523627,
  author       = {Pith},
  title        = {Pith review of: Prompting Whisper for Improved Verbatim Transcription and End-to-end Miscue Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCZGVNF2}},
  note         = {Machine review of arXiv:2505.23627}
}
read the original abstract

Identifying mistakes (i.e., miscues) made while reading aloud is commonly approached post-hoc by comparing automatic speech recognition (ASR) transcriptions to the target reading text. However, post-hoc methods perform poorly when ASR inaccurately transcribes verbatim speech. To improve on current methods for reading error annotation, we propose a novel end-to-end architecture that incorporates the target reading text via prompting and is trained for both improved verbatim transcription and direct miscue detection. Our contributions include: first, demonstrating that incorporating reading text through prompting benefits verbatim transcription performance over fine-tuning, and second, showing that it is feasible to augment speech recognition tasks for end-to-end miscue detection. We conducted two case studies -- children's read-aloud and adult atypical speech -- and found that our proposed strategies improve verbatim transcription and miscue detection compared to current state-of-the-art.

Figures

Figures reproduced from arXiv: 2505.23627 by the authors.

Figure 1
Figure 1. Proposed architecture for end-to-end miscue detec￾tion by augmenting Whisper to include miscue tokens and in￾corporate reading text via prompting [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Example of a target reading text prompt and potential miscue events defined in Section 2. 2. Method We propose modifying the canonical ASR task for MD with two architectural changes. Our first change fine-tunes the pre￾trained ASR while prompting with the target reading text. This prompting is a systematic way to incorporate additional relevant context from the reading text during speech recognition and should help … view at source ↗
Figure 3
Figure 3. Examples of ground truth transcripts, predicted transcripts, and processing applied to evaluate MD perfor￾mance given a reading text prompt. Predicted transcripts are processed in terms of predicted miscue event tokens and ‘CORRECT’ tokens. Note that ‘<CORRECT>’ is shown as ‘<C>’ for readability and incorrectly predicted miscue event tokens are highlighted in red. F1 is computed using the pre￾dicted and ground truth… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Transcription Policy as a Latent Variable: Activating Controllable Verbatim ASR with Word-Level Timing

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Mode-tag conditioning on paired verbatim/intended data makes Whisper produce either verbatim or intended transcripts on demand, with cross-lingual disfluency control and improved word timestamps.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [17]

    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,” inProc. In- terspeech 2024, 2024, pp. 4014–4018

    work page 2024

  2. [12]

    Sparsely shared lora on whisper for child speech recognition,

    W. Liu, Y . Qin, Z. Peng, and T. Lee, “Sparsely shared lora on whisper for child speech recognition,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 11 751–11 755

    work page 2024

  3. [13]

    Shared-adapters: A novel transformer- based parameter efficient transfer learning approach for children’s automatic speech recognition,

    T. Rolland and A. Abad, “Shared-adapters: A novel transformer- based parameter efficient transfer learning approach for children’s automatic speech recognition,” inInterspeech 2024, 2024, pp. 2370–2374

    work page 2024

  4. [1]

    Introduction The practice of annotating reading errors (i.e., miscues) in read-aloud text has applications in education, speech-language pathology, and speech data collection, and can be used to bench- mark, track progress, or identify areas for intervention. How- ever, annotating reading errors is traditionally a time-consuming manual process that involv...

    work page Pith review arXiv 2025

  5. [2]

    Our first change fine-tunes the pre- trained ASR while prompting with the target reading text

    Method We propose modifying the canonical ASR task for MD with two architectural changes. Our first change fine-tunes the pre- trained ASR while prompting with the target reading text. This prompting is a systematic way to incorporate additional relevant context from the reading text during speech recognition and should help improve both transcription per...

    work page

  6. [3]

    ASR model We ran experiments with various sizes of Whisper{tiny.en, small.en, medium.en}to better understand the effect of model size on the performance of the proposed approach

    Experiments 3.1. ASR model We ran experiments with various sizes of Whisper{tiny.en, small.en, medium.en}to better understand the effect of model size on the performance of the proposed approach. We use English-only models as all our datasets are in English. 3.2. Data To examine our approach, we consider two case studies where speakers were tasked with re...

    work page

  7. [4]

    Results and Discussion We compare our approach to the verbatim transcription and MD baselines specified in Section 3.3 in Tables 1 and 2, respectively. 4.1. V erbatim transcription Starting with untuned, unprompted performance, which repre- sents the de facto application of ASR, we observe that WERs are much higher than 15% for children and atypical speec...

    work page

  8. [5]

    Conclusion Our experiments covered a range of approaches towards reading error annotation through modifying a large-scale speech foun- dation model. Extending prior methods that primarily focus on updating the language model or fine-tuning to improve recogni- tion performance, we propose two novel strategies for training an E2E model to jointly predict ve...

    work page

Show all 26 references
  1. [6]

    A flexible recogniser architecture in a reading tutor for children,

    J. Duchateau, M. Wigham, K. Demuynck, and H. v. Hamme, “A flexible recogniser architecture in a reading tutor for children,” in Speech Recognition and Intrinsic Variation Workshop, 2006

    2006

  2. [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,”arXiv preprint arXiv:2406.07060, 2024

    2024 arXiv

  3. [8]

    Automatic detection and classification of disfluent reading mis- cues in young children’s speech for the purpose of assessment

    M. Black, J. Tepperman, S. Lee, P. Price, and S. S. Narayanan, “Automatic detection and classification of disfluent reading mis- cues in young children’s speech for the purpose of assessment.” in Proc. Interspeech 2007, 2007, pp. 206–209

    2007

  4. [9]

    Mispronunciation detection in children’s reading of sentences,

    J. Proenc ¸a, C. Lopes, M. Tjalve, A. Stolcke, S. Candeias, and F. Perdigao, “Mispronunciation detection in children’s reading of sentences,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 26, no. 7, pp. 1207–1219, 2018

    2018

  5. [10]

    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 Acous- tics, Speech and Signal Pr...

    2024

  6. [11]

    Mixed chil- dren/adult/childrenized fine-tuning for children’s asr: How to re- duce age mismatch and speaking style mismatch,

    T. Graave, Z. Li, T. Lohrenz, and T. Fingscheidt, “Mixed chil- dren/adult/childrenized fine-tuning for children’s asr: How to re- duce age mismatch and speaking style mismatch,” inInterspeech 2024, 2024, pp. 5188–5192

    2024

  7. [14]

    From user perceptions to technical improvement: Enabling people who stutter to better use speech recognition,

    C. Lea, Z. Huang, J. Narain, L. Tooley, D. Yee, D. T. Tran, P. Georgiou, J. P. Bigham, and L. Findlater, “From user perceptions to technical improvement: Enabling people who stutter to better use speech recognition,” inProceedings of the 2023 CHI Conference on Human Factors in...

    2023

  8. [15]

    Towards a single asr model that generalizes to disordered speech,

    J. Tobin, K. Tomanek, and S. Venugopalan, “Towards a single asr model that generalizes to disordered speech,” inInternational Conference on Acoustics, Speech, & Signal Processing (ICASSP), 2024

    2024

  9. [16]

    Automatic speech recognition of conversational speech in individuals with disordered speech,

    J. Tobin, P. Nelson, B. MacDonald, R. Heywood, R. Cave, K. Seaver, A. Desjardins, P.-P. Jiang, and J. R. Green, “Automatic speech recognition of conversational speech in individuals with disordered speech,”Journal of Speech, Language, and Hearing Research, vol. 67, no. 11, pp....

    2024

  10. [18]

    Why and how our automated reading tutor listens,

    J. Mostow, “Why and how our automated reading tutor listens,” inProceedings of the International Symposium on Automatic De- tection of Errors in Pronunciation Training (ISADEPT). KTH Stockholm, Sweden, 2012, pp. 43–52

    2012

  11. [19]

    Automatic assessment of oral reading accuracy for reading di- agnostics,

    B. Molenaar, C. Tejedor-Garcia, C. Cucchiarini, and H. Strik, “Automatic assessment of oral reading accuracy for reading di- agnostics,” inProc. Interspeech 2023, 2023, pp. 5232–5236

    2023

  12. [20]

    Sed-mdd: Towards sen- tence dependent end-to-end mispronunciation detection and diag- nosis,

    Y . Feng, G. Fu, Q. Chen, and K. Chen, “Sed-mdd: Towards sen- tence dependent end-to-end mispronunciation detection and diag- nosis,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 3492–3496

    2020

  13. [21]

    A full text-dependent end to end mispronunciation detection and diag- nosis with easy data augmentation techniques,

    K. Fu, J. Lin, D. Ke, Y . Xie, J. Zhang, and B. Lin, “A full text-dependent end to end mispronunciation detection and diag- nosis with easy data augmentation techniques,”arXiv preprint arXiv:2104.08428, 2021

    2021 arXiv

  14. [22]

    End- to-end word-level disfluency detection and classification in chil- dren’s reading assessment,

    L. Venkatasubramaniam, V . Sunder, and E. Fosler-Lussier, “End- to-end word-level disfluency detection and classification in chil- dren’s reading assessment,” inICASSP 2023-2023 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  15. [23]

    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,” inProceedings of Machine Learning Research (PMLR), vol. 202, 23–29 Jul 2023, pp. 28 492–28 518

    2023

  16. [24]

    Speech Accessibility Project,

    “Speech Accessibility Project,” https://speechaccessibilityproject. beckman.illinois.edu, accessed: 2024-08-30

    2024

  17. [25]

    The CMU kids corpus,

    M. Eskenazi, J. Mostow, and D. Graff, “The CMU kids corpus,” https://catalog.ldc.upenn.edu/LDC97S63, accessed: 2024-08-30

    2024

  18. [26]

    Disordered Speech Data Collec- tion: Lessons Learned at 1 Million Utterances from Project Eu- phonia,

    R. L. MacDonald, P.-P. Jiang, J. Cattiau, R. Heywood, R. Cave, K. Seaver, M. A. Ladewig, J. Tobin, M. P. Brenner, P. C. Nelson, J. R. Green, and K. Tomanek, “Disordered Speech Data Collec- tion: Lessons Learned at 1 Million Utterances from Project Eu- phonia,” inProc. Interspe...

    2021

Pith tools

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