Pith. sign in

REVIEW 3 major objections 4 minor 47 references

CHSER: A Dataset and Case Study on Generative Speech Error Correction for Child ASR

T0 review · 3 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 A 200K-pair child speech dataset makes generative error correction cut word error rate by 28.5%.

desk verdict A genuinely useful new child-speech error-correction dataset with plausible WER gains; the unverified reference transcriptions are the main thing to fix before the benchmark numbers are fully trusted. read the letter →

arxiv 2505.18463 v1 pith:C4XCYB7C submitted 2025-05-24 eess.AS

Natarajan Balaji Shankar , Zilai Wang , Kaiyuan Zhang , Mohan Shi , Abeer Alwan This is my paper · ORCID
classification eess.AS
keywords childautomaticspeechrecognitiongenerativeerrorcorrectionASRdatasetdisfluencydetectionFlanT5fine-tuningwordratebenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that generative error correction, which has worked for adult speech, can be made to work for child speech if a large enough paired dataset is available. It builds CHSER, 200K ASR hypothesis–transcription pairs from four child corpora spanning ages 4–14 and scripted and spontaneous speech, and shows that fine-tuning Flan T5 on these pairs reduces word error rate by 28.5% relative to a zero-shot Whisper baseline. The same fine-tuned model also improves transcriptions from other ASR architectures, and multi-domain training generalizes better than single-domain training. The paper further claims that correction shifts the error mix: substitutions and deletions drop, while insertions and repetition/revision disfluencies remain difficult. If these claims hold, CHSER gives the field a benchmark and a training resource for a previously underserved population.

What carries the argument

The central object is the paired hypothesis–transcription corpus generated by running Whisper-base.en beam search over four child speech corpora and filtering to the top five hypotheses, length at least three words, and tokenizer-vocabulary membership. The correction mechanism is instruction fine-tuning of Flan T5 large with LoRA on these pairs, and the evaluation machinery includes traditional reranking LMs, oracle N-best bounds, in-context learning with GPT-4o mini, and an error-type analysis by substitution/insertion/deletion, POS tags, and five disfluency categories.

What would settle it

Take a held-out sample of child speech with transcripts independently verified by a second annotator, generate hypotheses with Whisper-base.en, apply the CHSER fine-tuned T5, and compare corrected output to the verified transcripts. If the relative WER reduction does not reproduce, or if the corrected output differs from verified transcripts on disfluent utterances, the central claim is undermined.

Watch

Extended reading notes

Core claim

The paper introduces CHSER, a dataset of 200K ASR hypothesis–transcription pairs assembled from four child speech corpora covering ages 4–14 and scripted and spontaneous speech. Using these pairs to instruction fine-tune Flan T5 large with LoRA adapters yields a 28.5% relative WER reduction over a zero-shot Whisper-base.en baseline on the combined test set (30.5 to 21.8 WER), and a 13.3% relative reduction when applied to other fine-tuned ASR models. The paper claims CHSER is the first large-scale generative error-correction benchmark for child speech, and its error analysis shows the corrections reduce substitution and deletion errors but leave insertion errors and certain child disfluencies (repetitions, revisions) partly uncorrected.

Load-bearing premise

The ground-truth transcriptions in the four source corpora are treated as correct correction targets, including for spontaneous child speech with disfluent repetitions and revisions; if those transcripts contain recognition errors or inconsistent disfluency annotation, the measured WER reductions and error-type shifts would not reflect true correction quality.

Editorial extensions

If this is right

  • A 200K-pair child speech benchmark now exists for training and evaluating generative error-correction models.
  • Fine-tuning a text-only language model on paired hypotheses and transcripts can reduce child ASR word error rates without changing the acoustic model.
  • Multidomain training across scripted and spontaneous child speech yields better out-of-domain correction than training on one corpus alone.
  • Error-correction gains transfer to ASR systems of different architectures (encoder-decoder and CTC-based) that were fine-tuned on child speech.
  • Future correction methods must target insertion errors and disfluency categories like repetitions and revisions to close the remaining gap.

Reading between the lines

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

  • A natural extension would be injecting acoustic or prosodic cues into the correction model, since the paper's error analysis shows text-only correction misses child disfluencies.
  • The dataset's paired hypotheses could also serve as a supervised resource for disfluency detection and removal, not just error correction.
  • The filtering steps (top five hypotheses, three-word minimum, vocabulary membership) may bias the benchmark toward longer, cleaner utterances; evaluating on very short or highly disfluent child utterances would test whether the reported gains hold.
  • If disfluency-preserving transcriptions are desired (e.g., for clinical or educational analysis), the current WER metric may reward corrections that silently delete disfluencies; a disfluency-aware evaluation would be a stronger test.
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 / 4 minor

Summary. This paper introduces CHSER, a dataset of approximately 200k (ASR hypothesis, reference transcription) pairs for child speech, built from MyST, CMU Kids, CSLU OGI Kids, and OCSC using Whisper-base.en N-best hypotheses. The authors benchmark traditional LM rescoring, HyPoradise-pretrained and CHSER fine-tuned Llama 2 and Flan T5 models, and GPT-4o mini in-context learning, reporting WER reductions. The best result is a CHSER fine-tuned Flan T5 large model with a 28.5% relative WER reduction over the zero-shot Whisper baseline, and additional experiments address cross-domain robustness, generalization to different fine-tuned ASR architectures, and shifts in error types and disfluency patterns. The dataset, code, and models are publicly released.

Significance. The dataset fills a genuine gap: no large public generative speech error correction benchmark exists for child speech, and the paper's held-out test design, paired significance tests, and error-type analysis are appropriate. The central result is internally consistent and not circular, since the error-correction model is not fitted to the test targets. The value of the benchmark, however, depends on the reliability of the source-corpus transcriptions and on the transparency of the filtering pipeline; both issues are addressable in revision.

major comments (3)
  1. [Section 2 and Section 3.4] The paper treats the source-corpus transcriptions as ground truth without reporting any normalization, tokenization, or quality control. The spontaneous subsets (MyST, OGI Spontaneous, OCSC) are likely to use different conventions for marking filled pauses, partial words, repetitions, and restarts; if these markers are scored as ordinary tokens, the same acoustic event can be a correct token in one corpus and a deletion or substitution in another. Because the same transcripts serve as both training targets and evaluation references, the headline WER reductions in Tables 3 and 5 and the error-type and disfluency analyses in Figures 1 and 3 may partly reflect alignment to corpus-specific annotation conventions rather than genuine transcription quality. This is load-bearing for the benchmark contribution. Please provide (i) a sample audit of reference transcripts from each source corpus, (ii) an explicit statement of how disfluency tokens are handled in WER computation, and (iii) a sensitivity analysis showing that the reported relative WER reductions are stable when references are normalized to a common convention.
  2. [Section 2] The dataset construction filtering steps—removing repeated utterances, excluding transcriptions shorter than three words, and discarding utterances containing words outside the Whisper tokenizer vocabulary—are reported only as aggregate pair counts in Table 1. These filters can preferentially remove the hardest utterances (short child utterances, rare child-specific words, heavily disfluent turns), which would make the benchmark easier than the original corpora and bias the error analyses. For a dataset paper, the authors should report how many utterances and pairs were removed at each filtering step per source corpus, and ideally report baseline WER on the unfiltered test sets for at least MyST and OCSC to quantify the selection bias.
  3. [Section 3.3] The conclusion that GenSEC is robust "irrespective of the underlying ASR model" is broader than the evidence. Table 6 evaluates only the MyST domain, with all three alternative ASR models fine-tuned on MyST and a corrector also fine-tuned on the MyST subset of CHSER. This is a reasonable first demonstration, but the design does not support a claim of architecture-independent robustness across domains or across zero-shot ASR systems. Please qualify the conclusion or add cross-domain evidence, such as a corrector trained on the full CHSER set and evaluated on hypotheses from a different ASR model applied to a different child-speech corpus.
minor comments (4)
  1. [Abstract and Section 4] The phrase "zero-shot setting" is ambiguous: the 28.5% relative WER reduction is achieved by a fine-tuned error-correction model applied to hypotheses from a zero-shot ASR system (Whisper-base.en), not by a zero-shot corrector. Please rephrase to distinguish the ASR baseline from the error-correction model.
  2. [Section 3.1 and Table 3] Only the T5 FT results are marked as statistically significant in Table 3; the text should state which comparisons involving Llama FT and T5 Hyp are not significant rather than implying all fine-tuned models are significant. Confidence intervals or bootstrap error bars on the reported WERs would also improve interpretability.
  3. [Section 3.1] The LoRA fine-tuning details are incomplete: the paper specifies 8-bit quantization and five epochs but does not report LoRA rank, alpha, learning rate, batch size, or the instruction template used. These details are needed to reproduce the central result.
  4. [Section 3.4 and Figures 1-3] The error-type and disfluency figures report percentages without raw counts or confidence intervals, making it difficult to assess whether the observed shifts are reliable for small categories such as restarts or revisions. Please report counts and, where possible, significance tests for the before/after differences.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CHSER training and evaluation use disjoint splits, and the headline WER reductions are measured against held-out references rather than fitted targets.

full rationale

No significant circularity. The paper's central claim is that fine-tuning Flan T5 on CHSER hypothesis–transcription pairs reduces WER on held-out child-speech test sets (Table 3), with the largest relative reduction being 28.5% over the Whisper-base.en baseline. The paired training data are generated by running zero-shot Whisper-base.en beam search on source corpora (MyST, CMU Kids, OGI, OCSC) and pairing hypotheses with the corpora's reference transcriptions; train/dev/test splits follow the original source corpus splits, and the reported WER is computed on test pairs not seen during fine-tuning. The cross-domain RQ2 experiment trains only on the MyST subset and evaluates on other domains, so the generalization results are empirical rather than constructed. RQ3 applies a MyST-fine-tuned model to fresh hypotheses from independently fine-tuned Whisper-tiny, Whisper-small, and WavLM ASR systems; those hypotheses were not used to fit the correction model. RQ4's error-type, POS, and disfluency analyses are descriptive measurements on system outputs, not part of the training objective. The only self-citations are background references to prior work on child-ASR difficulty ([9], [10]); they motivate the study but are not load-bearing premises, and the paper invokes no uniqueness theorem or unverified result from the authors' own prior work. A possible concern that source-corpus reference transcriptions may be inconsistent or imperfect for spontaneous child speech is a data-quality and evaluation-validity risk, not circularity: the reported WER numbers could be inaccurate, but they are not true by construction.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No invented entities. The paper introduces a dataset, not a new theoretical construct. The central claim rests on dataset construction choices and the reliability of the underlying corpora.

free parameters (3)
  • Beam search temperature for Whisper-base.en = not reported
    Set to encourage diverse N-best hypotheses; determines the error distribution in the dataset.
  • LoRA rank and learning rate = not reported
    Fine-tuning hyperparameters chosen by hand; not derived from data.
  • Number of fine-tuning epochs = 5
    Chosen by hand; affects model performance.
assumptions (6)
  • domain assumption Whisper-base.en zero-shot hypotheses are representative of child ASR errors
    Dataset construction (Section 2) uses them as the sole source of hypothesis-transcription pairs.
  • domain assumption Original corpus transcriptions are accurate ground truth
    All WER computations in Section 3 compare corrected outputs against these transcripts.
  • domain assumption N-best list generation is a suitable framework for generative error correction
    Adopted from HyPoradise; underpins the dataset format and evaluation.
  • standard math WER and paired significance testing via NIST sctk are appropriate
    Used to compare models in Section 3.1.
  • domain assumption The disfluency annotation procedure applies to child speech
    Used in Section 3.4 for disfluency error analysis.
  • domain assumption spaCy POS tagging is reliable on child speech transcripts
    Used in Section 3.4 for syntactic error analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CHSER: A Dataset and Case Study on Generative Speech Error Correction for Child ASR." pith.science (2026). https://pith.science/paper/C4XCYB7C

@misc{pith2026250518463,
  author       = {Pith},
  title        = {Pith review of: CHSER: A Dataset and Case Study on Generative Speech Error Correction for Child ASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4XCYB7C}},
  note         = {Machine review of arXiv:2505.18463}
}
read the original abstract

Automatic Speech Recognition (ASR) systems struggle with child speech due to its distinct acoustic and linguistic variability and limited availability of child speech datasets, leading to high transcription error rates. While ASR error correction (AEC) methods have improved adult speech transcription, their effectiveness on child speech remains largely unexplored. To address this, we introduce CHSER, a Generative Speech Error Correction (GenSEC) dataset for child speech, comprising 200K hypothesis-transcription pairs spanning diverse age groups and speaking styles. Results demonstrate that fine-tuning on the CHSER dataset achieves up to a 28.5% relative WER reduction in a zero-shot setting and a 13.3% reduction when applied to fine-tuned ASR systems. Additionally, our error analysis reveals that while GenSEC improves substitution and deletion errors, it struggles with insertions and child-specific disfluencies. These findings highlight the potential of GenSEC for improving child ASR.

Figures

Figures reproduced from arXiv: 2505.18463 by the authors.

Figure 1
Figure 1. Comparison of error types (substitutions, deletions, inser￾tions) before and after error correction. Errors are categorized by com￾paring ASR hypotheses to ground truth transcriptions [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Syntactic errors categorized by part-of-speech (POS) tags (NOUN, PRON, VERB, DET, AUX) before and after error correction. Errors are identified by comparing ASR hypotheses to ground truth transcriptions [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Disfluency errors - filled pauses (FP), partial words (PW), repetitions (RP), revisions (RV), and restarts (RS), before and after er￾ror correction. Disfluencies are identified in the ground truth transcrip￾tions, ASR hypotheses, and corrected outputs to track correction rates. • Syntactic Errors: To evaluate whether the fine-tuned T5 model introduces syn￾tactic shifts, we categorize substitution-based errors by par… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages

  1. [1]

    CHSER: A Dataset and Case Study on Generative Speech Error Correction for Child ASR

    Introduction Speech Foundation Models (SFMs) have advanced Automatic Speech Recognition (ASR), demonstrating state-of-the-art per- formance and strong zero-shot generalization [1, 2, 3, 4, 5]. Their success stems from large-scale pretraining and transformer-based architectures, yet they struggle with child speech, exhibiting significantly high error rates...

    work page Pith review arXiv 2025

  2. [2]

    CHSER Dataset To construct a benchmark for child speech ASR error cor- rection, we compile a dataset from multiple publicly avail- able child speech corpora, ensuring diversity in age ranges and speaking styles. Our dataset consists of 200K ASR hypothesis-transcription pairs sourced from four major child speech datasets: MyST [30], CMU Kids [31], CSLU OGI...

    work page

  3. [3]

    Experimental Setup and Results 3.1. RQ1: Do existing GenSEC methods perform well on child speech? Table 2:WER of Traditional Reranking Models (Trigram LM, Trans- former LM) and oracle WER (N-best Oracle (onb), and Compositional Oracle (o cp)) on the CHSER dataset. Baseline refers to WER before error correction. Dataset Baseline Trigram Transformero nb ocp...

    work page

  4. [4]

    Conclusion This study presents CHSER, a Generative Speech Error Correction (GenSEC) dataset comprising 200K hypothesis- transcription pairs spanning diverse age groups and speaking styles specifically designed to address the challenges of ASR transcription for child speech. Through a comprehensive eval- uation, we systematically investigated four key rese...

    work page

  5. [5]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chenet al., “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  6. [6]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsuet al., “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 29, pp. 3451–3460, 2021

    work page 2021

  7. [7]

    Wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevskiet al., “Wav2vec 2.0: A framework for self-supervised learning of speech representations,”Advances in Neural Informa- tion Processing Systems, vol. 33, pp. 12 449–12 460, 2020

    work page 2020

  8. [8]

    Robust speech recognition via large-scale weak supervision,

    A. Radfordet al., “Robust speech recognition via large-scale weak supervision,” inInternational Conference on Machine Learning, ICML 2023, vol. 202. PMLR, 2023, pp. 28 492–28 518

    work page 2023

Show all 47 references
  1. [9]

    Less is more: Accurate speech recognition & translation without web-scale data,

    K. C. Puvvadaet al., “Less is more: Accurate speech recognition & translation without web-scale data,” inProc. Interspeech 2024, 2024

    2024

  2. [10]

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

    R. Fanet al., “Benchmarking children’s asr with supervised and self-supervised speech foundation models,” inProc. Interspeech 2024, 2024

    2024

  3. [11]

    Analyzing children’s speech: An acoustic study of consonants and consonant-vowel transition,

    M. Gerosaet al., “Analyzing children’s speech: An acoustic study of consonants and consonant-vowel transition,” in2006 IEEE In- ternational Conference on Acoustics Speech and Signal Process- ing Proceedings, vol. 1, 2006

    2006

  4. [12]

    Acoustics of children’s speech: Developmental changes of temporal and spectral parameters,

    S. Leeet al., “Acoustics of children’s speech: Developmental changes of temporal and spectral parameters,”The Journal of the Acoustical Society of America, vol. 105, no. 3, pp. 1455–1468, 03 1999

    1999

  5. [13]

    On the difficulties of automatic speech recognition for kindergarten-aged children,

    G. Yeung and A. Alwan, “On the difficulties of automatic speech recognition for kindergarten-aged children,” inProceedings of In- terspeech 2018. Hyderabad, India: ISCA, 2018, pp. 1661–1665

    2018

  6. [14]

    Challenges remain in building ASR for sponta- neous preschool children speech in naturalistic educational envi- ronments,

    S. Duttaet al., “Challenges remain in building ASR for sponta- neous preschool children speech in naturalistic educational envi- ronments,” inProceedings of Interspeech 2022. Incheon, Korea: ISCA, 2022, pp. 2706–2710

    2022

  7. [15]

    Childhood stuttering and speech disfluencies in relation to children’s mean length of utter- ance: A preliminary study,

    C. T. Zackheim and E. G. Conture, “Childhood stuttering and speech disfluencies in relation to children’s mean length of utter- ance: A preliminary study,”Journal of Fluency Disorders, vol. 28, no. 2, pp. 115–142, 2003

    2003

  8. [16]

    Estimation of probabilities from sparse data for the lan- guage model component of a speech recognizer,

    S. Katz, “Estimation of probabilities from sparse data for the lan- guage model component of a speech recognizer,”IEEE transac- tions on acoustics, speech, and signal processing, vol. 35, no. 3, pp. 400–401, 2003

    2003

  9. [17]

    Finding consensus in speech recognition: word error minimization and other applications of confusion networks,

    L. Manguet al., “Finding consensus in speech recognition: word error minimization and other applications of confusion networks,” Computer Speech & Language, vol. 14, no. 4, pp. 373–400, 2000

    2000

  10. [18]

    Recurrent neural network based language model

    T. o. Mikolov, “Recurrent neural network based language model.” inInterspeech, vol. 2, no. 3. Makuhari, 2010, pp. 1045–1048

    2010

  11. [19]

    Neural error corrective language models for au- tomatic speech recognition,

    T. Tanakaet al., “Neural error corrective language models for au- tomatic speech recognition,” inInterspeech 2018, 2018, pp. 401– 405

    2018

  12. [20]

    Asr error correction and domain adaptation using machine translation,

    A. Maniet al., “Asr error correction and domain adaptation using machine translation,” in2020 IEEE International Conference on Acoustics, Speech and Signal Processing), 2020, pp. 6344–6348

    2020

  13. [21]

    Improving readability for automatic speech recog- nition transcription,

    J. Liaoet al., “Improving readability for automatic speech recog- nition transcription,”ACM Trans. Asian Low-Resour. Lang. Inf. Process., vol. 22, no. 5, May 2023

    2023

  14. [22]

    Multi-modal asr error correction with joint asr error detection,

    B. Lin and L. Wang, “Multi-modal asr error correction with joint asr error detection,” in2023 IEEE International Conference on Acoustics, Speech and Signal Processing, 2023, pp. 1–5

    2023

  15. [23]

    Cross-modal asr post-processing system for error correction and utterance rejection,

    J. Duet al., “Cross-modal asr post-processing system for error correction and utterance rejection,”arXiv preprint arXiv:2201.03313, 2022

    2022 arXiv

  16. [24]

    Hyporadise: An open baseline for generative speech recognition with large language models,

    C. Chenet al., “Hyporadise: An open baseline for generative speech recognition with large language models,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  17. [25]

    Large language models are efficient learners of noise-robust speech recognition,

    Y . Huet al., “Large language models are efficient learners of noise-robust speech recognition,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  18. [26]

    Lora: Low-rank adaptation of large language models,

    E. J. Huet al., “Lora: Low-rank adaptation of large language models,”International Conference on Learning Representations, 2021

    2021

  19. [27]

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

    H. Touvronet al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  20. [28]

    Scaling instruction-finetuned language mod- els,

    H. W. Chunget al., “Scaling instruction-finetuned language mod- els,”Journal of Machine Learning Research, vol. 25, no. 70, pp. 1–53, 2024

    2024

  21. [29]

    Flanec: Exploring flan-t5 for post-asr error correction,

    M. La Quatraet al., “Flanec: Exploring flan-t5 for post-asr error correction,” in2024 IEEE Spoken Language Technology Work- shop (SLT). IEEE, 2024, pp. 608–615

    2024

  22. [30]

    Whispering LLaMA: A cross-modal generative error correction framework for speech recognition,

    S. Radhakrishnanet al., “Whispering LLaMA: A cross-modal generative error correction framework for speech recognition,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Dec. 2023

    2023

  23. [31]

    Asr error correction using large language models,

    R. Maet al., “Asr error correction using large language models,” arXiv preprint arXiv:2409.09554, 2024

    2024 arXiv

  24. [32]

    Failing forward: Improving generative error cor- rection for asr with synthetic data and retrieval augmentation,

    S. Ghoshet al., “Failing forward: Improving generative error cor- rection for asr with synthetic data and retrieval augmentation,” arXiv preprint arXiv:2410.13198, 2024

    2024 arXiv

  25. [33]

    Neko: Toward post recognition generative cor- rection large language models with task-oriented experts,

    Y .-T. Linet al., “Neko: Toward post recognition generative cor- rection large language models with task-oriented experts,”arXiv preprint arXiv:2411.05945, 2024

    2024

  26. [34]

    My science tutor (MyST)–a large corpus of children‘s conversational speech,

    S. Pradhanet al., “My science tutor (MyST)–a large corpus of children‘s conversational speech,” inProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), May 2024, pp. 12 040–12 045

    2024

  27. [35]

    The cmu kids corpus,

    M. Eskenaziet al., “The cmu kids corpus,”Linguistic Data Con- sortium, vol. 11, 1997

    1997

  28. [36]

    The ogi kids’ speech corpus and recognizers,

    K. Shobakiet al., “The ogi kids’ speech corpus and recognizers,” Proc. of ICSLP, pp. 564–567, 2000

    2000

  29. [37]

    The ohio child speech corpus,

    L. Wagneret al., “The ohio child speech corpus,”Available at SSRN 4846086, 2024

    2024

  30. [38]

    Kids: A database of children’s speech,

    M. S. Eskenazi, “Kids: A database of children’s speech,”The Journal of the Acoustical Society of America, vol. 100, no. 4, pp. 2759–2759, 1996

    1996

  31. [39]

    Generative speech recognition error correc- tion with large language models and task-activating prompting,

    C.-H. H. Yanget al., “Generative speech recognition error correc- tion with large language models and task-activating prompting,” in2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  32. [40]

    SCTK: The NIST Scoring Toolkit,

    J. Fiscus, “SCTK: The NIST Scoring Toolkit,” National Institute of Standards and Technology, 2007, [Software]

    2007

  33. [41]

    Librispeech: An ASR corpus based on pub- lic domain audio books,

    V . Panayotovet al., “Librispeech: An ASR corpus based on pub- lic domain audio books,”2015 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 5206–5210, 2015

    2015

  34. [42]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacobet al., “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 2704–2713

    2018

  35. [43]

    An explanation of in-context learning as im- plicit bayesian inference,

    S. M. Xieet al., “An explanation of in-context learning as im- plicit bayesian inference,” inInternational Conference on Learn- ing Representations, 2022

    2022

  36. [44]

    Gpt-4o system card,

    A. Hurstet al., “Gpt-4o system card,”arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  37. [45]

    Kid-whisper: Towards bridging the per- formance gap in automatic speech recognition for children vs. adults,

    A. A. Attiaet al., “Kid-whisper: Towards bridging the per- formance gap in automatic speech recognition for children vs. adults,” inProceedings of the AAAI/ACM Conference on AI, Ethics, and Society, vol. 7, 2024, pp. 74–80

    2024

  38. [46]

    spacy: Industrial-strength natural language processing in python,

    M. Honnibalet al., “spacy: Industrial-strength natural language processing in python,” 2020, available at https://spacy.io/

    2020

  39. [47]

    Automatic disfluency detection from untran- scribed speech,

    A. Romanaet al., “Automatic disfluency detection from untran- scribed speech,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

Pith tools

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