Pith. sign in

REVIEW 4 major objections 4 minor 36 references

Adapting Whisper for Lightweight and Efficient Automatic Speech Recognition of Children for On-device Edge Applications

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A fine-tuned Whisper tiny.en reaches 15.9% word error rate on children's speech and runs in real time on a Raspberry Pi.

desk verdict Solid, honest empirical study of Whisper tiny.en on children's speech with useful edge measurements; the 'real-time' claim rests on only three long utterances and needs tempering. read the letter →

arxiv 2507.14451 v1 pith:IWMOV5MT submitted 2025-07-19 eess.AS cs.HCcs.SD

Satwik Dutta , Shruthigna Chandupatla , John Hansen This is my paper · ORCID
classification eess.AScs.HCcs.SD
keywords childspeechrecognitionWhisperlightweightASRefficienton-devicetranscriptionlow-rankcompressionRaspberryPiMySTcorpus
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

What the paper tries to establish is that the smallest Whisper model, fine-tuned on a children's speech corpus, can be accurate and fast enough to run entirely on a low-cost edge device, so voice data never has to leave the device. The authors report a 15.9% word error rate on the original MyST test set (11.8% on a filtered version) and show that low-rank compression of the encoder removes 0.51M parameters, cuts about 2 GFLOPS on a Raspberry Pi, and speeds up GPU inference about 1.26 times while raising WER by roughly 11% relative. On-device inference runs faster than real time, with real-time factors between 0.23 and 0.41, and the tiny models keep RAM under 15% and CPU temperature below the Pi's throttling range. The motivation is that child-centered voice products face privacy regulations that make cloud transcription risky, and a local ASR backend sidesteps that exposure. If the claim holds, on-device transcription for children could be delivered with an off-the-shelf 39M-parameter model rather than a large cloud service.

What carries the argument

The load-bearing mechanism is the Whisper tiny.en end-to-end transformer, adapted by fine-tuning on four data versions of the MyST corpus and then compressed by a low-rank approximation of the encoder's self-attention and MLP weight matrices, calibrated on 500 samples with a 0.999 cumulative-energy threshold. The filtering strategies F1 (drop utterances with more than 50% WER under Whisper Large V2), F2 (drop utterances with fewer than three words), and F3 (concatenate utterances into 25-30 second clips) form the second mechanism, because they determine which utterances the model sees and thereby which tiny.en variant performs best. The third mechanism is the 30-second receptive field of Whisper, which makes the interaction between utterance length and padding behavior central to the result.

What would settle it

Manually correct the ground-truth transcriptions and remove clipped or silent utterances from the original MyST test set, then re-run the fine-tuned tiny.en model; if the corrected WER is much closer to the 11.8% filtered number than to 15.9%, the headline accuracy is an artifact of test-set noise.

Watch

Extended reading notes

Core claim

The paper's central claim is that a fine-tuned Whisper tiny.en is the smallest and best lightweight model for the MyST corpus, reaching a 15.9% test WER on the original split and 11.8% on the filtered split, and that a low-rank-compressed version of this model is still accurate enough for real-time edge use. Fine-tuning narrows the gap to larger models: relative to zero-shot, fine-tuning cuts WER by 55% for tiny.en, 49% for base.en, and 36% for small.en. The filtering experiments show that discarding utterances with high WER under Whisper Large V2 and utterances with fewer than three words helps short utterances, whereas training on concatenated 25-30 second clips does not. Compressing the fine-tuned encoder from 7.63M to 7.08-7.12M parameters raises WER by 11.4% (for the B-trained variant) to 19.3% (for the A-trained variant) relative and shortens GPU inference time by a factor of 0.79. On the Raspberry Pi, the compressed tiny models use about 2 GFLOPS less than the uncompressed one, keep RTF between 0.25 and 0.41, and stay within the Pi's thermal and memory limits, while the larger small.en model triggers thermal throttling.

Load-bearing premise

The load-bearing premise is that the original MyST test set, despite the transcription and audio-quality problems the authors found in it, is a fair benchmark for comparing models; if those problems skew WER unevenly across models, the headline 15.9% and the compression comparison would shift.

Editorial extensions

If this is right

  • A 39M-parameter model can deliver 15.9% WER on children's conversational speech, so the accuracy gap that previously required much larger ASR models may be mostly closed by targeted fine-tuning.
  • On-device inference with RTF between 0.23 and 0.41 means a Raspberry Pi can transcribe children's speech in real time without cloud connectivity, which directly addresses privacy constraints like COPPA.
  • Low-rank compression of the encoder gives a concrete trade-off: about 11% relative WER increase for a 0.51M parameter reduction and roughly 2 GFLOPS fewer per utterance on the edge device.
  • Training on natural-length utterances rather than 25-30 second concatenated clips better preserves accuracy on the short utterances that dominate the MyST corpus.
  • The small.en model's thermal throttling on the Pi marks a practical upper bound on model size for passively cooled edge devices.

Reading between the lines

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

  • If the MyST test set's transcription and audio errors were corrected, the fine-tuned tiny.en's true error rate is likely closer to the 11.8% filtered number than to the headline 15.9%, so the authors' primary benchmark probably understates the model's real accuracy.
  • The same fine-tuning plus low-rank compression recipe could transfer to younger children (preschoolers) whose speech differs more from adult speech, though the MyST corpus only covers grades 3-5.
  • The reported real-time margin depends on the specific Raspberry Pi 5 configuration with a cooling fan; on cheaper or passively cooled boards the thermal throttling seen for small.en might also affect the tiny models during sustained use.
  • Because the F1 filter removes utterances that Whisper Large V2 already transcribes poorly, the filtering step may bias the training data toward utterances that fit adult-like speech patterns, potentially capping the ceiling for child-specific adaptation.
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 adapts OpenAI Whisper models, specifically tiny.en, for children's speech recognition on the MyST corpus, with the goal of achieving a lightweight, privacy-preserving on-device ASR system for a Raspberry Pi. The authors evaluate several data filtering strategies (removing poor transcriptions, removing short utterances, and concatenating to 25--30 second segments), fine-tune Whisper models on the resulting data versions, report WER on the original and filtered MyST test sets, apply a low-rank compression scheme to the encoder, and measure GFLOPS, real-time factor (RTF), RAM usage, and CPU temperature on a Raspberry Pi 5. The headline results are a 15.9% WER on the original test set and 11.8% on the filtered test set for fine-tuned tiny.en, a 0.51M parameter reduction with an 11.36% relative WER increase after compression, and RTF values between 0.23 and 0.41 on three test utterances.

Significance. If the claims hold, the paper offers a useful empirical datapoint for building privacy-preserving child ASR on low-cost edge hardware, and it compares favorably with prior work on MyST. The strengths are that the WER numbers are measured on a held-out test set rather than fitted, the training/evaluation recipes and test splits are promised to be released on GitHub, and the authors report both original and filtered test results, which is more transparent than most corpus studies. The efficiency comparison with low-rank compression is also relevant. The main limitations are that the central on-device real-time claim rests on only three utterances, the RTF measurements lack confidence intervals, and the 'faster inference' claim is not consistently supported by the Pi measurements.

major comments (4)
  1. [Section 3.2, Table 4] The claim that the system is 'practical, real-time on-device' is supported only by RTF measurements on three utterances of 9.00s, 14.28s, and 29.73s. The paper itself states in Section 2 that the majority of MyST utterances are under ten seconds, yet Table 4 shows RTF decreasing monotonically with duration (0.39, 0.36, 0.23 for fine-tuned tiny.en-A), which indicates a fixed overhead that penalizes short inputs. Extrapolating to typical 2--3 second child utterances could easily push RTF above 1.0. To support the central claim, the authors should evaluate on a representative sample covering the corpus duration distribution and report RTF with confidence intervals per duration bin.
  2. [Section 3.3.1, Table 2 and Table 3] The paper acknowledges transcription and audio quality errors in MyST and constructs a filtered version B(Fil), but then selects A(Org) as the primary benchmark because it is deemed 'more robust and challenging.' If the corpus errors systematically inflate WER, the headline 15.9% figure and the compression comparison in Table 3, which uses A(Org), may not reflect true system accuracy. The authors should justify this preference with quantitative evidence about the nature and frequency of errors, and report both test sets consistently with the number of utterances in each, so readers can judge the sensitivity of the conclusions to the filtering choice.
  3. [Section 4.0.1, Table 3 vs. Section 4.1.2, Table 4] The abstract and conclusion state that low-rank compression yields '1.26x faster inference,' which is derived from the normalized time of 0.79 in Table 3 on GPU. However, the Raspberry Pi RTF measurements in Table 4 do not consistently support faster inference: for the 29.73s utterance, compressed tiny.en-A is actually slower than uncompressed tiny.en-A (0.25 vs. 0.23), and compressed tiny.en-B is not consistently faster than uncompressed tiny.en-A either. The 'faster inference' claim must be qualified by platform and input duration, and the conditions under which the 1.26x speedup was obtained should be stated explicitly.
  4. [Section 3.2, Table 4] The paper says the Raspberry Pi analyses were repeated ten times, but Table 4 reports only single RTF values and states that variance is shown on GitHub. For a load-bearing efficiency comparison, in-text reporting of the mean and standard deviation (or at least a range) is necessary, and the same applies to WER numbers throughout Section 3.3, which are given without confidence intervals. Without this information, it is impossible to tell whether the reported differences between models or compression variants are within run-to-run noise.
minor comments (4)
  1. [Table 2] There is a typo in the model size column: 'base 74MM' should be 'base 74M'.
  2. [Section 3] The subsection heading 'Model Training and Evaluation (on GPU)' appears twice (Sections 3.1 and 3.3), which is likely an editorial error and should be fixed.
  3. [Figure 5] The y-axis labels for RAM usage and CPU temperature are not visible in the figure caption; please ensure the axes are legible and clearly state whether RAM usage is a percentage of total memory.
  4. [Section 2, filtering strategies] The description of F1 says utterances with more than 50% WER are discarded, but it is not immediately clear whether the WER is computed against the provided ground truth or against a reference transcription produced by Whisper Large V2; please clarify this in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all headline results are independent empirical measurements against the MyST corpus, with self-citations only in background.

full rationale

No circularity found. The core results are empirical measurements against a fixed, external benchmark (the MyST corpus): fine-tuned Whisper tiny.en achieves a 15.9% WER on the original A(Org) test set, and the low-rank compression results (WER increase, parameter reduction, inference speedup) are measured quantities reported in Tables 3 and 4. The only place a model is used to alter data is the F1 filtering step, where Whisper Large V2 is used to discard training/development utterances with more than 50% WER; this is a data-cleaning step and does not define the evaluation, because the headline WER is reported on the unfiltered A(Org) test set. No parameter is fitted to the test set and then reported as a prediction. The model selection uses test-set WER to choose among variants, which is a methodological concern about test-set reuse, not a circular reduction covered by the defined patterns. The self-citations [6, 7, 11, 12] are background citations and are not load-bearing for the central derivation. The low-rank compression method is imported from external work [29] and independently evaluated. The derivation chain is self-contained, so the appropriate circularity score is 0.

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

The paper does not derive a mathematical result; it reports measurements. The free parameters are hand-chosen experimental thresholds and compression settings, not fitted model parameters. The axioms are standard domain assumptions about the corpus, pretrained models, and generalization of small-sample measurements.

free parameters (4)
  • F1 WER filtering threshold = 50%
    Utterances with more than 50% WER under Whisper Large V2 are discarded from training and development splits; chosen by hand after manual inspection, not optimized on the test set.
  • F2 minimum word count = 3 words
    Utterances with fewer than 3 words are discarded to preserve context for homophone disambiguation; chosen by hand as a reasonable threshold.
  • Low-rank compression threshold = 0.999
    The energy threshold for retaining components in the low-rank compression is set to 0.999 for both self-attention and MLP layers, following the prior LiteASR method.
  • Calibration set size = 500 samples
    The compressed models are calibrated with 500 samples; the selection process for these samples is not detailed, which could affect the compressed model evaluation.
assumptions (4)
  • domain assumption OpenAI Whisper pretrained weights are a valid starting point for child speech recognition.
    The paper relies on transfer learning from adult-oriented Whisper models without validating this assumption independently.
  • domain assumption MyST ground truth transcriptions are sufficiently accurate for evaluation of the original test set.
    The authors themselves report transcription errors in MyST (Section 2), yet use A(Org) as the primary benchmark; if errors are frequent, the WER values may be inflated.
  • domain assumption Low-rank compression from LiteASR [29] preserves the trained model's behavior as described.
    The paper applies the compression method from prior work and measures its effect, but does not re-derive the method's guarantees.
  • domain assumption Measurements on three utterances generalize to typical usage of the MyST corpus and the Raspberry Pi setup.
    The on-device efficiency and recording quality experiments use only 3 randomly chosen test utterances; the paper does not establish statistical significance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adapting Whisper for Lightweight and Efficient Automatic Speech Recognition of Children for On-device Edge Applications." pith.science (2026). https://pith.science/paper/IWMOV5MT

@misc{pith2026250714451,
  author       = {Pith},
  title        = {Pith review of: Adapting Whisper for Lightweight and Efficient Automatic Speech Recognition of Children for On-device Edge Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IWMOV5MT}},
  note         = {Machine review of arXiv:2507.14451}
}
read the original abstract

Reliability on cloud providers for ASR inference to support child-centered voice-based applications is becoming challenging due to regulatory and privacy challenges. Motivated by a privacy-preserving design, this study aims to develop a lightweight & efficient Whisper ASR system capable of running on a Raspberry Pi. Upon evaluation of the MyST corpus and by examining various filtering strategies to fine-tune the `tiny.en' model, a Word Error Rate (WER) of 15.9% was achieved (11.8% filtered). A low-rank compression reduces the encoder size by 0.51M with 1.26x faster inference in GPU, with 11% relative WER increase. During inference on Pi, the compressed version required ~2 GFLOPS fewer computations. The RTF for both the models ranged between [0.23-0.41] for various input audio durations. Analyzing the RAM usage and CPU temperature showed that the PI was capable of handling both the tiny models, however it was noticed that small models initiated additional overhead/thermal throttling.

Figures

Figures reproduced from arXiv: 2507.14451 by the authors.

Figure 1
Figure 1. Raspberry PI and accessories [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of A(Org) Test WER by audio duration [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Comparison of WER (%) for the tiny.en-‘A’ model: Original vs. Recorded. NIST STNR [30] and the WER based on the best lightweight model output were compared with that of the original input au￾dio. The recorded audio was downsampled to 16 kHz, simi￾lar to the original input audio. For each model and input audio pair, we calculate the Transcription Processing Time or Latency. For our case, this also includes feature ex… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: RAM Usage Percentage and CPU Temperature (this includes pre-processing steps like loading of model, audio, etc.) [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 35 canonical work pages

  1. [1]

    This tech- nology is at the core of various ‘AI-based products’ for chil- dren, at home or in the classroom

    Introduction The focus on developing Automatic Speech Recognition (ASR) systems for children has accelerated over the years, given the enormous benefit of its application in educational technology [1, 2, 3], spoken language learning [4], games/toys/robots [5], and speech/language development [6, 7] of children. This tech- nology is at the core of various ...

    work page

  2. [2]

    Adapting Whisper for Lightweight and Efficient Automatic Speech Recognition of Children for On-device Edge Applications

    MyST: The My Science Tutor Children’s Conversational Speech Corpus The ‘My Science Tutor Children’s Conversational Speech Cor- pus’ or MyST corpus [18], contains approximately 210 hrs of transcribed children’s utterances (470 hrs in total) during con- versation between 1371 students in grades 3-5 (ages 8-11) and a virtual science tutor on various topics s...

    work page Pith review arXiv 2025

  3. [3]

    Model Training and Evaluation (on GPU) All training (fine-tuning) and evaluation recipes were adapted from HuggingFace and OpenAI

    Experimental Details 3.1. Model Training and Evaluation (on GPU) All training (fine-tuning) and evaluation recipes were adapted from HuggingFace and OpenAI. For zero-shot evaluation, all models were downloaded from OpenAI’s Hugging Face model checkpoints [26]. These include both English and multilingual versions of tiny, base, small, and medium, and versi...

    work page

  4. [4]

    Evaluation of Low-Rank Compression All results for this section are shown in Tab.3

    Results and Discussion 4.0.1. Evaluation of Low-Rank Compression All results for this section are shown in Tab.3. By comparing OpenAI’s multilingual tiny model with the compressed version [29], the relative test WER increases by 8.24% for a reduction of 0.22M encoder parameters and quicker inference time on GPU. Particularly for fine-tunedtiny.en, two com...

    work page

  5. [5]

    Our lightweight fine-tuned ‘tiny.en’ model achieves a 15.9% WER for the MyST corpus

    Conclusion In this study, we developed a lightweight and efficient Whis- per ASR model for running on-device on a Raspberry Pi. Our lightweight fine-tuned ‘tiny.en’ model achieves a 15.9% WER for the MyST corpus. It also demonstrates our analysis on var- ious filtering strategies, with a conclusion that using only 25- 30s utterances for training/developme...

    work page 2025

  6. [6]

    Acknowledgements This work is supported by NSF Grants 1918032, 2234916, and 2341384; and HPC resources provided by the Texas Advanced Computing Center (TACC)

    work page

  7. [7]

    Lalilo: A Reading Assis- tant for Children Featuring Speech Recognition-Based Reading Mistake Detection,

    C. Hembise, L. Gelin, and M. Daniel, “Lalilo: A Reading Assis- tant for Children Featuring Speech Recognition-Based Reading Mistake Detection,” in Proc. Interspeech 2021, 2021, pp. 4860– 4861

    work page 2021

  8. [8]

    End- to-end evaluation of a spoken dialogue system for learning basic mathematics,

    E. Okur, S. Sahay, R. Fuentes Alba, and L. Nachman, “End- to-end evaluation of a spoken dialogue system for learning basic mathematics,” in Proceedings of the 1st Workshop on Mathematical Natural Language Processing (MathNLP) , D. Ferreira, M. Valentino, A. Freitas, S. Welleck, and M. Schubotz, Eds. Abu Dhabi, United Arab Emirates (Hybrid): Association fo...

    work page 2022

Show all 36 references
  1. [9]

    SoapBox Labs Fluency Assessment Platform for Child Speech,

    A. C. Kelly, E. Karamichali, A. Saeb, K. Vesel ´y, N. Parslow, G. M. Gomez, A. Deng, A. Letondor, N. Mullally, A. Hempel, R. O’Regan, and Q. Zhou, “SoapBox Labs Fluency Assessment Platform for Child Speech,” in Proc. Interspeech 2020, 2020, pp. 488–489

    2020

  2. [10]

    Using Commercial ASR Solutions to As- sess Reading Skills in Children: A Case Report,

    T. Piton, E. Hermann, A. Pasqualotto, M. Cohen, M. Magimai.- Doss, and D. Bavelier, “Using Commercial ASR Solutions to As- sess Reading Skills in Children: A Case Report,” inProc. INTER- SPEECH 2023, 2023, pp. 4573–4577

    2023

  3. [11]

    Child speech recog- nition in human-robot interaction: evaluations and recommenda- tions,

    J. Kennedy, S. Lemaignan, C. Montassier, P. Lavalade, B. Irfan, F. Papadopoulos, E. Senft, and T. Belpaeme, “Child speech recog- nition in human-robot interaction: evaluations and recommenda- tions,” in Proceedings of the 2017 ACM/IEEE international con- ference on human-robot...

    2017

  4. [12]

    Challenges remain in building asr for spontaneous preschool children speech in naturalistic educa- tional environments,

    S. Dutta, S. A. Tao, J. C. Reyna, R. E. Hacker, D. W. Irvin, J. F. Buzhardt, and J. H. Hansen, “Challenges remain in building asr for spontaneous preschool children speech in naturalistic educa- tional environments,” in Interspeech 2022, 2022, pp. 4322–4326

    2022

  5. [13]

    Activ- ity focused speech recognition of preschool children in early childhood classrooms,

    S. Dutta, D. Irvin, J. Buzhardt, and J. H. Hansen, “Activ- ity focused speech recognition of preschool children in early childhood classrooms,” in Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022) . Seattle, Washington: ...

    2022

  6. [14]

    F. T. Commission et al., “Ftc staff report finds large social media and video streaming companies have engaged in vast surveillance of users with lax privacy controls and inadequate safeguards for kids and teens,” 2024

    2024

  7. [15]

    Case proceedings,

    ——, “Case proceedings,” www.ftc.gov/enforcement/ cases-proceedings/terms/875, accessed: 2025-05-29

    2025

  8. [16]

    Amazon.com (alexa), u.s. v

    ——, “Amazon.com (alexa), u.s. v.” www.ftc.gov/legal-library/ browse/cases-proceedings/192-3128-amazoncom-alexa-us-v, accessed: 2025-05-29

    2025

  9. [17]

    Navigating the united states legislative landscape on voice privacy: Existing laws, proposed bills, protec- tion for children, and synthetic data for ai,

    S. Dutta and J. H. Hansen, “Navigating the united states legislative landscape on voice privacy: Existing laws, proposed bills, protec- tion for children, and synthetic data for ai,” in Synthetic Data’s Transformative Role in Foundational Speech Models , 2024, pp. 91–95

    2024

  10. [18]

    Exploring discrete speech units for privacy-preserving and efficient speech recognition for school-aged and preschool children,

    S. Dutta, D. Irvin, and J. H. Hansen, “Exploring discrete speech units for privacy-preserving and efficient speech recognition for school-aged and preschool children,” International Journal of Human-Computer Studies , vol. 199, p. 103460, 2025. [On- line]. Available: https://w...

    2025

  11. [19]

    On-device streaming transformer-based end-to-end speech recognition

    Y . R. Oh and K. Park, “On-device streaming transformer-based end-to-end speech recognition.” in Interspeech, 2021, pp. 967– 968

    2021

  12. [20]

    Speech understanding on tiny devices with a learning cache,

    A. Benazir, Z. Xu, and F. X. Lin, “Speech understanding on tiny devices with a learning cache,” in Proceedings of the 22nd An- nual International Conference on Mobile Systems, Applications and Services, ser. MOBISYS ’24. New York, NY , USA: Asso- ciation for Computing Machiner...

    2024

  13. [21]

    General data protection regulation,

    E. Union, “General data protection regulation,” https://gdpr.eu/, accessed: 2025-05-29

    2025

  14. [22]

    Uk’s data protection legislation,

    G. of the United Kingdom, “Uk’s data protection legislation,” https://www.gov.uk/data-protection, accessed: 2025-05-29

    2025

  15. [23]

    Acoustic variability and automatic recognition of children’s speech,

    M. Gerosa, D. Giuliani, and F. Brugnara, “Acoustic variability and automatic recognition of children’s speech,” Speech Communica- tion, vol. 49, no. 10-11, pp. 847–860, 2007

    2007

  16. [24]

    My science tutor and the myst corpus,

    W. Ward, R. Cole, and S. Pradhan, “My science tutor and the myst corpus,” Boulder Learning Inc, 2019

    2019

  17. [25]

    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

  18. [26]

    Adaptation of whisper models to child speech recognition,

    R. Jain, A. Barcovschi, M. Yiwere, P. Corcoran, and H. Cucu, “Adaptation of whisper models to child speech recognition,” in Interspeech 2023, 2023, pp. 5242–5246

    2023

  19. [27]

    Kid-whisper: Towards bridging the performance gap in auto- matic speech recognition for children vs. adults,

    A. A. Attia, J. Liu, W. Ai, D. Demszky, and C. Espy-Wilson, “Kid-whisper: Towards bridging the performance gap in auto- matic speech recognition for children vs. adults,” in Proceedings of the 2024 AAAI/ACM Conference on AI, Ethics, and Society, ser. AIES ’24. AAAI Press, 2025...

    2024

  20. [28]

    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

  21. [29]

    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 Proceedings of the 40th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, A. Krause, E. Brunsk...

    2023

  22. [30]

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

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet 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

  23. [31]

    Whisper tokenizer,

    Huggingface, “Whisper tokenizer,” https://huggingface. co/docs/transformers/en/model doc/whisper#transformers. WhisperTokenizer, accessed: 2025-05-29

    2025

  24. [32]

    ——, “openai,” https://huggingface.co/openai, accessed: 2025- 05-29

    2025

  25. [33]

    Distil-whisper: Robust knowledge distillation via large-scale pseudo labelling,

    S. Gandhi, P. von Platen, and A. M. Rush, “Distil-whisper: Robust knowledge distillation via large-scale pseudo labelling,” arXiv preprint arXiv:2311.00430, 2023

    2023 arXiv

  26. [34]

    Why is distil-small.en slower than distil- large-v2?

    Huggingface, “Why is distil-small.en slower than distil- large-v2?” https://huggingface.co/distil-whisper/distil-small. en#why-is-distil-smallen-slower-than-distil-large-v2, accessed: 2025-05-29

    2025

  27. [35]

    Liteasr: Effi- cient automatic speech recognition with low-rank approximation,

    K. Kamahori, J. Kasai, N. Kojima, and B. Kasikci, “Liteasr: Effi- cient automatic speech recognition with low-rank approximation,” arXiv preprint arXiv:2502.20583, 2025

    2025 arXiv

  28. [36]

    Speech signal to noise ratio measurements,

    NIST, “Speech signal to noise ratio measurements,” www.nist. gov/itl/iad/mig/nist-speech-signal-noise-ratio-measurements, 2016, accessed: 2025-05-29

    2016

Pith tools

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