Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

BabyHuBERT: Multilingual Self-Supervised Learning for Segmenting Speakers in Child-Centered Long-Form Recordings

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Pre-training on 13,000 hours of children's recordings lets a speech model identify who is speaking in daylong audio at near-human accuracy.

desk verdict A useful multilingual child-centered SSL model with a real but fixable evaluation flaw: the headline numbers come from selecting the best run on the test set. read the letter →

arxiv 2509.15001 v3 pith:FWTTS7TE submitted 2025-09-18 eess.AS cs.LGcs.SD

classification eess.AScs.LGcs.SD
keywords self-supervisedlearningchild-centeredrecordingsvoicetypeclassificationspeakersegmentationmultilingualspeechHuBERTlong-formaudiolanguagedevelopment
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 tries to show that self-supervised pre-training on domain-matched, multilingual child-centered audio solves the speaker-segmentation problem for naturalistic daylong recordings. It introduces BabyHuBERT, trained on 13,164 hours of audio from 19 corpora spanning more than 40 languages, and reports that fine-tuning it for voice type classification yields an average F-score of 64.6%, beating the best open prior systems by 11–14 points and coming within 5 points of human annotator agreement (69.8%). The result matters because automated labeling of who speaks when is a prerequisite for studying language development at scale, especially in underrepresented languages. The paper also establishes that full fine-tuning is necessary—frozen features are not enough—and that the largest gains appear in the hardest speaker classes: other children and male adults.

What carries the argument

The central object is BabyHuBERT, a HuBERT-style self-supervised model trained with masked prediction of k-means cluster targets. Pre-training uses speech segments extracted from raw daylong audio by an automatic voice-activity detector (segments under 2 seconds are padded, gaps under 2 seconds merged, capped at 30 seconds), then two clustering iterations: the first uses features from WavLM-base-plus layer 6, the second uses BabyHuBERT-1 layer 7. For voice type classification, four independent linear heads predict key child, other child, male adult, and female adult from the final encoder layer; the transformer is fully fine-tuned while the convolutional encoder stays frozen. The VTC task it

What would settle it

Compare BabyHuBERT-VTC's F-score on a set of manually transcribed long-form recordings against an identical model pre-trained on the same audio without VAD filtering or with oracle speech boundaries; if the oracle version outperforms the VAD-filtered version by several points, the central claim that the pre-training recipe is the source of the gains is weakened. A simpler check: measure the recall of the VAD specifically on target-child vocalizations and see whether corpora with lower VAD recall also show larger BabyHuBERT error.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a HuBERT-style representation model pre-trained on child-centered long-form recordings can serve as a strong foundation for voice type classification. BabyHuBERT's two-iteration masked-prediction pre-training on 13,164 hours of multilingual daylong audio produces representations that, after full fine-tuning, reach 64.6% average F-score across six corpora—outperforming W2V2-LL4300 (English-only child audio) and HuBERT (clean adult speech) everywhere, with particular gains on underrepresented languages, such as 14.0 and 18.3 absolute F1 points over HuBERT on Vanuatu and Solomon Islands. The gap to a second human annotator is only 5.2 absolute points

Load-bearing premise

The pre-training corpus is built from speech segments chosen by an automatic voice-activity detector, so if that detector systematically misses or distorts the target child's speech, the learned representations start from a biased sample of what actually happens in daylong recordings.

Editorial extensions

If this is right

  • Automated voice type classification can replace manual annotation for large-scale studies of child language environments, including in low-resource languages.
  • The 51.4% F-score on the Other Children class, up from roughly 20–30% in prior systems, may open the way to studying sibling and peer speech as factors in development.
  • Frozen feature extraction is not viable for VTC; future model use and downstream tasks should plan for full fine-tuning.
  • The consistent ordering of models across corpora suggests that domain-matched pre-training transfers across acoustic and linguistic conditions, so BabyHuBERT may benefit other child-speech classification tasks.

Reading between the lines

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

  • The VAD-based pre-training preparation may bias representations toward speech that an automatic detector already recognizes; a direct test is to pre-train on the same raw audio using oracle speech boundaries and compare fine-tuned VTC scores.
  • The paper's finding that performance varies little with language composition suggests that channel and speaker variability are the main difficulty; adding more recording devices and acoustic diversity may yield larger gains than adding more languages.
  • The low human agreement on Other Children (60.4%) means that downstream studies of sibling interaction should treat VTC labels as noisy, possibly using multi-annotator or device-based ground truth.
  • Since the paper selects the best fine-tuning run on the test set, the reported numbers are likely an optimistic bound; validation-selected runs would give a more conservative estimate.
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 / 5 minor

Summary. The paper introduces BabyHuBERT, a HuBERT-style self-supervised speech model pre-trained on approximately 13,164 hours of VAD-filtered multilingual child-centered audio from 19 corpora spanning over 40 languages. The model is evaluated on voice type classification (VTC), i.e., multi-label detection of key child, other child, male adult, and female adult speech in daylong recordings. On an English standard test set, the best BabyHuBERT-VTC model is reported to achieve 64.6% average F-score, compared to 58.7% for W2V2-LL4300 and 51.4% for HuBERT-base, and approaching the Human 2 topline of 69.8%. The paper also reports per-corpus results on six BabyTrain-2025 test sets, claiming consistent gains over both baselines, and releases code and models.

Significance. If the reported gains are robust, this is a valuable contribution to speech processing for developmental research: it is the first large-scale multilingual self-supervised model trained specifically on child-centered recordings, and the released code/models enable broader evaluation and use. The comparison with W2V2-LL4300 and HuBERT is a useful controlled test of pre-training data effect. However, the central claim depends on methodological choices that currently weaken the evidence: test-set-based model selection, unvalidated VAD preprocessing of the pre-training corpus, and missing statistical significance testing for cross-corpus comparisons. The paper ships reproducible artifacts (code and model checkpoints), which is a clear strength and makes the concerns checkable.

major comments (4)
  1. [§3.3] The paper states: 'Across all BabyHuBERT-1 and BabyHuBERT-2 training runs, we select the best performer on the test set as our final BabyHuBERT-VTC model.' This uses the test set for model selection, so the headline 64.6% in Table 3 is the maximum over six runs, not the expected performance of the modeling approach. The average fine-tuned BabyHuBERT-2 is 64.0 ± 0.7 in Table 2, which is close, but the selection rule still inflates the reported number and makes the comparison to baselines (which are not selected in the same way) unfair. Select the final model on the validation set and report the test-set performance of that model, or report the full distribution of fine-tuning runs.
  2. [§3.1] The pre-training data are not raw child-centered recordings but segments produced by PyanNet-VTC voice activity detection, with sub-2-second segments extended and gaps under 2 seconds merged. The paper reports no accuracy, bias, or error analysis of this VAD on the 19 corpora, nor any per-speaker-class breakdown. If the VAD systematically misses short, quiet, overlapping, or high-pitched child speech, the masking targets during HuBERT pre-training are drawn from a distribution skewed toward adult and/or cleaner speech. This would confound the interpretation that the observed gains come from 'child-centered multilingual pre-training' rather than from the VAD-induced filtering. Please provide VAD performance metrics (at least on a held-out subset) and discuss the potential bias, or perform a sensitivity analysis by pre-training on a subset with raw audio segments.
  3. [Figure 1 and Table 3] The claim that BabyHuBERT 'consistently outperforms' W2V2-LL4300 and HuBERT across all six corpora is based on single point estimates from the best-run model, with no variance, confidence intervals, or significance testing. The per-corpus differences in Figure 1 are plausible, but without run-level statistics it is not possible to assess whether the gains are reliable, particularly for corpora where the gap is small (e.g., English and Tsimane). Report the mean and standard deviation across fine-tuning runs for each corpus and baseline, or provide paired significance tests (e.g., bootstrap over test segments) for the primary comparisons.
  4. [Abstract (arXiv version) vs. Full-text abstract and Figure 1] The arXiv abstract reports F1-scores from 55.0% to 76.1% and gains of 14.0 and 18.3 points over HuBERT on Vanuatu and Solomon Islands, while the full-text abstract reports 52.1% to 74.4% and gains of 13.2 and 15.9 points. Figure 1, however, shows BabyHuBERT-VTC scores ranging from 60.9% to 74.4% across the six corpora, which matches neither range. These numeric inconsistencies affect the paper's central quantitative claims and must be reconciled before publication.
minor comments (5)
  1. [Table 2] The row 'BabyHuBERT 170.048.0' is a formatting error; it should read 'BabyHuBERT 1 70.0 48.0 62.8 73.9 63.7 ± 2.1'.
  2. [Introduction / Abstract] The claim of being 'the first self-supervised speech representation model trained on 13,000 hours of child-centered recordings' is imprecise because W2V2-LL4300 is also a self-supervised model trained on child-centered recordings. Clarify that BabyHuBERT is the first multilingual, large-scale model of this kind.
  3. [§3.1] Table 1 lists 'Total Duration' and 'Effective Duration' but the latter is the amount of audio after VAD preprocessing. The abstract and introduction repeatedly state pre-training on '13,000 hours of child-centered recordings,' which could be misread as raw audio; consider explicitly saying '13,000 hours of VAD-selected speech segments.'
  4. [§5] The phrase 'decreasing the model-human gap by a shocking 11%' is informal for a journal report; also specify whether this is an absolute or relative percentage. Later text uses 'absolute' but the sentence is ambiguous.
  5. [Figure 1] The corpus labels and bar grouping are not immediately readable in the text; please add a legend or clarify the ordering so readers can map each triplet of bars to BabyHuBERT-VTC, W2V2-LL4300, and HuBERT.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BabyHuBERT's VTC gains are empirically evaluated on held-out data against external and prior baselines, with no prediction reducing to a fitted input or self-citation chain.

full rationale

The paper's central claim is that self-supervised pre-training on 13,164 hours of multilingual child-centered recordings improves voice type classification. This is supported by fine-tuning on BabyTrain-2025, with test splits that are explicitly filtered out of pre-training, and by comparisons to W2V2-LL4300, HuBERT, PyanNet-VTC, and Whisper-VTC. The use of PyanNet-VTC as a voice activity detector to build the pre-training corpus is a preprocessing choice that could bias the training distribution, but it does not make the downstream F-scores equivalent to PyanNet-VTC's own outputs or to any fitted parameter by construction. The Human 2 topline and the Whisper-VTC/PyanNet baselines come from the authors' prior work, but they are used as empirical benchmarks, not as theoretical premises that force the result. Selecting the best performing run on the test set and choosing layer 7 based on VTC validation are methodological concerns about test-set overfitting, not circularity in the derivation chain. No self-definitional relationship, no renamed known result, and no imported uniqueness theorem are present. The empirical comparisons are self-contained and externally checkable, so the paper does not exhibit circular reasoning.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the quality and representativeness of the pre-training corpus, the VAD preprocessing, and the evaluation protocol. Most load-bearing items are empirical domain assumptions, not mathematical axioms; the only 'fitted' elements are standard hyperparameters plus a task-specific layer choice and a test-set model-selection rule.

free parameters (5)
  • BabyHuBERT-2 layer index = 7
    Layer 7 chosen instead of standard layer 9 'based on empirical validation on VTC' (§3.2), a task-specific hyperparameter selection.
  • Number of k-means clusters = 500
    Set to 500 following standard HuBERT; not optimized for this corpus (§3.4).
  • Preprocessing segment rules = 2s extension, 2s merge threshold, 30s cap
    Hand-set in §3.1 to balance context and non-speech; affects all pre-training data.
  • Fine-tuning hyperparameters = lr 1e-5, dropout 0.5, batch 128x4s
    Same hyperparameters for all models, not tuned per model (§3.4).
  • Model selection rule = best F1 on test set
    §3.3: 'we select the best performer on the test set as our final BabyHuBERT-VTC model', a selection on the evaluation target that inflates reported scores.
assumptions (4)
  • domain assumption HuBERT masked-prediction pretraining generalizes to noisy child-centered audio
    Invoked in §3.2 citing [20,21]; no experiment in this paper isolates this choice from alternative SSL objectives.
  • domain assumption PyanNet-VTC voice activity detection provides unbiased speech segments for pre-training
    §3.1 uses PyanNet-VTC to extract and extend speech segments; no VAD accuracy diagnostic is reported on the 19 corpora.
  • domain assumption BabyTrain-2025 annotations and the ACLEW standard test set are representative of VTC performance
    Evaluation relies on these sets (§3.1, §4); the 20-hour English standard test is a convenience benchmark from [15].
  • standard math K-means with 500 clusters yields meaningful HuBERT targets at both iterations
    Adopted from HuBERT [22]; no analysis of cluster quality on child speech.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BabyHuBERT: Multilingual Self-Supervised Learning for Segmenting Speakers in Child-Centered Long-Form Recordings." pith.science (2026). https://pith.science/paper/FWTTS7TE

@misc{pith2026250915001,
  author       = {Pith},
  title        = {Pith review of: BabyHuBERT: Multilingual Self-Supervised Learning for Segmenting Speakers in Child-Centered Long-Form Recordings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FWTTS7TE}},
  note         = {Machine review of arXiv:2509.15001}
}
read the original abstract

Child-centered daylong recordings are essential for studying early language development, but existing speech models trained on clean adult data perform poorly due to acoustic and linguistic differences. We introduce BabyHuBERT, a self-supervised speech model trained on 13,000 hours of multilingual child-centered recordings from 40+ languages. Evaluated on voice type classification, the task of identifying who produces speech and when in child-centered recordings (key child, other children, male, and female adults), BabyHuBERT-VTC achieves F1-scores from 55.0% to 76.1% across six corpora, consistently outperforming W2V2-LL4300 and HuBERT (pretrained on English daylongs and clean adult speech, respectively). Notable gains include 14.0 and 18.3 absolute F1 points over HuBERT on Vanuatu and Solomon Islands, demonstrating effectiveness on underrepresented languages. We share code and models to support researchers working with child-centered recordings across diverse linguistic contexts.

Discussion (0). Continue with ORCID 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. Deriving Benchmarking Datasets from Long-Form Recordings: Challenges and Opportunities

    eess.AS 2026-07 accept novelty 6.0 of 10

    A mutually dependent framework standardizes 27 LFR child-speech corpora, builds four benchmarks, and embeds ELSI governance, with a VTC case study showing private data is needed for competitive performance.

Reference graph

Works this paper leans on

35 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    INTRODUCTION Despite decades of advances in automatic speech processing, speech science and technology has had relatively limited im- pact on language development research due to a fundamental mismatch: existing speech models, trained predominantly on clean adult speech, fail catastrophically on the complex and di- verse acoustic environments where childr...

  2. [2]

    RELATED WORK 2.1. Child Speech Processing Given the scarcity of self-supervised speech representation models trained on child-centered naturalistic recordings, pre- vious work has primarily relied on fine-tuning adult-trained models (for example, see [ 6, 7] for automatic speech recog- nition). While these approaches show improvements over adult baselines...

  3. [3]

    Datasets Our pre-training dataset comprises 19 diverse datasets spanning mul- tiple continents over 40 languages (see Table 1)

    METHODS 3.1. Datasets Our pre-training dataset comprises 19 diverse datasets spanning mul- tiple continents over 40 languages (see Table 1). Some datasets were obtained from scientific archives like HomeBank [13], while others were accessed through direct data-sharing agreements with research groups. To avoid contamination we filter out audio files used i...

  4. [4]

    BabyHuBERT-2 achieves 64.0% average F-score, substan- tially outperforming both W2V2-LL4300 (58.7%) and HuBERT base (51.4%)

    RESULTS Fine-tuning experiments: Fine-tuning drastically improves perfor- mance in the standard test set for all voice types, as shown in ta- ble 2. BabyHuBERT-2 achieves 64.0% average F-score, substan- tially outperforming both W2V2-LL4300 (58.7%) and HuBERT base (51.4%). This demonstrates the value of large-scale, multilingual Table 2. F-scores (%) obta...

  5. [5]

    DISCUSSION Child-centered recordings represent an outstanding challenge for speech science and technology, and scarcity of human annotations has held back progress. The difficulty of such data is obvious in our human annotator topline: When comparing two humans’ decisions on when different voice types (Key Child, Other Children, Female Adults, Male Adults...

  6. [6]

    ED: ERC (InfantSimulator); AC and TK: ERC (ExELang, 101001095)

    ACKNOWLEDGMENTS This work was performed using HPC resources from GENCI-IDRIS (Grant 2024-AD01101545 and 2025-AD011016414) and was sup- ported in part by the Agence Nationale pour la Recherche (ANR- 17-EURE-0017 Frontcog, ANR10-IDEX-0001-02 PSL*). ED: ERC (InfantSimulator); AC and TK: ERC (ExELang, 101001095). ML: Simons Foundation International (034070-00...

  7. [7]

    Self-supervised models for phoneme recognition: Applica- tions in children’s speech for reading learning,

    Lucas Block Medin, Thomas Pellegrini, and Lucile Gelin, “Self-supervised models for phoneme recognition: Applica- tions in children’s speech for reading learning,”arXiv preprint arXiv:2503.04710, 2025

  8. [8]

    Long-form recordings to study children’s language input and output in under-resourced contexts,

    Alejandrina Cristia and Joseph Coffey, “Long-form recordings to study children’s language input and output in under-resourced contexts,” inThe Fifth Workshop on Resources for African In- digenous Languages@ LREC-COLING 2024. ELRA and ICCL, 2024

Show all 35 references
  1. [9]

    Fifteen years of child-centered long-form recordings: Promises, resources, and remaining challenges to validity,

    Loann Peurey, Marvin Lavechin et al., “Fifteen years of child-centered long-form recordings: Promises, resources, and remaining challenges to validity,”arXiv preprint arXiv:2506.11075, 2025

  2. [10]

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

    Sungbok Lee, Alexandros Potamianos, and Shrikanth Narayanan, “Acoustics of children’s speech: Developmental changes of temporal and spectral parameters,”The Journal of the Acoustical Society of America, 03 1999

  3. [11]

    Acous- tic variability and automatic recognition of children’s speech,

    Matteo Gerosa, Diego Giuliani, and Fabio Brugnara, “Acous- tic variability and automatic recognition of children’s speech,” Speech Communication, 2007, Intrinsic Speech Variations

  4. [12]

    Systematic inequalities in language technology performance across the world’s languages,

    Damian Blasi, Antonios Anastasopoulos, and Graham Neubig, “Systematic inequalities in language technology performance across the world’s languages,”arXiv preprint arXiv:2110.06733, 2021

  5. [13]

    Introduction To Partial Fine-tuning: A Comprehensive Evaluation Of End-to-end Chil- dren’s Automatic Speech Recognition Adaptation,

    Thomas Rolland and Alberto Abad, “Introduction To Partial Fine-tuning: A Comprehensive Evaluation Of End-to-end Chil- dren’s Automatic Speech Recognition Adaptation,” inInter- speech, 2024

  6. [14]

    A thorough eval- uation of the language environment analysis (lena) system,

    Alejandrina Cristia, Marvin Lavechin et al., “A thorough eval- uation of the language environment analysis (lena) system,” Behavior Research Methods, Apr 2021

  7. [15]

    Towards robust family-infant audio analysis based on unsu- pervised pretraining of wav2vec 2.0 on large-scale unlabeled family audio,

    Jialu Li, Mark Hasegawa-Johnson, and Nancy L. McElwain, “Towards robust family-infant audio analysis based on unsu- pervised pretraining of wav2vec 2.0 on large-scale unlabeled family audio,” inInterspeech, 2023

  8. [16]

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

    Alexei Baevski, Yuhao Zhou et al., “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato et al., Eds. 2020, vol. 33, Curran Associates, Inc

  9. [17]

    Employing self- supervised learning models for cross-linguistic child speech maturity classification,

    Theo Zhang, Madurya Suresh et al., “Employing self- supervised learning models for cross-linguistic child speech maturity classification,” inInterspeech, 2025

  10. [18]

    Reverse en- gineering language acquisition with child-centered long-form recordings,

    Marvin Lavechin, Maureen de Seyssel et al., “Reverse en- gineering language acquisition with child-centered long-form recordings,”Annual Review of Linguistics, 2022

  11. [19]

    Brian MacWhinney,The CHILDES project, Psychology Press, London, England, 3 edition, June 2014

  12. [20]

    Homebank: An online repository of daylong child-centered audio recordings,

    Mark VanDam, Anne S. Warlaumont et al., “Homebank: An online repository of daylong child-centered audio recordings,” Semin Speech Lang, Apr 2016

  13. [21]

    mhubert-147: A compact multilingual hubert model,

    Marcely Zanon Boito, Vivek Iyer et al., “mhubert-147: A compact multilingual hubert model,” 2024

  14. [22]

    For BabyHuBERT-1, we extract features from the 6th layer of WavLM-base-plus

    using the torchaudio implementation [25]. For BabyHuBERT-1, we extract features from the 6th layer of WavLM-base-plus. For BabyHuBERT-2, we use features from BabyHuBERT-1’s 7th trans- former layer. We apply k-means clustering to these extracted features using MiniBatchKMeans w...

  15. [23]

    An open-source voice type classifier for child-centered daylong recordings,

    Marvin Lavechin, Ruben Bousbib et al., “An open-source voice type classifier for child-centered daylong recordings,” in Interspeech, 2020

  16. [24]

    Signal processing for young child speech language development.,

    Dongxin Xu, Umit H Yapanel et al., “Signal processing for young child speech language development.,” inWocci, 2008

  17. [25]

    Speaker recognition from raw waveform with sincnet,

    Mirco Ravanelli and Yoshua Bengio, “Speaker recognition from raw waveform with sincnet,” in2018 IEEE Spoken Language Technology Workshop (SLT), 2018

  18. [26]

    Challenges in Automated Processing of Speech from Child Wearables: The Case of V oice Type Classifier,

    Tarek Kunze, Marianne M´etais et al., “Challenges in Automated Processing of Speech from Child Wearables: The Case of V oice Type Classifier,” inInterspeech, 2025

  19. [27]

    Developing a cross-cultural annotation system and metacorpus for studying infants’ real world language experience,

    Melanie Soderstrom, Marisa Casillas et al., “Developing a cross-cultural annotation system and metacorpus for studying infants’ real world language experience,”Collabra: Psychology, 2021

  20. [28]

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

    Sanyuan Chen, Chengyi Wang et al., “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing, 2022

  21. [29]

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

    Wei-Ning Hsu, Benjamin Bolte et al., “Hubert: Self-supervised speech representation learning by masked prediction of hid- den units,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021

  22. [30]

    Superb: Speech processing universal performance benchmark,

    Shu wen Yang, Po-Han Chi et al., “Superb: Speech processing universal performance benchmark,” inInterspeech, 2021

  23. [31]

    pyannote. metrics: A toolkit for reproducible evaluation, diagnostic, and error analysis of speaker diarization systems.,

    Herv´e Bredin, “pyannote. metrics: A toolkit for reproducible evaluation, diagnostic, and error analysis of speaker diarization systems.,” inInterspeech, 2017

  24. [32]

    Torchaudio 2.1: Advancing speech recognition, self-supervised learning, and audio process- ing components for pytorch,

    Jeff Hwang, Moto Hira et al., “Torchaudio 2.1: Advancing speech recognition, self-supervised learning, and audio process- ing components for pytorch,” 2023

  25. [33]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa, G. Varoquaux et al., “Scikit-learn: Machine learning in Python,”Journal of Machine Learning Research, 2011

  26. [34]

    Child-directed and overheard input from different speakers in two distinct cultures,

    Georgia Loukatou, Camila Scaff et al., “Child-directed and overheard input from different speakers in two distinct cultures,” Journal of Child Language, 2022

  27. [35]

    Putting the child in the driver’s seat: insights into language development from children’s interactions in preschool classrooms,

    Lynn K Perry, Sophia A Meibohm et al., “Putting the child in the driver’s seat: insights into language development from children’s interactions in preschool classrooms,”Advances in child development and behavior, 2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.