Pith. sign in

REVIEW 3 major objections 4 minor 42 references

Challenges in Automated Processing of Speech from Child Wearables: The Case of Voice Type Classifier

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 For child-worn audio, the paper claims, model upgrades add little once enough labeled data are available: pretrained Whisper features and new architectures converge with the older baseline near a 48% average F-score, so the field's…

desk verdict Useful negative result undermined by inconsistent baselines; deserves revision, not rejection. read the letter →

arxiv 2506.11074 v1 pith:ZG6MCTLP submitted 2025-06-04 eess.AS cs.LGcs.SD

Tarek Kunze , Marianne Métais , Hadrien Titeux , Lucas Elbert , Joseph Coffey , Emmanuel Dupoux , Alejandrina Cristia , Marvin Lavechin This is my paper · ORCID
classification eess.AScs.LGcs.SD
keywords voicetypeclassificationchild-wornaudiolong-formrecordingsWhisperPyanNetmulti-labelhumanannotationagreementdatascaling
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

Voice type classification labels who is speaking in long-form recordings from child-worn devices, a task that underpins most automated measures of children's language environments. This paper summarizes three years of experiments concluding that pretrained speech representations (Whisper), architectural changes, and parameter search improve performance only marginally over the open-source PyanNet-VTC baseline, while the quantity and relevance of training data is the main driver. In low-resource settings Whisper-VTC helps (42% vs 35% F-score at 10% of training data), but beyond roughly 420 hours of audio both approaches converge near 48% average F-score. The authors argue that the path forward is larger, better-annotated, shareable datasets, and they show that even two human annotators disagree substantially on the hardest class, so the task's ceiling is partly set by label noise.

What carries the argument

The central object is Whisper-VTC, a voice type classifier that replaces the SincNet encoder of PyanNet-VTC with a frozen Whisper speech encoder. Input audio becomes an 80-channel log-magnitude Mel spectrogram, passes through two convolutional layers into the frozen encoder, and the encoder's layer outputs are combined with a learnable weighted sum; shared bidirectional LSTMs then feed four independent binary classification heads, one per voice type (key child, other child, female adult, male adult). The learning-curve experiment, which trains both VTC variants on increasing fractions of the same 669-hour BabyTrain-2025 corpus and evaluates on the fixed 20-hour hold-out set, is the mechanism that carries the argument: the gap between models closes as data grows, showing diminishing returns from representation upgrades.

What would settle it

Train both VTC variants on 2,000+ hours of annotated child-worn audio from diverse languages and recording hardware, and evaluate on a multi-annotator adjudicated gold set; if the two models' F-scores diverge or average scores exceed roughly 60%, the plateau was a data-size or label-noise artifact rather than a fundamental modeling limit.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the choice of feature representation and model architecture is not what limits voice type classification in child-worn audio. Replacing PyanNet's SincNet raw-waveform encoder with a frozen Whisper encoder and four independent binary heads yields a marked gain only when training data is scarce (42% vs 35% F-score with 10% of data); by 70% utilization both systems converge to about 48% average F-score. Whisper's largest class-level benefit is on the rare male-adult class, and Whisper-VTC is appreciably more stable across runs. Against a second human annotator, who averages 69.8% F-score on the same hold-out set, neither automated system comes close, especially on other-child speech. The authors conclude that scaling annotated data, not further modeling sophistication, is the productive direction.

Load-bearing premise

The conclusion assumes that the 20-hour BabyTrain-2020 hold-out set represents the full variety of child-wearable recordings, and that F-scores against one reference annotator are an unbiased measure of true performance.

Editorial extensions

If this is right

  • If the plateau holds, further architecture or pretraining research on this task is unlikely to move the average; the field's next credible investment is annotated data, not new models.
  • For new recording environments with few labels, Whisper-based VTC is the better starting point (42% vs 35% F-score at 10% of training data), so pretrained features can serve as a bootstrap before in-domain data accumulates.
  • The male-adult class benefits most from Whisper representations, so studies focused on fathers or male caregivers should prefer the Whisper-VTC over PyanNet-VTC.
  • Because automated scores sit near 50% average F-score while two human annotators agree around 70%, any downstream measure (word counts, speaker-segmented analyses) carries a sizable error floor that should be reported alongside results.
  • Human disagreement on the other-child class (60% F-score) means the reference labels themselves are noisy; evaluations that ignore annotator disagreement overstate the gap between machines and the true underlying voices.

Reading between the lines

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

  • The 48% convergence may be partly an artifact of scoring against a single reference annotation; with adjudicated or majority-vote labels, models might prove closer to truth than their F-scores suggest.
  • A testable extension is to compute multi-annotator agreement on the same hold-out recordings and train on re-annotated, adjudicated labels; if the plateau moves, label noise rather than model capacity was the binding constraint.
  • The paper's experience that data sharing yields only 14% more annotated audio suggests that investment in standardizing recording and annotation formats could unlock more value than further modeling work.
  • More generally, for hard ecological audio tasks, pretrained speech models act as an initial head start that vanishes once in-domain data grows; the authors' 'bitter lesson' is drawn from a single task and dataset family, so its breadth is untested.
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. The paper reports three years of experiments aimed at improving Voice Type Classification (VTC) on child-worn audio. The authors introduce Whisper-VTC, which combines frozen Whisper representations with a bi-LSTM and per-class binary heads, and compare it against PyanNet-VTC, a retrained SincNet+LSTM baseline. Experiments on a large pooled dataset (BabyTrain-2025, 669h) show that Whisper representations help mainly in low-resource settings and that beyond roughly 70% data utilization both models converge to about 48% mean F-score. The paper also compares system performance against a second human annotator, demonstrating a substantial human ceiling gap, and reports a series of unsuccessful attempts (oversampling, undersampling, GradNorm, powerset, additional CHiME5 data) to improve performance. The authors conclude that data quantity and relevance, not architecture or pretrained representations, are the current bottleneck.

Significance. If the central claim holds, the paper makes a valuable meta-point for the child-worn audio community: further progress in VTC is likelier to come from larger, more diverse labeled corpora than from further model engineering. The paper is refreshingly honest about null results, provides an open-source implementation, and benchmarks against a human ceiling, which is a useful reference for future work. However, the evidence for the plateau is not yet airtight, and the manuscript's internal baseline inconsistency weakens the main conclusion.

major comments (3)
  1. [Section 3.2, Figure 2] The central plateau claim rests on a retrained PyanNet-VTC, but Table 2 reports the original PyanNet-VTC from [13] at 50.9% average F-score on the same 20h hold-out after training on only 159h of BabyTrain-2020, whereas the retrained curve in Figure 2 reaches approximately 48% even at 100% data (669h). This discrepancy suggests the retrained baseline may be undertrained or that the added CHILDES data actually hurts performance. Because Figure 1 compares Whisper models against the original [13] PyanNet while Figure 2 uses a different, retrained baseline, the 'convergence' shown in Figure 2 may not reflect a fundamental equivalence of the two modeling approaches but rather a degraded baseline. The authors must reconcile these numbers, e.g., by plotting the original [13] baseline as a horizontal reference in Figure 2 and by reporting the retrained baseline's hold-out F-score at full data.
  2. [Section 3.2, Figure 2] The assertion that 'Whisper-VTC's performance is much more stable across runs than PyanNet-VTC' is not supported by the reported evidence. Figure 2 shows single curves without error bars, confidence intervals, or the number of runs. Without repeated-run statistics, the stability claim is unverifiable. Please provide mean and variance (e.g., across at least three random seeds) for the key data utilizations, or remove the stability claim.
  3. [Section 3.4, Table 2] The comparison against human performance uses a single reference annotation and a single second annotator, and the systems are scored against the same reference used to score human 2. The reported human-human F-scores of 60% for OCH and 80% for KCHI indicate substantial reference-label noise, but no inter-annotator reliability statistic (e.g., kappa or agreement per segment) is reported, and the effect of reference noise on system F-scores is not quantified. This is not merely a presentation issue: it directly affects how the human ceiling should be interpreted. I recommend reporting reliability statistics and, if possible, evaluating with multiple references or a soft-label protocol.
minor comments (4)
  1. [Table 2] There is a formatting error in the Whisper-VTC (tiny) row: '62,6' should presumably be '62.6', and the OCH value of '1.34' appears implausibly low; if it is correct, please explain it, and if it is a typo, correct it.
  2. [Section 3.2] The parenthetical 'showing only Whisper-VTC base to facilitate inspection' seems to contradict the figure caption, which describes curves for both PyanNet-VTC and Whisper-VTC; please clarify what exactly is shown in Figure 2.
  3. [Section 2.2] The sentence 'BabyTrain-2021 highlighted here is a corrected version' is vague; specify what correction was applied relative to earlier usage of this corpus.
  4. [Section 3.1] The statement that larger Whisper sizes 'yield similar performance' is made without showing the data; a brief sentence quantifying the observed differences would make the claim more transparent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported F-scores are measured on a held-out set against human annotations, with no fitted parameter or self-defined target feeding the evaluation.

full rationale

The paper's central claim is empirical: pretrained representations and architectural changes yield only marginal gains relative to training-set size for voice type classification. The evaluation is straightforwardly external: models are trained on human-labeled audio and scored with pyannote.metrics F-scores on a fixed 20-hour hold-out set, with a second human annotator providing an independent ceiling. No equation in the paper defines the target metric in terms of a fitted parameter, and no 'prediction' is constructed from the data used to fit a constant. The PyanNet-VTC baseline comes from the authors' prior work [13], but it is used as a benchmark model, not as a theorem or an imported uniqueness result; comparing against one's own prior system is normal and does not make the measured differences circular. The addition of CHILDES datasets was selected partly because they 'were found to improve performance during piloting,' which is a potential selection-on-validation concern rather than circularity: the final reported F-scores are still produced by models trained on labeled data and evaluated on an untouched hold-out set. Similarly, the human-human agreement numbers provide an external, non-circular reference point. The paper's plateau conclusion may be vulnerable to methodological critique (e.g., baseline retraining details or lack of error bars), but such concerns do not amount to the derivation reducing to its own inputs. Under the stated rules, with no quote demonstrating that Eq. X equals Eq. Y by construction or that a fitted parameter is renamed as a prediction, the appropriate finding is 'no significant circularity.'

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

No free parameters are fitted to the data in a way that supports the central claim; the paper introduces no new physical or conceptual entities. It relies on standard supervised learning assumptions and the representativeness of its evaluation data.

assumptions (3)
  • domain assumption DARCLE annotation labels are treated as ground truth for voice type categories.
    Used throughout as training targets and evaluation reference; if the annotation scheme is inconsistent or biased, all F-scores are affected (Sections 2.2 and 3.4).
  • domain assumption Frozen Whisper representations are a fair representative of pretrained speech features for this task.
    The comparison relies on Whisper as the chosen pretrained representation; other foundation models could behave differently, so conclusions are scoped to the tested families (Section 2.1).
  • standard math The pyannote.metrics F-score implementation matches standard precision and recall definitions.
    Underlying evaluation metric used without derivation; standard and unlikely to be load-bearing (Section 2.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Challenges in Automated Processing of Speech from Child Wearables: The Case of Voice Type Classifier." pith.science (2026). https://pith.science/paper/ZG6MCTLP

@misc{pith2026250611074,
  author       = {Pith},
  title        = {Pith review of: Challenges in Automated Processing of Speech from Child Wearables: The Case of Voice Type Classifier},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZG6MCTLP}},
  note         = {Machine review of arXiv:2506.11074}
}
read the original abstract

Recordings gathered with child-worn devices promised to revolutionize both fundamental and applied speech sciences by allowing the effortless capture of children's naturalistic speech environment and language production. This promise hinges on speech technologies that can transform the sheer mounds of data thus collected into usable information. This paper demonstrates several obstacles blocking progress by summarizing three years' worth of experiments aimed at improving one fundamental task: Voice Type Classification. Our experiments suggest that improvements in representation features, architecture, and parameter search contribute to only marginal gains in performance. More progress is made by focusing on data relevance and quantity, which highlights the importance of collecting data with appropriate permissions to allow sharing.

Figures

Figures reproduced from arXiv: 2506.11074 by the authors.

Figure 1
Figure 1. F-score (%) averaged across speaker categories for PyanNet (results from [13]) vs. Whisper-VTC (using frozen fea￾tures from Whisper tiny, base, or small). Performance is com￾puted on the hold-out set [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. F-score (%) averaged across speaker categories as a function of training set size for PyanNet-VTC and Whisper￾VTC (using frozen features from Whisper base). Performance is computed on the hold-out set. 3.3. Segmentation errors: Similarities and differences [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Miss (%), false alarm (%), confusion (%), and correct (%) obtained by PyanNet-VTC and Whisper-VTC (base) as a function of speech-to-noise ratio (SNR) and C50 (estimated by Brouhaha [8]). Each point represents the average performance over all audio from a given child. Solid lines show linear regression fits with shaded 95% confidence intervals. data, we reasoned that the model could benefit from robust rep￾resentatio… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 36 canonical work pages

  1. [13]

    Alice: An open-source tool for automatic measurement of phoneme, syllable, and word counts from child-centered daylong recordings,

    O. R ¨as¨anen, S. Seshadri, M. Lavechin, A. Cristia, and M. Casil- las, “Alice: An open-source tool for automatic measurement of phoneme, syllable, and word counts from child-centered daylong recordings,” Behavior Research Methods , vol. 53, pp. 818–835, 2021

    work page 2021

  2. [1]

    They enable effortless, long-duration recordings, offering unprecedented insights into children’s language environments [1]

    Introduction Child-worn recording devices have transformed both funda- mental and applied speech sciences, capturing egocentric audio (or audio-video) that represents speech from the child’s perspec- tive. They enable effortless, long-duration recordings, offering unprecedented insights into children’s language environments [1]. Unlike laboratory studies,...

    work page 2025

  3. [2]

    Methods 2.1. End-to-end voice type classification As in [13], we framed the voice type classification problem as a multi-label classification problem, where the input is the au- dio stream divided into N frames S = {s1, s2, . . . , sN } and the expected output is the corresponding sequence of labels y = {y1, y2, . . . ,yN } where each yi is of dimension K...

    work page 2020

  4. [3]

    Experimenting with different Whisper sizes Figure 1: F-score (%) averaged across speaker categories for PyanNet (results from [13]) vs

    Results 3.1. Experimenting with different Whisper sizes Figure 1: F-score (%) averaged across speaker categories for PyanNet (results from [13]) vs. Whisper-VTC (using frozen fea- tures from Whisper tiny, base, or small). Performance is com- puted on the hold-out set. Figure 1 shows a clear trend: larger Whisper-VTC models achieve better performance, with...

    work page 2025

  5. [4]

    solved”. These attempts echo Sutton’s

    Discussion As alluded to in the Introduction, we summarize here some of the extensive experiments we did with null improvements. In one, we found no improvement when replacing the SincNet ar- chitecture (PyanNet-VTC) processing raw waveforms with a stack of convolutional layers processing spectrograms. Later, our colleagues raised valid concerns about cla...

    work page

  6. [5]

    Despite these challenges, this technology holds transformative potential for educational interventions and insights into children’s lan- guage learning mechanisms

    Conclusions We report on over three years of work improving V oice Type Classification in long-form, child-centered recordings—a chal- lenging task due to the realistic recording conditions. Despite these challenges, this technology holds transformative potential for educational interventions and insights into children’s lan- guage learning mechanisms. Gi...

    work page

  7. [6]

    Acknowledgements This work was performed using HPC resources from GENCI- IDRIS (Grant 2024-AD011015450)

    work page 2024

  8. [7]

    V o- cal development in a large-scale crosslinguistic corpus,

    M. Cychosz, A. Cristia, E. Bergelson, M. Casillas, G. Baudet, A. S. Warlaumont, C. Scaff, L. Yankowitz, and A. Seidl, “V o- cal development in a large-scale crosslinguistic corpus,”Develop- mental science, vol. 24, no. 5, p. e13090, 2021

    work page 2021

Show all 42 references
  1. [8]

    Reverse engineering language acquisition with child- centered long-form recordings,

    M. Lavechin, M. De Seyssel, L. Gautheron, E. Dupoux, and A. Cristia, “Reverse engineering language acquisition with child- centered long-form recordings,” Annual Review of Linguistics , vol. 8, no. 1, pp. 389–407, 2022

    2022

  2. [9]

    Day by day, hour by hour: Naturalistic language input to infants,

    E. Bergelson, A. Amatuni, S. Dailey, S. Koorathota, and S. Tor, “Day by day, hour by hour: Naturalistic language input to infants,” Developmental science, vol. 22, no. 1, p. e12715, 2019

    2019

  3. [10]

    Cognitive science in the era of artificial intelligence: A roadmap for reverse-engineering the infant language-learner,

    E. Dupoux, “Cognitive science in the era of artificial intelligence: A roadmap for reverse-engineering the infant language-learner,” Cognition, vol. 173, pp. 43–59, 2018

    2018

  4. [11]

    When cultural norms dis- courage talking to babies: Effectiveness of a parenting program in rural senegal,

    A. Weber, A. Fernald, and Y . Diop, “When cultural norms dis- courage talking to babies: Effectiveness of a parenting program in rural senegal,” Child Development, vol. 88, no. 5, pp. 1513–1526, 2017

    2017

  5. [12]

    Shifting parental be- liefs about child development to foster parental investments and improve school readiness outcomes,

    J. A. List, J. Pernaudet, and D. L. Suskind, “Shifting parental be- liefs about child development to foster parental investments and improve school readiness outcomes,” Nature communications , vol. 12, no. 1, p. 5765, 2021

    2021

  6. [14]

    Lavechin, L

    M. Lavechin, L. R. Hamrick, B. Kelleher, and A. Seidl, “Perfor- mance and biases of the lena® and aclew algorithms in analyzing language environments in down, fragile x, angelman syndromes, and populations at elevated likelihood for autism,” 2025

    2025

  7. [15]

    First, we demonstrate that pre-trained representations learned by Whisper [16] are most effective for smaller training sets

    investigated foundation models with English long-form data, we contribute novel data using a more diverse dataset and benchmark against the current open-source state-of-the-art [13]. First, we demonstrate that pre-trained representations learned by Whisper [16] are most effect...

    2025 arXiv

  8. [16]

    Brouhaha: multi-task training for voice activity detection, speech-to-noise ratio, and c50 room acoustics estimation,

    M. Lavechin, M. M ´etais, H. Titeux, A. Boissonnet, J. Copet, M. Rivi`ere, E. Bergelson, A. Cristia, E. Dupoux, and H. Bredin, “Brouhaha: multi-task training for voice activity detection, speech-to-noise ratio, and c50 room acoustics estimation,” in Automatic Speech Recognitio...

    2023

  9. [17]

    Who said what? an automated approach to analyzing speech in preschool classrooms,

    A. Sun, J. J. Londono, B. Elbaum, L. Estrada, R. J. Lazo, L. Vitale, H. G. Villasanti, R. Fusaroli, L. K. Perry, and D. S. Messinger, “Who said what? an automated approach to analyzing speech in preschool classrooms,” in International Conference on Develop- ment and Learning (...

    2024

  10. [18]

    The babyview dataset: High-resolution egocentric videos of in- fants’ and young children’s everyday experiences,

    B. Long, V . Xiang, S. Stojanov, R. Z. Sparks, Z. Yin, G. E. Keene, A. W. Tan, S. Y . Feng, C. Zhuang, V . A. Marchmanet al., “The babyview dataset: High-resolution egocentric videos of in- fants’ and young children’s everyday experiences,”arXiv preprint arXiv:2406.10447, 2024

    2024 arXiv

  11. [19]

    Signal processing for young child speech language development

    D. Xu, U. H. Yapanel, S. S. Gray, J. Gilkerson, J. A. Richards, and J. H. Hansen, “Signal processing for young child speech language development.” in WOCCI, 2008, p. 20

    2008

  12. [20]

    Mapping the early language environment using all-day recordings and automated analysis,

    J. Gilkerson, J. A. Richards, S. F. Warren, J. K. Montgomery, C. R. Greenwood, D. Kimbrough Oller, J. H. Hansen, and T. D. Paul, “Mapping the early language environment using all-day recordings and automated analysis,” American journal of speech- language pathology, vol. 26, n...

    2017

  13. [21]

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

    M. Lavechin, R. Bousbib, H. Bredin, E. Dupoux, and A. Cristia, “An open-source voice type classifier for child- centered daylong recordings,” in Interspeech, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:218889736

    2020

  14. [22]

    Enhancing child vocalization classification with phonetically-tuned embed- dings for assisting autism diagnosis,

    J. Li, M. Hasegawa-Johnson, and K. Karahalios, “Enhancing child vocalization classification with phonetically-tuned embed- dings for assisting autism diagnosis,” in Interspeech, 2024, pp. 5163–5167

    2024

  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,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

    2023

  16. [24]

    pyannote. metrics: A toolkit for reproducible eval- uation, diagnostic, and error analysis of speaker diarization sys- tems

    H. Bredin, “pyannote. metrics: A toolkit for reproducible eval- uation, diagnostic, and error analysis of speaker diarization sys- tems.” in Interspeech, 2017, pp. 3587–3591

    2017

  17. [25]

    Speaker recognition from raw wave- form with sincnet,

    M. Ravanelli and Y . Bengio, “Speaker recognition from raw wave- form with sincnet,” in Spoken Language Technology workshop (SLT). IEEE, 2018, pp. 1021–1028

    2018

  18. [26]

    Exploring speech foundation models for speaker diarization in child-adult dyadic interactions,

    A. Xu, K. Huang, T. Feng, L. Shen, H. Tager-Flusberg, and S. Narayanan, “Exploring speech foundation models for speaker diarization in child-adult dyadic interactions,” in Interspeech, 2024, pp. 5193–5197

    2024

  19. [27]

    The childes system,

    B. MacWhinney, “The childes system,” Handbook of child lan- guage acquisition, pp. 457–494, 1998

    1998

  20. [28]

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

    M. Soderstrom, M. Casillas, E. Bergelson, C. Rosemberg, F. Alam, A. S. Warlaumont, and J. Bunce, “Developing a cross- cultural annotation system and metacorpus for studying infants’ real world language experience,” Collabra: Psychology , vol. 7, no. 1, p. 23445, 2021

    2021

  21. [29]

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

    M. VanDam, A. S. Warlaumont, E. Bergelson, A. Cristia, M. Soderstrom, P. De Palma, and B. MacWhinney, “Homebank: An online repository of daylong child-centered audio recordings,” in Seminars in speech and language , vol. 37, no. 02. Thieme Medical Publishers, 2016, pp. 128–142

    2016

  22. [30]

    The talkbank project,

    B. MacWhinney, “The talkbank project,” in Creating and dig- itizing language corpora: Volume 1: Synchronic databases . Springer, 2007, pp. 163–180

    2007

  23. [31]

    Appropriating cultural conceptions of child- hood: Participation in conversation,

    M. A. Forrester, “Appropriating cultural conceptions of child- hood: Participation in conversation,” Childhood, vol. 9, no. 3, pp. 255–276, 2002

    2002

  24. [32]

    Two-year-old chil- dren’s production of multiword utterances: A usage-based anal- ysis,

    E. Lieven, D. Salomo, and M. Tomasello, “Two-year-old chil- dren’s production of multiword utterances: A usage-based anal- ysis,” 2009

    2009

  25. [33]

    Acoustical cues and grammatical units in speech to two preverbal infants,

    M. Soderstrom, M. Blossom, R. Foygel, and J. L. Morgan, “Acoustical cues and grammatical units in speech to two preverbal infants,” Journal of Child Language, vol. 35, no. 4, pp. 869–902, 2008

    2008

  26. [34]

    Lena recordings gathered from chil- dren growing up in rossel island. osf,

    A. Cristia and M. Casillas, “Lena recordings gathered from chil- dren growing up in rossel island. osf,” 2023

    2023

  27. [35]

    Vandam cougar homebank corpus,

    M. VanDam, “Vandam cougar homebank corpus,” 2018

    2018

  28. [36]

    Survey on deep learning with class imbalance,

    J. M. Johnson and T. M. Khoshgoftaar, “Survey on deep learning with class imbalance,”Journal of big data, vol. 6, no. 1, pp. 1–54, 2019

    2019

  29. [37]

    Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks,

    Z. Chen, V . Badrinarayanan, C.-Y . Lee, and A. Rabinovich, “Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks,” in International conference on ma- chine learning. PMLR, 2018, pp. 794–803

    2018

  30. [38]

    The fifth ’chime’ speech separation and recognition challenge: Dataset, task and baselines,

    J. Barker, S. Watanabe, E. Vincent, and J. Trmal, “The fifth ’chime’ speech separation and recognition challenge: Dataset, task and baselines,” in Interspeech, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:4397499

    2018

  31. [39]

    Powerset multi-class cross entropy loss for neural speaker diarization,

    A. Plaquet and H. Bredin, “Powerset multi-class cross entropy loss for neural speaker diarization,” in Interspeech, 2023, pp. 3222–3226

    2023

  32. [40]

    Explorations with a voice-type classifier for child- centered audio recordings,

    L. Elbert, “Explorations with a voice-type classifier for child- centered audio recordings,” Cognitive Machine Learning, INRIA Paris (M2 report), Tech. Rep., 2020

    2020

  33. [41]

    Objectively measured teacher and preschooler vocalizations: Phonemic diversity is associated with language abilities,

    S. G. Mitsven, L. K. Perry, Y . Tao, B. E. Elbaum, N. F. John- son, and D. S. Messinger, “Objectively measured teacher and preschooler vocalizations: Phonemic diversity is associated with language abilities,” Developmental science , vol. 25, no. 2, p. e13177, 2022

    2022

  34. [42]

    Reciprocal patterns of peer speech in preschoolers with and without hearing loss,

    L. K. Perry, S. G. Mitsven, S. Custode, L. Vitale, B. Laursen, C. Song, and D. S. Messinger, “Reciprocal patterns of peer speech in preschoolers with and without hearing loss,” Early childhood research quarterly, vol. 60, pp. 201–213, 2022

    2022

Pith tools

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