Fifteen Years of Child-Centered Long-Form Recordings: Promises, Resources, and Remaining Challenges to Validity

Alejandrina Cristia; Emmanuel Dupoux; Loann Peurey; Lucas Gautheron; Manel Khentout; Marvin Lavechin; Tarek Kunze

REVIEW 4 major objections 4 minor 1 cited by

Fifteen Years of Child-Centered Long-Form Recordings: Promises, Resources, and Remaining Challenges to Validity

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper argues that automated voice-type metrics from child-centered long-form recordings face a wide array of validity threats that automated quality indicators such as C50 and SNR only partially detect.

desk verdict Useful framework and honest null results, but the troubleshooting recipe has an internal SNR contradiction and rests on thin, private data. read the letter →

arxiv 2506.11075 v1 pith:WC3CJTLI submitted 2025-06-04 eess.AS cs.LGcs.SD

Loann Peurey , Marvin Lavechin , Tarek Kunze , Manel Khentout , Lucas Gautheron , Emmanuel Dupoux , Alejandrina Cristia This is my paper

classification eess.AScs.LGcs.SD

keywords long-formrecordingschildlanguagevoicetypeclassifierdaylongaudioecologicalvalidityC50speech-to-noiseratioautomatedqualitycontrol

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 makes the case that long-form, child-worn audio recordings—the gold standard for observing children's natural language environment—produce metrics whose accuracy cannot be assumed. It maps the sources of error into five families: hardware, how the device is worn, recording circumstances, family environment, and the child's own voice. Testing these hypothesized risks on private datasets, the authors find that some predicted failures (e.g., urban vs. rural performance) do not show up, while automated estimates of reverberation (C50) and background noise (SNR) do predict classification errors for adult voices but not for other children. The upshot is a call for caution: automated quality indicators are useful but not yet sufficient, so researchers must combine them with practical collection strategies and normed distributions of vocalization counts.

What carries the argument

The central machinery is the voice-type classifier pipeline applied to daylong child-worn audio: recordings are segmented into silence and four talker classes (key child, adult male, adult female, other children), producing metrics like vocalization counts. The paper's novel analytic tools are two automated acoustic-quality estimates—C50, an estimate of reverberation, and SNR, the speech-to-noise ratio—provided by the Brouhaha system. These serve as candidate troubleshooting indicators: the paper shows that segments flagged with high reverberation (C50 below 30 dB) and low SNR (below 0 dB) yield significantly lower correct-detection rates for FEM and MAL classes, though not for CHI and OCH, thereby defining both the promise and the limits of automated quality control.

What would settle it

A pre-registered, large-scale study that applies VTC and Brouhaha to diverse, publicly shared daylong recordings with human annotations, testing whether vocalization-count errors rise in the segments flagged with C50 below 30 dB or SNR below 0 dB; if the thresholds stop predicting error, the proposed troubleshooting approach fails.

Watch

Extended reading notes

Core claim

The central discovery, in the authors' own framing, is that fifteen years of long-form recording data are not a silver bullet: the validity of automated voice-type metrics is threatened by a wide array of sources ranging from cheap recorders and AI denoising to family routines and the child's own voice. The authors instantiate this claim with a taxonomy of hypothesized effects, then test several of them on private daylong corpora. The tests yield a mixed result: surprising null effects (no uniform urban-rural disadvantage, no sibling-driven drop in key-child detection) coexist with reliable effects of acoustic quality, where C50 and SNR estimates predict lower detection for adult female and male voices. On this basis they argue that while automated quality indicators are promising ingredients for troubleshooting, a fully automated quality-control system is not currently feasible.

Load-bearing premise

The main conclusions assume that the handful of private daylong recordings analyzed here, annotated by people, represent the full range of conditions where these classifiers will be used; if they do not, the patterns found—and the null results—may not generalize.

Editorial extensions

If this is right

Researchers using long-form recordings should not discard data simply because it was collected with cheaper hardware; performance differences across devices were not consistently aligned with price.
Segments with automated C50 below 30 dB or SNR below 0 dB should be treated with caution when reporting adult-directed speech counts, since these sections show lower classifier accuracy for adult voices.
The absence of significant urban-rural or sibling-based performance differences means the classifier may be more robust across settings than expected, but it also means simple demographic heuristics cannot be used for quality triage.
Efforts to build norms for vocalization counts—per class, per child age—are a necessary precondition for outlier-based troubleshooting, since individual counts alone did not reveal quality problems.

Reading between the lines

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

A testable extension would be to measure whether discarding low-SNR and low-C50 segments actually improves the accuracy of downstream vocalization-count estimates at the session level, which the paper leaves open.
The same C50 and SNR indicators could be repurposed as data-collection feedback: worn-device placement could be monitored in near-real time if the indicators track microphone obstruction, potentially flagging off-child recording before the session ends.
Because the null results come from a convenience sample with limited statistical power, the paper's 'surprising' findings should be read as hypotheses for a preregistered multi-site replication rather than as settled facts.
If quality indicators are to become normative, they will likely need to be language- and culture-specific, since the acoustics of households, outdoor noise, and caregiver voice patterns all shape the C50 and SNR distributions.

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.

Desk Editor's Note

read the letter

Short version: this is a review-plus-empirical-checks paper that gives the child-language recording community a reasonable, though not deeply original, framework for thinking about why automated voice-type classifiers underperform. The real value is in Table 1 and in the honest reporting of null results that contradict the authors' own hypotheses. The Bergelson comparison across seven datasets is new to me and useful. But I would not lean on the troubleshooting guidance as printed. Section 5.2.3 says correct detection for female and male adult classes is lower when lower signal-to-noise is detected; the Figure 3b caption says performance is significantly higher for sections with low (estimated) SNR. Those cannot both be right. This is not a wording quibble—the whole suggested workflow is to filter out problematic sections, and the direction determines which sections to exclude. The C50 and SNR thresholds (30 dB, 0 dB) come with no sensitivity analysis, so we cannot rule out post-hoc selection. The empirical base is also thin: private datasets, small convenience samples, and an urban-rural comparison that confounds setting with hardware. Those are real limits, and the paper acknowledges some of them. The null results for siblings and urban-rural are interesting but not essential to the central framework. On the whole, the conceptual message—automated metrics are not a silver bullet, and users should be cautious—is defensible and well-supported by the literature review. The paper is useful as an entry point and a cautionary summary. It deserves peer review, but the SNR contradiction and the threshold rationale need to be fixed before the specific troubleshooting claims are relied upon. I would bring it to a reading group as a discussion piece, and I would cite Table 1 or the Bergelson numbers, but I would wait for a corrected version before citing the C50/SNR results.

Referee Report

4 major / 4 minor

Summary. This paper provides an entry point to the use of child-centered long-form audio recordings in child language research, summarizing existing resources, annotation schemes, and ethical considerations. Its novel contribution is a taxonomy of factors (hardware, operating error, setting, family, child) hypothesized to affect the performance of automated voice-type classifiers, along with exploratory analyses on private datasets intended to evaluate these hypotheses. The analyses compare VTC performance across hardware, urban versus rural settings, children with and without siblings, and audio sections partitioned by automated estimates of reverberation (C50) and signal-to-noise ratio (SNR). The paper concludes that automated quality indicators are promising but not yet sufficient, and that many hypothesized negative effects were not verified.

Significance. The paper addresses a practical and timely problem for the growing community using long-form recordings. Its taxonomy in Table 1 is a useful organizing framework, and the paper is transparent about null results and the limitations of its private datasets. If the empirical analyses were internally consistent and appropriately hedged, the paper would be a valuable resource for researchers deciding how to assess automated metrics. The open discussion of hardware variability and the use of open-source tools (VTC, Brouhaha) are strengths. However, the internal contradiction in the SNR analysis and the presence of confounds in the urban-rural comparison currently limit the reliability of the paper's central practical guidance.

major comments (4)

[5.2.3 and Fig. 3b caption] The direction of the SNR effect is contradictory between the text and the figure caption. Section 5.2.3 states that correct detection for female and male adult voice classes is lower when lower signal-to-noise is detected, while the Fig. 3b caption states that VTC performance is significantly higher in the FEM and MAL categories for audio sections with high C50 and low (estimated) SNR. If 'low SNR' is interpreted literally, one passage says low SNR degrades performance and the other says it accompanies higher performance. Because the proposed troubleshooting workflow uses SNR to identify problematic sections, the sign of this association determines which segments a user would filter. Please correct the text or figure, and state the direction unambiguously.
[5.2.3] The analysis uses hand-chosen thresholds (C50 < 30 dB versus >= 30 dB; SNR < 0 dB versus >= 0 dB) without any sensitivity analysis or evidence that these thresholds were not selected post hoc from the same data. The reported differences in Figure 3 could depend heavily on these cutoff choices. Please either provide a sensitivity analysis across a range of thresholds or explicitly label the thresholds as exploratory and refrain from giving them the status of practical guidance.
[5.2.2 and Fig. 2b] The urban-rural comparison confounds community type with recording hardware. The rural sample (tsimane2017) contains both LENA and USB recordings, while the urban datasets (bergelson, lucid, warlaumont, winnipeg) are LENA-only. The null result for 'no urban-rural difference' cannot be attributed to setting alone, since hardware is known to affect VTC performance (as shown in Fig. 2a). Please restrict the comparison to a single hardware type or state this confound explicitly as a severe limitation of the analysis.
[5.2.2] The sibling analysis is underpowered and the significant OCH result appears to be driven by a near-floor effect: average F-score is 10% for children without siblings versus 40% with siblings, with the OCH label being extremely rare in the former. The null result for CHI (p = .52, N = 18 vs 20) has low statistical power to detect a meaningful difference. The text should temper the claim that 'sibling presence does not affect CHI performance' and should discuss the base-rate issue for OCH before interpreting the significant p-value.

minor comments (4)

[5.2.3] Typo: 'promissing' should be 'promising'.
[Fig. 3 caption] The caption 'low (estimated) SNR' is confusing and appears to contradict the main text. Please rephrase to make clear whether the condition is low estimated SNR (i.e., noisy) or high estimated SNR.
[5.2.1] The 'unpublished analysis' comparing USB devices and LENA is presented with minimal methodological detail. Please add a sentence describing the audio material, annotation procedure, and the number of recordings, or refer readers to the linked repository more explicitly.
[5.2.4] The final paragraph of Section 5.2.4 is presented as a failed exploratory idea. Consider shortening it or moving it to the Discussion to avoid an anticlimactic section ending.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical claims rest on comparing an automated classifier to human annotations, not on self-referential definitions or fitted inputs.

full rationale

The paper's central empirical contribution (Section 5) is an evaluation of an open-source voice-type classifier (VTC) against human annotations, with automated C50 and SNR estimates from Brouhaha used as candidate troubleshooting indicators. This is a standard external-validation setup: VTC outputs are scored against human labels, and those scores are correlated with acoustical estimates. No quantity is defined in terms of the quantity it is claimed to predict, and no fitted parameter is renamed as a prediction. The thresholds (C50<30, SNR<0) are presented descriptively in Figure 3, not as parameters fitted to force the reported differences. The paper's self-citations to VTC [15], Brouhaha [26], ACLEW [12], and the LENA review [19] are references to tools and prior work, but they are not load-bearing in a circular way: VTC is independently evaluated here against human ground truth, and Brouhaha provides features rather than the target outcome. The authors also explicitly flag limitations, stating 'we have not been able to prove this with the data available to us' (Section 5.2.4) and 'it would remain to be shown that more accurate measures ... may ensue if sections ... are excluded' (Section 5.2.3), which further weakens any claim that the proposed indicators were validated by construction. One internal inconsistency exists: Section 5.2.3 states detection is lower with lower signal-to-noise, while the Figure 3b caption states performance is higher with low estimated SNR. That is a correctness and consistency defect in the SNR guidance, but it is not a circularity, because the contradiction does not make the result equivalent to its inputs. Overall, the derivation chain is not circular; the paper's conclusions are appropriately hedged and depend on empirical comparison rather than self-definition.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper makes no parametric derivation, so the free-parameter count is limited to the two hand-chosen thresholds used in the troubleshooting analysis. The central claims rest on assumptions about dataset representativeness, the validity of human annotation as ground truth, and the trustworthiness of automated audio-quality estimates.

free parameters (2)

C50 threshold for low/high reverb = 30 dB
Hand-chosen cutoff to split audio windows into C50 < 30 and C50 >= 30; not derived from data or theory (Section 5.2.3).
SNR threshold for low/high noise = 0 dB
Hand-chosen cutoff to split audio into SNR < 0 and SNR >= 0; not fitted to the data (Section 5.2.3).

assumptions (3)

domain assumption Human annotations serve as ground truth for evaluating VTC.
F-scores are computed against human annotations from datasets described in Section 5.2, using the annotation scheme in reference [12].
domain assumption The private datasets are representative enough to support the observed null and positive effects.
The analyses rely on convenience corpora shared with the authors, with unknown sampling properties (Section 5.2).
domain assumption Brouhaha C50 and SNR estimates are valid proxies for reverberation and noise.
The paper uses automated C50 and SNR as troubleshooting indicators without external validation of these estimates on child-centered recordings (Section 5.2.3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Fifteen Years of Child-Centered Long-Form Recordings: Promises, Resources, and Remaining Challenges to Validity." pith.science (2026). https://pith.science/paper/WC3CJTLI

@misc{pith2026250611075,
  author       = {Pith},
  title        = {Pith review of: Fifteen Years of Child-Centered Long-Form Recordings: Promises, Resources, and Remaining Challenges to Validity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WC3CJTLI}},
  note         = {Machine review of arXiv:2506.11075}
}

read the original abstract

Audio-recordings collected with a child-worn device are a fundamental tool in child language research. Long-form recordings collected over whole days promise to capture children's input and production with minimal observer bias, and therefore high validity. The sheer volume of resulting data necessitates automated analysis to extract relevant metrics for researchers and clinicians. This paper summarizes collective knowledge on this technique, providing entry points to existing resources. We also highlight various sources of error that threaten the accuracy of automated annotations and the interpretation of resulting metrics. To address this, we propose potential troubleshooting metrics to help users assess data quality. While a fully automated quality control system is not feasible, we outline practical strategies for researchers to improve data collection and contextualize their analyses.

Figures

Figures reproduced from arXiv: 2506.11075 by the authors.

**Figure 2.** Surprising (non-)effects of hardware and setting (urban versus rural). Each point indicates one child. Data for the left panel: tsimane2017. Data for the right panel: tsimane2017 (rural, LENA and USB); bergelson, lucid, warlaumont, and winnipeg (all urban, LENA only) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** VTC performance is significantly higher in the FEM and MAL categories for audio sections with high C50 (low estimated reverberation) and low (estimated) SNR. Bars indicate averages across all annotated 1-second windows (the Y variable is 1 if the speaker was correctly detected, 0 otherwise). Data: tsimane2017, bergelson, warlaumont, and winnipeg. researchers could doubt the quality of a recording when the child had… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

BabyHuBERT: Multilingual Self-Supervised Learning for Segmenting Speakers in Child-Centered Long-Form Recordings
eess.AS 2025-09 conditional novelty 6.0 of 10

Pre-training HuBERT on 13,164 hours of multilingual child-centered audio and fine-tuning for voice type classification yields 64.6% average F1, beating English-only and adult-speech baselines by 5.9 and 13.2 points.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages · cited by 1 Pith paper

[1]

Challenges in Speech Data Collection, Curation, and Annotation

Introduction Audio-recordings collected with a child-worn device are a fun- damental tool in research on child language [1]. The last decade has seen increasing use of long-form recordings, collected as children wear a device typically over a whole day, to capture what children hear and what they say [2, 3, 4, 5, 6, to cite just a few]. In the context of ...

work page
[2]

Fifteen Years of Child-Centered Long-Form Recordings: Promises, Resources, and Remaining Challenges to Validity

Definition and Key Uses of Long-Form Recording Data Compared to short-form recordings, which often take place during a specific activity, long-form recordings aim to capture speech behavior “in the wild”: Families are asked to go about their normal day as much as possible. The hope is that the par- ticipation burden is lowered for families in this way, an...

work page Pith review arXiv 2025
[3]

In addition to introductory papers [10, 11], there exist also some video tutorials 1 and structured documentation 2

Extant Resources for Facilitating Data Collection, Human Annotation and Sharing A range of resources exist to help researchers navigate the collection, annotation, and sharing of long-form recordings. In addition to introductory papers [10, 11], there exist also some video tutorials 1 and structured documentation 2. A sum- mer school with spin-offs in sev...

work page 2025
[4]

The LENA software system [16, 17, 18], a widely used proprietary tool for long-form recordings, has been the focus of extensive ac- curacy evaluations

Accuracy of Automated Voice Type Classifiers for Long-Form Recordings In many cases, researchers using this technique must rely on automated metrics rather than human annotations. The LENA software system [16, 17, 18], a widely used proprietary tool for long-form recordings, has been the focus of extensive ac- curacy evaluations. A 2020 review of 33 bench...

work page 2020
[5]

Brouhaha

Challenges to the Interpretation of Automated Metrics, and Potential Solutions Having set the background on this technique for Special Ses- sion attendees, in this section we move on to the major novel contribution of this paper, which is providing a framework for potentially understanding (and eventually correcting) issues that complexify the interpretat...

work page
[6]

In this paper, we provide readers with an entry point to this tech- nique and the many resources that are accumulating around it

Discussion Long-form recordings aspire to give us a truthful view of the speech spontaneously produced around and by young children. In this paper, we provide readers with an entry point to this tech- nique and the many resources that are accumulating around it. Long-form recording data are not a silver bullet. Fifteen years of experience suggest there co...

work page
[7]

https://youtube.com/playlist?list= PLExuQICGVy3gMQllaZ5OLDOQ1KDMuhoMr

work page
[8]

doi:10.5281/zenodo.6685828

work page doi:10.5281/zenodo.6685828

Show all 37 references

[9]

https://lfraz2025.sciencesconf.org/
[10]

homebank.talkbank.org
[11]

https://gin.g-node.org/LAAC-LSCP/ longform-hardware-audio-test
[12]

The childes system,

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

1998
[13]

Bursty, irregular speech input to children predicts vocabulary size,

M. Cychosz, R. R. Romeo, J. R. Edwards, and R. S. Newman, “Bursty, irregular speech input to children predicts vocabulary size,” Developmental Science, vol. 28, no. 1, p. e13590, 2025

2025
[14]

Informing mothers about the benefits of conversing with infants: Experimental evidence from ghana,

P. Dupas, C. Falezan, S. Jayachandran, and M. P. Walsh, “Informing mothers about the benefits of conversing with infants: Experimental evidence from ghana,” National Bureau of Eco- nomic Research, Tech. Rep., 2023. [Online]. Available: https:// www.nber.org/system/files/workin...

2023
[15]

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
[16]

Everyday language input and production in 1,001 children from six continents,

E. Bergelson, M. Soderstrom, I.-C. Schwarz, C. F. Rowland, N. Ram´ırez-Esparza, L. R. Hamrick, E. Marklund, M. Kalash- nikova, A. Guez, M. Casillas et al. , “Everyday language input and production in 1,001 children from six continents,” Proceed- ings of the National Academy of...

2023
[17]

Characteriza- tion of children’s verbal input in a forager-farmer population us- ing long-form audio recordings and diverse input definitions,

C. Scaff, M. Casillas, J. Stieglitz, and A. Cristia, “Characteriza- tion of children’s verbal input in a forager-farmer population us- ing long-form audio recordings and diverse input definitions,” In- fancy, vol. 29, no. 2, pp. 196–215, 2024

2024
[18]

Modeling early phonetic acquisition from child-centered audio data,

M. Lavechin, M. de Seyssel, M. M ´etais, F. Metze, A. Mohamed, H. Bredin, E. Dupoux, and A. Cristia, “Modeling early phonetic acquisition from child-centered audio data,” Cognition, vol. 245, p. 105734, 2024

2024
[19]

Semi-automatic as- sessment of vocalization quality for children with and without an- gelman syndrome,

L. R. Hamrick, A. Seidl, and B. L. Kelleher, “Semi-automatic as- sessment of vocalization quality for children with and without an- gelman syndrome,” American Journal on Intellectual and Devel- opmental Disabilities, vol. 128, no. 6, pp. 425–448, 2023

2023
[20]

Using big data from long-form recordings to study development and optimize societal impact,

M. Cychosz and A. Cristia, “Using big data from long-form recordings to study development and optimize societal impact,” in Advances in child development and behavior. Elsevier, 2022, vol. 62, pp. 1–36

2022
[21]

A step-by-step guide to collecting and analyzing long-format speech environment (lfse) recordings,

M. Casillas and A. Cristia, “A step-by-step guide to collecting and analyzing long-format speech environment (lfse) recordings,” Collabra: Psychology, vol. 5, no. 1, p. 24, 2019

2019
[22]

Daylong egocentric recordings in small-and large-scale language communities: A practical intro- duction,

M. Casillas and K. Casey, “Daylong egocentric recordings in small-and large-scale language communities: A practical intro- duction,” Advances in child development and behavior , vol. 66, pp. 29–53, 2024

2024
[23]

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
[24]

Longform recordings of everyday life: Ethics for best practices,

M. Cychosz, R. Romeo, M. Soderstrom, C. Scaff, H. Ganek, A. Cristia, M. Casillas, K. De Barbaro, J. Y . Bang, and A. Weisleder, “Longform recordings of everyday life: Ethics for best practices,” Behavior research methods , vol. 52, pp. 1951– 1969, 2020

1951
[25]

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
[26]

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

2020
[27]

A guide to understanding the design and purpose of the LENA® system,

J. Gilkerson and J. A. Richards, “A guide to understanding the design and purpose of the LENA® system,” LENA Foundation: Boulder, CO, 2020

2020
[28]

The LENA natural language study,

J. Gilkerson and J. A. Richards, “The LENA natural language study,” Boulder, CO: LENA Foundation. Retrieved March, vol. 3, p. 2009, 2008

2009
[29]

Reliability of the LENA™ lan- guage environment analysis system in young children’s natural language home environment (technical report LTR-05-2),

D. Xu, U. Yapanel, and S. Gray, “Reliability of the LENA™ lan- guage environment analysis system in young children’s natural language home environment (technical report LTR-05-2),” 2008

2008
[30]

Accuracy of the lan- guage environment analysis system segmentation and metrics: A systematic review,

A. Cristia, F. Bulgarelli, and E. Bergelson, “Accuracy of the lan- guage environment analysis system segmentation and metrics: A systematic review,” Journal of Speech, Language, and Hearing Research, vol. 63, no. 4, pp. 1093–1105, 2020

2020
[31]

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

J. Li, M. Hasegawa-Johnson, and K. Karahalios, “Enhancing child vocalization classification with phonetically-tuned embeddings for assisting autism diagnosis,” inProceedings of the Annual Con- ference of the International Speech Communication Association, INTERSPEECH, 2024, pp...

2024
[32]

The language 0-5 project,

C. F. Rowland, S. Durrant, M. Peter, A. Bidgood, J. Pine, and L. S. Jago, “The language 0-5 project,” Jun 2024. [Online]. Available: osf.io/kau5f

2024
[33]

San joaquin valley homebank corpus (formerly the warlaumont homebank corpus),

A. S. Warlaumont, G. M. Pretzer, S. Mendoza, S. Schneider, J. Mutrie, L. Lopez, E. A. Walle, and C. T. Kello, “San joaquin valley homebank corpus (formerly the warlaumont homebank corpus),” 2024. [Online]. Available: https://doi.org/10.21415/ T54S3C

2024
[34]

Homebank english mcdivitt corpus,

K. McDivitt and M. Soderstrom, “Homebank english mcdivitt corpus,” 2016. [Online]. Available: https://homebank.talkbank. org/access/Secure/McDivitt.html

2016
[35]

Vandam cougar homebank corpus,

M. VanDam, “Vandam cougar homebank corpus,” Home- bank, 2018, available at: https://homebank.talkbank.org/access/ Password/Cougar.html

2018
[36]

Vandam public 5-minute homebank corpus,

M. VanDam, “Vandam public 5-minute homebank corpus,” Homebank, 2018, available at: https://homebank.talkbank.org/ access/Public/VanDam-5minute.html

2018
[37]

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,”ASRU, 2023

2023

Pith tools

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

[1] [1]

Challenges in Speech Data Collection, Curation, and Annotation

Introduction Audio-recordings collected with a child-worn device are a fun- damental tool in research on child language [1]. The last decade has seen increasing use of long-form recordings, collected as children wear a device typically over a whole day, to capture what children hear and what they say [2, 3, 4, 5, 6, to cite just a few]. In the context of ...

work page

[2] [2]

Fifteen Years of Child-Centered Long-Form Recordings: Promises, Resources, and Remaining Challenges to Validity

Definition and Key Uses of Long-Form Recording Data Compared to short-form recordings, which often take place during a specific activity, long-form recordings aim to capture speech behavior “in the wild”: Families are asked to go about their normal day as much as possible. The hope is that the par- ticipation burden is lowered for families in this way, an...

work page Pith review arXiv 2025

[3] [3]

In addition to introductory papers [10, 11], there exist also some video tutorials 1 and structured documentation 2

Extant Resources for Facilitating Data Collection, Human Annotation and Sharing A range of resources exist to help researchers navigate the collection, annotation, and sharing of long-form recordings. In addition to introductory papers [10, 11], there exist also some video tutorials 1 and structured documentation 2. A sum- mer school with spin-offs in sev...

work page 2025

[4] [4]

The LENA software system [16, 17, 18], a widely used proprietary tool for long-form recordings, has been the focus of extensive ac- curacy evaluations

Accuracy of Automated Voice Type Classifiers for Long-Form Recordings In many cases, researchers using this technique must rely on automated metrics rather than human annotations. The LENA software system [16, 17, 18], a widely used proprietary tool for long-form recordings, has been the focus of extensive ac- curacy evaluations. A 2020 review of 33 bench...

work page 2020

[5] [5]

Brouhaha

Challenges to the Interpretation of Automated Metrics, and Potential Solutions Having set the background on this technique for Special Ses- sion attendees, in this section we move on to the major novel contribution of this paper, which is providing a framework for potentially understanding (and eventually correcting) issues that complexify the interpretat...

work page

[6] [6]

In this paper, we provide readers with an entry point to this tech- nique and the many resources that are accumulating around it

Discussion Long-form recordings aspire to give us a truthful view of the speech spontaneously produced around and by young children. In this paper, we provide readers with an entry point to this tech- nique and the many resources that are accumulating around it. Long-form recording data are not a silver bullet. Fifteen years of experience suggest there co...

work page

[7] [7]

https://youtube.com/playlist?list= PLExuQICGVy3gMQllaZ5OLDOQ1KDMuhoMr

work page

[8] [8]

doi:10.5281/zenodo.6685828

work page doi:10.5281/zenodo.6685828