CHILDES-Aligned is derived from the CHILDES database (TalkBank) and is distributed under TalkBank's terms of use (https://talkbank.org/share/rules.html). Access is reviewed manually. By requesting access you agree to:

Use the data for non-commercial research only.
Cite the BEACON paper and every source CHILDES corpus you use (see corpora.tsv).
Follow the TalkBank Ground Rules, including respecting participant privacy and not redistributing the audio.

How to access this dataset: the official public release is hosted by TalkBank at https://talkbank.org/childes/access/Derived/CHILDES-Aligned.html (audio archives + CSV/JSONL metadata, CC BY-NC-SA 4.0). Please obtain the dataset there. This Hugging Face copy is retained gated, for internal use; access requests are approved manually and general requests may be declined — use the TalkBank release instead.

CHILDES-Aligned: Curated Child-Speech Dataset (BEACON)

English child-speech audio clips with corrected utterance-level timestamps, curated from CHILDES with BEACON (Boundary Estimation via Alignment CONsensus): multiple ASR systems' word timestamps are aligned to the human CHAT transcripts and fused by consensus to recover reliable utterance boundaries, without altering the original transcripts.

Speakers: target child (CHI) only.
Audio: 16 kHz mono WAV, one clip per (merged) child turn.
License/terms: CC BY-NC-SA 4.0 (attribution, non-commercial, share-alike), inherited from CHILDES, under the TalkBank ground rules. Cite the source corpora you use — see corpora.tsv.

Two released versions (paper Table 1)

	General (raw CHAT)	ASR (verbatim)
CHILDES corpora	49	49
Recordings (sessions)	5,533	5,344
Child age range	0;6–14;3	0;6–14;3
Utterances (clips)	501,880	310,983
Audio (hours)	413.3	282.6
Words / vocabulary	2.18 M / 20,581	1.63 M / 18,168
Transcript	raw CHAT	verbatim-normalized

Ages use CHILDES years;months notation (0;6 = 0 years 6 months).

general/ — lossless, general-purpose. All clips keep their raw CHAT text (unintelligibility xxx, pauses, &=events, terminators, etc.), so it is not tied to any single downstream task: distribution, child-language analysis, custom pipelines.
asr/ — a stricter subset of General for ASR training: transcripts are verbatim-normalized and each clip must pass an independent ASR agreement check (drops clips whose audio and transcript disagree). The ASR version is a labeled subset of the same audio — no separate audio store.

The ASR subset is derived from General in two passes: (1) verbatim normalization of the CHAT text (text_normalization.md), dropping empty / unintelligible (xxx/yyy/www) references; (2) an ASR-agreement filter — each clip is transcribed by a fixed out-of-ensemble ASR and dropped if it disagrees with the normalized reference beyond a threshold.

Label schema (`{version}/manifest.jsonl`, one JSON object per clip)

Fields per row, in order:

field	meaning
`audio`	the clip audio, 16 kHz mono, embedded (plays in the viewer)
`text`	transcript — raw CHAT (General) or verbatim-normalized (ASR)
`text_original`	(ASR only) the raw CHAT text before normalization
`start_sec`, `end_sec`	the clip's recovered span on the original recording timeline (seconds)
`duration`	clip length in seconds (equals `end_sec − start_sec`)
`recording_id`	source CHILDES recording/session the clip was cut from
`corpus`	source CHILDES corpus — cite it (see `corpora.tsv`)
`child_age`	child's age at recording in CHILDES `years;months.days` notation, e.g. `2;00.29` = 2 years, 0 months, 29 days (`0;6` = 6 months). Empty where the source did not record it.
`child_sex`	child's sex (`male` / `female`); empty where unspecified
`child_group`	child's group / clinical status where annotated, e.g. `TD` = typically developing; empty (~80% of clips) where unspecified

(child fields come from the source .cha @ID header.) Labels carry the transcript plus minimal participant/provenance metadata only — no per-clip quality/confidence fields. All clips were recovered with the same BEACON configuration (quad_all ensemble, conservative 0.5/0.9 consensus, union spans, ≤1 s same-speaker turn merge, +0.5 s padding).

Method / provenance

English CHILDES recordings → recording-level QC (drop mislabeled non-English, silent, all-model-untranscribable) → per-recording multi-ASR word timestamps (WhisperX, Parakeet, Canary, Qwen3) → BEACON chunk-DP alignment to the CHAT reference utterances → IoU-consensus fusion → adjacent CHI turns with ≤1 s gaps merged (+0.5 s padding) → clip cutting. The ASR version adds normalization + the ASR-agreement filter. Text normalization spec: text_normalization.md.

Limitations

Child-speaker only; adult turns are not released as clips.
Timestamps are model-recovered, not hand-aligned; the hardest child utterances are disproportionately abstained on (no clip) or filtered.
The ASR subset is biased toward judge-transcribable speech; do not use it as the sole source of hard child speech.
English CHILDES only.

Citation

Please cite the paper and the individual source CHILDES corpora you use, listed in corpora.tsv (TalkBank requires citing the corpora):

@article{zheng2026childesaligned,
  title   = {CHILDES-Aligned: A Curated Children's Speech Dataset via Multi-Model Timestamp Ensembling},
  author  = {Zheng, Haolong and Hu, Yuanzhuo and Liang, Xinyu and Sunder, Vishal and Liu, Dancheng and Xiong, Jinjun and Thomas, Samuel and Kingsbury, Brian and Wu, Zhizheng and Hasegawa-Johnson, Mark A.},
  journal = {arXiv preprint arXiv:2607.03670},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.03670}
}

Downloads last month: 55

Total file size:

80.3 GB

Paper for MagicLuke/CHILDES-Aligned

CHILDES-Aligned: A Curated Children's Speech Dataset via Multi-Model Timestamp Ensembling

Paper • 2607.03670 • Published 26 days ago

Access CHILDES-Aligned