CHILDES-Aligned Derived Corpus

Haolong Zheng
University of Illinois

Mark Hasegawa-Johnson
University of Illinois

The CHILDES-Aligned corpus is downloadable here.

Citation

Zheng, H., Hu, Y., Liang, X., Sunder, V., Liu, D., Xiong, J., Thomas, S., Kingsbury, B., Wu, Z., & Hasegawa-Johnson, M. A. (2026). CHILDES-Aligned: A Curated Children's Speech Dataset via Multi-Model Timestamp Ensembling. arXiv preprint arXiv:2607.03670. https://arxiv.org/abs/2607.03670 with PDF available here.

When using this database, please cite the above paper.

Project Description

CHILDES-Aligned is a corpus of **501,880 transcript-paired English child-speech clips (413.3 hours)** with corrected utterance-level timestamps, derived from CHILDES. Utterance boundaries were recovered with BEACON (Boundary Estimation via Alignment CONsensus): word-level timestamps from four ASR systems (WhisperX, Parakeet, Canary, Qwen3) are aligned to the original human CHAT transcripts and fused by consensus voting, without altering the transcripts. Adjacent target-child utterances separated by at most 1 s are merged into single continuous clips.

All clips are target-child (CHI) speech from 49 English CHILDES corpora (child ages 0;6–14;3), cut as 16 kHz mono WAV.

There are two versions of the database.

General is lossless: clips keep their raw CHAT text (unintelligibility, pause, and event markers included), suiting child-language analysis and custom pipelines. It includes 501,880 clips with 413.3 hours from 5,533 recording sessions.
ASR is a stricter subset of the same audio for speech-recognition training: transcripts are verbatim-normalized (fillers, contractions, and spoken number forms preserved; CHAT markers removed) and every clip passes an independent ASR agreement check. It includes 310,983 clips with 282.6 hours from 5,344 recording sessions.

Metadata per clip (CSV and JSONL provided) provide: relative audio path, transcript (text; the ASR version also carries text_original raw CHAT), the recovered span on the original recording (start_sec, end_sec), duration, recording_id, source corpus, and child_age (CHILDES years;months.days notation, e.g. 2;00.29 = 2 years, 0 months, 29 days), child_sex, child_group. corpora.tsv lists all 49 source corpora with PIDs and clip/hour counts.

Downloads

CHILDES-Aligned-metadata.zip — general.csv, asr.csv (+ JSONL equivalents), dataset card, datasheet, text-normalization spec, corpora.tsv, checksums.
CHILDES-Aligned-audio (45 GB total) — the 501,880 clip WAVs; all archives unpack into one shared audio/ tree that the metadata paths resolve against.

Copyright is CC BY-NC-SA 4.0 (attribution, non-commercial, share-alike), inherited from CHILDES, under the TalkBank ground rules.