CHILDES-KWS Derived Corpus

Hüseyin Karaca Bilkent University Middle East Technical University huseyin.karaca@bilkent.edu.tr

Citation

Karaca, H. (2026). CHILDES-KWS: Children's mama, papa and dada for keyword spotting [derived corpus]. TalkBank.

When using this database, please also cite MacWhinney (2000), the PhonBank reference if you use the French clips, and the reference of every source corpus you use. The corpora, their DOIs and their required references are listed in metadata/corpora.csv and metadata/references.bib.

Project Description

Keyword spotters are usually trained on adult speech. A device that should react when a toddler calls mama or papa needs the opposite: short, unclear words from children who are just learning to say them. Recordings like that already sit in CHILDES and PhonBank, but they are buried inside hour-long sessions and time-stamped by whole utterance rather than by word.

CHILDES-KWS is those words, located. In the 20 CHILDES and PhonBank corpora it draws on, the child utterances containing mama, papa or dada were found in the transcripts and cut from their recordings, and 1,424 clips of the word itself were cut from those utterances for the voice subsystem of a student smart-baby-monitor project at METU in 2024.

All audio is 16 kHz mono WAV. The corpus has two layers over the same recordings.

• Utterances — 2,957 complete child utterances containing one of the three words, from 54 children, 101 minutes in all, each with its CHAT transcript line. This is everything the transcript search found, whether or not it was later used, so it also serves as the pool the keyword clips were selected from.
• Keywords — 1,424 clips of about a second holding the word itself, from 49 children, 22 minutes in all, with train / validation / test splits assigned by child so that no child appears in two splits.

Checking status and use

The keyword metadata distinguishes 1,103 clips marked by_ear, 110 marked whole_utterance, and 211 marked unchecked. These are different levels of evidence; the complete collection is not listener-verified. The transcript-label field is yes for 1,405 clips, no for 14 and unknown for five. Filter these fields according to your task. Five keyword rows have no utterance ID; the audio is included and the missing historical link is left blank.

The supplied train, validation and test groups contain 1,137, 132 and 155 keyword clips. Do not train on the utterance layer and test on keyword clips drawn from the same recordings. The supplied speaker groups have no identifier overlap; this is not an independent validation of speaker identity.

The contributor reports that the utterance layer is reproducible from a TalkBank download: extract_utterances.py repeats the CLAN search (kwal +sdada +smama +spapa +t*CHI +f) by reading the CHAT files directly, and reports agreement with the CLAN output saved in 2024 on all 2,999 utterance/keyword pairs. cut_utterances.py then cuts the spans; the contributor reports that 99.9% of them line up with the 2024 clips to within a millisecond at a median correlation of 1.000. The keyword layer is the edited one: a listener trimmed and edited those clips in 2024, and they are the clips the project trained on, so they ship as audio.

Metadata (CSV) gives, per utterance: a stable id, the keywords the line contains, the collection, corpus, child id, age and sex, the recording and the span inside it, the CHAT transcript line, the TalkBank PID of the transcript, the rate the recording was made at, loudness, and how many keyword clips came out of the utterance. Per keyword clip: the label, the split, the utterance it came from, its own span, how it was segmented, how the word in it was checked, whether the label agrees with the transcript, how the clip was traced back to its source, usable bandwidth, loudness and duration. corpora.csv lists all 20 source corpora with DOIs, required references, what each contributed, and source-specific notes.

Contents

• metadata/ — utterances.csv, keywords.csv, corpora.csv, splits.json, references.bib, checksums.sha256
• audio/utterances/ and audio/keywords/ — the audio, about 234 MB
• scripts/ — the search, the cutting, the linking and the splits

Note. papa is almost entirely French while mama and dada are almost entirely English, and the corpora differ in microphones, rooms and decades. These factors may create an imbalance or a bias in your training.

Copyright is CC BY-NC-SA 4.0 (attribution, non-commercial, share-alike), inherited from CHILDES, under the TalkBank ground rules. The scripts are Apache-2.0.