This dataset is a table of pointers into recordings of young children in the CHILDES and PhonBank databases of TalkBank; it contains no audio. Access is granted automatically; by requesting it you agree to the licence (CC BY-NC-SA 4.0) and to the TalkBank ground rules (https://talkbank.org/0share/rules.html).

MaPaDa

Children's mama, papa and dada for keyword spotting — as pointers, not audio. 1,424 hand-curated clips of 49 young children saying their first kinship words, each one located in a CHILDES or PhonBank recording by corpus, child, recording and time span. The splits never share a child.

TalkBank does not permit redistribution of media from its corpora, so this dataset ships the time stamps and scripts/rebuild_from_childes.py cuts the clips from your own copy of the recordings. A free TalkBank account gets you the media from media.talkbank.org.

Project page and data card: https://hkaraca.com/mapada/
Code and metadata: https://github.com/huseyin-karaca/mapada
Archive with DOI: Zenodo, doi:10.5281/zenodo.23648586

	mama	papa	dada	children
train	462	415	260	34
validation	49	52	31	7
test	73	52	30	8

Getting the clips

# 1. download the recordings clips.csv names into MEDIA/ (free TalkBank account, ffmpeg needed)
# 2. cut the clips
python scripts/rebuild_from_childes.py --media-root MEDIA --out clips

That writes clips/<label>/<clip_id>.wav at 16 kHz mono, plus a manifest.csv with the split of each one. The pointer table itself loads without any audio:

from datasets import load_dataset

ds = load_dataset("huseyin-karaca/mapada")
ds["train"][0]  # {'clip_id': ..., 'label': ..., 'corpus': ..., 'media_id': ..., 'clip_start_ms': ..., ...}

Each row gives label, corpus, child, age_months, media_id, the utterance's and the clip's time spans in the source recording (utt_start_ms, utt_end_ms, clip_start_ms, clip_end_ms), segmentation (auto, manual_trim or manual_edit), rebuildable, attribution, word_check, bandwidth_khz and rms_dbfs. The columns are described on the data card.

1,198 of the 1,424 clips have a time span and rebuild exactly (rebuildable = yes). The other 226 were cut and edited by hand in 2024 and no time span survives for them, so they cannot be reproduced; their utterance spans are in the table, and --manual-edit approx writes the last second of the utterance instead, which is not the clip that was curated.

Read before training on it

papa comes almost entirely from French-speaking children and mama and dada almost entirely from English ones, and the corpora differ in recording setup. A model trained only on these clips can separate the classes by language or channel instead of by word. Use the clips as positives inside a keyword spotter that also sees other speech, noise and silence, and check performance per corpus. Most clips are one second long; short utterances are shorter and some hand-edited clips run to two seconds, so pad or crop to your model's input length. 211 dada clips are the last second of a longer utterance and are not known to have been listened to (word_check = unchecked); filter them out if you need clean positives.

Licence and citation

The pointer table is under CC BY-NC-SA 4.0, the terms TalkBank shares its corpora under. The recordings themselves stay with TalkBank: using them means following the ground rules, which do not allow you to redistribute the media or clips cut from it. Utterances from the Braunwald and Nelson corpora, which ask for permission before their audio is used publicly, are listed separately in metadata/excluded_utterances.csv.

Cite the dataset, MacWhinney (2000), and the reference of every corpus you use; all are in references.bib.

@misc{karaca2026mapada,
  author       = {Karaca, H{\"u}seyin},
  title        = {{MaPaDa}: Children's \emph{mama}, \emph{papa} and \emph{dada} for Keyword Spotting},
  year         = {2026},
  version      = {1.1.0},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23648586},
  howpublished = {\url{https://hkaraca.com/mapada/}}
}

CHILDES is supported by NICHD grant HD082736 and PhonBank by NIH-NICHHD grant RO1-HD051698. The recordings belong to the families who shared them and to the researchers who contributed them to TalkBank.

Downloads last month: -

Total file size:

674 kB