Skip to content

Data card

This card follows the questions of Datasheets for Datasets (Gebru et al., 2021). Numbers refer to version 1.0.0.

Motivation

Why was the dataset created? For the voice subsystem of a smart baby monitor built as a METU Electrical and Electronics Engineering capstone project in 2023–24 (the project). The monitor had to notice, on a microcontroller and without the cloud, when a toddler called mama or papa, and when a baby cried. Public keyword-spotting data is adult speech; child-produced keywords had to be collected.

Who created it? Hüseyin Karaca, then an undergraduate at METU, built the extraction scripts and curated the clips (March–May 2024), and prepared this release (October 2026). The capstone had no external funding.

Composition

What are the instances? Short WAV clips (16 kHz, mono, 16-bit) of a single child saying mama, papa or dada, each cut from a recording in a CHILDES or PhonBank corpus. There are 1,424 clips from 49 children in 16 corpora.

collection corpus mama papa dada children
Eng-UK MPI-EVA-Manchester 300 13 172 4
French KernFrench 4 383 4 4
Eng-NA Brent 207 0 103 15
French GoadRose 2 77 0 2
Eng-NA Bloom 40 0 2 1
Eng-NA Weist 2 26 1 4
Eng-UK Edinburgh 6 0 23 6
Eng-NA Gleason 7 4 2 4
Eng-NA MacWhinney 2 7 3 1
Eng-UK Sekali 5 5 0 2
Eng-NA Snow 1 2 5 1
Eng-UK Thomas 4 0 2 1
Eng-NA Sachs 3 2 0 1
Eng-NA VanHouten 1 0 1 1
Eng-UK Forrester 0 0 2 1
Eng-NA Soderstrom 0 0 1 1
total 584 519 321 49

MacWhinney's two children cannot be told apart from the file names and count as one speaker; so do the single-child corpora.

What does each instance consist of? The audio and one row of metadata/clips.csv:

column meaning
clip_id corpus-child-uttstart-n, unique
label mama, papa or dada
file path of the WAV file inside the release
collection, corpus CHILDES / PhonBank collection and corpus; DOIs and citations in metadata/corpora.csv
child the child's identifier as used in the corpus (a pseudonym or code chosen by the corpus)
age_months age at the recording, from the CHAT age code in the file name; empty when the name carries none
media_id the recording, as <child folder>_<file> or <file>
utt_start_ms, utt_end_ms the utterance's time stamps in the recording, from the transcript
clip_start_ms, clip_end_ms the clip's time span in the recording; empty for manual_edit clips
duration_s clip length
segmentation auto: an automatic window of the utterance; manual_trim: a window chosen by hand; manual_edit: audio edited by hand (see below)
attribution how the source utterance is known: filename, matched or matched_probable (how it was made)
word_check how the word is known to be in the clip: by_ear (listened to in 2024), whole_utterance (the clip spans the whole transcribed utterance) or unchecked (see below)
bandwidth_khz usable bandwidth: half the source recording's sample rate, at most 8 (4 for the 8 clips from 8 kHz recordings), and 4 for manual_edit clips
rms_dbfs loudness of the clip
legacy_file the clip's file name in the 2024 working folders, for traceability
split train, validation or test

Is there a label for each instance? Yes: the word the CHILDES transcriber wrote on the child's tier of the transcript. Clips were cut around that word; the mama and papa clips were then screened by ear, and the dada clips may not have been (see word_check).

Are there recommended splits? Yes. Each child is assigned to exactly one split, for every label, so the splits measure generalisation to unseen children. The assignment is a seeded search (seed 494) for the proportions closest to 80/10/10 in every class; the speakers of each split are listed in metadata/splits.json.

mama papa dada children
train 462 415 260 34
validation 49 52 31 7
test 73 52 30 8
total 584 519 321 49

The test set has eight children (Brent i1, q1 and s1; Edinburgh Amy; KernFrench Baptiste; Weist Ben, Emily and Jillian), so test scores have wide error bars. Its papa clips come from one French child (27) and two English children (25), which makes the test set less exposed to the language confound below than the training set.

Is anything missing? Utterances from the Braunwald (101) and Nelson (6) corpora are held back because their terms ask for permission before audio is shared publicly; their time stamps are in metadata/excluded_utterances.csv. Material the project used that was not from CHILDES or PhonBank is not included (see how it was made).

Are there errors or noise?

  • A clip contains the target word but may also contain the end of the rest of the utterance, an adult talking over the child, toys, or room noise. Nothing was denoised.
  • 211 dada clips (word_check = unchecked), 169 of them from Manchester, are the last second of a longer utterance and are not known to have been listened to, so in some of them the word may fall outside the clip. The other 110 dada clips span their whole utterance. Drop the unchecked ones if you need clean positives.
  • Transcribers write what they hear; a young infant's mama or dada may be babbling rather than a word for a parent.
  • manual_edit clips (226) were edited on an 8 kHz working copy, so they carry nothing above 4 kHz, and they were peak-normalised; the other clips keep the level of the recording.
  • 14 clips are very quiet (below −50 dBFS), mostly from Brent and Manchester, and some Brent recordings have a steady background hum louder than the child. A check that does not involve listening (tools/check_activity.py) found, in every clip, frequency bands whose level rises and falls well beyond what steady noise does, so none is silent or hum only; all were kept.
  • Durations vary: most clips are exactly 1 s, short utterances are shorter, and some manual_edit clips run to 2.1 s.

Does it contain confidential data or data about individuals? It contains recordings of children, contributed to TalkBank by the researchers who collected them, with the families' consent as described in each corpus's documentation. Children are identified only by the identifiers their corpus uses. No information was added beyond what TalkBank already publishes.

Known biases

Label and language are confounded. papa is almost all French; mama and dada are almost all English:

mama papa dada
English 578 59 317
French 6 460 4

The corpora also differ in microphones, rooms and decades (from the 1970s to the 2010s). A classifier trained only on this dataset can separate the classes by language or recording setup. Use the clips as positives inside a keyword spotter that also sees other speech, noise and silence, and check per-corpus performance before trusting a score.

A few children dominate. Manchester, KernFrench and Brent contribute 83% of the clips, and four Manchester children 485 of them. Ages are skewed young: half the clips come from children younger than 23 months, and the range is 8 months to 6 years.

Collection process

The clips come from CHILDES and PhonBank, downloaded in March 2024. Corpora in the Eng-NA, Eng-UK and French collections with transcripts time-aligned to audio were searched with CLAN (kwal +smama +spapa +sdada +t*CHI +f), which returns the child's utterances containing the words, together with the time stamps that link each utterance to the recording. The steps from there are on how it was made.

No new recordings were made, and no one was recorded for this dataset.

Preprocessing and cleaning

In 2024 the clips were cut at the last second of each utterance, screened by ear, and the ones where the word fell outside that second were re-cut by hand in Audacity. Every mama and papa clip went through that screening. The 359 dada clips of 2024 were chosen from 1,026 candidates, and how they were chosen was not recorded: possibly by ear, possibly to keep the classes balanced (most of the dropped candidates came from one Manchester child). The working copies were 8 kHz, the rate of the on-device model.

For this release (2026), each 2024 clip was located inside its source recording by cross-correlation, which recovered the source of 208 clips whose file names had been lost and showed which hand-made clips are plain windows of the recording. Every clip that is a window of its recording was re-cut from the original-rate audio at 16 kHz; only the 226 clips whose audio was changed by hand keep their edited audio. 142 duplicates, 8 clips that matched no recording, and 17 clips whose source could not be established were removed.

Uses

Intended: training and evaluating small keyword spotters that should respond to children's mama, papa or dada; studying how well such models generalise across children, ages and recording conditions.

Used so far: the 2024 capstone trained TensorFlow Lite Micro models on an earlier version of these clips plus non-CHILDES material, with random (not child-disjoint) splits. Those results are not comparable with these splits.

Not suitable for: commercial products (the licence forbids it); identifying, tracking or profiling children; speaker recognition; claims about word recognition across languages without controlling the confound above; clinical or developmental assessment of any child.

Distribution

The audio and metadata are on Zenodo (with a DOI) and mirrored on Hugging Face; the code and metadata are on GitHub. The data are licensed CC BY-NC-SA 4.0 and the code Apache-2.0. TalkBank's ground rules apply in addition; see Licence and citation.

Maintenance

Hüseyin Karaca maintains the dataset. Report problems through GitHub issues. Clips will be removed promptly if a contributing researcher, a family or TalkBank asks for it, and clips from the held-back corpora may be added in a later version if their owners permit it. Versions follow semantic versioning, and each has its own DOI.