Transcripts were manually generated by a team of expert linguists as part of DrivenData's On Top of Pasketti: Children's Speech Recognition Challenge. Only the child was transcribed, neither the adult facilitator nor the social robot were transcribed. Transcribers marked unintelligible speech as xxx. Utterances that contain personally identifiable information were dropped from the transcripts. Transcripts include both orthographic (per-word) and phonetic (per-sound) transcription.
The dataset includes multi-lingual learners. Approximately 40% of the participating children reported exposure to additional languages – predominantly Spanish – at home. Nearly one-third of the children were enrolled in a Spanish-English dual language program.
Audio data was collected through a series of sessions between a social robot and a child (see the JIBO Kids Corpus). Children participated in a letter and digit identification task, and extended oral discourse tasks requiring explanation skills, including playing with blocks. Transcripts were also generated alongside the audio data (available via Github), but those transcripts contain only orthographic transcription.
Opt-out was available
Yes
Data processing
One transcript file was generated per audio file. Both raw and normalized transcriptions are available.
Tips for using the data
Technical tips
The transcripts shared on this site can be joined to the audio from Github using each transcripts file name (raw_file_name in the included metadata.csv). The raw file name is the task plus the name of the raw audio file. For example, the transcript file called blocks_T2027214.jsonl describes the audio saved at blocks/T2027214/T2027214.wav. This is because T2027214 (audio file name) alone is not enough to identify individual recordings in cases where the same child completed multiple tasks.
The key files shared are:
metadata.csv: one row per recording with metadata
transcripts/{raw_file_name}.jsonl: one transcript file for each recording
README.md: Comprehensive documentation of each field in the metadata and each field in the transcripts
Appropriate uses and limitations
The JIBO Kids Corpus Transcripts are best suited to studying children's automatic speech recognition (ASR) and human-robot interaction.
Massachusetts Institute of Technology
UCLA