BabbleCor

What is BabbleCor?

BabbleCor is a crosslinguistic corpus of infant and child vocalizations from 52 children exposed to five different languages: English, Spanish, Tsimane', Yêlí-Dnye, Tseltal Mayan, and bilingual Quechua-Spanish.

How was BabbleCor created?

BabbleCor consists of very short audio clips (approximately 400ms) of child vocalizations. To generate these clips, each child first completed a daylong audio recording, between 6 and 16 hours in length, where a small, lightweight recorder was worn inside of a clothing pocket designed for the device.

From these daylong recordings, child vocalizations were either identified by the proprietary Language ENvironment Analysis algorithm, which assigns utterances to speakers in naturalistic audio recordings (e.g. Female Adult, Child) or the vocalizations were identified by hand. 100 of the utterances identified as child vocalizations were randomly selected and chopped into the smaller clips in BabbleCor.

Where do the BabbleCor clip annotations come from?

Each short clip (~400ms) was categorized according to a 5-way scheme by citizen science annotators on the iHEARu PLAY platform (https://www.ihearu-play.eu/). Annotators classified clips as 1) canonical - containing a consonant to vowel transition, 2) non-canonical - not containing a consonant to vowel transition, 3) crying, 4) laughing, or 5) junk.

For further details on corpus creation, please see Methods described in Cychosz et al. (2021).

What are the metadata?

There are two metadata components in BabbleCor: Annotation_Tags and Public_Metadata. As the name suggests, Public_Metadata includes corpus metadata that is publicly available to all corpus users: child ID, child age, child's assigned gender, corpus of origin, and clip ID. Annotation_Tags contains the annotation tags for each clip ID, such as canonical babble, laughing, etc. For access to the annotation tags, please sign, scan, & email the data sharing agreement to babblecorpus@gmail.com (see Data_Sharing_Agreement).