VocabStudy: A remote collection of naturalistic topic-structured parental speech and toddlers' vocabularies using a...

The current work presents VocabStudy, a collection of natural language samples and children's vocabularies collected remotely by parents via a mobile phone application. The corpus contains 567,003 word tokens and represents 144 hours of speech over a period of six months from the language environment of 63 British toddlers aged 13 to 28 months. The corpus incorporates labeled speech samples of five typical routines: mealtime, bedtime, playtime, bathtime, and nappytime (i.e., diaper). To explore consistency and variability across these five linguistic contexts, topic modeling was employed. The topic most successfully detected as having a unique structure was mealtime, which was identified as such nearly 100% of the time; bathtime, nappytime, and bedtime were found to cluster together most of the time, suggesting that they have a similar language structure; playtime was correctly identified as such about 14% of the time. To validate the accuracy of parents marking the words that their child produced, the child's utterances found in the audio recordings were examined. About 18% of the vocabulary reported by parents appeared in the transcripts, and the reported vocabulary sizes were highly correlated with the number of unique words uttered by the children (ρ = .72, p < 001). The results suggest that most parents marked the words soon after their children start producing them (p < .001, d = 0.9). I discuss the advantages and disadvantages of using a mobile phone application as a method to collect children's data remotely, what worked to keep participants entering data, and what could have been done to avoid some issues encountered.