Examining variation in early language environments using child-centered long-form audio-recordings

The present project seeks to answer the question: which factors predict variation in children's spoken input? Previous research has shown that children from different backgrounds are exposed to different amounts of child-directed speech (e.g. Cristia, 2022, Developmental Science). Bunce et al. (2020) studied similarities as well as differences across corpora in quantity and sources of children's input based on human annotations of 69 children across 7 corpora from various cultural contexts. Their analyses suggested that several factors explained variance in the amount of speech afforded to infants, including the child's age and the type of speakers. Surprisingly, the authors concluded that there was little cross-population variation in the quantity of speech directed to children, counter Cristia, 2022. Their analyses also suggest that there is important cross-population variation in total quantity of speech afforded to young children, due to cross-population variation in the quantity of overheard speech. Given the importance of better understanding the extent of population variation, in this paper, we revisit the same corpora included in Bunce's paper: Bergelson/SeedLings, Warlaumont, Winnipeg/McDivitt, LuCiD, Rosemberg, Tseltal, and Rossel. We want to pre-register the present analyses to test hypotheses based both on previous literature and Bunce et al. We believe our results may be different from those found in Bunce et al for several reasons: Here we will include all children, and not just the 9-10 in each corpus that had human annotation, which means that we have greater statistical power. Here, we employ automated annotation, rather than human annotation, allowing us to represent each child through all of their recorded data, rather than the 30 minutes that were human-annotated. This may mean greater accuracy (based on more, and potentially more representative of a day’s experience, data) or not (because we rely on automated annotations, which may be less accurate than human annotation). Here, we need to adapt the definition of key child directed speech (KCDS) and overhearable speech (OHS) to our automated annotations. KCDS is the speech specifically directed towards the key child; OHS is the speech that is not directed towards the key child, but available around the key child, which can be distant or background. While Bunce et al. defined KCDS based on content and context, our study relies on automated annotations. Therefore, we cannot utilize content-based features. Instead, we will use the human-annotated data to set parameters for defining KCDS based on temporal information alone. We describe how we will do this in the variables section below.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.