Modeling non-uniformities in infants' everyday speech environments

• In Event: 1-162 - New models of variation in everyday infancy: Insights from dense audiovisual corpora in six cultures

Integrative Statement

Infants do not re-live the same hour all day or the same day all week. Yet few aspects of developmental science rigorously model the non-uniformity of lived experience. Uniformity assumptions pervade practices in estimating cumulated sensory input (e.g., by observing a rate at one timescale then linearly extrapolating by a relevant number of waking hours), in tracking behavior change (e.g., by measuring behavioral snapshots before and after a supportive parenting intervention, on the logic that any randomly chosen snapshot is equally representative of global patterns), and in making recommendations to families (e.g., urging predictable home routines, without guidance about supportive parameters between monotony and chaos). We suggest that advancing beyond these practices requires large datasets with enough observations to discover subset-superset relations and quantitative measures well-suited to describe variations in experiences at multiple timescales. Here, we do so for everyday speech.

The FauseyTrio corpus (Fausey & Mendoza, 2018; doi:10.21415/T5JM4R) consists of three day-long recordings per week for 35 infants ages 6-12 months (LENA; Ford et al., 2008). Families were assigned recording days and were told that researchers were interested in the full range of infant experiences. Each recording thus contains at least 10 total hours (M=13.55 hours), at least 4 of which were continuously recorded, yielding a corpus of over 1400 recorded hours of everyday infancy captured at home without experimenters present. We analyze variation in the quantity and timing of infant-available speech within and across days.

One commonly reported index of infant-available speech in day-long recordings is "Adult Word Count" (AWC) as automatically estimated by LENA (Oller et al., 2010). Each infant encountered substantial variation across the days of their week, with the difference between their most and least verbose days (Maximum-Minimum Daily AWC; Mdn= 5,549 daily words) nearing 21% of reported variation across families (Weisleder & Fernald, 2013). Within-day variation consisting of verbose periods interleaved with lulls was also evident (Fig.1; see also Tamis-LeMonda et al., 2017).

Word quantity per unit time is a summary measure that arises from real-time caregiver decisions to talk at particular moments. To more directly model non-uniformities in this rhythm, we computed the Coefficient of Variation to reveal the timescale and extent to which infants encounter similar speech quantities at similar times of day across their three recorded days (Fig.2). Lower coefficients indicate more predictable speech quantities. We also computed the Allan Factor of speech onset timing to quantify the extent to which speech bouts cluster separately from silence bouts (see also Abney et al., 2014; 2016). More clustering signals more predictable onsets and offsets of speech. Analyses are ongoing; preliminary analyses suggest individual differences in these indices of predictability in early language environments.

The combination of multi-day recordings and a suite of measures to quantify variation at multiple timescales offers unprecedented opportunities to model the non-uniformities of early language environments. We will discuss the implications of these insights for guiding methods of sampling input, subsetting data to transcribe and annotate, and theorizing about the role of predictability in early learning.

Authors

• Heather T Anderson, University of Oregon
  Presenting Author

• Caitlin Fausey, University of Oregon
  Non-Presenting Author