Bridging Machine Learning and Event-Based Analyses to Assess Maternal Contingent Responses to Infant Nondistress and...

Caregiver-infant vocal interactions are foundational for early development, yet most evidence is derived from brief laboratory or home-visit observations and assessing caregiver-infant vocal contingency in everyday settings requires scalable methods. To this end, we integrated machine learning (ML) with traditional event-based analyses to measure maternal contingent vocal responsiveness assessed from daylong audio recordings in the home. In Study 1 (N = 61 infants; 1-18 months, M_age = 9.43 months), we evaluated a wav2vec2-based audio tagging algorithm to detect infant nondistress and distress vocalizations and caregiver vocalizations. Performance was high when tested against human annotations. In Study 2 (N = 56 infants; 1-10 months; M_age = 6.61 months), we applied the algorithm to daylong home audio recordings. Maternal response rates based on ML-generated labels showed strong correspondence with those based on human-annotated labels, although a consistent positive bias was observed (ML > human). Response-rate distributions tended toward higher values in the 20 most voluble segments relative to all available mother-infant interaction segments. Additionally, on average, the 20 most voluble segments (vs all available segments) produced higher maternal response rates but lower base-rate-adjusted estimates of contingency. Response rates increased with lag window, and in most cases, the sharpest increase was observed between 1 and 2 s, with diminishing returns thereafter. Together, these findings support ML-derived vocalization labels for scalable contingency estimation and underscore analytic decisions regarding sampling and lag window when estimating maternal contingent responsiveness from daylong home recordings.