Psychometric Validation of a Deep Learning-based Foreground Speech Detection Algorithm for Everyday Conversation Detection

Passive ambient audio sampling bears great potential for objectively measuring daily social activity and its association with wellbeing and health. However, detecting true conversations via human annotation is labor-intensive, and automatic labeling via audio signal processing has thus far only received proof-of-concept validation. Here, we conduct a comprehensive psychometric validation of a deep learning-based foreground speech detection algorithm for conversation activity detection (CAD) from ambient audio sampled with the Electronically Activated Recorder (EAR) method (Hebbar et al., 2021). We assess the CAD algorithm’s validity as an objective measure of conversation activity using four archival EAR datasets with human ground-truth conversation annotations (N = 566 participants, n = 167,539 audio recordings). Specifically, we evaluate, across the four samples, the degree to which the CAD algorithm converges with human-annotated conversation activity, yields temporal stability estimates of conversation activity, and replicates patterns of external correlates (with demographic, wellbeing, and personality measures) comparable to those derived from human ground-truth annotations. We further compare the distributional properties of conversation activity derived from the CAD algorithm and human ground-truth annotations and use this information for thresholding the algorithm’s continuous conversation activity estimates. Overall, the CAD algorithm evidences strong psychometric properties for estimating conversation activity across a range of participants and study characteristics, suggesting that it is suitable for at-scale deployment to objectively measure daily socializing from passively sampled ambient audio.