A Comprehensive Framework for F0 Estimation and Sampling in Modeling Prosodic Variation in Infant-Directed Speech

Accurate estimates of F 0 are essential for modeling how pitch variation is used as an informative cue to linguistic structure. Multiple challenges exist for estimation and valid statistical modeling of F 0 variation. First, certain speech styles, such as infant-directed speech, can involve dramatic pitch variation across utterances. Second, non-modal phonation can cause spurious F 0 values. Third, F 0 samples are not independent of one another, leading to issues with validity in applying generalized linear mixed effect models (GLMMs). To address these problems, we propose a comprehensive framework for accurate F 0 estimation and sampling to model prosodic variation. Our method involves segmentation of speech into utterances, followed by determination of speaker- and utterance-specific pitch range parameters. Regions of non-modal phonation are identified, ensuring that portions of speech leading to spurious F 0 values are rejected early. Next, F 0 stylization at the utterance level ensures robustness to microprosodic variation. Finally, F 0 turning points (e.g., local F 0 minima and maxima) are extracted; these are linguistically significant “control points” in F 0 contours connected by monotonic interpolations. This overall approach not only ensures accurate F 0 estimates, but critically overcomes the problem of non-independence of successive samples for valid statistical treatments within GLMMs.