Speaker assessment: The impact of environment on speech systems and individuals.

Abstract

Assessing speaker variability is critical in developing a scientific understanding or model of human speech production. The environmental context plays a significant role in how this variability plays out. In this study, recent findings are presented on the variability of speech production due to environmental factors that influence man-machine interaction as well as human-to-human interaction. Speech produced under task stress, emotional stress, and background noise (resulting in Lombard effect) all cause speech production changes. This impacts both speech processing algorithms intended for speech recognition/technology and human-to-human interaction. Specifically, two speech production domains are briefly considered including Part-1: speech production under varying types/levels of background noise and how this produces flavors of Lombard effect and impacts speaker recognition systems, and Part-2: assessing the stress/emotional state of parents/care-givers in quantifying the language learning exposure of children (ages 10–36 months). In Part-1: the UT-Scope corpus is employed with speech from 30 subjects (19M,11F) for analysis of duration and spectral tilt as well as developing an automatic Lombard effect classification scheme which is incorporated into speaker recognition. Next, Part-2: considers how neutral versus stressed/emotional state of adults impacts conversational turns and adult word-count in a child language learning environment (20 child-parent interactions).