Validating acoustic predictors of sentence intelligibility using a racially diverse corpus

Abstract

The goal of the present work is to validate a set of widely used acoustic measures that have been found to predict talker intelligibility using a new, racially diverse corpus of native speakers of American English. Previous work by Bradlow et al (1996, Speech Communication) predicted sentence intelligibility on a set of 20 demographically unspecified talkers from the Midwest, where they found changes in F0 and larger vowel spaces to be associated with higher sentence intelligibility. Here, we examine whether these same acoustic features can be generalized to a more racially diverse population of talkers, of which only 9 are white and non-Latine. The current corpus comprises 28 talkers, where each talker produced 60 sentences. Here, we focus on a core set of sentences, 10 Havard/IEEE and 10 BEL sentences, that all 28 talkers produced. Key acoustic features such as F0, speaking rate, and first and second formants of vowels, were extracted to associate with intelligibility measurements for these individuals. Intelligibility measures are taken from a previous study (Tripp et al., 2022, J. Acoust. Soc. Am.). Results from the present work are important in generalizing which acoustic features impact intelligibility in a racially diverse corpus.