PERCEPT-US: A Multimodal American English Child Speech Corpus Specialized for Articulatory Feedback
Abstract
We present PERCEPT-US, a multimodal corpus of audio and ultrasound data from 126 American English-speaking children ages 8 to 17. Collected during clinical trials investigating biofeedback in speech therapy for residual speech sound disorder (RSSD), it includes 80 children with no history of speech-language-hearing challenges and 46 children with no challenges other than RSSD. The corpus is balanced by sex (58 females, 68 males) and stratified by age and speech therapy history. Participants completed syllabic, word, and sentence level speech production tasks, totaling 24,699 utterances, with a focus on American English rhotics. The corpus demonstration uses mixed-effects linear regression on segmented acoustic and labeled ultrasound data from 69 children to show that rhotic tongue shape categories significantly predict the acoustics of higher formant frequency values - the first study to demonstrate this acoustic-articulatory relationship in a large pediatric sample.