Using automatic acoustic analysis to reveal disruptions to speech articulation in individuals at risk for psychosis

Individuals who are experiencing attenuated psychosis symptoms (placing them at clinical high risk for developing psychosis) exhibit disruptions to cortico-cerebellar circuits that manifest as difficulties in motor control of the face and limbs. These motor dysfunctions are predictive of the onset and progression of psychosis, making them a potential biomarker for this devastating disease. We examine whether these motor abnormalities disrupt speech production, leading to greater variability in speech acoustics. Clinical high risk (CHR) individuals and matched healthy control (HC) individuals produced diadochokinetic speech (rapid, repeated syllable production, e.g., papapa…, pataka…) and read aloud a paragraph. Consonant and vowel onsets and offsets were automatically segmented from diadochokinetic speech using a deep-learning model trained on human annotators. Read speech was automatically segmented at the word level using forced alignment. Vowel segmentation from read speech was based on forced aligner output; stop consonant onset/offsets were segmented by a deep-learning model. The segmentations from each task were then used to estimate speech rate, voice onset time (VOT), vowel durations and formant trajectories. CHR individuals produced more variable VOTs and exhibited greater speech rate variability than HC individuals in both tasks. This suggests that speech acoustics may provide a window into disruptions in this at-risk population.