Monitoring laryngeal and respiratory adjustments during voice production using a simulation-based voice production inversion neural network

Abstract

Previously we reported a simulation-based neural network for estimating vocal fold properties and subglottal pressure from the produced voice. In this study, we evaluate the feasibility of using this neural network to monitor laryngeal and respiratory adjustments during speech production in individual speakers. Acoustic and aerodynamic data were collected in human subjects while producing utterances of five repetitions of the syllable /pa/ at different loudness levels. Voice features were then extracted and used as input to the neural network to estimate changes in vocal physiology and subglottal pressure. The results showed all subjects increased the subglottal pressure and vocal fold adduction when producing a louder voice, although the degrees of laryngeal and respiratory adjustments were speaker-specific. The neural network also estimated on average shorter and thinner vocal folds in female subjects than male subjects. These results demonstrate the potential of this neural network toward monitoring and identifying potentially unhealthy vocal behaviors.