Automatic voice disorder recognition using acoustic amplitude modulation features

Abstract

An automatic dysphonia recognition system is designed that exploits amplitude modulations (AM) in voice using biologically-inspired models. This system recognizes general dysphonia and four subclasses: hyperfunction, A-P squeezing, paralysis, and vocal fold lesions. The models developed represent processing in the auditory system at the level of the cochlea, auditory nerve, and inferior colliculus. Recognition experiments using dysphonic sentence data obtained from the Kay Elemetrics Disordered Voice Database suggest that our system provides complementary information to state-of-the-art mel-cepstral features. A model for analyzing AM in dysphonic speech is also developed from a traditional communications engineering perspective. Through a case study of seven disordered voices, we show that different AM patterns occur in different frequency bands. This perspective challenges current dysphonia analysis methods that analyze AM in the time-domain signal.