Effects of simulated cochlear-implant processing on voice quality distinction: Evidence from analysis of disordered voices

Abstract

ABSTRACTThe ability to discern variations in voice quality from speech is important for effective talker identification and robust speech processing; yet, little is known about how faithfully acoustic information relevant to variations in talkers’ voice quality is transmitted through a cochlear implant (CI) device. The present study analyzed unprocessed and CI-simulated versions of sustained /a/ vowel sounds from two groups of individuals with normal and disordered voice qualities in order to explore the effects of CI speech processing on acoustic information relevant for the distinction of voice quality. The CI-simulated voices were created by processing the vowel sounds along with 4-, 8-, 12-, 16-, 22-, and 32-channel noise-vocoders. The variations in voice quality for each voice sound was characterized by calculating mel-frequency cepstral coefficients (MFCCs). The effects of simulated CI speech processing on the acoustic distinctiveness between normal and disordered voices were then measured by calculating the Mahalanobis distance (MD) metric, as well as accuracy of support vector machines (SVMs) applied to MFCC features. The results showed that CI speech processing, as simulated by noise vocoding, is highly detrimental to the acoustic information involved in conveying voice quality distinctions. This supports the view that listeners with CIs will likely experience difficulties in perceiving voice quality variations due to the reduced spectral resolution, shedding light on challenges listeners with CIs may face for effective recognition and processing of talkers’ voices.