Neural-scaled entropy predicts the effects of nonlinear frequency compression on speech perception

Abstract

Signal processing schemes used in hearing aids, such as nonlinear frequency compression (NFC) recode speech information by moving high-frequency information to lower frequency regions. Perceptual studies have shown that depending on the dominant speech sound, compression occurs and the amount of compression can have a significant effect on perception. Very little is understood about how frequency-lowered information is encoded by the auditory periphery. We have developed a measure that is sensitive to information in the altered speech signal in an attempt to predict optimal hearing aid settings for individual hearing losses. The Neural-Scaled Entropy (NSE) model examines the effects of frequency-lowered speech at the level of the inner hair cell synapse of an auditory nerve model [Zilany et al. 2013, Assoc. Res. Otolaryngol.]. NSE quantifies the information available in speech by the degree to which the pattern of neural firing across frequency changes relative to its past history (entropy). Nonsense syllables with different NFC parameters were processed in noise. Results are compared with perceptual data across the NFC parameters as well as across different vowel-defining parameters, consonant features, and talker gender. NSE successfully captured the overall effects of varying NFC parameters across the different sound classes.