Benefits of the WaveNet-Based Speech Intelligibility Enhancement for Normal and Hearing Impaired Listeners
* Presenting author
Speech perception becomes challenging in adverse listening conditions, hence increasing mental effort. Modification of recorded clean speech with the goal to increase its intelligibility is one approach for improving listening experience. Recently, we suggested a data-driven WaveNet-like speech intelligibility enhancement method which is based on the Spectral Shaping and Dynamic Range Compressions (SSDRC) approach. Both approaches achieve intelligibility gains by relocating energy in the time-frequency plane. In this study we have assessed the performance of the WaveNet intelligibility enhancer for both normal and hearing impaired participants using formal listening tests, in terms of intelligibility and sound quality. Intelligibility was measured as a percentage of correct words recalled (CWR) and quality was assessed via a comparative mean opinion score (CMOS). Furthermore, we compared performance among native and non-native listeners using English and Greek stimuli. We observed that modified speech was significantly more intelligible than plain across all listener groups. Hearing impaired candidates rated modified speech 30% higher on average (on a CMOS scale) than normal hearing subjects. The positive impact for the hearing impaired listeners might be attributed to the reallocation of energy to perceptually relevant frequency bands in which the hearing impaired are more sensitive.