Speech enhancement for hearing-impaired listeners using deep neural networks with auditory-model based features

Speech understanding in adverse acoustic environments is still a major problem for users of hearing-instruments. Recent studies on supervised speech segregation show good promise to alleviate this problem by separating speech-dominated from noise-dominated spectro-temporal regions with estimated time-frequency masks. The current study compared a previously proposed feature set to a novel auditory-model based feature set using a common deep neural network based speech enhancement framework. The performance of both feature extraction methods was evaluated with objective measurements and a subjective listening test to measure speech perception scores in terms of intelligibility and quality with 17 hearing-impaired listeners. Significant improvements in speech intelligibility and quality ratings were found for both feature extraction systems. However, the auditory-model based feature set showed superior performance compared to the comparison feature set indicating that auditory-model based processing could provide further improvements for supervised speech segregation systems and their potential applications in hearing instruments.