On nonintrusive speech quality estimation for hearing aids

Assessment and monitoring of speech quality in assistive hearing devices is of great interest to Audiologists, hearing researchers, device manufacturers and consumers. While subjective assessment of device quality is the “gold standard”, objective speech quality evaluation methods that emulate subjective perception are attractive for their efficiency and ease of use. This paper focuses on data-driven, non-intrusive speech quality metrics for potential application in hearing aids (HAs), using a custom database of HA speech recordings acquired in different noisy and reverberant environments. A large set of speech specific features were extracted from HA recordings, followed by a dimensionality reduction procedure to identify the most relevant features for quality prediction. The selected features are subsequently mapped to predicted quality scores and validated with subjective ratings obtained from hearing impaired listeners. Results showed that the non-intrusive metrics perform well in HA speech quality prediction.