Targeted Voice Enhancement by Bandpass Filter and Composite Deep Denoising Autoencoder

In many hearing-aids systems, background noise degrades the speech quality and intelligibility. In this paper, we propose a hybrid system for hearing-aids application, which works to separates the target voice from the noisy signal and then enhance the speech based on the user's hearing loss. We achieve this by using two stages: (1) A bandpass filter to filter out the unwanted noise which is followed by (2) composite of two-level of multi-layers deep denoising autoencoder, each which specialized for specific enhancement task of a complete set of tasks. We evaluated the improvement of the speech quality using two typical hearing loss audiograms. For evaluation, hearing-aid speech perception index (HASPI), hearing-aid sound quality index (HASQI), and perceptual evaluation of speech quality (PESQ) used in two types audiograms of high-frequency hearing loss (HFHL). The results for the experiments show that the proposed method achieved significant results compared with the individual deep denoising autoencoder.