Speech Enhancement for Hearing Impaired Based on Bandpass Filters and a Compound Deep Denoising Autoencoder

Deep neural networks have been applied for speech enhancements efficiently. However, large variations of patterns and noisy environments, an individual network with a fixed number hidden layers causes strong interference, which can lead to slow learning process, poor generalisation in unknown signal-to-noise ratio new inputs, some residual noise the enhanced output. In this paper, we present approach hearing impaired based on combining two stages: (1) set bandpass filters that split up signal into eight separate bands each performing frequency analysis signal; (2) multiple deep denoising autoencoder networks, working small specific enhancement task handle subset whole training set. To evaluate performance approach, hearing-aid perception index, aid sound quality perceptual evaluation were used. Improvements intelligibility evaluated using seven subjects sensorineural loss audiogram. We compared proposed three five layers. The experimental results showed yielded higher was more intelligible