Neural speech enhancement in the time-frequency domain

Abstract

A pulsed neural network approach for the enhancement of speech is presented. It employs a recently introduced pulse-induced masking principle on a joint time-frequency representation of signals. Based on the notion of a spectro-temporal receptive field, the principle provides signal-dependent dictionary elements, which are integrated into an enhancement process in the time-frequency domain. Experiments on signals with different types of noise are conducted with the resulting neural time-frequency domain method.