Speech enhancement and the instantaneous phase signal

Abstract

An approach to speech enhancement based on a representation of the speech signal by an envelope and instantaneous phase model is proposed. Envelope modification for the removal of room reverberation and noise from signals has already been studied and found to produce signal‐to‐noise ratio but not intelligibility improvements. The envelope alone does not account sufficiently for the detailed structure of speech. A noisy instantaneous phase has been found to have a perceptually degrading effect even with a perfectly restored envelope. Thus it is proposed to enhance speech by also replacing the measured instantaneous phase signal (in a number of contiguous frequency bands) by a modeled version. To determine the characteristics of the instantaneous phase function both synthetic and recorded speech have been studied. A simple model for the instantaneous phase of speech is proposed and an analytical solution is developed relating the digital model of speech production to an envelope and instantaneous phase model. Enhancement of a noisy instantaneous phase function can be achieved by using a compensating frequency shift, low‐pass filtering and linear prediction.