Single Channel Speech Separation with Frame-based Summary Autocorrelation Function Analysis

Abstract

Single channel speech separation system with frame based pitch range estimation is presented. The system decomposes the input mixture speech into frames and pitch of the speech is estimated in each frame of modulation spectrum by using summary autocorrelation function (SACF) analysis. In this method, each frame is divided into two, above and below 1 KHz and compute generalized autocorrelation for periodicity detection. By using this pitch range, a mask is created to filter the target speech from noisy mixture speech. Performance evaluation of the proposed system shows a better response compared to the existing methods.