Blind speech signal separation based on non-stationary and colored characteristics

Abstract

Some algorithms based on Second Order Statistics (SOS) succeed in separating the non-stationary or colored mixing signals. Among those algorithms, the nonstationary signals are blocked, or the time delay is considerable for colored signals. The speech signal is non-stationary and colored. Based on the autocorrelation matrix of the delayed mixing signals in each block, a new algorithm to infer mixing speech signals is formulated. Since our algorithm covers both charactes of speech, the convergence of our algorithm needs fewer steps than those algorithms with only one characteristic; what’s more, the speed of our algorithm for separation is even faster than FastICA. Blind Signal Separation (BSS) experiment on speech signals under instantaneous mixing proves the effectiveness of our algorithm.