Speech separation by kurtosis maximization

We present a computationally efficient method of separating mixed speech signals. The method uses a recursive adaptive gradient descent technique with the cost function designed to maximize the kurtosis of the output (separated) signals. The choice of kurtosis maximization as an objective function (which acts as a measure of separation) is supported by experiments with a number of speech signals as well as spherically invariant random processes (SIRPs) which are regarded as excellent statistical models for speech. Development and analysis of the adaptive algorithm is presented. Simulation examples using actual voice signals are presented.