Experiments on Blind Speech Separations

Abstract

Intelligent voice recognitions have been attracting many interests these days. Acoustic array based on beamforming have been successfully used in suppressing noises and enhancing the voice of interest. But that cannot work when multiple voices need to be recognized simultaneously. Blind source separation in complex frequency domain can separate mixed voice signal in practical condition. Compared to beamforming, we can get multiply channels output signal. In this work, we build a speech signal acquiring, transferring and separating system with microphone array, FPGA, WiFi and host computer. The separation algorithm based on IVA in which unmixing matrix iterating according to Newton's method until the increment of likelihood contrast less than the threshold. The experimental result measurements validated the proposed approach.