Real-time blind source separation system with applications to distant speech recognition

Abstract

A real-time BSS system based on DUET was developed and implemented in order to assess its potential as the front-end for a DSR engine. The system uses only two closely-spaced standard omni-directional microphones and a computer soundcard and was developed for low reverberation environments with several human speakers and different noise sources.A novel multi-source real-time audio streaming module was developed, with arbitrary statistics, movement tracking, continuity cues such as position and cross-correlation, a spurious peak classifier stage based on kurtosis, and spectral subtraction post-processing.Two intrinsic error causes for the binaural attenuation and delay estimators were identified, due to FFT spectral leakage and to sibilants, which violate the taken for granted DUET assumptions. A comprehensive study on time windows was done and new window types proposed in order to minimize the DUET assumptions violations.The implemented system correctly identifies the clusters in the binaural estimators’ space for the case of a real room with two human speakers, up to distances of 2m from the two microphones, although for distances greater than 1m the separation quality quickly degrades.