Simultaneous Adaptive Co-Channel Speaker Separation

Abstract

Abstract : In this research, we have developed several unique techniques used in the separation of a speech signal corrupted by another talker's speech recorded over a single channel. Historically, this has been referred to as the cocktail party problem. Our work is useful in such applications as separating the speech signals recorded onto an in-flight voice data recording box from the cockpit of an airplane, enhancing the quality of speech transmitted through a hearing aid, and in the enhancement of speech transmitted over a noisy communication channel. We have made significant contributions to the field of speaker separation. We have developed and tested an adaptive co-channel speaker separation system which can simultaneously estimate the speech of two speakers recorded onto a single channel. We have developed and tested several methods to estimate the voicing state of a co-channel speech segment. We have developed and tested a technique to estimate the fundamental frequency and pitch contour of each speaker. This technique is based on the maximum likelihood pitch estimator and harmonic magnitude suppression. Using the estimate of the fundamental frequency, we have developed a technique to estimate the harmonic parameters of overlapping voice speech segments. Finally, we have developed and tested an innovative technique to simultaneously estimate overlapping voiced speech segments using a constrained nonlinear least squared optimization algorithm. These techniques have been integrated into end to end speaker separation system to separate co-channel speech.