Single Channel Speaker Segregation using Sinusoidal Residual Modeling

In this paper we address the two speaker segregation problem in a single channel paradigm using sinusoidal residual modeling. An appropriate selection of the number of sine waves, window length and hysteresis threshold, is done so as to model and synthesize the underlying signal corresponding to the speaker with the lower pitch period, using an amplitude only sine wave synthesis. The sinusoidal residual is then computed after restimating the phases with known amplitudes, by minimizing a criterion function. This residual corresponds to the the speaker with the higher pitch period. But such a residual consists of harmonic components of the speaker with the lower pitch period. We therefore estimate a binary mask from the spectrograms of the synthesized signal and the residual using a min-max technique to further improve the quality of the segregated speech. This segregation technique is then integrated into a co-channel speaker identification system, at various target to interference ratios. Reasonable improvements in identification performance are noted from these experiments.