Speaker discrimination in a conversation

A method to automatically discriminate between speakers in a conversation between two people, with no prior training have been developed. It is intended as a preprocessing stage in a speaker recognition system. The speech is preprocessed by extracting the syllable nuclei and discarding fricatives and noise. Next, a multidimensional feature set is calculated for the whole conversation and an axis in the feature space is obtained that gives good discrimination between the two speakers. Speech that reaches either extreme on this axis is choosen as belonging to the two speakers respectively. The axis is found by taking a moving average of the feature vectors over a short time interval. This has the effect of reducing the variance in all directions in the feature space. Because the averaging relatively rarely goes over two speakers, the variance changes least in the direction in the direction separating the means of the speakers. This direction of least change of variance is found by eigenvector analysis of the covariance matrices before and after averaging