Detail publikace

Abstrakt

The automatic speaker recognition made a signicant progress in the last two decades. Huge speech corpora containing thousands of speakers recorded on several channels are at hand, and methods utilizing as much information as possible were developed. Nowadays state-of-the-art methods are based on Gaussian mixture models used to estimate relevant statistics from feature vectors extracted from the speech of a speaker, which are further concatenated into a high dimensional vector  supervector. Methods concerning the extraction of high dimensional supervectors along with techniques capable to build a speaker model in such a high dimensional space are described in depth and links between these methods are found. The main emphasize is laid on the analysis of these methods and an ecient implementation in order to process huge amounts of development data to train the speaker recognition system. Also the inuence of development corpora on the recognition performance is experimentally tested.