Factor analysis based channel compensation in speaker verification

This report describes a powerful channel compensation method for the text-independent speaker verification task. This powerful method is developed in the LRDE Speaker Verification framework. The purpose of a text-independent speaker verification system is to check whether a hypothesised speaker is really the author of a speech utterance. The channel compensation problem arises when training data and test data come from different channels. This report first focuses on the current state-of-the-art system based on a probabilistic approach of the acoustic event distribution with Gaussian mixture models. Unfortunately, this baseline system does not cope with channel effects. Then, the method developed in the LRDE Speaker Verification framework is introduced. This one is based on a factor analysis which enables to deal with the channel compensation problem by taking advantage of the limits of the state-of-the-art system. The Factor Analysis model considers the variability of the Gaussian mixture model as a linear combination of the variabilities of the speaker and channel unobservable components. This decomposition is based under the classical MAP and the eigenchannel MAP assumptions. Finally, the results obtained on the NIST-2006 Speaker Recognition Evaluation (female trials) will be presented.