Speaker recognition in reverberant environments

Abstract

The comparison of the methods used for the purpose of speaker recognition and parametrizations, when there is a mismatch between training and test conditions due to reverberation, was discussed. Gaussian mixture models (GMM), covariance models, and AR-vector models were used for this purpose. It was fond that the performance of all the methods degrades under reverberation. It was also found that recognition accuracy improved for all the methods when training was performed with reverberant speech prior to testing with minor reverb or major reverb.