Incremental speaker adaptation with minimum error discriminative training for speaker identification

The minimum classification error (MCE) has been shown to be effective in improving the performance of a speaker identification system. However, there are still problems to solve, such as the variability of the voice characteristics of a particular speaker through time. In this paper, we analyze the degradation of a Gaussian mixture model (GMM) based text-independent speaker identification system when using test data recorded over six months after the training session, and, in an attempt to avoid this degradation, we study the use of supervised adaptation based on maximum a posteriori (MAP) estimation and MCE. These techniques have been shown to provide good results for speaker adaptation in speech recognition. The major result we have obtained is that, by starting with GMM models trained with only speech from session 1, similar identification results can be obtained for all the other sessions using an incremental adaptation using only 2.5 seconds of speech per speaker and session as data for the MCE training adaptation procedure. We have also found that, in our extreme experimental setup, MAP becomes unhelpful when combined with MCE adaptation.