Noise robust speaker identification using Bhattacharyya distance in adapted Gaussian models space

This is a study on the issue of noise robustness of text independent Speaker Identification (SID). Over the past years, SID technology has emerged as extremely important tool with applications in security and authentication. The current technology works well in presence of matched acoustic conditions for training and testing but the performance shows immediate loss in mismatched conditions. Our broad approach in this work is to map features to models and then do classification in the space of models. In particular, our algorithm is works in the space of adapted Gaussian Mixture Models, where we use Bhattacharyya Shape to measure closeness of models. We show our approach to be robust to noise in SID evaluations. We tested our approach on speech corrupted by white and music noise and found it to be very advantageous in low SNR conditions.