Automatic Text-independent Speaker Tracking System using Gaussian Mixture Models (GMMs)

Speaker tracking is the process of identifying who says something in a given speech signal. In this paper, we propose a new set of robust source features for Automatic Text-Independent speaker tracking system using Gaussian Mixture Models (GMMs). LP analysis is used to extract the source information from the speech signal, which is speaker specific. The time varying speaker-specific source characteristics are captured using Linear Prediction (LP) residual signal of the given speech signal. Further, MFCC features are extracted from the source speech signal, which contains prosody and speaker specific information. These source features which are extracted are proven to be robust and insensitive to channel characteristics and noise. In this paper, experiments were conducted for varying train and test durations.