A distance metric based outliers detection for robust Automatic Speaker Recognition applications

Outlier detection in Automatic Speaker Recognition (ASR) context is a task to detect those points in feature space which are less representative of a speaker. The existence of outliers is related to handset, noise or speaker´s non-intrinsic characteristics. So detection and removal of outliers is useful in robust speaker recognition. The detection can be done in training phase or in testing phase or both. In this paper, we try to investigate the outliers in testing phase using three different distance measures with the databases, one is microphone speech, YOHO and the other is telephone speech, POLYCOST. The experiment is conducted on Mel-Frequency Cepstral Coefficients (MFCC) features with Gaussian Mixture Model (GMM) based speaker model. The results show that distance metric based outlier removal can remove maximum 29.43% of outliers in YOHO and 22.86% for POLYCOST while the accuracy improves or remain same as baseline depending on distance metric used.