Fuzzy Vector Quantization for speaker recognition under limited data conditions

Abstract

This work focuses on the task of speaker recognition under limited data conditions. In case of limited data, the amount of available training and testing data will be few seconds. Under such conditions the conventional classifiers will have very few feature vectors for modelling. This work performs an experimental evaluation of three simple modelling techniques namely, direct template matching (DTM), crisp vector quantization (CVQ) and fuzzy vector quantization (FVQ). Among these FVQ shows significant improved performance compared to DTM and CVQ. For about 3 s of training and testing data the performance for DTM, CVQ and FVQ are 76.67, 73.33, and 86.67, respectively, for a set of first 30 speakers taken from the YOHO database.