Perceptual MVDR-based unsupervised built-in speaker normalization for Kazakh speech recognition

Abstract

In this work we present a novel approach to unsupervised speaker normalization on top of the Perceptual MVDR-based Built-in Speaker Normalization technique. We showed that the proposed method can be efficient for the task of phonetic recognition on TIMIT and then applied it to Kazakh speech recognition. From the experiments, we see that this method is able to improve the relative performance of ASR systems up to 20% The analysis of the optimal warp factor selection by the algorithm revealed a nice gender separation ability which may be used for gender/speaker classification tasks.