I-Vector Extraction Using Speaker Relevancy for Short Duration Speaker Recognition

摘要

This paper presents a novel scheme for considering the frame-level speaker relevancy during i-vector extraction for speaker recognition. In the proposed system, the frame-level point-wise mutual information is utilized to directly modify the Baum-Welch statistics in order to extract a robust i-vector. Furthermore, a method for computing the frame-level speaker relevancy using deep neural network (DNN) analogous to the DNN used in robust automatic speech recognition (ASR) is proposed. The results show that the modified i-vectors obtained using the proposed methods outperformed the conventional i-vectors.