Distance Caluculation Method for Speakers Based on a Structure of Hidden Markov Network

Abstract

Most of large vocabulary continuous speech recognition system employ speaker adaptation or speaker selection method. Many speaker adaptation and speaker selection methods require distance between speakers. However, there are many variations in definition of distance between speakers, and their performances are not investigated. In this paper, we propose a new distance calculation method for distance between speakers based on a structure of Hidden Markov Network(HMnet)made by SSS-free. A structure of HMnet represents speaker distribution because SSS-free constructs an HMnet based on acoustical feature of training samples. From the experimental results, the new distance showed better correspondence with speaker distribution than traditional distances.