Talker Identification using Reflection Coefficients

Abstract

In this paper, we propose a method of identifying the talker based on modeling the voice for reading free sentences by AR(auto regressive) models. By remarking the similarity of the AR models and the no loss sound-tube models for vocal tracts, we assume the features of the talker voice are the higher reflection coefficients of AR models. In the simulation, there are 6 talkers' voice as speech examples from ”Japanese speech sound database of ATR.” As a result, we get that average recognition rate is 90.67 %.