Whispered Speaker Recognition Based On Factor Analysis And SVM

Whispered speech has been widely used in speech communication in public, identification, telephone network query and telephone banking, etc. As a result, more attention has been paid to the whispered speaker recognition.Due to the special way of whispering pronunciation, whispered speech speaker recognition is seriously affected by the speaker’s health status, psychological change, the pronunciation of status and channel environment factors. Speaker recognition system based on normal speech is not suitable to whispered speech.Joint factor analysis(JFA) has got good recognition accuracy in whispered speaker recognition under mismatch channel environment, which can separate speaker and channel information in model domain to engage in channel compensation. This paper adopts one part of JFA, only estimates speaker space, extracts internal speaker factor and combines speaker factor with support vector machine(SVM) to form a factor analysis-support vector machine(FA-SVM) whispered speech speaker recognition system.This paper adds class covariance normalization(WCCN), linear discriminant analysis(LDA) and nuisance attribute projection(NAP) to the basis of system. The speaker factors which are processed by three kinds channel compensation technique are input into the SVM system. The result proves that the system that uses channel compensation has better recognition result than baseline system.Because of the problem that speaker factor still includes channel information, this paper makes corresponding improvement on the above methods. For within class covariance normalization, the paper replaces the original matrix with smooth normalization matrix. For nuisance attribute projection, the paper analyzes the relationship between NAP and baseline system, and proposes a variability compensation-nuisance attribute projection(VCNAP) system in between. This system realizes a tunable channel compensation which could avoid the channel compensation error caused by whole NAP. The experiment proves that this theory is correct and gets better result than NAP.