Joint Factor Analysis of Channel Mismatch in Whispering Speaker Verification

A speaker recognition system based on joint factor analysis (JFA) is proposed to improve whisper- ing speakers’ recognition rate under channel mismatch. The system estimated separately the eigenvoice and the eigenchannel before calculating the corresponding speaker and the channel factors. Finally, a channel-free speaker model was built to describe accurately a speaker using model compensation. The test results from the whispered speech databases obtained under eight different channels showed that the correct recognition rate of a recognition system based on JFA was higher than that of the Gaussian Mixture Model–Universal Background Model. In particular, the recognition rate in cellphone channel tests increased significantly.