Biometric identification through speaker verification over telephone lines

In this paper, the identity of a remote user is verified through his voice by means of a simple telephone in order to gain access to a specific system or service. We have used state-of-the-art text-independent speaker modeling algorithms, likelihood normalization in the verification process, and channel normalization techniques. Several experiments are presented showing the negative effects of channel variability and temporal lapse between training and testing recordings, but we will show that using the appropriate parameterization, channel compensation and multisession training will allow us to obtain less than 1% of joint false acceptances and false rejections from single utterances of the speaker passport number.