Abstract |
Automatic speaker veri―cation and identi―cation are probably the most natural and eco- nomical methods for solving the problems of unauthorized use of computer and communi- cations systems and multilevel access control. With the ubiquitous telephone network and microphones embedded into computers, the cost of a speaker recognition system might only be for the software of the recognition algorithm. Biometric systems automatically recognize a person using distinguishing traits. Speaker recognition is a performance biometric i.e. you perform a task to be recognized. Your voice, like other biometrics, cannot be forgotten or misplaced, unlike knowledge-based (e.g. password) or possession-based (e.g. key) access control methods. Due to the inherent variability of the speech signal, as far as the iden- tity of a person is concerned, we emphasize the use of statistical approaches in this thesis for speaker recognition. Speaker recognition is based on appropriate probabilistic models. The form of the Gaussian mixture model (GMM) motivates its use as a representation of speakers for text-independent speaker recognition. The most popular method for auto- matic speaker recognition uses the cepstrum, with nonlinear frequency axis following the Bark or mel scale. Using these features the performance of the system is quite satisfactory. One step forward is the use of sub-cepstrum which gives even better results with the same computational cost.
|