Automatic speaker recognition using Gaussian mixture models

We describe the automatic identification of speakers from their voices. This process has an application in forensics and in voice actuated security systems. The implementation and theoretic underpinnings of a statistical based speaker recognition system are presented in addition to the performance of the system on standard speech corpora. In a speaker verification experiment, the system yielded an equal error rate of under 5% when identical microphones are used for testing and training.