Text dependent speaker verification using dynamic time warping and vector quantization of LSF
Résumé
This paper describes a text-dependent speaker verification system based on the use of DTW and VQ connected in series. It describes the algorithms and the extraction of the used LSF parameters. This way of using two classical algorithms makes it possible to reach high discrimination rates and particularly to eliminate the most dangerous category of impostors: those who could enter without knowing someone else's key. By this means, the system performances obtain a 95% recognition rate versus 2.3% false admission rate for a database recorded over telephone lines.