Augmented phonetic map for voice verification

A perceptually based model for speaker identity verification (SIV) using derivative of phase spectrum (DPS) as the primary identity-bearing feature to model individual speakers´ vocal tract dynamics is presented. The basic technique used to model a speaker is to create a two-dimensional trajectory of changing vocal tract based on formant movement and pitch information. The map is further augmented with both instantaneous and dynamic feature parameters of DPS as well as with conventional energy-based acoustic features. A series of verification experiments was conducted, using a three-layer artificial neural network as a classifier, with an isolated digit database recorded over 11 different telephone handsets. The preliminary testing results suggest that this system performs significantly better than a baseline system using a standard cepstrum front-end