A new nonlinear speaker parameterization algorithm for speaker identification

In this paper we propose a new coding algorithm based on nonlinear prediction: the Neural Predictive Coding model which is an extension of the classical LPC one. The features performances are estimated by two different methods: the ArithmeticHarmonic Sphericity (AHS) and the Auto-Regressive Vectorial Models (ARVM). Two different methods are proposed for the coding method based on the Neural Predictive Coding (NPC): classical neural networks initialization and linear initialization. We applied these two parameters to speaker identification. The fist model obtained smaller rates. We show for the first model how it can be combined with the classical feature extractors (LPCC, MFCC, etc.) in order to improve the results of only one classical coding (MFCC provides 97.55% and MFCC+NPC 98.78%). For the linear initialization, we obtain 100% which is a great improvement. This study opens a new way towards different coding schemes that offer better accuracy on speaker recognition tasks.