New Issues on Nonlinear Discriminant Analysis for Speaker Recognition

摘要

In this paper, we study a nonlinear discriminant analysis (NLDA) technique that extracts a speaker-discriminant feature set. We present different system architectures to extract features that are more invariant to non-speakers-related conditions such as handset types and channel effects: (a) the first approach uses a Time Delay Neural Network (TDNN). (b) The second approach uses the Dynamic Decay Adjustment (DDA) algorithm for Radial basis Function (RBF) method. (c) We finally combined linearly the normalized scores TDNN/RBF architectures to maximize the separation between speakers by nonlinearly projecting a large set of acoustic features to a lower-dimensional feature set. The architecture proposed takes into account both the temporal changing of the speech signal and the powerful of the neural networks (NN). The extracted features are optimized to discriminate between speakers and to be robust to mismatched training and testing conditions. The transformed features are used to train a GMM-based speaker identification system. We have trained and tested the different proposed architectures on 45 speaker's SPIDRE corpus of telephone conversations. The results show an improvement of more than 12% compared to our standard system.