On neural networks-based preprocessing for speaker identification

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 Dynamic Programming Vector Quantization (DPVQ) Algorithm. (b) The second approach uses a simple Multi Layer Perceptron (MLP) network. (c) Finally the MLP uses DPVQ as an input to the network 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. 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 an HMM-based Speaker Identification system. The proposed architectures have been trained and tested over 60 speaker's Yoho corpus. The results have shown an average improvement of Speaker Identification performance by more than 18% compared to an HMM-Based reference system.