Pi-sigma and hidden control based self-structuring models for text-independent speaker recognition

Two text-independent speaker recognition methods based on Selfstructuring Hidden Control (SHC) neural models and Self-structuring Pi-Sigma (SPS) neural models are proposed. We have designed the self-structuring models to achieve better model architectures i.e. data determined architectures instead of a priori determined architectures. PS and HC neural models for speaker recognition are also proposed. Each of the four methods require typically 75% less neural models compared to the predictive neural network based text-independent speaker recognition method [I] i.e. the latter contains an ergodic M-state model using M neural models (M = 4) for each speaker; each of our speaker recognition systems uses only one neural model to realize an ergodic M-state model. The Pi-Sigma models [2] have been modified to obtain Self-structuring PS models and our speech recognition SHC models which we presented at ICASSP92 [3] have been changed to fit into speaker recognition systems.