An automatic text-free speaker recognition system based on an enhanced Art 2 neural architecture

The adaptive-resonance-theory (ART) architectures comprise a family of neural networks that efficiently self-organize stable recognition codes in response to arbitrary sequences of input patterns. This paper focuses on the ART 2 network architecture that is designed for the processing of both binary- and analog-valued patterns. The problems encountered in implementing the primary ART 2 architecture as originally presented by Gail Carpenter and Stephen Grossberg are discussed. An enhanced ART 2 architecture that receives its input from a functional-link preprocessor is proposed. Experimental results that demonstrate this architecture's superior performance over the previous ART 2 architecture are provided. Finally, a text-free speaker recognition system that employs the enhanced ART 2 architecture as its classifier is described.