Adaptive spatiotemporal neural network for speech recognition

Abstract

An adaptive spatiotemporal neural network (ASTNN) capable of performing a near Bayesian speech recognizer is proposed. A tuning algorithm is introduced in the ASTNN to adjust the threshold parameters in order to maximize the discrimination distance between competing patterns. The highly confusable English E-set (b, c, d, e, g, p, t, v, z) is used to demonstrate the efficiency of the ASTNN. We show that the recognition rate can reach 80% for single-speaker and 68% for multispeaker respectively.