A spatio-temporal pattern recognition approach to word recognition

Abstract

A limited-vocabulary, speaker-independent, isolated-word recognition system has been built. This system recognizes digitized isolated words without performing segmentation, phoneme identification, or dynamic time warping. It uses a static pattern recognition approach to recognize a spatio-temporal pattern. The preprocessing only includes endpoint identification, preceding and trailing silence removal and word length determination. A fourth-order linear prediction coding (LPC) analysis is preformed on each of 32 equally spaced frames. The four LPC coefficients plus four other features from each frame are input to a neural network consisting of 50 hidden-layer units and four output-layer units. The authors have trained the system for four words spoken by one single speaker. The system is then able to recognize four words from each of five other speakers.< >