We purpose to demonstrate the feasibility of speech recognition on a VLSI neurocomputer. The speech recognition system performs speaker-independent recognition of spoken English letters. The current implementation of the system, now running on a sun 4 workstation, classifies letters of the English alphabet at 94% accuracy - the best reported performance of any system on this difficult task. The high level of accuracy is obtained by training neural networks to make the important classification decisions at each level of the system. Neural networks are used to track pitch, to locate speech boundaries, and to classify letters. The goal of the Phase I research is to implement a complete recognition system in which neural network classification is performed in real time on the ASI board. The research consists of: (a) experiments needed to modify the current recognition system to meet the computational requirements of the board; (b) training the neural classifiers on the ASI simulator; and (c) implementing the classification modules on the ASI neurocomputer. Anticipated benefits/potential commercial applications - there are a variety of near-term applications for speaker-independent recognition of letters and digits, including credit card verification and automatic directory retrieval. A statistic that is commonly mentioned by researchers at telephone companies is that every second removed from an average interaction involving a human operator saves the company approximately $10 million per year.Key words: neurocomputers, neural networks, speech recognition, signal processing, classification
