FEASIBILITY OF SPOKEN LETTER RECOGNITION ON A VLSI NEUROCOMPUTER

WE PROPOSE TO DEMONSTRATE THE FEASIBILITY OF SPEECH RECOGNITION ON A VLSI NEUROCOMPUTER. THE SPEECH RECOGNITIONSYSTEM WILL PERFORM SPEAKER-INDEPENDENT RECOGNTION OF SPOKENENGLISH LETTERS. THE CURRENT IMPLEMENTATION OF THE SYSTEM, NOW RUNNING IN THE CONTEXT OF A DIRECTORY ASSISTANCE APPLICATION ON A SUN4 WORKSTATION, CLASSIFIES ISOLATED LETTERS OF THE ENGLISH ALPHABET AT 96% 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 TRACKPITCH, TO LOCATE SPEECH BOUNDARIES, AND TO CLASSIFY LETTERS. THE GOAL OF THE PHASE 1 RESEARCH IS TO IMPLEMENT ACOMPLETE RECOGNITION SYSTEM IN WHICH NEURAL NETWORK CLASSIFICATION IS PERFORMED IN REAL TIME ON THE ADAPTIVE SOLUTIONS BOARD. THE RESEARCH CONSISTS OF: (A) EXPERIMENTS NEEDED TO MODIFY THE CURRENT RECOGNITION SYSTEM TO MEET THE COMPUTATIONAL REQUIREMENTS OF THE BOARD; (B) IMPLEMENTING THE CLASSIFICATION MODULES ON THE NEUROCOMPUTER; AND (C) TRAINING THE NEURAL CLASSIFIERS.