APPLICATION OF NEURAL NETS TO STAGED SPEECH RECOGNITION

NEURAL NETWORK COMPUTATION IS PARTICULARLY SUITED TO FINDING OPTIMAL SOLUTIONS TO PROBLEMS THAT HAVE A LARGE NUMBER OF POSSIBLE SOLUTIONS. THE TASK OF SPEECH RECOGNITION IS TO TRANSFORM ONE OF AN INFINITELY VARYING SET OF ACOUSTICS WAVE-FORMS INTO ONE ORTHOGRAPHIC TEXT STRING OUT OF THE INFINITE SET OF ENGLISH SENTENCES. GIVEN THE STRUCTURE OF THE PROBLEM OF SPEECH RECOGNITION, IT MAKES SENSE TO LOOK FOR CONBERGENCE STRATEGIES. FURTHERMORE, SINCE HUMAN BRAINS ARE MADE OF NEURONS AND BRAINS ARE THE ONLY PROCESSES THAT DO SPEECH RECOGNITION WELL, IT IS WISE TO APPLY NEURONAL COMPUTATIONAL MODELS TO THE TASK. WE PROPOSE TO EXPLORE THE APPLICATION OF NEURAL NETWORK PROCESSING STRATEGIES TO THE LATER STAGE OF SSI'S SPEECH RECOGNITION SYSTEM, WITH THE GOAL OF IMPLEMENTING A SECTION OF THIS COMPONENT AS A NEURAL NETWORK. THE FINAL REPORT OF THE PHASE I WORK WILL SHOW ANY POSSIBLE GAINS IN RECOGNITION SPEED, ACCURACY AND SPEAKER INDEPENDENCE.