NEURAL NETWORK DESIGNATED SPEAKER EXTRACTION FROM NOISE

THIS RESEARCH WILL DEVELOP AND DEMONSTRATE INNOVATIVE ARTIFICIAL NEURAL NETWORK (ANN) TECHNIQUES FOR EXTRACTING THE VOICE OF DESIGNATED SPEAKERS FROM VARIABLE NOISE/INTERFERENCE BACKGROUNDS, AND TRANSCRIBING THIS SPEECH TO TEXT. THE TECHNICAL APPROACH IS BASED ON: (1) INNOVATIVE METHODS USING HIDDEN-LAYER IDENTITY-MAPPING ANNS TO CREATE TRAINABLE NONLINEAR TRANSFORMATIONS FROM CORRUPTED SPEECH INPUT PARAMETERS TO CLEAN SPEECH OUTPUT PARAMEATERS; (2) THE USE OF OPTIMAL SPEECH PARAMETERS FOR SUCH MAPPINGS, BASED ON MODELS OF VOCAL AND AUDITORY PROCESSING; AND (3) SEAMLESS INTEGRATION OF THESE TECHNIQUES INTO AN EXISTING "CLEAN SPEECH" RECOGNITION DESIGN. IN PHASE I, INITIAL ALGORITHM DEVELOPMENT AND DEMONSTRATION WILL BE ACCOMPLISHED, USING TWO SPEAKERS, A SMALL VOCABULARY, AND A LIMITED RANGE OF INTERFERENCE BACKGROUNDS. PHASE I WILL VALIDATE BASIC FEASIBILITY OF THE APPROACH. A NON-REAL-TIME NOISY SPEECH EXTRACTION AND TRANSCRIPTION CAPABILITY WILL ALSO BE DEMONSTRATED ON LIMITED PROBLEM SETS, AND SPECIFICATIONS FOR A REAL-TIME SYSTEM WILL BE PROVIDED.