ADAPTIVE WAVELET PARAMETERIZATION OF SPEECH SIGNALS

FASTMAN PROPOSES TO UTILIZE THE ADAPTIVE WAVELET TRANSFORM (AWT) TO DECOMPOSE SPEECH SIGNALS INTO THEIR INTEGRAL, INFORMATION-BEARING COMPONENTS FOR USE AS A PREPROCESSOR FOR PATTERN RECOGNIZERS. FOR A GIVEN SIGNAL, FAST-MAN'S ADAPTIVE WAVELET TRANSFORM SELECTS THE BASIS WHICH PRODUCES THE MOST CONCENTRATED REPRESENTATION OF THE SIGNAL IN WHICH MOST OF THE INFORMATION IN A SIGNAL IS "COMPRESSED" INTO A FEW BASIS ELEMENTS ALLOWING THE IMPORTANT INFORMATION TO BE READILY EXTRACTED. IN THIS PROPOSAL WE DEMONSTRATE THE ABILITY OF THE AWT TO ISOLATE THE INFORMATION BEARING ELEMENTS OF SPEECH SIGNALS BY SHOWING THAT MOST OF THE INFORMATION IN COMPLEX SPEECH SIGNALS CAN BE COMPRESSED INTO 1/100TH THE NUMBER OF ORIGINAL COEFFICIENTS. WE BELIEVE THAT OUR APPROACH WILL MAXIMIZE SPEECH RECOGNITION AND TALKER IDENTIFICATION SCORES WHILE MINIMIZING THE EFFECTS OF NOISE AND CHANNEL DISTORTIONS. IN PHASE I WE WILL: 1) PRODUCE AN ANALYTICAL STUDY AND SIMULATION WHICH COMPARES THE AWT TO CURRENTLY-USED TECHNIQUES AND DEMONSTRATES ITS ADVANTAGES OVER THOSE TECHNIQUES, 2) DETERMINE THE ROBUSTNESS OF THE DECOMPOSITION PRODUCED BY THE AWT IN THE PRESENCE OF NOISE AND CHANNEL DISTORTIONS AND 3) DEVELOP PATTERN RECOGNITION METHODS WHICH USE THE INFORMATION-BEARING SPEECH COMPONENTS EXTRACTED BY THE AWT TO IMPROVE SPEECH RECOGNITION ALGORITHMS. ANTICIPATED BENEFITS/POTENTIAL APPLICATIONS - THE ADAPTIVE WAVELET TRANSFORM CAN PROVIDE THE BASIS FOR NEW DATA COMPRESSION AND PROCESSING ALGORITHMS FOR: VOICE MAIL, TEXT-TO-VOICE, VOICE ANNOTATED RECORDS, VOICE RECOGNITION, VOICEPRINTING, VIDEOCONFERENCING AND OTHERS.