ACOUSTIC CLASSIFICATION WITH NEURAL NETWORKS

A TIME DELAY NEURAL NETWORK (TDNN) IS A TYPE OF BACK PROPAGATION NEURAL NETWORK THAT HAS BEEN USED TO RECOGNIZE TRANSIENT SPEECH PHONEMES. WE PROPOSE TO DEVELOP A TDNN FOR THE RECOGNITION OF NONSPEECH ACOUSTICAL SIGNALS WHERE THE SOUNDS MAY RUN TOGETHER AND MAY BE IN BACKGROUND NOISE. WHEN TRAINED TO RECOGNIZE PHONEMES, TDNN'S HAVE SHOWN A REMARKABLE ABILITY TO LEARN TO SEGMENT CONNECT PHONEMES. WE BELIEVE THAT THIS IS A GENERAL PROPERTY OF THESE NETWORKS THAT CAN BE USED WITH NON-SPEECH CONNECTED SOUNDS. IF THIS PROVES THAT FOR THE TDNN THAT WE WILL DEVELOP, WE WILL HAVE SOLVED A MAJOR (AND OFTEN INTRACTABLE PROBLEM IN AUTOMATIC SOUND CLASSIFICATION. BACK PROPAGATION NETWORKS FREQUENTLY RECALL EXCESSIVE TRAINING TIMES OR DO NOT CONVERGE AT ALL WHEN THE TRAINING SET IS LARGE AND COMPLEX. A NEW CLASS OF NEURA; NETWORKS CALLED ADAPTIVE TREES THAT REQUIRES NO MULTIPLICATION AND THAT HAS TRAINING CONVERGENCE RATES THAT ARE ONE OR MORE MAGNITUDES GREATER THAN THE RATES OF EQUIVALENT BACK PROPAGATION NETWORKS HAS BEEN DEVELOPED. WE WILL CONSTRUCT AN ADAPTIVE TREE NETWORK THAT ALSO WILL DO SEGMENTATION, BUT WHICH WE BELIEVE WILL ADAPT MUCH FASTER. THE INTERACTION BETWEEN THE NETWORKS AND A HUMAN OPERATOR DURING AND AFTER TRAINING WILL BE STUDIED.