A Signal Processing Method Based on a Homotopic Correlation Product Applied to Speech Recognition Problems

摘要

The assumptions used to formulate the processing method, the proposed algorithm, and phoneme recognition test results of a homotopic signal processing method are presented. The hearing system is considered as a box with one imput, that applies a signal whose information content = 500 Kbit/sec, and many thousand outputs, the nerve fibers, having a transmission rate variable between 30 and 400 bit/sec. The signal transmitted by any one fiber is a series of equal impulse. Homotopic representation of a phoneme is available in steady state after 2 to 3 msec. The phoneme patterns are very different, although patterns for the same phoneme from different speakers are similar. Transition patterns between phonemes change rapidly. Recognition rate, using a minicomputer, of all possible combinations of 'a', 'e', 'r' and 'm' is 95.2%.