Speech recognition system in which voiced intervals are broken into segments that may have unequal durations

摘要

In a speech recognition system which performs a frequency analysis of speech input and determines its spoken interval, a time-normalized pattern (matching pattern) is prepared for use in similarity calculations. In the time-normalization, the time axis is segmented in such a manner as to cancel the effects of differences in the length of syllables in the input speech. With such normalization, it is possible to achieve high performance in linear matching of the input speech pattern against reference patterns using simple processing and a small memory space even when there are variations in the speaking rate.