Synthesis of speaker-adaptive word templates by concatenation of the monosyllabic sounds

This paper describes a new method for synthesizing speaker-adaptive templates in an isolated word recognition system based on the time-warping algorithm. We prepare in advance a template for each word class by averaging utterances spoken by a few talkers, and use it as a guide to excerpt the subpattern of an isolated monosyllable that is best matched against a syllable embedded in the word. The excerpted subpatterns are concatenated and smoothed at their boundaries to create speaker-adaptive word templates. The method was evaluated through the recognition of 44 Japanese city names spoken by 19 male talkers. The synthesized templates for each talker gave the average recognition rates of 95.2%. This shows that the proposed procedure is promising although the problem of compensating coarticulation effects is left unsolved.