Speaker Adaptation in an Isolated Word Recognition System .

Abstract

A speaker-dependent isolated word recognition system can attain a high recognition rate. It is, however, not easy to train such a system with a large vocabulary, and is necessary to adapt the system to a specific speaker with a small set of his training samples. We present a new method for the speaker adaptation in an isolated word recognition system and give some preliminary experimental results of its application. The word recognition system is based upon the time-warping algorithm. The reference patterns used in the matching process are adapted to a new speaker in the following two steps. (1) The utterances of a word class spoken by multiple speakers are averaged to create a common reference of the class. (2) This common refernece is transformed so that its base points (usually selected from characteristic parts of vowels and consonants) coincide with speaker-dependent spectra extracted from the training utterances of the speaker. Some experiments were conducted to recognize 65 Japanese city names spoken by 20 adult male speakers. The common references were made from the utterances spoken by five seperate speakers. The recognition rate was 96.6% for the speaker-adapted references, while it was 96.4% for the speaker-dependent references and 95.7% for the common references. These results show that the proposed method is effective for the speaker adaptation.