A Pre-Matching Method for a Real Time Spoken Word Recognition System and a Learning Procedure of Speaker Differences.

Abstract

If we enlarge the vocabulary size of the word recognition system to about several hundreds, we are afraid that the recognition time becomes not only very long by increasing an amount of processing but also the correct rate of recognition decreases. To cope with these weak points, we adopted the method which reduced candidate words in the vocabulary by means of pre-matching using both local and global features of a spoken word. That is, to eliminate the most unlike group of candidates using the measurements of both features from the vocabulary list was tried to reduce the recognition time, and this operation also eliminated the misleading candidates to make increase the correct rate of recognition. Furthermore, to add the measurement to the final judgement made increase the correct rate. Moreover, in order to absorb the influence of speaker differences, we added the capability of learning to the system. In an experiment on name recognition using 100 Japanese-city names, the system recognized the names correctly at the rate of 83% for unspecific speakers and 93% after learning, using a mini-computer in real time. The number of candidate words was reduced to one tenth by pre-matching.