A real time phonetically based spoken word recognition system and recognizer for unspecified speakers

Abstract

We have developed a real time phonetically based spoken word recognition system on a mini-computer and arecognition equipment which consisted of a filter bank and amicrocomputer. If the vocabulary size of the word recognition system is enlarged to about several hundreds, not only the recognition time becomes 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. Our system is based on the phoneme classification, that is, hierarchical linguistic knowledge. To extend this approach to connected word srecognition, we developed a word boundary detection method by using the pitch and energy contours. In an experiment on the recognition using 100 Japanese city names or ten digits, the system recognized the names (digits) correctly at the rate of 84%(98%) for unspecific speakers and recognized two connected digits at the rate of 83% in real time. Furthermore, we constructed a special equipment of speech recognition which was based on the system. In experiments on word recognition of 32 city names, the recognizer recognized them correctly at the rate of 93.4-94.4% for unspecific male and female speakers.