A Recognition System of Connected Spoken Words Based on Word Boundary Detection.

Abstract

We have developed a real time spoken words recognition system on a minicomputer.1), 2) This system is based on the phoneme classification, that is, hierachical linguistic knowledge. To extend this approach to connected words, we developed a ward boundary detection method by using the pitch contour and energy envelope. In this paper, we discuss the possibility of word boundary detection by the prosodic information from view points of listening and reading. Further we describe the detection algorithm by machine. From experiments on connected digits and city names, the listener could detect correctly the number of words in connected spoken words at the rate of 95%, and reader could locate correctly the word boundary at the rate of 70%~100%. The detection algorithm was applied to connected spoken digits. The number detected as candidates of word boundary was 1.2~1.7 times of the actual number and the boundary was detected correctly at the rate of 85~95%. The recognition rate of two connected digits was about 83% for unspecific speakers.