Word boundary detection by pitch contours in an artificial language

The pitch contour plays an important role of marking syntactic boundaries in the natural language. In order to improve the continuous speech recognition system we have developed, we have carried out experiments on how the pitch contour reflects word boundaries in an artificial language-continuously read material of a programming language. On the analysis of 173 sentences of 5 male speakers including 1655 word boundaries, 67.1 % of the total word boundaries were correctly detected and the false alarm is 28.8 %. The preceding and the following boundaries of keywords were found at the correct rate of 94.2 % and 85.0 % respectively. The application of boundary information to the recognition of continuous speech is also discussed in this paper.