A telephone speech recognition system using word spotting technique based on statistical measure

A telephone speech recognition system for an isolated word which has sufficient performance in practical use is reported here. Word spotting technique is applied to the system so as to keep it relatively immune from noise. Word spotting is performed by a new time normalization algorithm based on linear time distortion pattern matching method named COLM(Continuous Linear Matching). The method includes only simple operations such as look up tables and additions. The entire system is implemented in a small board including a single digital signal processor. An experiment proved 96.4% average recognition rate. The experiment was carried out using 10 Japanese numerals pronounced by 240 males and females through telephone lines.