Speech recognizer for voice control of mobile telephone

Infovox is marketing a speaker-dependent, pattern-matching word recognition system, developed at KTH. The algorithms in the system have been modified for noise immunity, and performance has been evaluated in moving cars. The main problems were word detection and noise compensation. After simulations we decided to use a close-talking microphone and a "noise addition" method, where we added the measured noise in the moving car to the reference patterns recorded in a parked car. Using this method, the recognition rate was improved from 69% to 97% on a ten-word vocabulary using the best microphone. A more extensive test was performed on the modified recognition system using two cars and twelve speakers, seven male and five female. Most of them were naive speakers. The twenty-word vocabulary contained some confusable words and was trained in a parked car. During 98 sessions, 1,960 words were read under different conditions with an average recognition rate of 86%. With closed windows at 90 km/h the mean was 91%. An open window at the same speed decreased the result to 82%.