SPEECH RECOGNITION IN THE NOISY CAR ENVIRONMENT

An algorithm for recognition of connected words has been adapted to an application for mobile radio telephony. For this purpose, several manners of generating feature vectors were evaluated using two databases collected in a small car moving at about 120 km/h. The databases contain digits and digit strings uttered via handset and in hands-free mode for each of 10 speakers. In the first phase, experiments were done using signal analysis resulting in linear scaled magnitude filterbank coefficients in the feature vector. This approach achieved a digit error rate of 6.0% on 7-digit strings collected via handset and employing isolated word training. The second approach replaced the magnitudes of the filterbank coefficients by coefficient energy, with the energy being scaled logarithmically. The digit error rate was reduced to 3.0% for handset data. For hands-free data, a digit error rate of 13.2% was achieved. The results derived in the simulation environment were verified in a car using the product hardware. © 1991.