Adaptation and use of an algorithm for recognition of connected words in an application for mobile radio telephony is described. Two data bases were collected in a compact car running about 120 km/h containing speech uttered via handset and in hands-free mode for each 10 speakers. In the first phase, a connected-words recognition algorithm was improved using the handset data base. Starting with a recognition mismatch rate of less than i% for undisturbed speech, 14.5% errors were measured for noisy speech and an unmodified algorithm. After several modifications and introduction of new modules for speech signal preprocessing, the error rate decreased to 3% for handset data and 13.2 % for hands-free data. Work on recognition in hands-free mode is still in progress.