Multiple approaches are needed to achieve robustness in speech recognition in an automobile due to various noise sources, relatively enclosed cabin acoustics, and the manner of speech of the driver under stress. A HMM-based word recognizer was developed for automobile applications, by integrating cepstral normalization into the front end and by training model parameters in multiple conditions according to a few driving scenarios mainly in city traffic. Speaker-dependent models were built and evaluated for a male and a female speaker using samples recorded in an automobile, and conducted a road test in city traffic. It was found that the system works well with several percentage points degradation in recognition accuracy compared with the laboratory experiments. The difference was attributed to the errors in word boundary detection.