This paper describes issues in developing a noise robust speaker-dependent isolated word recognizer usedin a running automobile. A new definition of signal-to-noise ratio is proposed to cope with the low-pass type auto noise and frequency characteristics of human hearing, and to evaluate speech power level properly. Noise robustness is realized by the improvement of end-point detection and sharpening of spectral peaks. To minimize memory size and computational load, cepstrum coefficients of the matching template are quantized with the steps designed according to the inverse of cepstrum weighting function. Recognition and ADPCM recording-synthesis are executed in parallel in one DSP. Road tests revealed that the recognition rate for 20 words vocabulary exceeds 95% with 2 successive utterances and 40cm of microphone distance.