The paper presents an efficient computation algorithm of the output probability for a continuous HMM speech recognizer using a rough and detail HMM combination. In general, the more number of mixtures or the more number of contextual classes, the better accuracy with the heavier computation, and vice versa. The proposed algorithm first estimates the state output probabilities using the rough HMMs and then re-estimates those of the probable states using detail HMMs. We proposed two realizations for the algorithm and carried out experiments for each. Both results showed about 60% or 70% reduction of the output probability calculation with no reduction of recognition accuracy.