We propose an efficient pre-classification algorithm extracting candidate words to reduce the recognition time in a large-vocabulary recognition system and also propose the use of spectral and temporal smoothing of the observation probability to improve its classification performance. The proposed algorithm computes the coarse likelihood score for each word in a lexicon using the observation probabilities of speech spectra and duration information of recognition units. With the proposed approach we could reduce the computational amount by 74% with slight degradation of recognition accuracy in a 1160-word recognition system based on the phoneme-level HMM.