PURPOSE: To reduce misrecognition due to the appearance of an incorrect word even with small throughput by finding the rate of a state wherein self-transition is absent from HMM(hidden Markov model) state transition at the time of probability calculation of each word and weighting the output probability of each word by using the value.

CONSTITUTION: This device is equipped with a feature extraction part 12 which divides a digital value obtained from a speech signal through conversion into frames and calculates the speech feature quantities of the respective frames, a standard data storage part 13 which stores HMM of a word speech that is already registered, and a probability calculation part 14 which calculates the output probability of each word in a frame section of each frame. Here, only some state continues long in the HMM state transition of an incorrect answer word due to appearance and in other states, features having no self transition are utilized. Namely, a weighting process part 15 weights the output probability of each candidate on the basis of the rate of the state having no self-transition to the HMM state transition and a recognition result decision part 16 excludes appearance candidates on the basis the weighing result.