To enable fast speech recognition with a high recognition rate by using the analytic result of an unspecified speaker model (unspecified speaker HMM) as estimated values of respective spekaer classes HMM and making the calculation of output probability and a language search efficeint.

The device consists of a speech input device 101 including a microphone and A/D, a sound analysis part 102, an output probability calculation part 103, a speaker class HMM 104, a language search part 105, a grammar dictionary 106, and a display part 107. An input speech is analyzed, the output probability of the analytic result of both speaker models clustered by speaker classes and the input speech is found and on the basis of the found output probability, the recognition result of the input speech is determined. Models by vocal sounds are preferably generated from the inputted speech, the speaker class that the inputted speech is determined according to the distance between the generated models, and the model of the inputted speech is stored as the determined speaker class.