PURPOSE: To prevent the output of a recognition result or response from being too fast and to improve a feeling of use by providing a delay part which delays the output of the recognition result or response by a specific time.

CONSTITUTION: A spectrum calculation part 12 divides a speech signal into about 20 frequency bands, reads squared mean values of respective band outputs in frame cycles of 10-20msec out as feature parameters, and obtains and supplies powers spectra of the speech to a DP matching part 14. This DP matching part 14 performs nonlinear matching between the feature parameters of the speech section obtained by the spectrum calculation part 12 and standard parameters stored previously in a word template 15 by dynamic programming(DP) technique and outputs the word of the standard pattern which is the least in distance obtained by similarity calculation after time base distortion is removed as the recognition output. Then the recognition result is delayed by the specific time at the delay part 16 and then supplied to a display part 17, where the recognition result of the input speech is displayed as characters.