PURPOSE: To provide a technique for automatically forming a phoneme dictionary which is used for a non-specific speaker speech recognizing device based on phonemes and in which the phonemes of many persons are previously stored.

CONSTITUTION: The data of the elementary piece waveforms of the many phonemes segmented from speech data of words, etc., uttered by many speakers of several tens persons or above are previously stored in an elementary piece waveform storage section 13. Speech synthesis is executed sequentially by each phoneme group in certain specified time (for example, at every other one second) in a phoneme synthesizing section 14. The speeches are outputted from a speaker 12. The speeches picked up by a microphone 20 are converted to digital data by an A/D conversion section 21. The digital data is subjected to calculation processing of recognition parameters in real time in a parameter conversion section 22 for recognition. The parameters of stationary parts between the speech sections for one second uttered in the speech synthesizing section are segmented by a phoneme segmenting section 23 and are stored into a phoneme dictionary storage section 34, by which the phoneme dictionary is formed.