PURPOSE: To enable accurate speech recognition by updating a reference vector only when it is decided that the reference vector is updated with information in a reference vector update category table.

CONSTITUTION: The speech signal inputted from an input device 1 such as a microphone is converted by a feature sequence conversion part 2 into speech patterns X=x1, x2...xI (I: number of frames of input speech) as a time series of feature vectors. Various feature spectra which are effective for speech recognition are known and the output of 15 band-pass filter group which have center frequencies of 250 - 6,300Hz at frame period intervals of, for example, 10ms is used. The input speech patterns for reference spectrum update are divided by a pattern division part 3, phoneme by phoneme. The divisions are, for example, frames wherein difference vectors of feature vectors become maximum.