Abstract: In speech recognition, an unknown input signal is analyzed for its feature pattern sequence A.sub.i which is compared to prestored reference patterns grouped in categories. Each category m has a representative reference pattern B.sup.m which represents the average value of the n patterns in the catagory B.sub.n.sup.m. A first comparison (prematching) to the representative patterns is followed by a second matching to the actual reference patterns in the most likely category. The representative pattern is used in pre-matching and matching, thereby simplifying calculation and increasing accuracy.