PURPOSE: To enhance performance by updating a corresponding parameter by learning when the parameter dispersion value of the standard patterns of many speakers is larger than the intended threshold at the time of adapting the previously formed standard patterns to learning data.

CONSTITUTION: The standard patterns (SP) of the many speakers (MP) are formed by an FB argorithm in a learning section 2 by using the learning data in a memory 1 and are stored in a memory 3. The learning section 4 executes the learning of the SP by AL with the SP of the MP of the memory 3 as an initial model by using the learning data of the specific speaker held in the memory 5. A comparing section 6 reads out the variance σext of the SP of the MP from the memory 3 and sends a control signal to the learning section 4 when this value is larger than the intended threshold c. The learning section 4 updates only the parameter to this control signal, holds the SP in the memory 7 and recognizes the input voice by prescribed system using this SP. The SP is adapted to the specific speaker by the less learning data. The higher performance is thus obtd.