PURPOSE: To provide the voice recognizing method with satisfactory recognizing performance for unspecified speakers.

CONSTITUTION: The plural voice samples of the same word is divided into acoustic events, and a first standard pattern is prepared 15 by calculating an average value for each partial block in the same acoustic event and covariance. With this standard pattern as a standard pattern for learning, frames are made correspondent between each sample for learning and the standard pattern, and the partial vector is segmented again. A second standard pattern is prepared and defined 16 as for recognition by calculating the statistic information of the partial vectors belonging to the same segmented class again. When preparing this standard pattern 16 for recognition, a statistic distance scale suitable for handling the data of unspecified speakers can be utilized, therefore, a lot of learning sampled can be divided into acoustic events with high accuracy, and the high-quality standard pattern can be prepared.