METHOD AND DEVICE FOR SPEECH RECOGNITION

摘要

PROBLEM TO BE SOLVED: To make higher the precision of recognition for a continuous speech of an unspecified speaker which has been deficient by a conventional recognizing method using an HMM (hidden Markov molded). SOLUTION: The recognition rate for the continuous speech of the unspecified person is improved by using correlation information between frames lasting for a long section which is not used for neither conventional DP matching nor the HMM. For the purpose, a speech recognition model of the HMM having a plurality of states is used to segment input speech frames in corresponding sections by the states by force alignment, patterns obtained by averaging input feature patterns included in the respective segmented frames are connected by as many as the states, and the vector obtained as a result is used to a new recognition parameter to perform recognition. COPYRIGHT: (C)2004,JPO