METHOD FOR SPEECH RECOGNITION USING HIDDEN TRACK, HIDDEN MARKOV MODEL

摘要

PROBLEM TO BE SOLVED: To provide a method for speech recognition that finds generation related values of states of individual frames, in concrete, vocal tube resonance frequencies according to generation related values related to two precedent frames by using a recursive method.;SOLUTION: Generation related values are used to find a probability distribution of observation feature vectors of their states. Then the probability of observed values received as to the frames is found from the probability distribution. In an embodiment, the generation related values are found by using recursive definitions of no noise for the values. Using the recursive method greatly improve a decoding speed. When decoding algorithm is applied to learning data including a known transcription, forcible matching that improves note segmentation obtained by conventional technology is generated.;COPYRIGHT: (C)2004,JPO&NCIPI