Output probability calculation method and speech recognition apparatus of HMM

摘要

PROBLEM TO BE SOLVED: To reduce the amount of calculation or an memory area to be used as much as possible and obtain high recognition performance in speech recognition by HMM conducting the speech recognition after conducting vector quantization and obtaining output probability in table reference. ;SOLUTION: A code book used in conducting the vector quantization is one of each phoneme, for example, as a code book CB1 is one of a phoneme/a/and a code book CB2 is one of a phoneme/i/when a sub word is regarded as the phoneme and they correspond to each phoneme HMM11 and 12. When feature vector making speech analysis is made into the vector quantiztion with the code book CB1 and a code (label) is outputted, speech recognition processing is done by referring to tables T1, T2 and T3 in each situation of the phoneme HMM11 corresponding to the code book CB1 respectively, obtaining situation output probability toward the label and using the situation output probability as one of parameters.;COPYRIGHT: (C)2003,JPO