To solve the problem wherein it is generally difficult to automatically estimate HMM as a sound standard pattern without speech data labeled whose speech waveform is labeled with a 'sub-speech segment (SPS)' symbol although speech recognition performance can be made more sophisticated by introducing the SPS being a kind of fine segment.

An HMM parameter of an XSAMPA symbol which can be obtained relatively easily is regarded as a starting point. A method for automatically estimating the HMM parameter of SPS using the HMM parameter is developed and mounted on a speech recognition system. This sound model standard pattern generating device can generate a proper sound standard pattern as to an SPS symbol system even when the data whose speech waveform is labeled with an SPS series are not available. To perform estimation calculation of HMM for the 'sub-speech segment SPS' for realizing high-performance speech recognition, an initial value of the HMM parameter has to be properly determined as to a necessary SPS symbol. Here, a method for setting the initial value of the HMM parameter as to SPS from the value of a parameter when HMM of an XSAMPA symbol is given provided.

COPYRIGHT: (C)2003,JPO