METHOD FOR SPEECH RECOGNITION USING VARIATIONAL INFERENCE WITH SWITCHING STATE SPATIAL MODEL

摘要

PPROBLEM TO BE SOLVED: To provide a training system that makes it possible to train the boundary of a hidden dynamic model together with other parameters and also overcome difficulty in handling accompanying such a training. PSOLUTION: (1) Defined is a switching state space model as to hidden production-related parameters with successive values and an observed speech acoustics. (2) Approximated is posterior probability providing the likelihood between a sequence of hidden production-related parameters and a sequence of speech units based upon a sequence of observed input values. When the posterior probability is approximated, boundaries between speech units are not fixed, but determined in the optimum form. In an embodiment, mixture of Gaussian approximation is employed. In another embodiment, posterior probability approximation of HMM is used. PCOPYRIGHT: (C)2005,JPO&NCIPI