SPEECH SYNTHESIS MODEL LEARNING DEVICE, METHOD, AND PROGRAM

摘要

PROBLEM TO BE SOLVED: To learn HMM using, as a voice feature amount, a CWM parameter for which the index of each Gaussian function is guaranteed so as to match in the same state.SOLUTION: An auxiliary variable update unit 262 updates an auxiliary variable λ. A CWM parameter update unit 264 updates a CWM parameter so as to reduce an auxiliary function. A state output distribution update unit 268 and a state series update unit 270 update a state output distribution and a state series, which are HMM parameters, so as to reduce the auxiliary function. An observation spectrum envelope series posterior probability update unit 272 updates an observation spectrum envelope series posterior probability by using the updated CWM parameter and HMM parameter. A second convergence determination unit 274 repeats the updating of the CWM parameter and HMM parameter when it determines that a convergence condition is not satisfied, and outputs the CWM parameter to a HMM learning unit 28 when it determines that the convergence condition is satisfied.