The acoustic model parameter learning device, acoustic model parameter learning manner, the characteristic quantity of the voice data for acoustic model

摘要

PROBLEM TO BE SOLVED: To provide a sound model parameter learning device using an object function for discrimination learning, which clearly has strong relation with the number of errors of words and phonemes.;SOLUTION: A sound model parameter learning device includes an error value count part, a differentiated value calculation part, a sound model parameter update part and a convergence determination part. A recording part records an object function with risk weighting. The error value count part calculates a difference degree from a teacher label with a predetermined particle size as a risk degree, for each symbol sequence. The differentiated value calculation part calculates a differentiated value or a difference value, when the risk weighting is 0. The sound model parameter update part updates a sound model parameter by an update amount according to the differentiated value or the difference value. When update of the sound model parameter satisfies predetermined condition, the updated sound model parameter is set to be the sound model parameter, and when it does not satisfy the condition, processing of the differentiated value calculation part and the sound model parameter update part are repeated, by the convergence determination part.;COPYRIGHT: (C)2011,JPO&INPIT