Acoustic feature quantity conversion apparatus, acoustic model adaptation apparatus, acoustic feature quantity conversion method, acoustic model adaptation method, and program

摘要

PROBLEM TO BE SOLVED: To improve a recognition ratio even when learning data are not sufficiently available.SOLUTION: A featured value extraction part 13 generates an object acoustic featured value system extracted from an object voice signal and a reference acoustic featured value system extracted from a reference voice signal. A featured value collation part 14 collates a correspondence relation between the object acoustic featured value system and the reference acoustic featured value system to generate a collated object acoustic featured value system and a collated reference acoustic featured value system. A conversion model generation part 15 converts the acoustic features of an input acoustic featured value on the basis of a correlation relation between the object acoustic featured value system and the reference acoustic featured value system by using the collated object acoustic featured value system and the collated reference acoustic featured value system to learn a conversion model for outputting an acoustic featured value system. A pseudo featured value generation part 16 converts the acoustic features of the acoustic featured value system extracted from a voice signal by using the conversion model to generate a pseudo acoustic featured value system.SELECTED DRAWING: Figure 1