摘要 |
PROBLEM TO BE SOLVED: To provide a method for deciding an intrinsic space when a model having a drastically large number of model parameters is adopted. SOLUTION: In the method, a set of specified speaker models is generated by using learning voice data for an individual learning speaker, more than one model parameters are respectively connected to matched super vectors, a connection model for each speaker is displayed in a high-dimensional model space and conversion using contraction standard based on vector fluctuation property is performed in order to obtain an intrinsic space base vector. At first the high-dimensional model space is contracted to a speaker part space where the whole learning speakers are expressed by using base conversion and, then, conversion is performed to the vector for expressing the learning speaker in order to obtain the intrinsic space base vector.
|