SPEAKER ADAPTATION METHOD BASED ON JOINT BILINEAR TRANSFORMATION SPACE AND APPARATUS THEREFOR

摘要

PURPOSE: it is a kind of to be provided the uniform propulsion that same pedestal or model are shared in the heterogeneous speaker adaptation method between removing based on combination speaker adaptation bilinearity conversion space and its device, when execution speaker adaptation. ;CONSTITUTION: speaker model generates unit (110) and generates speaker SI (independence) model. Speaker model generates unit generation MLLR (transition matrix that maximum likelihood linearly returns) and passes through loudspeaker. Calculation matrix generation unit (120) generates two kinds of calculation matrix. Bilinearity converts spatial model and generates unit (132) generation bilinear transformation BITS (space) based on singular value decomposition SVD (pedestal). A kind of speaker adaptation model generation unit (140) generates speaker adaptation model according to the form of the SF of estimation. ;The 2013 of copyright KIPO submissions;[Reference numerals] (110) block 1; (111) speaker; (112) SI models; (113) modeling difference loudspeaker; (120) block 2; (130) block 3; (131) it generates bilinearity and converts space; (132) bilinearity converts spatial model; (140) block 4; (141) application combination converts space loudspeaker based on bilinearity; (142) speaker's adaptive pattern; (150) block 5; (AA) training data; (BB) 1 MLLR transition matrixes of speaker; (CC) speaker S MLLR transition matrixes; (DD) calculation matrix is generated; (EE) voice speaker adaptation; (FF) Te Site voices; (GG) speech recognition device; (HH) recognition result