摘要 |
PROBLEM TO BE SOLVED: To make compatible the reduction of an operation processing amount required for generating separation matrices and an improvement in the accuracy of a sound-source separation. SOLUTION: A signal separator 24 generates separation signals Y1(t) and Y2(t) by applying the separation matrix W(fk) to observation signals V1(t) and V2(t) collecting the mixed sound of sounds SV1 and SV2 by sound collectors M1 and M2. A frequency selector 44 selects a frequency fk to a first frequency fA and a second frequency fB. A first learning portion 54 generates the separation matrix WA(fk) of the first frequency fA by a primary learning processing utilizing learning data D(fk) of the first frequency fA. A direction estimator 62 estimates directions θ1 and θ2 of each sound source from the separation matrix WA(fk). An initial matrix setter 52 generates an initial separation matrix WB<SP>"0"</SP>(fk) so that the dead angle of sound collection is formed in the directions θ1 and θ2. A second learning portion 66 generates the separation matrix WB(fk) of the second frequency fB by carrying out secondary learning processing applying the learning data D(fk) of the second frequency fB by the number of occurrence smaller than the primary learning processing while using the initial separation matrix WB<SP>"0"</SP>(fk) as an initial value. COPYRIGHT: (C)2011,JPO&INPIT |