| 摘要 |
<p><P>PROBLEM TO BE SOLVED: To reduce a storage capacity required for generation (learning ) of a separation matrix. <P>SOLUTION: A storage device 14 stores observed data D(fk) for indicating intensive time sequence at each frequency fk in each of the plurality of observation signals V, in which mixed sound of multiple sound is collected, for each of K pieces of frequencies f1 to fK. An index calculator 52 calculates an index value of significance of learning of the separation matrix W(fk) which uses the observation data D(fk) of each frequency fk from the observed data D(fk). A frequency selector 54 selects at least one frequency fk in K pieces of frequencies f1 to fK according to the index value of each frequency fk calculated by the index calculator 52. A learning processor 44 determines the separation matrix W(fk) of the frequency fk by learning an initial separation matrix W0(fk) using the observed data D(fk) of the frequency fk which is selected by the frequency selector 54. <P>COPYRIGHT: (C)2010,JPO&INPIT</p> |
