TARGET SOUND SOURCE ESTIMATION DEVICE, TARGET SOUND SOURCE ESTIMATION METHOD, AND TARGET SOUND SOURCE ESTIMATION PROGRAM

To estimate a target sound source even under a condition that the number of sound sources is unknown.SOLUTION: A target sound source estimation device 1 has: a time frequency analysis unit 11 which configures observation vectors of M observation signals recorded when acoustic signals corresponding to N target sound sources coexist together; a feature quantity extraction unit 12 which extracts acoustic features quantity from the observation vectors; an embedded vector estimation unit 13 which receives the acoustic feature quantities and outputs a D-dimensional embedded vector for every time frequency point through a neural network; a sound source number estimation unit 14 which finds an estimated value ^N of the number of sound sources based upon the embedded vector; a clustering unit 15 which clusters the embedded vector into ^N clusters each corresponding to a cluster of the target sound source based upon the embedded vector and the estimated value of the number of sound sources; a mask estimation unit 16 which calculates an estimated value of a mask corresponding to the ^N target sound sources based upon a cluster index; and a target sound source estimation unit 17 which estimates a target sound source using the observation vector and mask.SELECTED DRAWING: Figure 1