Demonstrations of Research

 Source Separation Using Supervised NMF with Basis Deformation

In a source separation using supervised NMF, a sample sound of the target source must be used to achieve the separation.
For example, if we want to separate only piano sounds from the music signal that contains many instrumental sounds, we must prepare the sample sound (its scale sounds or something) of the same timbre.
However, in a real situation, it is impossible to prepare the perfect sample sounds in advance because the timbre depends on many factors, e.g., the recording environments, playing styles, individuality of the instruments.
The sample sounds only user can prepare are just an artificially synthesized signals made by DAW or the scale sounds played by user's own instrument.
To solve this problem, we proposed a new separation method called "Supervised NMF with basis deformation."
This method can separate only the components that have the similar timbre with the given sample sounds.
In this demonstration, the input signal is created by a real MIDI tone generator, and we prepare the sample sounds, which are synthesized using another MIDI tone generator.
Supervised NMF with basis deformation can separate only the target source even if the timbre of the sample sound is slightly different.

Iutput signal that contain five sources
Input signal
(Sax, A.Guitar, E.Guitar, Bass, Drums)

Output signals by supervised NMF with basis deformation
Given supervised signal
(Only sax scale synthesized by
another MIDI tone generator)
Separated signal
Residual signal
(A.Guitar, E.Guitar, Bass, Drums)

Given supervised signal
(Only bass scale synthesized by
another MIDI tone generator)
Separated signal
Residual signal
(Sax, A.Guitar, E.Guitar, Drums)

Copyright c 2014 Yamaha Corporation. All rights reserved.

    Top     Research Interests     Academic Background     Publications     Demonstrations     Program Codes     Work History     Awards     Contact me     Links