davidkant / sndsplt

happy valley source separation

#sndsplt

happy valley source separation

###what

• really just a few conveniences for handling PLCA sound decomposition (all of the heavy lifting is done by Bregman Toolkit). sndsplit manages training and selective fitting of components. it allows you to train a collection of sources and decompose a mixture in terms of these sources + unknown components. this is particularly effective for sound separation when components can be isolated for training.

###depends on

• my fork of Bregman Toolkit modified to allow selective updating of PLCA components

sndsplt's Issues

ValueError?

I just changed the iterations to 10, but also tried with the default values you have and it gives me the same error, I don't understand:

fitting source models...
channel 0...
Traceback (most recent call last):
File "demo.py", line 61, in
sources, residual, mix = splt.fit(mix_info, sources, classifier=classifier,
priors=priors, **params2)
File "C:\soundsplit\sndsplt.py", line 297, in fit
initW=W_trained, updateW=W_update, win=win, **priors)
File "C:\Python27\lib\site-packages\bregman\plca.py", line 284, in analyze
logprob, WZH = params.do_estep(W, Z, H)
File "C:\Python27\lib\site-packages\bregman\plca.py", line 359, in do_estep
WZH = self.reconstruct(W, Z, H)
File "C:\Python27\lib\site-packages\bregman\plca.py", line 321, in reconstruct

return norm * np.dot(W * Z, H)

ValueError: operands could not be broadcast together with shapes (2049,10) (20,)

If you could help me with this, would be awesome.