Background Ok so I am implementing a BSS algorithm to separate vocals from a hip hop song using the NMF method. I have done so by implementing the following paper by Feng: https://hal.archives-ouvertes.fr/hal-01547459/document
I create and cluster a source data matrix, R, just like in Feng's paper in equation (8) and assign each column/row of my factorised matrices V and H to a particular cluster. Multiplying together gives me the estimated STFT of my separated source. I have obtained this successfully and can hear the slight suppression of the background instrumental.
Problem/Question However, now that I have qualitatively deduced my algorithm is working I want to get a quantitative assessment using evaluation metrics: SDR, SIR & SAR. I am doing so using the paper at the top and with the 2.1 version downloaded MATLAB files. On page 11 it gives the following decomposition initially required before you calculate the metric ratios:
[s target,e interf,e artif] = bss decomp gain(se,index,S)
where se is a row vector containing the separated source and S is a matrix containing each of the sources in its rows.
My problem/question is that I do not have the sources in this format; I have them in their own unique spectrograms (TF view). How can I get my sources into this format in order to decompose the sources and then calculate the metric ratios?