Progress Report
November 15, 2024
Plots of the whole track and its FFT
Plots of the vocal track and its FFT
Nonnegative Matrix Factorization (NMF) is a method commonly used in source separation. It approximates a nonnegative matrix as the product of two smaller nonnegative matrices, i.e. V = W*H. In our case, V would be the spectrogram of our audio. Below is an attempt to decompose the audio into two sources.
As there are four channels in the mixture audio (vocal, drum, bass, guitar), we also tried decomposing into four components:
Apparently, this method doesn't work quite well. We plan to try NMF on other audio tracks to see how it works, experiment different NMF settings in the coming weeks (like increasing the number of components), try different source separation method (like matched filter).
​
We learned that we can use nnmf function in Matlab to perform NMF. This helps us skip the technical algorithm details and focus on the testing of this method.
​