Time-Frequency Masks for Reconstruction

  • synopsis: TF masks allow high quality reconstruction for decomposed components when you still have access to the original mixture

  • Assumptions: Additive mixture;

  • Many spectral decomposition techniques (including NMF and HPSS) work only with the magnitude of the spectrum, and ignore the phase

  • Resynthesis under these conditions is hard

  • However, if we still have the original spectrum, we can use our magnitude-only thing as a mask, essentially re-using the phase information from the original

  • An important difference between doing this directly when we have a collection of masks is that we scale each one by the total contribution of all of them

  • This allows us to null sum w/r/t the original

  • OTOH, it makes for a less pronounced separation

  • An alternative is to make a winner-takes-all binary mask
  • But this introduces more artefacts