Skip to content

Add support for Cocktail-Fork model

StefanUhlich requested to merge StefanUhlich/sdx-2023-cinematic-sound-demixing-starter-kit:master into master

Hello @dipam and @gordon_wichern, here is a MR to add the cocktail-fork model as a baseline to the starter-kit.

Currently, the scores are not so high (although the score for the test sample was very high):

  • Test sample:
Evaluation Results
sdr_dialog 11.773962020874023
sdr_effect 1.1484180688858032
sdr_music 8.998164176940918
mean_sdr 7.306848049163818
  • Evaluation on Split 1:
private_sdr_dialog  
4.578439905400412
private_sdr_effect  
-2.2845692574590832
private_sdr_music  
-1.9252570330032281
private_mean_sdr  
0.122871212866883
sdr_dialog
4.980081608633091
sdr_effect
-0.7943571978339962
sdr_music
-3.446344327985072
mean_sdr
0.24646003369022818

I am not sure why the scores are so low - maybe there is a problem with the output scaling which is needed due to using SI-SDR? I also saw that the SI-SDR implementation subtracts the mean (https://github.com/merlresearch/cocktail-fork-separation/blob/main/si_snr.py#L12), which could be another problem.

Ping: @GiorgioFabbro

Merge request reports