I’m currently setting up a system and had a question I wanted to sort out before I made any more decisions. The MUSDB18 dataset has dual channel audio for the input audio files. It would obviously be simpler to just mix this into a single source, but that may not make sense if the evaluation is checking the ability to reproduce the de-mixed dual channel audio. So is the evaluation ran on the dual channel audio?
Yeah, the evaluation is run on dual channel audio, you can check the starter kit for details.