On improving problematic output and iterative usage of RAVE models.

[OlaFosheimGrostad]

Some inputs that the model doesn't really account well for create some problems. For instance, for the vintage model a cymbal crash can cause the output the blow up and therefor the normalized output will make for a very low signal for the rest of the song. I was a bit surprised that the output could go so high. Maybe the high frequency content is the problem? What can be done to correct this? I think this would be problematic in a live performance (although you should use some sort of compressor).

I've also noticed that input like a loud perfect square wave form (not bandlimited) which typically isn't found in recorded sound will cause a bad output. I've noticed that this output can be improved upon iteratively, so it converges towards something less ugly after a few iterations. Are there any examples of people using RAVE in an iterative fashion? I'd like to explore this approach more, and would like to see what others have tried out before continuing with this approach.