I tried it with their examples, trying to isolate speech only or background music only, but it seems to produce audio that is near identical to the original.