Actually, I’d defer to @groma for a more authoratative answer on this: my inkling is that, in practice, it probably doesn’t matter so much, as the NMF normalises everything internally.
If you were going to try and use an ensmble of rank-1 decompositions on the same sound, I’d start by just summing the dictionaries (but I’d also check to make sure that there was some degree of consistency between them, so that summing wasn’t just taking you towards a flat spectrum!)