It's a definite memory issue. The player has to read ahead and decode the audio first, then mesh it and store the result in memory (or something like that. That sounds a bit like how it was explained).