When we have anough compute or the models get much smaller. At this point it seems wasteful to utilize gaming level GPU to decode audio stream.