Improves quality by not squeezing 32 bit samples down to 16 bit, and
then back to 32 bit floating point.  Reduces CPU usage by skipping a
conversion step.