To fix  issue 673782 , essentially two copies of the decoded audio data are held in memory.  The first is a list of decoded packets, converted to float data, and the second is the final destination to which the decoded packets are copied.

To reduce memory usage, it's possible to keep the decoded packets in their "native" format.  For some codecs this format is 16-bit signed integers, which take up half as much space as the final float destination format.  

We still end up using more data than before, but the total memory usage can be reduced.