hmm apparently libopus preferably encodes to floating point sample format, interesting. The fixed point format is said to be worse quality and meant for embedded software and might require recompiling libopus to do it.

Cool.