[FFmpeg-soc] Instability in AMR-NB

[Colin McQuillan]

2009/8/4 Reimar Döffinger <Reimar.Doeffinger at gmx.de>:
> On Tue, Aug 04, 2009 at 01:32:00PM +0100, Colin McQuillan wrote:
>> I could set the excitation vector to zero whenever it is close it
>> zero, but this could be brittle. Should I just leave this as a bug?
>
> Is there any hint as to a recommended precision for the excitation vector?
> At the very least, treting denormals as 0 should be quite risk-free I guess...

Yes - it should be integral. And I can get rid of the buzz on this
particular test by adding:

        for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
            exc_feedback[i] = truncf(exc_feedback[i]);

But since the decoder is not bit-exact, how can I know that this has
solved the problem? I'd need a bound on the size of the error between
this decoder and the reference.