[opus] Alleged bug in Silk codec
jmvalin at jmvalin.ca
Mon Jun 16 14:23:07 PDT 2014
Thanks for the info and the proposed fixes. I'm currently investigating
what's going on here before deciding on the best way to fix the problem.
Have you been able to figure out why it doesn't work for rshifts >= 3?
On 13/06/14 12:28 PM, Marcello Caramma (mcaramma) wrote:
> Hi Jean Marc,
> please find attached the audio file (mono 16khz). I shortened it to about
> 10 seconds. I also add 2 patches that worked for me. Further info that
> might help:
> - The problem seems to be related to silk_burg_modified not reaching the
> maximum gain, so the actual filter order is 16 rather than 2 (which is
> what would be expected with a sine wave).
> - The problem seems to happen when rshifts >= 3
> - when pre-scaling the signal to be < 16384 the problem goes away (patch
> - When calculating C0 and rshifts based on a 64 bits correlation instead
> of using silk_sum_sqr_shift the problem also goes away (patch
> I suspect that for very high prediction gain the fixed point
> implementation becomes very sensitive to numerical error, but I am not too
> sure why the new versions work better.
> I favour the version with the new C0 calculation, as it avoids rescaling
> the input.
> I also played around with a version (not attached) that prescales the
> input by rshifts/2 - this might be considering as it simplifies the code.
> PS: I am using 1.1 but the same issue is present with the latest code well.
> On 13/06/2014 06:05, "Jean-Marc Valin" <jmvalin at jmvalin.ca> wrote:
>> Hi Marcello,
>> Thanks for the report. It's hard to debug this without the actual file.
>> Can you please post the sweep_in.raw file you used?
>> On 11/06/14 04:46 AM, Marcello Caramma (mcaramma) wrote:
>>> Apologies if this is a known issues, but I have found what I believe is
>>> a bug in the fixed point implementation of the Silk codec and could not
>>> find any mention on this in the archives.
>>> The bug can be easily reproduced with the fixed point demo program
>>> (./configure ‹enable-fixed-point ‹disable-float-api && make) using the
>>> following command:
>>> ./opus_demo voip 16000 1 23000 sweep_in.raw sweep_out.raw
>>> Where sweep_in.raw is a 30 seconds full scale frequency sweep from 0 to
>>> 8kHz sampled at 16kHz.
>>> The first 6 seconds of audio after transcoding sound Ok. After that
>>> artefacts are introduced all the way to the end of the file.
>>> The floating point version does not have the issue (even though the
>>> quality is subjectively worse roughly from the same point).
>>> I believe I narrowed down the problem to the file burg_modified_FIX.c
>>> if I make sure the input signal is scaled down to 14 bits before
>>> processing the coefficients of the predictor are calculated correctly
>>> and no artefact is introduced.
>>> Is anyone experiencing the same problem or has a proper fix for this? (I
>>> can work around the bug with input scaling for now).
>>> Thanks and best regards,
>>> Marcello Caramma
>>> opus mailing list
>>> opus at xiph.org
More information about the opus