[CELT-dev] llcon software using CELT

[Volker]

Gregory Maxwell schrieb:
> On Fri, Aug 21, 2009 at 1:12 AM, Gregory Maxwell<gmaxwell at gmail.com> wrote:
>   
>> On Fri, Aug 21, 2009 at 1:07 AM, Volker<v.fischer at nt.tu-darmstadt.de> wrote:
>>     
>>> Gregory Maxwell schrieb:
>>>       
>>>> Hm? If you call celt_decode with a null pointer in place of the data
>>>> it should fade out the audio after consecutive losses.
>>>>
>>>> The relevant code is around line 1286 in celt.c:
>>>>   for (i=0;i<C*N;i++)
>>>>      freq[i] = ADD32(EPSILON, MULT16_32_Q15(QCONST16(.9f,15),freq[i]));
>>>>
>>>>         
>>> If I interpret your code correctly, you use an exponential decay for the
>>> fade out. In previous software projects I did something similar and got
>>> strange effects when I applied a multiplication to very small floating point
>>> values. I guess the same happens here, too. You should introduce a bound for
>>> the floating point values. If the signal is below the bound, set the
>>> floating point value to zero and the problems should disapear (I guess ;-)
>>> ).
>>>       
>> The addition of EPSILON prevents the creation of denormals and is more
>> efficient than the compare and branch required for zeroizing.
>>     
>
> This makes me think however.. are you applying any gain control the
> the output which might be making a quiet tone loud?
I finally understand the code you posted above :-). I was a bit confused 
by the ADD32 operation but with your explaination it makes sense now. 
Putting your explaination as a comment in the code would maybe help 
others, too.

In llcon, I do not apply any gain to the decoded audio signal. The mono 
signal is just copied in both stereo channels of the sound card and then 
played by the sound card.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.xiph.org/pipermail/opus/attachments/20090821/205030a8/attachment-0002.htm