[Speex-dev] speex-1.2beta1 AEC garbles up audio unless compiled with --enable-fixed-point

[Jean-Marc Valin]

Thanks for the sample. I can reproduce it here, so I'll be able to
investigate. I'll let you know when I find something.

	Jean-Marc

Andras Kadinger a écrit :
> I'll try to make some shorter samples later, but for now here are the
> ones I have tried with:
> 
> http://www.surfnonstop.com/~bandit/speex/1.2beta1_AEC_garble/
> 
> The original recordings are in mic.raw and spk.raw.
> 
> Jean-Marc Valin wrote:
>> You may have triggered an instability problem. Can you upload your files
>> somewhere so I can have a look at them?
>>
>>     Jean-Marc
>>
>> Andras Kadinger a écrit :
>>  
>>> Greetings everyone,
>>>
>>> I was about to compare AEC performance between 1.1.12 and 1.2beta1 when
>>> I noticed something.
>>>
>>> If I configure (and compile) speex-1.1.12 with
>>>
>>> ./configure --enable-shared=no --enable-static=yes
>>>
>>> it compiles and works as expected: I can run a mic and speaker signal
>>> through testecho, it runs in a reasonable amount of time (about 23 secs
>>> for 3 minutes of audio) and I get back good audio.
>>>
>>> If I configure (and compile) speex-1.2beta1 with
>>>
>>> ./configure --enable-shared=no --enable-static=yes
>>>
>>> it compiles, but with the same input files testecho runs very long (say
>>> about 21-24 minutes, probably data-dependent), and the output audio is
>>> garbled up.
>>>
>>> It does the same with --enable-sse.
>>>
>>> But with --enable-fixed-point, AEC returns to normal.
>>>
>>> This is happening on a Gentoo stable (1.12.4) system, gcc 4.1.1, glibc
>>> 2.4, processor is "Intel(R) Pentium(R) 4 CPU 3.00GHz" (with
>>> HyperThreading disabled).
>>>
>>> My data is real-world audio captured from a physical speaker and mic
>>> setup in a room, by the left and right channels of the same soundcard
>>> input, and is sampled at 32 kHz. 1.2beta1 AEC without
>>> --enable-fixed-point garbles the audio up whether I claim it is sampled
>>> at 8 kHz or 32 kHz in testecho.c, it just does so differently. I can't
>>> hear anything of my original audio, instead I hear a very loud
>>> (full-scale), regularly repeating noise. When I look at the waveform, it
>>> looks a bit like 1/x for x>0, where overflowed samples values are not
>>> clipped but wrapped back down; that is, the first few dozen samples are
>>> wildly swinging between -32768 and 32767, later slow down but still
>>> full-scale, even later the sample values asymptotically approach zero
>>> from above; then suddenly the cycle begins again. (I can provide samples
>>> of input and output on request.)
>>>
>>> Is anyone else seeing anything like this?
>>>
>>> Thanks in advance.
>>>
>>> Best Regards,
>>> Andras Kadinger
>>> _______________________________________________
>>> Speex-dev mailing list
>>> Speex-dev at xiph.org
>>> http://lists.xiph.org/mailman/listinfo/speex-dev
>>>
>>>
>>>     
> 
> _______________________________________________
> Speex-dev mailing list
> Speex-dev at xiph.org
> http://lists.xiph.org/mailman/listinfo/speex-dev
> 
>