[Speex-dev] Fwd: Re: Fixed Point on wideband-mode: Single Frame loss on 2000 Hz sine causes "freak off"

Thu Jan 14 04:15:31 PST 2010

What happens if you change that line:

if (cumul_gain > 262144)

to use a smaller value? What value works OK (if any)?

One more thing, when things go wrong, do they eventually go back to 
normal or does the codec never recover? It's unavoidable that the audio 
goes bad for a short period of time because of the long-term predictor.

	Jean-Marc

On 2010-01-14 05:57, Frank Lorenz wrote:
> Hi Jean-Marc,
>
> yes, problem exists in narrowband-mode, too.
>
> I already twiddled with max_gain, but did not have real success. I
> changed line 337 of ltp.c (function pitch_gain_search_3tap_vq)
>
> if (sum>best_sum&&  gain_sum<=max_gain) {
>
> to if (sum>best_sum&&  gain_sum<max_gain) {
>
> -- that stabilizes speex for 2000 Hz and 2200 Hz input on quality
> setting 7 (23800 bits/s) complexity 2, while it does not help much
> for quality setting 9 ((34400 bits/s) (just slightly reduces the
> duration of the "freak out").
>
>
> best regards, Frank
>
>
>
> Jean-Marc Valin<jean-marc.valin at usherbrooke.ca>  hat am 14. Januar
> 2010 um 02:27 geschrieben:
>
>> Hi,
>>
>> That's very useful information. From what you tell me, it may be
>> related to the handling of max_gain in pitch_gain_search_3tap()
>> (ltp.c). Also, are you able to reproduce the problem in narrowband
>> mode?
>>
>> Cheers,
>>
>> Jean-Marc
>>
>> On 2010-01-13 07:34, Frank Lorenz wrote:
>>> Hi Jean-Marc,
>>>
>>> yes, I tested with floating point. It is only a fixed point
>>> problem, floating point works fine. Even a fixed point decoder is
>>> o.k. -- problem is inside encoder.
>>>
>>> I had the idea of an "above unity" pitch gain, too, but I did not
>>> manage to modifiy the encoder in a way to fix it. Because the
>>> encoder's pitch handling is some kind of "feedback system" (i.e.
>>> a change in output parameters of frame N will influence input
>>> parameters in frame N+1), it's not so easy for me to reduce gain
>>> to test for stability. All modifications I tried made the
>>> situation worse, so I need some adivce how to proceed...
>>>
>>> The complexity parameter has some influence, too: For a sine
>>> input of 2200 Hz, the instability is present only for complexity
>>> values up to 3, but for the 2000 Hz input, instability can be
>>> observed even for the highest complexity values.
>>>
>>> As already written, I observed that the pitch and pitch gain
>>> values computed with the fixed point encoder seem quite odd and
>>> are absolutely different from the ones computed with floating
>>> point. This seems to stabilize for higher complexity values, so
>>> I'm not sure if this is the cause of the problem or only a side
>>> effect.
>>>
>>> best regards,
>>>
>>> Frank
>>>
>>>
>>> Jean-Marc Valin<jean-marc.valin at usherbrooke.ca>  hat am 13.
>>> Januar 2010 um 12:48 geschrieben:
>>>
>>>> Hi Frank,
>>>>
>>>> Sorry, I *do* care about the problem and just happen to be
>>>> overworked at the moment. What I suspect is that the pitch gain
>>>> gets close enough to unity that the loss makes it bust. Did you
>>>> test with the floating-point code?
>>>>
>>>> Jean-Marc
>>>>
>>>> On 2010-01-13 03:45, Frank Lorenz wrote:
>>>>> Hi,
>>>>>
>>>>> is no one willing to spent some effort on this topic? At
>>>>> least it would be good to know if anybody can reproduce this
>>>>> problem.
>>>>>
>>>>> I did further investigation on the topic and found out that
>>>>> the encoder in fixed point mode produces absolutely different
>>>>> pitch and pitch gain
>>> parameters
>>>>> than in floating point mode (for a 2000 Hz input signal). The
>>>>> codec breaks on frame loss if the encoder is in fixed point
>>>>> mode, decoder doesn't matter, so I assume the pitch
>>>>> parameters are "bad". Maybe its caused by accuracy/rounding
>>>>> problems for this periodic
>>> signal?
>>>>>
>>>>> Because I do not know much about the internals of Speex, it's
>>>>> hard
>>> for me to
>>>>> analyse the pitch / pitch gain computation. Some help would
>>>>> be really nice...
>>>>>
>>>>> best regards, Frank
>>>>>
>>>>>
>>>>>
>>>>> ---------- Ursprüngliche Nachricht ---------- Von: Frank
>>>>> Lorenz<Frank_wtal at web.de> An: speex-dev at xiph.org Datum: 4.
>>>>> Januar 2010 um 09:57 Betreff: Re: [Speex-dev] Fixed Point on
>>>>> wideband-mode: Single Frame
>>> loss
>>>>> on 2000 Hz sine causes "freak off"
>>>>>
>>>>>
>>>>> Hi Jean-Marc, all,
>>>>>
>>>>> I didn't get any response to my issue up to now and would
>>>>> like to
>>> now if
>>>>> anyone can reproduce this behaviour and if there is some idea
>>>>> what
>>> happens. I am
>>>>> willing to fix this issue, but because I do not know the
>>>>> internals of speex, I need some advice on how to proceed...
>>>>>
>>>>> best regards, Frank
>>>>>
>>>>>
>>>>>
>>>>> Frank Lorenz<Frank_wtal at web.de>  hat am 21. Dezember 2009 um
>>>>> 10:49 geschrieben:
>>>>>
>>>>>> An update:
>>>>>>
>>>>>> I found that the "ADD32: output is not int:" messages are
>>>>>> caused
>>> by the
>>>>>> enhancer. When I turn the enhancer off, the messages
>>>>>> vanish, but the
>>>>> "freak
>>>>>> out" of the codec is still there - so the problem seems not
>>>>>> to be
>>>>> related to
>>>>>> the overflow messages.
>>>>>>
>>>>>> best regards, Frank
>>>>>>
>>>>>>
>>>>>> ---------- Ursprüngliche Nachricht ---------- Von: Frank
>>>>>> Lorenz<Frank_wtal at web.de> An: speex-dev at xiph.org Datum: 18.
>>>>>> Dezember 2009 um 13:36 Betreff: [Speex-dev] Fixed Point on
>>>>>> wideband-mode: Single Frame loss
>>>>> on 2000
>>>>>> Hz sine causes "freak off"
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> I'm quite sure we found a bug inside the codec (1.2rc1,
>>>>>> wideband
>>>>> mode). It
>>>>>> only occurs when speex is compiled with the FIXED_POINT
>>>>>> flag.
>>>>>>
>>>>>> I feed in a 2000 Hz sine wave (-15 dBFS). When I simulate a
>>>>>> single
>>>>> frame loss
>>>>>> (i.e. pass a null pointer to speex_decode_int as second
>>>>>> parameter for
>>>>> just one
>>>>>> frame), the decoder starts to freak out a second later: It
>>>>>> produces horrible noise with 0
>>>>> dBFS.
>>>>>> This horrible noise will not vanish until you turn off the
>>>>>> input signal.
>>>>>>
>>>>>> I did some investigation on this with FIXED_DEBUG. The
>>>>>> error
>>> messages are
>>>>>> (lots of this): ADD32: output is not int: 1702286034 in
>>>>>> ltp.c: line 68 ADD32: output is not int: 1455757562 in
>>>>>> ltp.c: line 69
>>>>>>
>>>>>> Call stack for this (only set the break point once, may be
>>>>>> there are
>>>>> other
>>>>>> call stacks on error, too):
>>>>>>
>>>>>> sb_decode ( ) at sb_celp.c:898 nb_decode ( ) at
>>>>>> nb_celp.c:1471 multicomb ( ) at filter.c:709 interp_pitch (
>>>>>> ) at filter.c:603 inner_prod ( ) at ltp.c:68
>>>>>>
>>>>>> Can you help please?
>>>>>>
>>>>>> best regards, Frank
>>>>>> ______________________________________________________
>>>>>> GRATIS für alle WEB.DE-Nutzer: Die maxdome Movie-FLAT!
>>>>>> Jetzt freischalten unter http://movieflat.web.de
>>>>> <http://movieflat.web.de/>
>>>>>>
>>>>>> _______________________________________________ Speex-dev
>>>>>> mailing list Speex-dev at xiph.org
>>>>>> http://lists.xiph.org/mailman/listinfo/speex-dev
>>>>> ______________________________________________________ GRATIS
>>>>> für alle WEB.DE-Nutzer: Die maxdome Movie-FLAT! Jetzt
>>>>> freischalten unter http://movieflat.web.de
>>> <http://movieflat.web.de/>
>>>>>
>>>>> _______________________________________________ Speex-dev
>>>>> mailing list Speex-dev at xiph.org
>>>>> http://lists.xiph.org/mailman/listinfo/speex-dev
>>>>>
>>>>> Web Bug from https://img.web.de/p.gif WEB.DE MillionenKlick:
>>>>> Kostenlos tippen, täglich 1 Million gewinnen!
>>>>> http://produkte.web.de/go/07/
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________ Speex-dev
>>>>> mailing list Speex-dev at xiph.org
>>>>> http://lists.xiph.org/mailman/listinfo/speex-dev
>>>>
>>>
>>>
>>>
>>> Web Bug from https://img.web.de/p.gif GRATIS für alle
>>> WEB.DE-Nutzer: Die maxdome Movie-FLAT! Jetzt freischalten unter
>>> http://movieflat.web.de
>>>
>>>
>>>
>>> _______________________________________________ Speex-dev mailing
>>> list Speex-dev at xiph.org
>>> http://lists.xiph.org/mailman/listinfo/speex-dev
>>
>
> ___________________________________________________________
> Preisknaller: WEB.DE DSL Flatrate für nur 16,99 Euro/mtl.!
> http://produkte.web.de/go/02/
>
>
>