[Clam-devel] MFCC Loops

Xavier Amatriain xavier at amatriain.net
Mon Sep 7 15:03:55 PDT 2009


Good point ;-) If you don't use ports, the configured hop size has no 
effect whatsoever.

paul simmons wrote:
> And what is the use for hopsize then at that situation?
>
>
>
> On Tue, Sep 8, 2009 at 12:42 AM, Xavier Amatriain<xavier at amatriain.net> wrote:
>   
>> The audio is simply "trimmed" to WindowSize. Look at line 141:
>>
>> http://clam-project.org/doc/CLAM-doxygen/SpectralAnalysis_8cxx-source.html#l00134
>>
>> paul simmons wrote:
>>     
>>> A more clear question, sorry for two posts; what happens when I send
>>> an audio of size larger than SpectrumConfig.WindowSize() ?
>>>
>>>
>>>
>>> On Tue, Sep 8, 2009 at 12:27 AM, paul simmons<pauls631 at gmail.com> wrote:
>>>
>>>       
>>>> Hi,
>>>> I think the problem is about my SpectralAnalysis config. Just to
>>>> clarify; I send an audio as the code below, at each while() iteration,
>>>> it reads a chunk of audio (in size of Audio.SetSize()) and returns
>>>> just one spectrum, melspectrum and mfcc vector (ok, let's put
>>>> overlapping for a moment). But if you say SpectralAnalysis handles
>>>> everything (and I configure right), I think I should get an array of
>>>> spectrum (and others) vector according to the HopSize right? What is
>>>> my mistake here?
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> On Tue, Sep 8, 2009 at 12:01 AM, Xavier Amatriain<xavier at amatriain.net>
>>>> wrote:
>>>>
>>>>         
>>>>> I don't think there is a simple example of this. However, if I remember
>>>>> right, there are two ways to do the loop if you want to have overlapping
>>>>> windows or frames:
>>>>>
>>>>> 1) Read the whole audio from the audio file into a CLAM::Audio and
>>>>> manage
>>>>> the "chunk" you send to the SpectralAnalysis Do by advancing your read
>>>>> pointer by "hopsize" (which, if overlapping, will be less than
>>>>> windowSize)
>>>>>
>>>>> 2) Use ports and let them manage this for you. This is what you can see
>>>>> if
>>>>> you analyze the code in SMSAnalysisCore.cxx
>>>>>
>>>>>
>>>>> http://clam-project.org/doc/CLAM-doxygen/SMSAnalysisCore_8cxx-source.html
>>>>>
>>>>> If you don't need overlap, SpectralAnalysis will take care of everything
>>>>> (windowing, etc...) as long as you configure the parameters right so you
>>>>> don't need to worry about 1 or 2.
>>>>>
>>>>> X
>>>>>
>>>>> paul simmons wrote:
>>>>>
>>>>>           
>>>>>> well, let me change the question then. is there a "code" sample for
>>>>>> doing spectral analysis with overlapping windows (ie using windowsize
>>>>>> and hopsize parameters)?
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Mon, Sep 7, 2009 at 8:24 PM, Pau Arumí<pau.arumi at barcelonamedia.org>
>>>>>> wrote:
>>>>>>
>>>>>>
>>>>>>             
>>>>>>> Hi Paul,
>>>>>>> I didn't understand exactly where the problem is.
>>>>>>> (However I think the most qualified to answer you is David Garcia who
>>>>>>> is
>>>>>>> currently on holidays. He'll be back in a week or two.)
>>>>>>>
>>>>>>> Cheers
>>>>>>> P
>>>>>>>
>>>>>>> El dl 07 de 09 de 2009 a les 16:54 +0200, en/na paul simmons va
>>>>>>> escriure:
>>>>>>>
>>>>>>>
>>>>>>>               
>>>>>>>> Hi,
>>>>>>>> I am trying to do MFCC extraction. As far as I get; the steps are
>>>>>>>>
>>>>>>>> Read an "Audio"
>>>>>>>> Run "SpectralAnalysis" on output of "Audio"
>>>>>>>> Run "MelFilterBank" on output of "SpectralAnalysis"
>>>>>>>> Run "CepstralTransform" on output of "MelFilterBank"
>>>>>>>>
>>>>>>>> right?
>>>>>>>>
>>>>>>>> If so; I cannot understand where to do looping, for example if I do:
>>>>>>>>
>>>>>>>>   while(mReader.Do(audio))
>>>>>>>>   {
>>>>>>>>       spectralAnalysis.Do(audio, spectrum);
>>>>>>>>       melFilterBank.Do(spectrum, melSpectrum);
>>>>>>>>       cepstralTransform.Do(melSpectrum, melCepstrum);
>>>>>>>>
>>>>>>>>       CLAM::XMLStorage::Dump(audio,"Audio",a);
>>>>>>>>       CLAM::XMLStorage::Dump(spectrum, "Spectrum",s);
>>>>>>>>       CLAM::XMLStorage::Dump(melSpectrum, "MelSpectrum",m);
>>>>>>>>       CLAM::XMLStorage::Dump(melCepstrum, "MelCepstrum",n);
>>>>>>>>   }
>>>>>>>>
>>>>>>>> it applies mfcc to each read audio patch (mentioned by Audio.SetSize)
>>>>>>>> and got one MFCC feature vector. But I cannot understand how
>>>>>>>> WindowSize or HopSize is applied in this fashion. Because since Audio
>>>>>>>> is read in non-overlapping windows with mReader.Do(), overlapped
>>>>>>>> windows are run only for current audio part
>>>>>>>>
>>>>>>>> I think there must be another , more correct way to get overlapping
>>>>>>>> windows continuesly in whole audio, but cannot find it. Any ideas?
>>>>>>>>
>>>>>>>> Thanks...
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> Clam-devel mailing list
>>>>>>>> Clam-devel at llistes.projectes.lafarga.org
>>>>>>>>
>>>>>>>>
>>>>>>>> https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
>>>>>>>>
>>>>>>>>
>>>>>>>>                 
>>>>>>> _______________________________________________
>>>>>>> Clam-devel mailing list
>>>>>>> Clam-devel at llistes.projectes.lafarga.org
>>>>>>>
>>>>>>> https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>               
>>>>>> _______________________________________________
>>>>>> Clam-devel mailing list
>>>>>> Clam-devel at llistes.projectes.lafarga.org
>>>>>>
>>>>>> https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
>>>>>>
>>>>>>
>>>>>>             
>>>>> _______________________________________________
>>>>> Clam-devel mailing list
>>>>> Clam-devel at llistes.projectes.lafarga.org
>>>>>
>>>>> https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
>>>>>
>>>>>
>>>>>           
>>> _______________________________________________
>>> Clam-devel mailing list
>>> Clam-devel at llistes.projectes.lafarga.org
>>> https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
>>>
>>>       
>> _______________________________________________
>> Clam-devel mailing list
>> Clam-devel at llistes.projectes.lafarga.org
>> https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
>>
>>     
>
> _______________________________________________
> Clam-devel mailing list
> Clam-devel at llistes.projectes.lafarga.org
> https://llistes.projectes.lafarga.org/cgi-bin/mailman/listinfo/clam-devel
>   





More information about the clam-devel mailing list