problem with nchroma in stChromaFeatures

[jmw67]

Hi,
I got the following error :
Traceback (most recent call last):
File "/home/jmw/Bureau/bacasable/musiques-stream/testpydub.py", line 42, in
F = pafe.stFeatureExtraction(mysample[:2*w], Fs, w, pas)
File "/home/jmw/Bureau/bacasable/musiques-stream/pyAudioAnalysis-master/pyAudioAnalysis/audioFeatureExtraction.py", line 591, in stFeatureExtraction
chromaNames, chromaF = stChromaFeatures(X, Fs, nChroma, nFreqsPerChroma)
File "/home/jmw/Bureau/bacasable/musiques-stream/pyAudioAnalysis-master/pyAudioAnalysis/audioFeatureExtraction.py", line 283, in stChromaFeatures
print "nFreqsPerChroma[nChroma]",nFreqsPerChroma[nChroma]
IndexError: index 56 is out of bounds for axis 1 with size 55
if the windows is small, e.g. below the max(nChroma), in my case the Win size is 110 (10ms with Fs=11000), half window is 55 then C /= nFreqsPerChroma[nChroma] crashes.

This error happens only with a small window, e.g. Fs=11000 with 10ms window so I set Win=110 samples and half window =55 for FFT.

I put a few prints in the code and saw that he nChroma table contains 92 samples) so that the code :
C /= nFreqsPerChroma[nChroma]
in stChromaFeatures crashes.
This is because the nFreqPerChroma array is of size nChroma (=55) but nChroma contains values larger than 55.
I think you should have a look at this code.
best regards

[tyiannak]

Thanx alot for the info.
Indeed there seems to be a problem with such small short-term windows, due to the chroma vector computation. Strange I have never faced this problem.
I am looking into it and will let you know as soon as possible.
However, 10ms window is very small, especially for Fs = 11KHz, since the frequency resolution will be almost 50Hz, which means that you won't be able to discrinate between different chroma classes.

[tyiannak]

Hi again. It should work now, please update and retry.
Regards

[jmw67]

Hi,
I will check. For the 10ms size I know it is small it was a mistake I wanted to set to 100ms. And this error led to the bug.
Best regards

[jmw67]

Hi,
I am puzzled by your bug fix. Correct me if I am wrong. But based on my example with 10 ms / 110 sample per frame I get 55 frequencies with midi values up to 94 so larger than 55 the max size.
Spec=X2 and nChroma have the same length (frequency axis)
So I would write the following code :
C=spec #X2
C/=nFreqPerChroma #normalize according to nb of frequency bin duplicates
Il will obtain 5 octaves (55/12 ~4+1==> 60=5*12) and sum the 12 chroma.
Why adding the max test at the beginning of stChromaFeature?
Best regards

[jmw67]

OK, I missed one part.
I think you should work with a C array of 120 (10*12 midi sounds) and sum up the duplicate bin frequencies you have normalized.

[jmw67]

Would the code below be correct ?
def stChromaFeatures(X, fs, nChroma, nFreqsPerChroma):
#TODO: 1 complexity
#TODO: 2 bug with large windows
chromaNames = ['A', 'A#', 'B', 'C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#']
spec = X**2
cN=int(numpy.ceil((nChroma.max()+1) / 12.0) * 12)
spec/=nFreqsPerChroma
C=numpy.bincount(nChroma, spec, cN)
finalC = numpy.zeros((12, 1))
C2 = C.reshape(cN/12, 12)

finalC = numpy.matrix(numpy.sum(C2, axis=0)).T
finalC /= spec.sum()


return chromaNames, finalC

[kacperbazylinski]

At last i found the solution. Thanks! Faced this problem week ago. when was training my SER model with shorten samples, and then wanted to test real-life, longer examples.