ChromaFeature indexing issues

[kkawabat]

I believe there's an error in the implementation of the chromaFeatures. In stChromaFeaturesInit if values of freq is smaller than Cp=27.5 the nChroma values will have negative index due to the log(freq/Cp) term on line 246 this will happen when fs/nfft < Cp

The down stream effect of this is that in stChromaFeatures, nChroma is used as indices to initialize C C[nChroma] = spec, which, if the indices are negative, will incorrectly index C from the end of the array. Also if nChroma has multiple of the same index it would override each other.

Is the ChromaFeature in pyAudioAnalysis a similar implementation to http://www.ctralie.com/Teaching/ECE381_DataExpeditions_Lab1/? if so there is a chroma matrix generator in librosa that might be useful or this custom implementation of chroma matrix by myself might also be helpful

def _stChromaFeaturesInit(nfft, fs):
    """
    This function initializes the chroma matrices used in the calculation of the chroma features
    """
    chroma_matrix = np.zeros((12, nfft))
    freqs = np.array([((f + 1) * fs) / (2 * int(nfft)) for f in range(int(nfft))])
    base_freq = 880  # reference hz for the A note
    freqs_halfstep_approx = (np.log2(freqs / base_freq) * 12)
    note_indices = (np.round(freqs_halfstep_approx) % 12).astype(int)
    halfsteps_freq = base_freq*2**(np.round(freqs_halfstep_approx)/12)
    for i in range(12):
        chroma_matrix[i, note_indices == i] = 1
    return chroma_matrix, halfsteps_freq

if __name__ == '__main__':
    nfft = 1000
    fs = 16000
    chroma_matrix, _ = _stChromaFeaturesInit(nfft, fs)
    import matplotlib.pyplot as plt
    plt.imshow(chroma_matrix, aspect='auto')
    plt.yticks(range(12), ['A', 'A#', 'B', 'C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#'], fontsize=9)
    plt.ylabel('frequency bins 8hz per bin')
    plt.show()
    pass

this will produce the following chroma matrix

which can then be used with spec to generate chromaF (assuming the implementation is the same as the link above).

[kkawabat]

Is the chromaF calculated in pyAudioAnalysis suppose to be a vector of the sum of the frame's chromagram energy at each of the 12 notes, normalized by the energy of the frame?