Problem with spectral_bandwidth having only one frame

Hello folks,

I try to use the librosa library for some library I want to implement, which essentially extracts some features from audio files. These files can be very short (including only one frame when transforming it with say stft)

So I stumpled about some unexpected behaviour when using the spectral_bandwith feature.

Let S be some spectrogram (S[bins][frames]) then the following calls will produce (for me) unexpected output:

res_expected_1 = spectral_bandwidth(sr=sr, S=S) res_expected_2 = spectral_bandwidth(sr=sr, S=S[:,0:2]) res_unexpected = spectral_bandwidth(sr=sr, S=S[:,0:1])

Expected results return for each frame a number. So for res_expected_2 exactly 2 and for res_expected_1 exactly len(S[0]) while res_unexpected produces bin_num results. Which I don’t think is what it is intented to do.

in addition I don’t know why the results are wrapped in an array like [[frames]] instead of just [frames]

I hope someone can help me out a bit; Thanks for your time meanwhile

[Mereep]

@bmcfee I don’t know exactly what I did, but that fixed for me 😉

if freq.ndim == 1:
    squeezed_centroid = np.squeeze(centroid)
    if squeezed_centroid.ndim == 0:
        squeezed_centroid = [squeezed_centroid]
    deviation = np.abs(np.subtract.outer(freq, squeezed_centroid))

[bmcfee]

Cool, glad it’s working for you.

I think we’ll need a slightly more general solution to handle the instantaneous frequency case properly, but it’ll for sure make it into the next point release.