We have several audio tracks, all of which are different versions of the same track under one type of distortion: the speed of the track is increased by a constant factor K (hence frequences are multiplied by K, timesteps are divided by K), K is relatively small (0.8< K < 1.2).
I want to know the best way to get a good spectrogram out of these tracks. Before defining what I mean by 'good', let me state my final goal: I want to retrieve points, à la Shazam, based on the maximum amplitudes (ie amplitudes > a certain threshold). A good threshold is to be defined too, but I was thinking of taking the average local RMS of the signal and its standard deviation and make the treshold = mean(localRMS) + (constant factor)*std(localRMS). Then I'd like to apply computer vision algorithms to the resulting image to measure similarity.
A good spectrogram is then defined as such:
the frequencies are logarithmically spaced the points of interest
for one track are a translation along the frequency axis of the points of interest of another track.
Do you know what type of spectrogram would be the best? (I'm thinking of using either a logarithmically spaced fourier transform or a constant Q Fourier transform) Would my threshold work or are there better ones? Any remark on the whole design would be truly appreciated.