From a bird's-eye perspective, do CQT, FFT, and NNLS algorithms provide different "perspectives" on the original signal? Does each preference some features of the original signal?
I'm using Sonic Visualizer ( https://www.sonicvisualiser.org/index.html ) with VAMP plugins to perform MIR (Music Information Research) on audio. You can find descriptions of the VAMP plugins here : https://www.vamp-plugins.org/download.html , and NNLS for reference, since that seems less common that CQT or FFT : http://www.isophonics.net/nnls-chroma.
There are a number of plugins that use different algorithms to yield similar output: for each time slice/frame/window, how much energy is present at each frequency, quantized into chromatic buckets (12 steps per octave).
So far, by trial and error outputting results into a spreadsheet, it seems that all three approaches yield similar but not identical results.
Ultimately, I'm going to want to continue processing the output to try to determine which pitches are present and in which octaves, and I'm trying to understand which algorithm will yield the best starting point for that.
So I ask, from a birds-eye view, do these algorithms provide different perspectives on the signal? Would musical genre or complexity (e.g. pop vs classical, homophonic vs polyphonic) suggest one vs another?