Long-term coherence: extending the spectrogram

[enn-nafnlaus]

I suspect you could get long-term coherence to tracks if you trained with and diffused the spectrograms containing a "thumbnail bar" of past spectrograms. That is to say, adding at the top or the bottom 32 individual 16-by-15 pixel images, each one being a thumbnail of the previous spectrograms. On diffusion, this "thumbnail bar" would be masked out and non-diffusable.

The diffusion net would thus have - at least at coarse scales - insight into what was played recently, nearly 3 minutes worth of history. The cost would be a bit over 3% of your spectral resolution.