I'm not sure - for me it's the random-colored blocks on the plot (because of uninitialized memory) that are the problem. As long as the plot behaves normally (that is, not randomly, like now), I'm ok.
Note that there are other items on the plot (besides the spectrogram) which should still work, that's why I picked a transparent image. Also, since NaN behaves like a transparent pixel, and if all my data is composed of NaNs (so the intervals are NaNs as well), I guess that's what I expected. If null image provides all that, I have nothing against it.
Thanks