Are the results wrong or just misleading? Warnings tend to not be read by users who should read them and just annoy users who read them, but know what they are doing.

I would prefer to either raise if the length of the data is less than NFFT or add a comment to the docstring.

Well with the default parameters and a small array the spectrogram is wrong since it shows a static distribution of power.

I don't know if there's ever a case when it's useful to have an NFFT that's bigger than your data, but an exception could work too. An alternative is have an 'auto' option by default that picks better values.

Thinking about this a bit more, the result is 'correct', in that there is only one window and it gives you back time point sample, just not terribly useful.

Looking at this problem, I agree with @tacaswell : the result is what the user asked the specgram function to produce.

Because when asking for
pyl.specgram(signal, Fs=fs, NFFT=64, noverlap=32, pad_to=1024) the result becomes what @roger- might have expected: an upward trending slope for frequency.

So the point is not that the spectrogram is wrong, rather that the default value for NFFT (which is 256) is not adapted to a case where you only have 200 points in the input signal, I would say.

Hope this clarifies the issue a bit.

Best, Florian

@flothesof: well I guess "wrong" in the sense that the default NFFT=256 is unable to produce a useful spectrogram if you have less than a certain number of points.

I think a simple check and a warning would be good in this case so it doesn't bite anyone with a misleading representation of their signal.

Looks like this is fixed and good to be closed!

Awesome!