Better error message when using wrong number of FFT samples

When I call a FFT-related function with incorrect parameters, I run into the same error message which I find hard to interpret: ValueError: range() arg 3 must not be zero

Reproducible example:

import librosa
import numpy as np
audio = np.random.random(10000)
fft = librosa.stft(audio.astype('float64'), n_fft=100000)

returns:

/usr/local/Caskroom/miniconda/base/envs/deep-music/lib/python3.6/site-packages/librosa/core/spectrum.py in stft(y, n_fft, hop_length, win_length, window, center, dtype, pad_mode)
    233                                           stft_matrix.itemsize))
    234 
--> 235     for bl_s in range(0, stft_matrix.shape[1], n_columns):
    236         bl_t = min(bl_s + n_columns, stft_matrix.shape[1])
    237 

ValueError: range() arg 3 must not be zero

I was also getting the same error message with librosa.cqt but I can’t remember what settings it was with. I was trying to maximise the frequency resolution (maximising n_bins and bins_per_octave) across a single time bin (hop_length=audio.shape[0]) so I imagine the cause of the error was the same as the one in my example above.

Could you please add an error message for this?

[lostanlen]

I vote for (1). I work with bird calls, which are very short sounds (3k-10k samples). If i were to accidentally set an n_fft above the duration of the sound while computing the STFT, i would prefer to be notified, update my code, and start again; rather than to have librosa perform some padding under the hood without telling me.

@JulianFerry cases where n_fft is exactly equal to the input length are best addressed by np.fft.fft https://numpy.org/doc/stable/reference/generated/numpy.fft.fft.html

[JulianFerry]

Having never contributed to open source it didn’t even occur to me I could PR this myself. I’m happy to give this a shot