stft changes time resolution when center=False and n_fft changes

I am having a problem performing two stft for a signal with the same hop and window lengths but with a different number of fft points. The problem occurs with the frames_to_time function too.

To Reproduce

import librosa
import numpy as np

# generates a signal with sr=100 and f0=10
sr=100
ts = np.arange(0, 1, 1/sr)
x = np.sin(2*np.pi*10*ts)

# computes the stft with n_fft=win_length=hop_length
n_fft = win_length = hop = 10
s1 = lb.stft(x, n_fft=n_fft, win_length=win_length, hop_length=hop,  center=False)
f1 = lb.fft_frequencies(sr, n_fft)
t1 = lb.frames_to_time(range(s1.shape[1]), sr=sr, hop_length=hop)
print(s1.shape, f1.shape, t1.shape)
#computes the stft with more points and fft but not changing win_length and hop_length
n_fft = 20
s2 = lb.stft(x, n_fft=n_fft, win_length=win_length, hop_length=hop,  center=False)
f2 = lb.fft_frequencies(sr, n_fft)
t2 = lb.frames_to_time(range(s2.shape[1]), sr=sr, hop_length=hop, n_fft=n_fft)
print(s2.shape, f2.shape, t2.shape)

I expected de output to be:

(6, 10) (6,) (10,)
(11, 10) (11,) (10,)

Since I’m still dividing the 100 samples into 10 windows of length 10, but instead I got:

(6, 10) (6,) (10,)
(11, 9) (11,) (9,)

if I inspect the content of t1, t2, f1 and f2:

f1 = [ 0. 10. 20. 30. 40. 50.]
f2 = [ 0.  5. 10. 15. 20. 25. 30. 35. 40. 45. 50.]
t1 = [0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95]
t2 = [0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9]

I can see that the frequency resolution has increased, but the time resolution has decreased instead of remaining the same.

I’ve tested in different environments but here’s where I’ve tested last

Linux-4.19.112+-x86_64-with-Ubuntu-18.04-bionic
Python 3.6.9 (default, Oct  8 2020, 12:12:24) 
[GCC 8.4.0]
NumPy 1.19.5
SciPy 1.4.1
librosa 0.8.0

I know that are other issues related to this, but they don’t include the frames_to_time part so I posted again.

[Feulo]

I see, so for the example where the signal y has length 40, n_fft = 20, hop_length = win_length = 10, and center=False the frames will look like this:

I have 3 frames starting at samples 0, 10 and 20. The frame that would start at sample 30 is not created because the signal ends at sample 39 and there is no padding at the end of the signal

Because win_length < n_fft the (n_fft - win_length)/2 = 5 samples in the beginning of the signal won’t be used, the same for the last (n_fft - win_length)/2 = 5`.

Finally, the n_fft parameter sets the frame size to be used and it is the number of points used in the DFT too, therefore I can’t have a number of points in the DFT greater than the frame size (using the stft function)

Thank you one more time for the clarifications, I will you the stft code as base for a custom function where I can change the number of points in the DFT without changing the frame size.

[bmcfee]

Seems like this one is resolved, so I’m closing it out. Feel free to reopen if it warrants further discussion.