The STFT results of single channel wav don't match with the results of librosa.stft

[Raise-me-up]

As mentioned, I set frame_size = 512, and shift_size = 256, and compare your results with that of the librosa.stft. However, I can't find the good match.

[Raise-me-up]

Can you provide a better test.c file please?

[kooBH]

Since I usually work on STFT to match with my MATLAB algorithms, this code fits with MATLAB routines.
I don't have better test code for librosa.stft .

But as I remember it matched with librosa.stft as well, with modification on args.
There are a few options for librosa.stft. Might be delay problem from padding options.

[Raise-me-up]

@kooBH
Could you please provide your m file so I can check my codes?

[Raise-me-up]

Or can you provide the args modification of librosa.stft so I can check the data please?

[Raise-me-up]

Hey @kooBH

I find there are lots of padding mode in the librosa.stft, such as constant, edge, linear_ramp, reflect, warp etc. Which one do you use? Does the 'Copy and Shift' part mean padding?

[kooBH]

manually shifted index due to differences in padding.

https://github.com/kooBH/STFT/blob/main/valdation/validation.ipynb

yet, there is 1e-7 error.

[Raise-me-up]

Hey @kooBH

I have seen your validation codes, where the center option of librosa is false. How can I add the padding codes to match the results of librosa when the center option is true?

[Raise-me-up]

Well, when I set shift_size = 256, the error become a little big. Why will you set cpp_stft = cpp_stft[:,7:-1] under the shift_size = 128?

[kooBH]

cpp_stft is for real-time processing
where samples for STFT are like

 0  0  0  0
 0  0  0  s1
 0  0  s1 s2
 0  s1 s2 s3
 s1 s2 s3 s4
 s2 s3 s4 s5
...

buf most of STFT routines are for wav reading processing where samples for STFT are like

s1 s2 s3 s4
s2 s3 s4 s5
...

for frame_size = 512, shift_size = 128, which is 25% overlap

when you set padding options on librosa.stft there will be zero samples in initial sequences.

and cpp_stft = cpp_stft[:,7:-1] is just skipping initial sequences where samples are different.

[Raise-me-up]

wo~ Thank you for your reply! That's a little sophisticated.

[Raise-me-up]

Sadly, I can't make the data match under the setting of

librosa.stft(audio, n_fft=512, hop_length=256, win_length=512, window='hann', center=True, pad_mode="reflect")

[kooBH]

https://github.com/kooBH/STFT/blob/main/valdation/validation.ipynb

  const int ch = 1;
  const int rate = 16000;
  const int frame = 512;
  const int shift = 256;  

  STFT process(ch,frame,shift);

rs_stft = rs.stft(input_wav, n_fft=512, hop_length=256, win_length=512, window='hann', center=True, pad_mode="reflect")

[Raise-me-up]

Hi @kooBH

Thank you very much for your patient reply!

I find the difference between your validation and mine. You can always find the right slice of cpp_stft, while I can't.

BTW, I find the part below can be removed to get the smaller error rate:

for (i = 0; i < frame_size; i++)
      tmp += hann[i] * hann[i];
    tmp /= shift_size;
    tmp = std::sqrt(tmp);

for (i = 0; i < frame_size; i++)
  hann[i] /= tmp;