librosa sfft with center=True ，value is different with tensorflow tf.signal.sfft

[Mddct]

librosa sfft with center=True ，value is different with tensorflow tf.signal.sfft。

librosa with center=true， it will pad the feame_length to n_fft with constant zero value 。It also pad signal with reflect mode。

The feature extraction shuold cover this difference。

[nglehuy]

@Mddct There will be differences due to different implementation between librosa and tf.signal. But changing center=False helps reduce it.

[Mddct]

In my experiment，center=true and frame_length = n_fft=2^m , the result is same。 Nguyễn Lê Huy <notifications@github.com> 于 2020年11月3日周二 下午10:34写道：

…

@Mddct <https://github.com/Mddct> There will be differences due to different implementation between librosa and tf.signal. But changing center=False helps reduce it. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#39 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABFN3Q47NV2ZPJOXULGHWBDSOAIJFANCNFSM4TIGJ45A> .

[nglehuy]

@Mddct what is m in your experiment?

[nglehuy]

@Mddct I don't know why but librosa splits frames using nfft instead of frame_length, this causes differences between librosa and tf.signal

[Mddct]

@Mddct I don't know why but librosa splits frames using nfft instead of frame_length, this causes differences between librosa and tf.signal

thx ， the torch.stft also splits frames using nff_t。

So I implemented a simple version of tensorflow, but it has not been tested in detail

def my_stft( input, center, frame_length, frame_step, n_fft):
  
  samples = input

  if center :
    tf_samples = tf.pad(samples, [[0,0], [n_fft//2, n_fft//2]], mode="reflect")
  
  ## window
  tf_window_fn = tf.signal.hann_window
  tf_window = tf_window_fn( window_length=frame_length,
                            periodic=True)
  # print(tf_window.shape) # (320, )
  if n_fft > frame_length:
    left_pad = (n_fft - frame_length) // 2 
    right_pad =  (n_fft - frame_length) - left_pad
    tf_window  = tf.pad(tf_window, [[left_pad, right_pad]])
 
  tf_window = tf.expand_dims(tf_window, 0)

  # frames
  tf_frames = tf.signal.frame( tf_samples, 
                  frame_length=n_fft,
                  frame_step = frame_step,
                  pad_end=False)


  tf_rfft = tf.signal.rfft(tf_frames*tf_window, 
                fft_length=[n_fft])

 
  return tf_rfft

[Mddct]

I found something new。The mel matrix in librosa and pytorch is also different from tensorflow。I will make a PR later。