The delta pitch feature seems to be strange

[cantabile-kwok]

Hi and thanks for the fascinating work! I am using Athena to extract MFCC and pitch without the use of Kaldi. It went smoothly, but when I inspect the value of delta pitch feature (the last dimension of the default 3-dim pitch feature), I got confused. Here is what I found:

1. Although the pitch-feature and the warped NCCF (POV feature) values seems a little bit different from those extracted by Kaldi, the trajectory is similar and the difference is not that huge. This is OK, but the delta-pitch feature is very different from Kaldi output, in default settings. I did not find any difference in Athena default pitch settings and Kaldi. The delta pitch extracted by Athena looks like a noise sequence.
2. Hence I lowered down the standard deviation of noise added to delta pitch even to 0. Then Athena outputs a sequence of almost all zeros. Therefore I believe there is something wrong with the source code.

Below is some results.

This plot is the 4-dim pitch extracted by Kaldi.

This plot is the 3-dim pitch extracted by Athena (in default settings).

After setting the delta_pitch_noise_stddev to 0, I get the result below.

[JianweiSun007]

Hi,
I tried extracting the pitch feature using the athena-transform example audio: 'examples/sm1_cln.wav'. there are some configurations as follows:
'window_length': 0.025,
'soft_min_f0': 10.0,
'delta_pitch_noise_stddev':0,
and the output looks like usual:
Transform: [[3.8811225e-02 3.0000305e-01 3.5762787e-07]
[6.7564729e-03 3.0000973e-01 3.5762787e-07]
[2.4553644e-02 3.0001450e-01 3.5762787e-07]
[2.4535857e-02 3.0002213e-01 3.5762787e-07]
[3.4553111e-02 3.0003071e-01 3.5762787e-07]
[4.2932931e-02 3.0004215e-01 3.5762787e-07]]

[cantabile-kwok]

@JianweiSun007 Thanks for experimenting this. I tried this as well, and the result is close to yours.

But there's still a problem in this result. As we can see, the last dimension of pitch output is very close to 0 (with 1e-7 order). As this wav obviously has pitch variations, the delta pitch feature should not be such a small value all the time. And this result is much different to Kaldi, in terms of delta pitch.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[stale]

This issue is closed. You can also re-open it if needed.