the difference between kaldifeat-kaldi and kaldifeat-python?

[mezhou]

May I ask you the numerical difference between kaldifeat by kaldi implementation and kaldifeat by your python implementation?
I have compared the two computed features, and I find it has some difference.
I wonder that the experiment results showed in D-TDNN master and D-TDNN-kaldifeat branch is absolutely the same.

Thanks~

[yuyq96]

Yes, the generated features can be different, but it merely affects the results. There are two reasons for the difference:

• Floating point precision, which can be ignored.
• Dither: when you turn it on, we add some randomness to the features, which changes the results. Fortunately, the differences are limited.

However, if you did find some implementation errors in kaldifeat-python, please let me know.

[mezhou]

I have turn off the dither and unified the configuration, maybe it because the floating point precision you mentioned.

BTW, is the performance results in your paper/github based on the feature used kaldi official implementation or your python implemention on the fly?

Anyhow, your kaldifeat-python implementation is really a nice work! And thank for your fast reply.

[yuyq96]

The performance in our paper is based on Kaldi official implementation to make sure it is exactly the same with previous papers.

[mezhou]

OK, I got it.
Thanks again~