How can I avoid interpolating across unvoiced regions when using kaldi pitch tracker

[Learn Micheal]

Hello everyone,

Now I am working with exploring tonal information for Lhasa dialect speech recognition. I have known that Kaldi pitch tracker (compute-kaldi-pitch-feats and process-kaldi-pitch-feats) treats all frames as voiced and interpolates across unvoiced regions. By adding the POV and Kaldi-pitch features, I have got relative 11% improvement than not using pitch-related features for DNN-HMM.

Now, I want to modify the pitch-related features. I want to make hard decisions about voicing on each frame and directly assign zero to those unvoiced frames. I have thought about Getf0 method, but there are two reasons to give up this method. One is that the Getf0 doesn't provide POV features. The other is that when I use Getf0 to extract pitch values, the number of frames is two more than Kaldi's and I haven't found out why it happens. 

To guarantee that other conditions are same, I want to use kaldi pitch tracker to do this. So I want to know if there is any way to shut down the kaldi pitch tracker's functionality of interpolating values for unvoiced frames? Any help will be highly appreciated. Thanks a lot.

Micheal

[Daniel Povey]

There isn't a way to stop it interpolating since that's in the core of
how it works-- it never makes a hard decision about voiced versus
unvoiced, the objective function chosen just naturally interpolates
across parts of the signal where it's not sure.
If you are just using this for features then I strongly recommend to
just use what we have, we did a lot of experiments and comparisons and
our method does work better than the commonly available ones.

Dan

> --
> You received this message because you are subscribed to the Google Groups
> "kaldi-help" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to kaldi-help+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Learn Micheal]

okay, thanks for your reply, Dan. That saves me a lot of time.

Micheal

在 2016年9月30日星期五 UTC+8上午2:44:25，Dan Povey写道：