how can i convert feature bank from kaldi to same output of python_speech_feature?

[Trương Trang]

hi all

I have sepearte module extract feature filterbank from kaldi c++ for another project, but when I try get output from filter bank kaldi, It's completely difference with filter bank from python_speech_feature (https://github.com/jameslyons/python_speech_features). The range value of kaldi is not in [-1, 1]. I wonder how i convert output kaldi like or near like python speech feature. Sorry because My model in another project using from python speech feature, but I can't convert direct from python_speech_feature to c++. Tks.

[Daniel Povey]

Check the documentation of that repo.  It mentions the filterbank energies are not in log space; kaldi's are. 

But there may be other differences.

On Wed, Jun 1, 2022 at 5:04 PM Trương Trang <truongtr...@gmail.com> wrote:

hi all

I have sepearte module extract feature filterbank from kaldi c++ for another project, but when I try get output from filter bank kaldi, It's completely difference with filter bank from python_speech_feature (https://github.com/jameslyons/python_speech_features). The range value of kaldi is not in [-1, 1]. I wonder how i convert output kaldi like or near like python speech feature. Sorry because My model in another project using from python speech feature, but I can't convert direct from python_speech_feature to c++. Tks.

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
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.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/ef2cd720-a72d-4835-998a-2587b4b9437bn%40googlegroups.com.

[Trương Trang]

I know we have a many difference between kaldi and that repo. But understand and change to matching 2 feature with me is hard. So, How can I convert feature output of kaldi to range -1, 1 make sense. Not just use normalization, can be change some parameter. My purpose is I want use feature filterbank kaldi ouput to train Wake word client. And my ww client is deep convolutional network.Can you give me some idea? Tks Dan.