nara_wpe icon indicating copy to clipboard operation
nara_wpe copied to clipboard

Are there any differences between these versions of WPE?

Open qinxiaoyi opened this issue 6 years ago • 1 comments

I notice there are three versions of example,tensorflow,numpy and NTT? Are there any differences between these versions of WPE? NTT is new version that has never been before?Can you help me to solve the difference of these three versions. Thank you!

qinxiaoyi avatar Jan 08 '19 03:01 qinxiaoyi

Dear @qinxiaoyi, we implemented the code in Numpy and Tensorflow to address the different needs of users. The results should be equal as there are quite a few test cases comparing the two.

Numpy Use the Numpy code if you just want to use Nara-WPE as a front-end. If you want to make changes or invent an improved-WPE it is easiest to use the Numpy code and extent on that.

Tensorflow If you already use Tensorflow, it may be easier to integrate the Tensorflow WPE implementation directly into your computational graph. This can be helpful if you want to backpropagate gradients through WPE.

NTT This is just meant to be a wrapper around a closed source implementation in Matlab. In the future, this can be used to compare our implementation against it.

At least in Numpy and Tensorflow we have an offline-WPE and a frame-online-WPE implementation. Obviously the offline-WPE has the better performance. If your application require online processing, you need to use the frame-online- or block-online-WPE.

LukasDrude avatar Jan 08 '19 08:01 LukasDrude