Nothing. I changed the value of VadThreshold 4.4 , but gives the same performance. Still showing when words are not said (although detects fewer words than 4.3 VAD ) .
With the wav you can see well the failures, but most of the problems gives voice is live .
There are still particular words that are better, but they are rare.
Up to this point we do not know very well to do.
If you need the project as I can go via dropbox ( our project and the modified example of openears ).