how to train the vad model in release

[kismeter]

In the release section, I noticed that there's a vad model, can I know how to training this model? or it's closed source

[kahrendt]

it's definitely not closed source, I just haven't had time to document it yet! The VAD model uses the same architecture as the v2 models with an initial convolution layer with a stride of 3 and then depthwise convolutions following it. I made it smaller for speed, so it has fewer layers and channels. It mainly uses the MUSAN dataset with additional voice samples from the Google Speech Commands and noises from the FDK50K set (after removing all noises with speech). They are trained using the current mWW workflow, but there are a few additional minor tweaks on the audio processing side that I haven't put together yet for release.

I have fallen way behind on getting mWW in a better state for training more wake words (or VAD models). It may be a couple of months before I can come back and release a newer version, but it is the next priority after working on Nabu Casa's upcoming voice hardware. I'm sorry for the delay!