How to implement VAD model with GMM?

[kareemamrr]

The appropriate paper Speaker Diarization with LSTM writes that the Voice Activity Detection model is a GMM model using the same PLP features as the i-vector and two full covariance Gaussians. How can I implement this using scikit-learn's GMM class?

[wq2012]

We used a pretrained ASR model to generate force alignment for the data to get per-frame speech vs non-speech ground truth. Then you can simply fit a Gaussian to all speech frames, then fit another Gaussian to all non-speech frames.