NeMo

NeMo

Two routes:

1. Automatic Speech Recognition i.e Whisper
2. VAD i.e MarbleNet $\to$ Diarizer i.e Speechbrain

Note: 2nd route is the familiar one where we have a VAD i.e. Silero or MarbleNet(NeMo) and a diarizer i.e Speechbrain, Pyannote.Audio or MSDD(NeMo)

Conclusion from NeMo Diarization

The segmentation model of NeMo is prone to cutting speakers off. Can be improved by using the neural diarizer.

Improved model idea: Take the segmentation approach of pyannote and combine it with the speaker detection of NeMo.