Best way to train two-speaker embeddings

[picheny-nyu]

I have an application in which I know a priori there are only two speakers, but the conditions are low snr and not much data, maybe an hour or so of total speech. Any suggestions for what sort of configuration to use to train embeddings?

Thanks
Michael

[hbredin]

I would fine-tune this segmentation model.
Note that this relies on upcoming pyannote.audio 2.0 currently available in develop branch.