Hervé BREDIN

Better: find a Python library that does this automatically :)

Any update on this? It is really difficult to modify the start/end time without keyboard shortcuts, using only the mouse.

Since it seems that we won't find a library that will work for any layouts, we will probably stick with options. What about replacing `--qwerty` by something like `--start-shortcut Q`...

AMI website provides an [official](http://groups.inf.ed.ac.uk/ami/corpus/datasets.shtml) train/dev/test split. For data preparation, I did most of the work in a `pyannote.audio` [tutorial](https://github.com/pyannote/pyannote-audio/tree/develop/tutorials/data_preparation) already.

I don't know. I am not the creator of the AMI corpus. Check the official AMI website.

Thanks for your feedback. The behavior of pyannote.database is indeed confusing. Actually, the first time `pyannote.database` is imported, it will look as the value of `PYANNOTE_DATABASE_CONFIG` and build its list...

I recommend you read [this](https://github.com/juanmc2005/StreamingSpeakerDiarization/blob/main/paper.pdf) paper, which is an online variant of what is currently implemented in `develop` branch.

Overlapping frames are only masked for computing the embeddings, not for the rest of the pipeline. Anyway, I plan to write a technical report describing the approach and will share...

Most likely a Windows / Python 3.10 / multiprocessing issue. Windows is not supported by pyannote.audio (though some people were succesful training pyannote models on Windows) Can you try again...

Possibly related: https://pytorch-lightning.readthedocs.io/en/latest/guides/speed.html#spawn