v2.0 diarization processing time

[martinjbaker]

I'm seeing way longer processing times with v2 compared with v1.

Same GPU hardware and same source file takes 3x longer to process with v2. I can't deny that the accuracy is improved but is this something that is expected to change or is it just an inevitable tradeoff for better accuracy?

[hbredin]

Thanks for your interest in pyannote.audio.

Switching to the (much larger) speechbrain speaker embedding is the main reason of this slow down (and for half of the accuracy improvement).

I can think of many ways of making things faster but this is definitely not my priority right now, especially since I am still working on the speaker diarization pipeline internals, trying to improve accuracy even more.

[pengyuan101]

Thanks for your work! it really helped me a lot , but can you give me some suggestions for speeding up?

[hbredin]

I can think of many ways of making things faster but this is definitely not my priority right now, especially since I am still working on the speaker diarization pipeline internals, trying to improve accuracy even more.

Latest commit in develop branch brings a 35% speed up when applied to AMI test set on one GPU (from 40min down to 26min).

[RYucel]

Sorry for noob question. How can I use develop version if available? I still have a very slow process for diarization. Any improvement welcome

[CompRhys]

When we use the pre-trained models on the AMI data set are their issues arising from the fact that we don't first break up the long meetings ~35 minutes into shorter segments for the model to process? I can't find any detailed documentation on the pre-trained model to work out whether/how they break up longer sequences.

[RezaBi]

Hi, is it possible to run diarization model on pyannote v2 on cpus?? I test some files and never got to end of it!

[RezaBi]

@hbredin

[hbredin]

Yes.