pyannote-audio
pyannote-audio copied to clipboard
pyannote
Embeddings takes 3x the length of the audio length
Tested versions
- pyannote-audio 3.1.1
System information
windows 10 - pyannote.3.1.1 - rtx 3070
Issue description
Diarization taking much longer than it should, using the progress hook see that getting embeddings is the slow part. The audio file I am using is a minute long, but the embeddings step alone is taking 3:14. Also when running it is using the majority of my vram, like upwards of 6.5GB. I can't find anything online that shows what the actual size of speaker-diarization-3.1 is but I didn't think it was suppose to be that big.
Also when I upload the wav file to Google Colab and run the same exact code it runs almost instant. So can't think of what locally is causing the embeddings to run so slowly.
ProgressHook output when ran locally: segmentation 100% 0:00:00 speaker_counting 100% 0:00:00 embeddings 100% 0:03:14 discrete_diarization 100% 0:00:00
NVIDIA GeForce RTX 3070
Minimal reproduction example (MRE)
https://colab.research.google.com/github/pyannote/pyannote-audio/blob/develop/tutorials/MRE_template.ipynb
Update: I just uninstalled pyannote.audio 3.1.1 and installed 3.1.0 instead and it fixed it. Running the same exact code with 3.1.0 took 3.06 seconds, but with 3.1.1 it took 3 minutes and 14 seconds
That's strange.
Looking at the diff between 3.1.1 and 3.1.0, I see no reason why this would happen: https://github.com/pyannote/pyannote-audio/compare/3.1.0...3.1.1
Also found very long runtime for embeddings in 3.1.1 on both Ryzen and Apple M1 CPU. Around 27 and 20 minutes for a 22 minute audio file.
@hbredin yeah I tried to look at the diffs between the versions and couldn't find anything that would cause the issue. I almost didn't even try the older version because there didn't appear to be much code difference, yet it still resulted in huge change in processing time
tried it, did not change; worked fine with 3.0.0; we will try to revert back to that
After reading further, I think I am actually experiencing issue https://github.com/pyannote/pyannote-audio/issues/1621