Variant of Issue with pyannote starting 'Error in step 2 - identifiying the speakers.'

[sagefuentes]

I'm having an issue where setting speaker detection has been largely inconsistent across attempts. Most attempts I have gotten the error 'Error in step 2 - identifiying the speakers.' after trying to load pyannote.

I have not gotten Speaker Detection to work on 0.6.2 when set to auto or set to the number of speakers. This with pyannote_xpu set to cuda or cpu and whisper_xpu set to cuda or cpu.

I have gotten 0.6.1 with pyannote_xpu set to cuda to run when the number of speakers was set to 2, but I get the same error as above when set to auto. Additionally, the model still only identifies one speaker even when the speakers are distinct, though this may be an issue with the recording and general difficulty with speaker identification.
I have also gotten 0.6.1 with pyannote_xpu set to cpu to run when number of speakers was set to auto, but not consistently across attempts.

There have been no issues when speaker detection is set to none in either 0.6.1 or 0.6.2

Any help is greatly appreciated, what I can get running is helping a lot with a few research projects.

System:

• OS: Windows 11

• GPU: Nvidia GeForce GTX 1660 Super w/ 6GB VRAM

• CUDA Toolkit installed and up-to-date 13.0.2

• noScribe Version: both CUDA 0.6.2 and CUDA 0.6.1

  • unrelated issue: normal version would not launch for me (can post another issue about this, but I think I didn't clear the Local files before I installed normal, so need to test that as well, but 0.6.1 is working fine enough for some reason)

• noScribe Settings:

  • default start and stop, but portions of audio also tested
  • Language: both auto and english tested
  • Model: both precise and fast model tested
  • Mark Pause: 1sec+ mark
  • Speaker detection: None, Auto, and 2 tested
  • Overlapping Speech, Disfluencies, and Timestamps check on

• Other steps taken

  • Tried changing both pyannote_xpu and whisper_xpu to cpu, individually and at the same time, in config.yaml while noScribe was not running.
  • Ran in admin mode, just incase it was a write permission issue with the diarize.exe as when I attempted to run just the diarization via console, it had issues with write permissions.
  • Ensured Microsoft Visual Studio was up to date as per the CUDA Toolkit install
  • Ensured GPU drivers were up to date

[kaixxx]

You have tried a lot of good things already, so I don't really know what I can suggest to you. There are some problems with CUDA support on the RTX/GTX1xxx generation of cards. But with pyannote_xpu set to cpu it should definitely run. Some ideas:

• Please check the log files in C:\Users\USERNAME\AppData\Local\noScribe\noScribe\log. There might be a more detailed error report in there.
• Are you capable to run noScribe from source? I'm in the process of preparing a new version of noScribe that also contains a major update of pyannote (3.1 to 4.1): https://github.com/kaixxx/noScribe/tree/pyannote4 If you don't want to run it from source, wait for the release in a few days. I cannot promise that this will fix your problems, but there is a good chance.