[Bug] `Pipeline.from_pretrained()` got an unexpected keyword argument `token`

[Tiritibambix]

Disclaimer: I got out of my depth pretty quickly with this one, and I'm not a coder at all. I used AI assistance to help me isolate the issue, so please take my analysis with a grain of salt. Happy to be corrected!

I ran into an issue with speaker diarization. Transcription works fine, but as soon as diarization kicks in, the job fails with this error:

TypeError: Pipeline.from_pretrained() got an unexpected keyword argument 'token'

Full traceback from the logs:

ERROR Job failed | ts=2026-05-14T11:07:56.642051+00:00 logger=amicoscript.worker job_id=79dbceb0-ee46-44e8-b5d5-131921ad781e traceback=Traceback (most recent call last):
  File "/app/backend/core/transcription.py", line 406, in _process_job
    speakers = _run_diarization_phase(job_id, segments_list, job)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/core/diarization.py", line 60, in _run_diarization_phase
    pipeline = _Pipeline.from_pretrained(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: Pipeline.from_pretrained() got an unexpected keyword argument 'token'
ERROR error: Pipeline.from_pretrained() got an unexpected keyword argument 'token' | ts=2026-05-14T11:07:56.642472+00:00 logger=amicoscript.worker job_id=79dbceb0-ee46-44e8-b5d5-131921ad781e 

My setup

• Self-hosted via Docker (cloned from main, built locally)
• HF token set via the UI (Settings)
• Token has read/write permissions
• Licenses accepted on both:
  • pyannote/speaker-diarization-3.1
  • pyannote/segmentation-3.0

What I found

The code in backend/core/diarization.py calls:

pipeline = _Pipeline.from_pretrained(
    "pyannote/speaker-diarization-3.1",
    token=opts["hf_token"],
)

But when I checked the actual signature of Pipeline.from_pretrained in the installed version of pyannote.audio (3.4.0), I got this:

(checkpoint_path, hparams_file=None, use_auth_token=None, cache_dir=...)

So pyannote 3.4.0 expects use_auth_token= and does not accept token= at all, which explains the error.

The fix would be to replace token= with use_auth_token= in diarization.py. Again, I may be wrong, but that's what the AI and I came up with!

Thanks for the great tool regardless.

[Tiritibambix]

Hey, just a quick follow-up on my previous issue.

After digging deeper, the real problem is a dependency conflict: pyannote.audio 3.4.0 calls hf_hub_download() internally with use_auth_token=, but huggingface_hub 1.14.0 (which your requirements.txt pulls in) has dropped that argument entirely. The two deps are incompatible with each other.

No workaround seems to stick. Swapping argument names just moves the error around. Needs a fix on the dependency side.

[sim186]

Thanks a lot for the precise reporting @Tiritibambix . Will take soon a look. In general how felt so far the experience with the software? Not considering of course the diarization issue?

[Tiritibambix]

Hey, glad you're looking into it!

For what it's worth, I managed to get diarization working by manually forcing these versions and changing one line:

• pyannote.audio == 3.4.0
• huggingface_hub == 0.24.7
• torch == 2.5.1
• torchaudio == 2.5.1
• diarization.py line 62: use_auth_token= instead of token=

As for the overall experience, honestly, I love it. I take remote classes, I record the videos, and being able to search directly through the transcripts saves me an incredible amount of time. It's exactly the kind of tool I was looking for. Thanks so much for building it, and I hope sorting out the dependency mess won't be too painful!

[sim186]

nice to hear 😄 . If you pull the latest main hopefully should get the fix.

[Tiritibambix]

Hey, diarization is still broken out of the box. After a clean clone + docker-compose build --no-cache, I get a chain of errors that only go away by manually forcing specific versions inside the running container.

The errors come in sequence:

1. hf_hub_download() got an unexpected keyword argument 'use_auth_token' — huggingface_hub latest drops that argument, so pyannote fails internally
2. Weights only load failed / TorchVersion was not an allowed global — PyTorch 2.6 changed the default of torch.load(weights_only=...) and the pyannote checkpoints aren't compatible with it

The combination that actually works for me:

• huggingface_hub==0.24.7
• torch==2.5.1
• torchaudio==2.5.1

These need to be pinned in requirements.txt (or capped with < constraints) so the build pulls compatible versions. Right now the unpinned build pulls versions that break pyannote internally regardless of what the app code does.

Hope that helps!

[sim186]

@Tiritibambix I think the issue should be fixed by now. Tested both in docker and venv.

I also took the chance to clean-up a bit the interface I hope is a little clearer now.

[Tiritibambix]

Hey, thanks for the fix, tested and working!

One small thing though: I notice you're now using speaker-diarization-community-1 but the "Accept license" link in the UI still points to pyannote/speaker-diarization-3.1. It should be updated to https://huggingface.co/pyannote/speaker-diarization-community-1

Thanks for the great work!