Model.from_pretrained() fails with local path in vad_utils.py (commit 8e78880)

[Reuelster]

After commit 8e78880 (Merge #54, Dec 26, 2025), the vad_utils.py module passes a local filesystem path to pyannote.audio.Model.from_pretrained(), which expects a HuggingFace repo_id (format: "owner/repo"), not a file path.
This causes the following error when using transcribe_longform():

ValueError: Repo id must be in the form 'repo_name' or 'namespace/repo_name':
'/root/.cache/huggingface/hub/models--pyannote--segmentation-3.0/snapshots/e66f3d3b9eb0873085418a7b813d3b369bf160bb'

Environment

• GigaAM version: 0.1.0 (commit 8e78880)
• pyannote-audio version: 3.3.x / 3.4.x
• Python: 3.11
• torch: 2.5.1

In commit 8e78880, the function load_segmentation_model() was added:
vad_utils.py (broken - commit 8e78880)

  def resolve_local_segmentation_path(model_id: str) -> str:                                                                                                         
      return snapshot_download(repo_id=model_id, ...)  # Returns LOCAL PATH

  def load_segmentation_model(model_id: str) -> Model:
      local_path = resolve_local_segmentation_path(model_id=model_id)
      return Model.from_pretrained(local_path)  # FAILS: local path not accepted

def get_pipeline(device, model_id="pyannote/segmentation-3.0"):
      model = load_segmentation_model(model_id)  # Calls broken function

pyannote.audio.Model.from_pretrained() does not accept local filesystem paths — it expects a HuggingFace repo_id string.

The previous implementation (commit 0a3f103) works correctly.

[georgygospodinov]

It looks like the issue comes from a dependency mismatch rather than the repository itself. Your environment uses pyannote-audio 3.3.x / 3.4.x, but GigaAM pins specific versions in setup.py, and the code is expected to run with those exact dependencies:

"longform": ["torch==2.8.*", "torchaudio==2.8.*", "pyannote.audio==4.0", "torchcodec==0.7", "numba>=0.62"]

Could you please try reinstalling the project exactly as described in the README to ensure the correct extras and pinned versions are applied?

pip install -e .[longform]