Cannot load models from the local directory

[Harshs-09]

I have tried to follow the various discussions threads on loading the pyannote/speaker-diarization-3.1 pipeline locally but I cannot still load the models successfully.

Python version: 3.10
pyannote.audio version:
torch version:

My directory structure looks like this:

root/
|-> models/
| |-> config.yaml
| |-> pytorch_model.bin
| |-> seg_pytorch_model.bin
|-> test.py

content of config.yaml:

version: 3.1.0

pipeline:
  name: pyannote.audio.pipelines.SpeakerDiarization
  params:
    clustering: AgglomerativeClustering
    embedding: models/pytorch_model.bin
    embedding_batch_size: 32
    embedding_exclude_overlap: true
    segmentation: models/seg_pytorch_model.bin
    segmentation_batch_size: 32

params:
  clustering:
    method: centroid
    min_cluster_size: 12
    threshold: 0.7045654963945799
  segmentation:
    min_duration_off: 0.0

from pyannote.audio import Pipeline

pipeline =  Pipeline.from_pretrained("models/config.yaml")
audio_file = "audio.wav"

segments = pipeline(audio_file)
for segment, _, speaker in segments.itertracks(yield_label=True):
        print(segment.start, segment.end, segment.speaker)

Errors:

    raise RuntimeError(
RuntimeError: A pipeline must be instantiated with `pipeline.instantiate(parameters)` before it can be applied.

[DanteTrb]

Hi @Harshs-09, great question, and I’ve seen this issue trip up many users when switching from HuggingFace-hosted models to local pipelines with pyannote.

Let me walk you through it..

Why you’re seeing pipeline.instantiate(parameters)

When you load a config.yaml via:

pipeline = Pipeline.from_pretrained("models/config.yaml")
you’re only parsing the config template, not the full pipeline object yet. That’s why pipeline(audio_file) throws:

RuntimeError: A pipeline must be instantiated with pipeline.instantiate(parameters)``

This happens because config.yaml doesn't contain the pre-frozen parameters — you still need to explicitly instantiate the pipeline.

Fix: Use .instantiate({}) after loading

Here’s how you should modify your code:

from pyannote.audio import Pipeline

pipeline = Pipeline.from_pretrained("models/config.yaml")

# 💡 This is the crucial step
pipeline = pipeline.instantiate({})  # or provide specific parameters if needed

# Then apply it
audio_file = "audio.wav"
segments = pipeline(audio_file)

for segment, _, speaker in segments.itertracks(yield_label=True):
    print(segment.start, segment.end, speaker)

This will correctly initialize and apply your local speaker diarization pipeline.

Bonus tips

Validate paths inside config.yaml:
Make sure paths like models/pytorch_model.bin are relative to the working directory, or use absolute paths.

If you have pretrained weights but not the YAML config, you can manually define the pipeline using:

from pyannote.audio.pipelines import SpeakerDiarization
pipeline = SpeakerDiarization(segmentation="models/seg.bin", embedding="models/emb.bin")
pipeline.instantiate({})

Let me know if that solves the issue, happy to help with further debugging.
And if it works, marking it as resolved might help future users navigating the same instantiate() pitfall 🙌

Good luck with your project! 👏🏽

[Harshs-09]

Hi,

Thank you for the help. It works properly now.