Question: How to load models from file?

[Kavi-Gupta]

I have download these two models from HF:

1. https://huggingface.co/pyannote/segmentation-3.0/blob/main/pytorch_model.bin (segmentation model)
2. https://huggingface.co/pyannote/wespeaker-voxceleb-resnet34-LM/blob/main/pytorch_model.bin (embedding model)

How do I load these models from files?

Trying segmentation = models.SegmentationModel.from_pretrained("PyAnnoteDiarization/pyannote_model_segmentation-3.0.bin", use_hf_token=False) still tries to access HF. I think this should be possible since ONNX models can be loaded from a file.

Thanks.

[Kavi-Gupta]

I have tried the following:

from diart import SpeakerDiarization, SpeakerDiarizationConfig
from diart.sources import MicrophoneAudioSource, FileAudioSource
from diart.inference import StreamingInference
from diart.sinks import RTTMWriter
import diart.models as models
from pyannote.audio import Model


# The Mode.from_pretrained call does not fail, but the output graph is blank

segmentation = Model.from_pretrained("./models/segmentation-3.0.bin") # pyannote/segmentation-3.0
embedding = Model.from_pretrained("./models/embedding.bin") # pyannote/wespeaker-voxceleb-resnet34-LM


# This works
# segmentation = models.SegmentationModel.from_pretrained("pyannote/segmentation-3.0", use_hf_token="valid_token")
# embedding = models.EmbeddingModel.from_pretrained("pyannote/wespeaker-voxceleb-resnet34-LM", use_hf_token="valid_token")

diarization_config = SpeakerDiarizationConfig(
  segmentation,
  embedding
)

pipeline = SpeakerDiarization(diarization_config)
file = FileAudioSource("./audio.wav", sample_rate=16e3)
# mic = MicrophoneAudioSource
inference = StreamingInference(pipeline, file, do_plot=True)
inference.attach_observers(RTTMWriter(file.uri, "./output/audio.rttm"))
prediction = inference()

It seems like the models may be loaded properly but the output graph is still blank, downloading the same models online from hugging face with my token does work.

[Kavi-Gupta]

The following code works:

# Copyright (c) 2025 Kavi Gupta. All rights reserved.
# This software is provided under the MIT License.

from diart import SpeakerDiarization, SpeakerDiarizationConfig
from diart.sources import MicrophoneAudioSource, FileAudioSource
from diart.inference import StreamingInference
from diart.sinks import RTTMWriter
from diart.models import PyannoteLoader, SegmentationModel, EmbeddingModel
from pyannote.audio import Model




segmentation = SegmentationModel(PyannoteLoader("./models/segmentation-3.0.bin")) # pyannote/segmentation-3.0
embedding = EmbeddingModel(PyannoteLoader("./models/embedding.bin")) # pyannote/wespeaker-voxceleb-resnet34-LM

diarization_config = SpeakerDiarizationConfig(
  segmentation,
  embedding
)

pipeline = SpeakerDiarization(diarization_config)
file = FileAudioSource("./audio.wav", sample_rate=16e3)
# mic = MicrophoneAudioSource
inference = StreamingInference(pipeline, file, do_plot=True)
inference.attach_observers(RTTMWriter(file.uri, "./output/audio.rttm"))
prediction = inference()

@juanmc2005 I noticed in previous issues you mentioned that Diart will only work online. I believe this is a way to make Diart work offline.

[Kavi-Gupta]

@juanmc2005 I propose creating a new method similar to from_onnx, but from_bin instead. Let me know if you agree and I can open a PR.

[juanmc2005]

Hi @Kavi-Gupta, the SegmentationModel class has a from_pyannote() method that you can use to force load from a pyannote model checkpoint by passing the path. It's simply a convenient way of running SegmentationModel(PyannoteModel(path)).

Any checkpoint accepted by pyannote.audio's Model/from_pretrained() will be able to load correctly.