Hervé BREDIN

Thanks for the heads-up @JaesungHuh. Switching reference labels from 0.2 to 0.3 did ["improve"](https://huggingface.co/pyannote/speaker-diarization/commit/e313d0fafef636b012065d5e7711227d327908a5) my baseline by a whooping 2.8% (relative) in terms of speaker confusion rate. That is not...

Please read the FAQ. The FAQtory bot did a good job pointing to the answer...

I guess you installed from `develop` branch. You can safely ignore those warnings for now.

@Filimoa you should send the pipeline to GPU. It relies on CPU by default. ```python import torch diarization_pipeline.to(torch.device("cuda")) ```

Yes, you would need to optimize thresholds for each version of the embedding network. However, I did not keep track of the optimized thresholds, sorry.

Grid search should be fine. For benchmarking, I guess you’d have to use data similar to the expected test/production data.

I'd gladly have a look at a PR facilitating the offline use of pyannote. Would be nice to also update the related part of the documentation.

`pyannote` models and pipelines have always been usable offline. The documentation is just... missing. * download the [segmentation model](https://huggingface.co/pyannote/segmentation-3.0/blob/main/pytorch_model.bin) * download the [embedding model](https://huggingface.co/hbredin/wespeaker-voxceleb-resnet34-LM/blob/main/speaker-embedding.onnx) * copy the parameters of the...

Do not mistake [pipelines](https://huggingface.co/models?other=pyannote-audio-pipeline) and [models](https://huggingface.co/models?other=pyannote-audio-model). Models are neural networks that come with `pytorch_model.bin` files (e.g. [`pyannote/segmentation-3.0`](https://hf.co/pyannote/segmentation-3.0)). Pipelines are piece of code that may (or may not) rely on models...

Is this ready for review?