Updating Finetuning Segmentation Tutorial for Pyannote 3.4

[alunkingusw]

I want to test training the diarisation on some data to see the improvements. The Adapting Pretrained Pipeline tutorial (https://github.com/pyannote/pyannote-audio/blob/main/tutorials/adapting_pretrained_pipeline.ipynb) seemed just the ticket, but it's written for pyannote 2.1, so I have set about updating it.

I've got the basic outline of what I have changed below, and would appreciate if there are any glaring errors. Currently I'm not seeing an improvement, but if I know it works, I can type it up and do a PR to update the tutorial.

The first change (aside from installing a newer version of pyannote is the pipeline

pretrained_pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token=TRUE)

Then downloading the segmentation model has also changed

model = Model.from_pretrained("pyannote/segmentation-3.0", token=TRUE)

The biggest change is in the training setup, I believe you now set the task differently, like this..

from pyannote.audio.tasks import SpeakerDiarization

task = SpeakerDiarization(
    protocol=ihm_dataset,
    duration=10.0,  # recommended fixed chunk length
    max_speakers_per_chunk=3,
    max_speakers_per_frame=2,
    batch_size=32,
    num_workers=2,
)

model.task = task
model.prepare_data()
model.setup()

I then believe you no longer need the optimization of segmentation threshold, please correct me if I am wrong, but the optimization of clustering threshold can stay. Subsequently, we can load the finetuned model into the pipeline and test it..

# Set up pipeline with the finetuned figures
finetuned_pipeline = SpeakerDiarization(
    segmentation=finetuned_model,
    embedding=pretrained_pipeline.embedding,
    embedding_exclude_overlap=pretrained_pipeline.embedding_exclude_overlap,
    clustering=pretrained_pipeline.klustering,
)
finetuned_pipeline.instantiate({
    "segmentation": {

        "min_duration_off": 0.0
    },
    "clustering": {
        "method": "centroid",
        "threshold": best_clustering_threshold,
        "min_cluster_size": 1
    }
})

Does this look right to you all? Let me know if there are any major problems. I am going to try it again on the small AMI dataset this week, and see if I can get an improvement. if I do manage it, I will write up the tutorial, I promise :-)
If not, I will post here to try and figure out why.
Many thanks!

[alunkingusw]

Just an update, I ran the tests again with the mini dataset (much faster than the full dataset that I have been training on!)
The DER gets worse, from 25%ish up to 50%.. so I'm not doing something right..