How to change parameters of segmentation model in diarization pipeline

[pmbaumgartner]

I'm really enjoying how easy it is to do diarization with the pipeline in pyannote, thanks so much for this awesome library. Currently I'm trying to do diarization on a meeting recording with 4 people, and the default pipeline is creating a lot of small segments where I know the speaker doesn't change. Currently I'm doing some post-processing steps on the output segments from the model, but I actually think it might be more productive to solve this problem using a parameter like min_duration_off in the segmentation model in the pipeline.

I couldn't find any examples on how to change these parameters in a pipeline. Is this something that's easily doable or do I essentially have to recreate the pipeline? Ideally I'm looking for something similar to the snippet below.

from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization")

# change things like `min_duration_off` for the segmentation part of the pipeline

diarization = pipeline(AUDIO_FILE, num_speakers=4)

[hbredin]

pipeline.segmentation.min_duration_off = FILL_GAPS_SHORTER_THAN_THAT_MANY_SECONDS should do the trick.

[pmbaumgartner]

I get AttributeError: 'str' object has no attribute 'min_duration_off' with that.

However, pipeline._segmentation (on the private attribute) doesn't error, but I think that's an instance of pyannote.audio.core.inference.Inference so it's just setting those attributes without them being used.

[hbredin]

Can you please let me know about your version of pyannote.audio?

[pmbaumgartner]

$  pip freeze | grep 'pyannote.audio'
pyannote.audio==2.0.1

[hbredin]

My bad. This won't work. You'd have to go with

diarization = pipeline("audio.wav")
updated_diarization = diarization.support(collar=FILL_GAPS_SHORTER_THAN_THAT_MANY_SECONDS)

See Annotation.support documentation .