This package contains the code used to perform the experiments conducted in the paper 'On the calibration of powerset speaker diarization models' published at Interspeech 2024.
This package builds upon the pyannote suite, and thus heavily depends on pyannote.audio
, pyannote.core
and pyannote.database
functionalities.
pip install powerset_calibration
git clone https://github.com/FrenchKrab/powerset_calibration
pip install -e powerset_calibration
Most of the functionalities of the library are easy to access and just require you to plug in the right parameters.
To learn how to use this library, please refer to the notebooks which should give you 90% of the informations you need.
- Essential features
- A1_model_inference: Generate and evaluate an 'inference file' from your segmentation model
- A2_active_learning_protocol: Create subsets from an existing protocol using active learning-like criterions (e.g. select the 10% least confident data)
- Advanced usage
- B1_subset_one_file: Manually do all the steps to find the regions of interest in one file (instead of relying on
ActiveLearningProtocol
).
- B1_subset_one_file: Manually do all the steps to find the regions of interest in one file (instead of relying on
If you want more detail about function/method arguments, please refer to the documentation: https://frenchkrab.github.io/powerset_calibration