Since pyannote/segmentation and pyannote/embedding seem to be aliases to the newest pyannote.audio releases, the instructions for reproducibility keep getting out of date and diart scripts don't allow you to choose a different checkpoint.
Currently the only option for users to choose a different model (even if it's only loading a different one from huggingface) is writing a custom script nearly identical to the existing ones.
Idea
Add arguments to diart scripts so that users can choose the pyannote models to load. For example, the segmentation model should be pyannote/segmentation@Interspeech2021 when reproducing the paper's results.
Problem
Since
pyannote/segmentation
andpyannote/embedding
seem to be aliases to the newest pyannote.audio releases, the instructions for reproducibility keep getting out of date and diart scripts don't allow you to choose a different checkpoint.Currently the only option for users to choose a different model (even if it's only loading a different one from huggingface) is writing a custom script nearly identical to the existing ones.
Idea
Add arguments to diart scripts so that users can choose the pyannote models to load. For example, the segmentation model should be
pyannote/segmentation@Interspeech2021
when reproducing the paper's results.Example