These scripts automatically download the AISHELL-4 dataset and set it up to be used with pyannote-database.
It will generate two subsets from the original train set : custom_train and custom_dev, as the original dataset only has training and test data.
Defaults are 12h for custom_dev, and what's left (~92h) for custom_train.
Out-of-the-box protocol for pyannote.audio training is AISHELL.SpeakerDiarization.Custom.
Run setup.sh to download and extract the files.
| subset | # files | total length |
|---|---|---|
| train | 191 | 104h46m |
| test | 20 | 12h34m |
- AISHELL-4 (CC BY-SA 4.0) :
- Dataset: https://www.openslr.org/111/
- Original website : http://www.aishelltech.com/aishell_4