This repository automatically downloads the MSDWILD dataset and set it up to be used with pyannote-database.
It will generate two subsets from the original few.train set : custom1_train and custom1_dev, as the original dataset only has training and test data.
Defaults are 6h for custom1_dev, and what's left (~60h) for custom1_train.
Out-of-the-box protocol for pyannote.audio training is MSDWILD.SpeakerDiarization.CustomFew.
Clone this repository, download the dataset zip at https://github.com/X-LANCE/MSDWILD#wavs and put it under the msdwild folder.
Then, run setup.sh in the msdwild directory to download/extract/generate the files (wav, rttm, uem, uris).
| subset | # files | total length |
|---|---|---|
| few.train | 2476 | 65h54m |
| few.val | 490 | 9h49m |
| many.val | 177 | 4h04 |