SSL-pretraining-separation
This is the official repository of Stabilizing Label Assignment for Speech Separation by Self-supervised Pre-training .
Corpus Preprocessing
WHAM! / WSJ0-mix
- Prepare your WSJ0 corpus and place under
./
- Run:
bash prepare_wham_data.sh
Libri2Mix
- Run:
bash prepare_librimix_data.sh --n_src 2
Train
Run scripts/*.sh
to reproduce experiments in the paper.
Models
- ConvTasNet
- DPRNNTasNet
- DPTNet
- SepFormerTasNet (my implementation of SepFormer)
- SepFormer2TasNet (my modification of SepFormer)
Note: our SepFormer does not include data augmentation and dynamic mixing, thus could not perform as well as the official results.
Reference
The codes were adapted from