A PyTorch implementation of the "USEV: Universal Speaker Extraction with Visual Cue"
/data: Scripts to pre-process voxceleb2 and IEMOCAP dataset.
/pretrained_networks: The pre-trained usev network on the voxceleb2 dataset.
/src: The pre-training and training scripts.
We provided the pre-trained weights from the Voxceleb2 dataset, you can just generate the iemocap-mix dataset and train from the pre-trained checkpoint; or you can generate the pre-trained voxceleb2-mix dataset and start from the pre-training stage.
For any questions, you can send an email to pan_zexu at u.nus.edu