I'm trying to duplicate the SVoice results here shared by Facebook:
https://github.com/facebookresearch/svoice/
They mentioned that their model was created using WSJ. They couldn't share the model due to legal concerns but the repo outlines how to replicate if you have access to this data. Would anyone know where I could find it?
This repository includes the scripts required to create the SMS-WSJ database, a spatial clustering baseline for separation, and a baseline ASR system using Kaldi.
In multi-speaker ASR the WSJ0-2MIX database and the spatialized version thereof are widely used. Observing that research in multi-speaker ASR is often hard to compare because some researchers pretrain on WSJ, while others train only on WSJ0-2MIX or create other sub-lists of WSJ we decided to use a fixed file list which is suitable for training an ASR system without additional audio data. Punctuation pronunciation utterances are filtered to further facilitate end-to-end ASR experiments.
