Code for the paper: "MULTI-CMGAN+/+: LEVERAGING SPEECH QUALITY METRIC PREDICTION FOR SPEECH ENHANCEMENT TRAINED ON REFERENCE-FREE REAL-WORLD DATA" by George Close, William Ravenscroft, Thomas Hain, and Stefan Goetze
Uses data format and dataloading from CHiME-7 UDASE task. See that task for data preperation guide.
Set variables in __config__.py
to point to required training data
For the HuBERT representation, edit HuBERT_wrapper.py
to point to the file hubert_base_ls960.pt
Conda environment for training is chime.yaml
To train the framework
python3 train.py hparams/hyperparams_chime_bak_ovr_pesq_1.0.yaml
or use one of the other provided hyperparameter files.
Use the eval_cmgan.py
script to evaluate a trained model. See the command line arguments for this script for details.