Source separation task using SpEx+ model trained on mixes based on Librispeech speakers. The task was to separate two speakers with a reference audio of a target speaker.
pip install -r ./requirements.txtBest model can be found here. All parameters of the model are set to the default values in a SpexPlus class. For the test you might need to set only num_speakers=251.
Configs can be found in hw_ss/configs folder. In particular, for testing use config_test.json.
Trained on this dataset.
python train.py -c CONFIGValidated on this dataset.
python test.py -c CONFIG -r CHECKPOINT -t TEST_DIRECTORY| SI-SDR | PESQ | |
|---|---|---|
| Val set | 11.23 | 2.14 |