Conv-TASNET:

Model with SDR = 16.7 (15.0 in the paper) on WJS0-2speaker dataset

You can find the oringnal paper TasNet: Surpassing Ideal Time-Frequency Masking for Speech Separation. The testing results are shown as follows:

Function

• dataset.py: read the data into the model
• Tasnet_model.py：the forword network
• Tasnet_train.py：the main function to run
• trainer.py：calculate the loss and for training and testing
• utils.py： process the raw audio and other useful functions
• train.yaml: all the parameters used in the model
• test.py：separate the mixed audio and calculate SDR
• loss/convTasnet_batch_12.file: the loss for each epoch during training stage
• loss/test_SDR.file: the SDR on the testing set (step by 10 samples)
• log/: the loss curves for tensorboard

Training stage：

• from the beginning: remove the line with "trainer.rerun" in Tasnet_train.py, use "trainer.run" instead
• from a trained model: remove the line with "trainer.run" in Tasnet_train.py，use "trainer.rerun" instead, and change the "model_path" in train.yaml/temp