TASNET: Time-domain Audio Separation Network

Two models are provided here, one with SDR = 11.7 and the other with SDR = 13.6

Model with SDR = 11.7

You can find the oringnal paper here. 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

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

Model with SDR = 13.6

You can find the oringnal paper here.