‼️new‼️: Updated model code, added code for skip connection section.

‼️notice‼️: Training Batch size setting 8/16

‼️notice‼️: The implementation of another article optimizing Conv-TasNet has been open sourced in "Deep-Encoder-Decoder-Conv-TasNet".

Demo Pages: Results of pure speech separation model

Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation Pytorch's Implement

Luo Y, Mesgarani N. Conv-TasNet: Surpassing Ideal Time–Frequency Magnitude Masking for Speech Separation[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2019, 27(8): 1256-1266.

• Pytorch 1.3.0
• TorchAudio 0.3.1
• PyYAML 5.1.2

• Support Multi-GPU Training, you can see the train.yml
• Use the Dataloader Method That Comes With Pytorch
• Provide Pre-Training Models

1. Generate dataset using create-speaker-mixtures.zip with WSJ0 or TIMI
2. Generate scp file using script file of create_scp.py

• If you want to adjust the network parameters and the path of the training file, please modify the option/train/train.yml file.
• Training Command

  python train.py ./option/train/train.yml

• Inference Command (Use this command if you need to test a large number of audio files.)

  python Separation.py -mix_scp 1.scp -yaml ./config/train/train.yml -model best.pt -gpuid [0,1,2,3,4,5,6,7] -save_path ./checkpoint

• Inference Command (Use this command if you need to test a single audio files.)

  python Separation_wav.py -mix_wav 1.wav -yaml ./config/train/train.yml -model best.pt -gpuid [0,1,2,3,4,5,6,7] -save_path ./checkpoint

• Currently training, the results will be displayed when the training is over.
• The following table is the experimental results of different parameters in the paper

Google Driver

• Luo Yi's Conv-Tasnet Code