Conv-TasNet

‼️new‼️: The modified training and testing code is now able to separate speech properly.

‼️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.

Requirement

• Pytorch 1.3.0
• TorchAudio 0.3.1
• PyYAML 5.1.2

Accomplished goal

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

Preparation files before training

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

Training this model

• 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 this model

• 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

Results

• 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

N	L	B	H	Sc	P	X	R	Normalization	Causal	Receptive field	Model Size	SI-SNRi	SDRi
128	40	128	256	128	3	7	2	gLN	x	1.28	1.5M	13.0	13.3
256	40	128	256	128	3	7	2	gLN	x	1.28	1.5M	13.1	13.4
512	40	128	256	128	3	7	2	gLN	x	1.28	1.7M	13.3	13.6
512	40	128	256	256	3	7	2	gLN	x	1.28	2.4M	13.0	13.3
512	40	128	512	128	3	7	2	gLN	x	1.28	3.1M	13.3	13.6
512	40	128	512	512	3	7	2	gLN	x	1.28	6.2M	13.5	13.8
512	40	256	256	256	3	7	2	gLN	x	1.28	3.2M	13.0	13.3
512	40	256	512	256	3	7	2	gLN	x	1.28	6.0M	13.4	13.7
512	40	256	512	512	3	7	2	gLN	x	1.28	8.1M	13.2	13.5
512	40	128	512	128	3	6	4	gLN	x	1.27	5.1M	14.1	14.4
512	40	128	512	128	3	4	6	gLN	x	0.46	5.1M	13.9	14.2
512	40	128	512	128	3	8	3	gLN	x	3.83	5.1M	14.5	14.8
512	32	128	512	128	3	8	3	gLN	x	3.06	5.1M	14.7	15.0
512	16	128	512	128	3	8	3	gLN	x	1.53	5.1M	15.3	15.6
512	16	128	512	128	3	8	3	cLN	√	1.53	5.1M	10.6	11.0

Pre-Train Model

‼️new‼️: Huggingface Pretrain Google Driver

Our Results Image

Reference

• Luo Yi's Conv-Tasnet Code