A PyTorch implementation of Conv-TasNet described in "TasNet: Surpassing Ideal Time-Frequency Masking for Speech Separation".
| From | Activatoin | Norm | Causal | batch size | #GPU | SI-SDRi(dB) | SDRi(dB) |
|---|---|---|---|---|---|---|---|
| Paper | Softmax | cLN | Yes | - | - | 10.8 | 11.2 |
| Mine | ReLU | cLN | Yes | 6 | 1 Tesla V100 | 11.3 | 11.6 |
| Mine | ReLU | cLN | Yes | 24 | 4 Tesla V100 | 11.2 | 11.5 |
| Paper | Softmax | cLN | No | - | - | 14.0 | 14.4 |
| Mine | ReLU | cLN | No | 24 | 4 Tesla V100 | 15.1 | 15.4 |
| Paper | Softmax | gLN | No | - | - | 14.6 | 15.0 |
| Mine | Softmax | gLN | No | 8 | 2 GTX 1080TI | 5.0 | 5.2 |
| Mine | Sigmoid | gLN | No | 8 | 2 GTX 1080TI | 14.6 | 14.9 |
| Mine | ReLU | gLN | No | 8 | 2 GTX 1080TI | 15.6 | 15.8 |
| Mine | ReLU | gLN | No | 20 | 4 Tesla V100 | 16.0 | 16.3 |
Seems increasing the batch_size will get better performance for non-causal convolution. SI-SDR and SI-SNR are the same thing (different name) in different papers.
- PyTorch 0.4.1+
- Python3 (Recommend Anaconda)
pip install -r requirements.txt
scripts/run_tasnet.sh