This is the unofficial implementation of MFNet, from paper''a Mask Free Neural Network for Monaural Speech Enhancement''
arxiv:https://arxiv.org/abs/2306.04286
I appreciate the guidance and assistance from the author. After the correction following our discussion:
1.The initial learning rate is 3e-4, correcting the value from 0.0034 in the paper.
2.The features input to the network are compressed spectra, i.e., input = sign(stdct) * sqrt(abs(stdct)).
3.DCT transformation without normalization.
I put the key code of STDCT, which may be useful for you.
The model used the code modified by the author in the issues of his repo, ioyy900205/MFNet#1
Experiments were conducted on the Voicebank+Demand, and the training parameters were set as follows:
- The warm-up strategy mentioned in the paper was not used
- Initial learning rate: 3e-4
- Batch size: 2
- Training for 100 epochs
- Cosine annealing with a period of 100 epochs
- Minimum value during cosine annealing: 1e-5
Performance of MFNet on the Voicebank+Demand test set:
| PESQ | STOI | CSIG | CBAK | COVL | SSNR | |
|---|---|---|---|---|---|---|
| Noisy | 1.97 | 92.1 | 3.35 | 2.44 | 2.63 | 1.73 |
| MFNet | 3.05 | 94.6 | 4.19 | 3.55 | 3.63 | 9.79 |
The training process is as follows:
