Yang, Haici, et al. "Source-Aware Neural Speech Coding for Noisy Speech Compression." ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021.
Python 3.6.8
torch 1.6.0
torchaudio 0.6.0
- Speech: TIMIT(https://www.ldc.upenn.edu)
- Noise: Duan stational noise (http://www2.ece.rochester.edu/~zduan/is2012/examples.html)
- Duan, Zhiyao, Gautham J. Mysore, and Paris Smaragdis. "Speech enhancement by online non-negative spectrogram decomposition in nonstationary noise environments." In Thirteenth Annual Conference of the International Speech Communication Association. 2012.
Main hyper-parameters and their default setting for model training:
| Symbol | Description |
|---|---|
| filters = 100 | Output channel size of encoder |
| d = 1 | Dimension of the codec |
| m = 32 | The number of codes in the code book |
| sr = True | To do super-resolution based downsampling or not |
| lr = 0.0001 | Learning rate |
| br = 8 | Bitrate(khz) |
| scale = 1000 | Scale to control the hardness of the softmax function. |
| label = time.strftime("%m%d_%H%M%S") | Model label |
| weight_mse = 30 | Loss weight for MSE(waveforms) term |
| weight_mel = 0.5 | Loss weight for mel-spectogram term |
| weight_qtz = 0.5 | Loss weight for quantization |
| weight_etp_total = 0.1 | Loss weight for the total entropy |
| weight_etp_ratio = 0.05 | Loss weight for the entropy ratio between source and noise |
| ratio = 1.0 | Ratio of assigned bitrate between source and noise |
| update_ratio = False | Whether update the ratio during training or not |
| db = 0 | Initial SDR of input data, 0 or 5 |
Train proposed model, python3 train_model.py.
Train baseline model, python3 train_base.py.