Speech Enhancement on DNS Challenge synthetic With reverb
2.71PESQConv-TasNet (*)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Conv-TasNet (*)# of params [million]=5.08, Time per frame [ms]=1, Time to infer a frame [ms]=351.9, Loss function=SNR, Reverberant augmentation=true2020.05 | 2.71 | 4.21 | 3.67 | 3.47 | |
| STFT-TCN# of params [million]=5.03, Time per frame [ms]=4, Time to infer a frame [ms]=353.8, Loss function=PCMSE, Reverberant augmentation=false2020.05 | 2.19 | 3.75 | 3.1 | 2.97 | |
| STFT-TCN# of params [million]=5.03, Time per frame [ms]=4, Loss function=PCMSE, Reverberant augmentation=true2020.05 | 2.11 | 3.77 | 3.02 | 2.92 | |
| Conv-TasNet# of params [million]=5.08, Time per frame [ms]=1, Time to infer a frame [ms]=355.5, Loss function=PASEMSE, Reverberant augmentation=false2020.05 | 1.85 | 3.25 | 2.78 | 2.53 | |
| Noisy (Raw)2020.05 | 1.82 | 3.5 | 2.8 | 2.64 | |
| Baseline (NSNet)# of params [million]=1.26, Time per frame [ms]=10, Loss function=[6]2020.05 | 1.52 | 2.71 | 2.51 | 2.09 | |
| STFT-TCN# of params [million]=5.03, Time per frame [ms]=4, Time to infer a frame [ms]=350.1, Loss function=PASEMSE, Reverberant augmentation=false2020.05 | 1.51 | 3.2 | 2.31 | 2.33 | |
| STFT-TCN# of params [million]=5.03, Time per frame [ms]=4, Time to infer a frame [ms]=348.1, Loss function=PASEMSE, Reverberant augmentation=true2020.05 | 1.46 | 2.89 | 2.45 | 2.15 |