-
Since the evaluation code for SNR contains typo 10 instead of 20, all of the paper's SNR values should be twice.
-
There is no difference that our model shows better results than other models on SNR.
-
page 4, Table 2
Before
| model | SingleSpeaker SNR | MultiSpeaker SNR |
|---|---|---|
| Linear x2 | 9.69 | 11.1 |
| U-Net x2 | 10.3 | 9.86 |
| MU-GAN x2 | 10.5 | 12.3 |
| NU-Wave x2 | 11.1 | 13.2 |
| Linear x3 | 8.04 | 8.71 |
| U-Net x3 | 8.81 | 10.7 |
| MU-GAN x3 | 9.44 | 11.7 |
| NU-Wave x3 | 9.62 | 12.0 |
After
| model | SingleSpeaker SNR | MultiSpeaker SNR |
|---|---|---|
| Linear x2 | 19.38 | 22.2 |
| U-Net x2 | 20.6 | 19.72 |
| MU-GAN x2 | 21.0 | 24.6 |
| NU-Wave x2 | 22.2 | 26.4 |
| Linear x3 | 16.08 | 17.42 |
| U-Net x3 | 17.62 | 21.4 |
| MU-GAN x3 | 18.88 | 23.4 |
| NU-Wave x3 | 19.24 | 24.0 |
- page 4, Section 5
Before
Our model improves SNR value by 0.18-0.9 dB from the
best performing baseline, MU-GAN.
After
Our model improves SNR value by 0.36-1.8 dB from the
best performing baseline, MU-GAN.