iSDR(0.07dB), iPESQ(1.514), iSTOI(0.734), iDNSMOS(2.356), results of three stage methods for Libri2Vox test set. Values in ‘()’ represent baseline performance metrics of original mixture signals compared WITH the clean target. “Real only” means use of only real data with cosine similarity less than 0.5 in 1st stage (about 71% of all data).
| Model | Method | Stage 1 | Stage 2 | Stage 3 | iSDR | iPESQ | iSTOI | iDNSMOS |
|---|---|---|---|---|---|---|---|---|
| Conformer | w/o CL (Real only) | ✓ | 15.42 | 1.325 | 0.198 | 0.650 | ||
| w/o CL (Real+SynVox2) | ✓ | 15.46 | 1.295 | 0.196 | 0.640 | |||
| w/o CL (Real+SALT) | ✓ | 15.43 | 1.287 | 0.194 | 0.635 | |||
| w/1-stage CL (Real only) | ✓ | 15.39 | 1.401 | 0.203 | 0.664 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 15.89 | 1.424 | 0.205 | 0.664 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 16.01 | 1.420 | 0.205 | 0.664 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | 16.20 | 1.470 | 0.206 | 0.692 | |
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 16.20 | 1.466 | 0.207 | 0.683 | |
| BLSTM | w/o CL (Real only) | ✓ | 12.34 | 0.852 | 0.157 | 0.495 | ||
| w/o CL (Real+SynVox2) | ✓ | 12.53 | 0.903 | 0.164 | 0.501 | |||
| w/o CL (Real+SALT) | ✓ | 12.54 | 0.885 | 0.161 | 0.491 | |||
| w/1-stage CL (Real only) | ✓ | 11.73 | 0.794 | 0.149 | 0.473 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 12.50 | 0.922 | 0.162 | 0.520 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 12.65 | 0.932 | 0.165 | 0.531 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | 13.07 | 0.997 | 0.171 | 0.554 | |
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 13.06 | 0.986 | 0.170 | 0.544 | |
| SpeakerBeam | w/o CL (Real only) | ✓ | 11.62 | 0.804 | 0.150 | 0.489 | ||
| w/o CL (Real+SynVox2) | ✓ | 12.01 | 0.915 | 0.156 | 0.503 | |||
| w/o CL (Real+SALT) | ✓ | 12.07 | 0.887 | 0.153 | 0.501 | |||
| w/1-stage CL (Real only) | ✓ | 11.17 | 0.753 | 0.144 | 0.473 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 11.76 | 0.869 | 0.159 | 0.512 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 11.88 | 0.899 | 0.163 | 0.519 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | 12.35 | 0.974 | 0.171 | 0.547 | |
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 12.34 | 0.959 | 0.168 | 0.533 | |
| Voicefilter | w/o CL (Real only) | ✓ | 11.92 | 0.818 | 0.158 | 0.453 | ||
| w/o CL (Real+SynVox2) | ✓ | 11.41 | 0.747 | 0.149 | 0.402 | |||
| w/o CL (Real+SALT) | ✓ | 11.46 | 0.728 | 0.147 | 0.398 | |||
| w/1-stage CL (Real only) | ✓ | 11.34 | 0.709 | 0.149 | 0.414 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 12.10 | 0.847 | 0.161 | 0.463 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 12.15 | 0.856 | 0.164 | 0.468 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | 12.39 | 0.927 | 0.169 | 0.493 | |
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 12.26 | 0.908 | 0.166 | 0.476 |
| Model | Method | Stage 1 | Stage 2 | Stage 3 | iSDR | iPESQ | iSTOI | iDNSMOS |
|---|---|---|---|---|---|---|---|---|
| Conformer | w/o CL (Real only) | ✓ | 15.42 | 1.325 | 0.198 | 0.650 | ||
| w/o CL (Real+SynVox2) | ✓ | 15.46 | 1.295 | 0.196 | 0.640 | |||
| w/o CL (Real+SALT) | ✓ | 15.43 | 1.287 | 0.194 | 0.635 | |||
| w/1-stage CL (Real only) | ✓ | 15.39 | 1.401 | 0.203 | 0.664 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 15.89 | 1.424 | 0.205 | 0.664 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 16.01 | 1.420 | 0.205 | 0.664 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | 0.206 | ||||
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 1.466 | 0.683 | |||
| BLSTM | w/o CL (Real only) | ✓ | 12.34 | 0.852 | 0.157 | 0.495 | ||
| w/o CL (Real+SynVox2) | ✓ | 12.53 | 0.903 | 0.164 | 0.501 | |||
| w/o CL (Real+SALT) | ✓ | 12.54 | 0.885 | 0.161 | 0.491 | |||
| w/1-stage CL (Real only) | ✓ | 11.73 | 0.794 | 0.149 | 0.473 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 12.50 | 0.922 | 0.162 | 0.520 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 12.65 | 0.932 | 0.165 | 0.531 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | |||||
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 13.06 | 0.986 | 0.170 | 0.544 | |
| SpeakerBeam | w/o CL (Real only) | ✓ | 11.62 | 0.804 | 0.150 | 0.489 | ||
| w/o CL (Real+SynVox2) | ✓ | 12.01 | 0.915 | 0.156 | 0.503 | |||
| w/o CL (Real+SALT) | ✓ | 12.07 | 0.887 | 0.153 | 0.501 | |||
| w/1-stage CL (Real only) | ✓ | 11.17 | 0.753 | 0.144 | 0.473 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 11.76 | 0.869 | 0.159 | 0.512 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 11.88 | 0.899 | 0.163 | 0.519 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | |||||
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 12.34 | 0.959 | 0.168 | 0.533 | |
| Voicefilter | w/o CL (Real only) | ✓ | 11.92 | 0.818 | 0.158 | 0.453 | ||
| w/o CL (Real+SynVox2) | ✓ | 11.41 | 0.747 | 0.149 | 0.402 | |||
| w/o CL (Real+SALT) | ✓ | 11.46 | 0.728 | 0.147 | 0.398 | |||
| w/1-stage CL (Real only) | ✓ | 11.34 | 0.709 | 0.149 | 0.414 | |||
| w/2-stage CL (Real only) | ✓ | ✓ | 12.10 | 0.847 | 0.161 | 0.463 | ||
| w/3-stage CL (Real only) | ✓ | ✓ | ✓ | 12.15 | 0.856 | 0.164 | 0.468 | |
| w/3-stage CL (Real + SynVox2) | ✓ | ✓ | ✓ | |||||
| w/3-stage CL (Real + SALT) | ✓ | ✓ | ✓ | 12.26 | 0.908 | 0.166 | 0.476 |
Sharing content requires targeting cookies to be enabled. Please update your cookie preferences to use this feature.