Audio-Visual Target Speaker Extraction on LRS2 2-mix (test)
3.16DNSMOSLibri2Mix GT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Libri2Mix GT2026.03 | 3.16 | — | 3.93 | |
| TF-GridNetBackbone=TF-GridNet, # params=14.4M2026.03 | 2.8 | 14.81 | 4.32 | |
| TF-GridNet + LSM (Ours)Backbone=TF-GridNet, AO backbone frozen=true, # params=2.0M2026.03 | 2.79 | 14.79 | 4.29 | |
| MossFormer2Backbone=MossFormer2, # params=55.7M2026.03 | 2.79 | 12.68 | 3.47 | |
| MossFormer2 + LSM (Ours)Backbone=MossFormer2, AO backbone frozen=true, # params=1.6M2026.03 | 2.79 | 12.65 | 3.47 | |
| AV-MossFormer2# params=57.3M2026.03 | 2.64 | 15.52 | 3.06 | |
| TF-GridNet + ResidualBackbone=TF-GridNet, AO backbone frozen=true, # params=2.6M2026.03 | 2.56 | 13.09 | 3.2 | |
| MossFormer2 + ResidualBackbone=MossFormer2, AO backbone frozen=false, # params=57.3M2026.03 | 2.51 | 15.54 | 3.02 | |
| AV-TFGridNet# params=9.6M2026.03 | 2.51 | 15.1 | 3.53 | |
| MossFormer2 + ResidualBackbone=MossFormer2, AO backbone frozen=true, # params=1.6M2026.03 | 2.44 | 14.02 | 2.95 | |
| AV-ConvTasNet# params=10.3M2026.03 | 2.44 | 11.51 | 2.49 | |
| AV-DPRNN# params=4.1M2026.03 | 2.4 | 11.97 | 2.58 | |
| LRS2-2mix GT2026.03 | 2.38 | — | 3.19 | |
| TF-GridNet + ResidualBackbone=TF-GridNet, AO backbone frozen=false, # params=17.0M2026.03 | 2.36 | 13.38 | 3.17 | |
| Conv-TasNetBackbone=Conv-TasNet, # params=5.1M2026.03 | 2.35 | 7.12 | 2.31 | |
| DPRNNBackbone=DPRNN, # params=2.6M2026.03 | 2.35 | 7.77 | 2.37 | |
| DPRNN + LSM (Ours)Backbone=DPRNN, AO backbone frozen=true, # params=2.0M2026.03 | 2.35 | 7.74 | 2.36 | |
| Conv-TasNet + ResidualBackbone=Conv-TasNet, AO backbone frozen=false, # params=6.6M2026.03 | 2.31 | 11.72 | 2.57 | |
| Conv-TasNet + LSM (Ours)Backbone=Conv-TasNet, AO backbone frozen=true, # params=1.5M2026.03 | 2.28 | 6.82 | 2.22 | |
| DPRNN + ResidualBackbone=DPRNN, AO backbone frozen=false, # params=5.2M2026.03 | 2.28 | 11.66 | 2.43 | |
| Conv-TasNet + ResidualBackbone=Conv-TasNet, AO backbone frozen=true, # params=1.5M2026.03 | 2.13 | 9.37 | 2.27 | |
| DPRNN + ResidualBackbone=DPRNN, AO backbone frozen=true, # params=2.5M2026.03 | 2.04 | 7.9 | 1.71 |