Universal Speech Enhancement on URGENT non-blind 2025 (test)
3.26DNSMOSShifted anechoic + GAN correction
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Shifted anechoic + GAN correctionTraining target=Shifted anechoic, GAN correction=true2026.03 | 3.26 | 4.12 | 2.8 | 2.38 | 76 | 8.18 | 3.73 | 3.51 | 89 | 84 | 83 | 89.88 | |
| Shifted anechoicTraining target=Shifted anechoic, GAN correction=false2026.03 | 3.25 | 3.85 | 2.76 | 2.41 | 77 | 8.23 | 3.63 | 3.28 | 89 | 84 | 82 | 89.41 | |
| Specialized (upper bound)Exit layer=8, Look-ahead=1, Parameters=2.9M, MACs (G/s)=25.41, Algo. latency (ms)=60, Comp. latency (ms)=18.312026.06 | 3.15 | 3.9 | 2.42 | 2.27 | 72 | — | — | — | 85 | 79 | — | 84.93 | |
| Multiple dec. stageExit layer=8, Look-ahead=1, Parameters=2.9M, MACs (G/s)=25.41, Algo. latency (ms)=60, Comp. latency (ms)=18.312026.06 | 3.13 | 3.74 | 2.37 | 2.24 | 72 | — | — | — | 84 | 79 | — | 84.62 | |
| Multiple dec. stageExit layer=12, Look-ahead=0, Parameters=3.7M, MACs (G/s)=31.67, Algo. latency (ms)=40, Comp. latency (ms)=25.052026.06 | 3.11 | 3.7 | 2.34 | 2.17 | 70 | — | — | — | 84 | 78 | — | 83.25 | |
| Specialized (upper bound)Exit layer=8, Look-ahead=0, Parameters=2.9M, MACs (G/s)=25.41, Algo. latency (ms)=40, Comp. latency (ms)=18.312026.06 | 3.1 | 3.77 | 2.36 | 2.19 | 71 | — | — | — | 84 | 78 | — | 83.71 | |
| Parallel conv. (MoE)Exit layer=8, Look-ahead=1, Parameters=2.9M, MACs (G/s)=25.41, Algo. latency (ms)=60, Comp. latency (ms)=18.312026.06 | 3.1 | 3.82 | 2.34 | 2.21 | 71 | — | — | — | 84 | 79 | — | 84.06 | |
| Specialized (upper bound)Exit layer=12, Look-ahead=0, Parameters=3.7M, MACs (G/s)=31.67, Algo. latency (ms)=40, Comp. latency (ms)=25.052026.06 | 3.1 | 3.76 | 2.37 | 2.21 | 71 | — | — | — | 84 | 78 | — | 84.24 | |
| Early-exitExit layer=12, Look-ahead=0, Parameters=3.7M, MACs (G/s)=31.67, Algo. latency (ms)=40, Comp. latency (ms)=25.052026.06 | 3.1 | 3.81 | 2.34 | 2.14 | 70 | — | — | — | 84 | 77 | — | 82.62 | |
| Early-exitExit layer=8, Look-ahead=0, Parameters=2.9M, MACs (G/s)=25.41, Algo. latency (ms)=40, Comp. latency (ms)=18.312026.06 | 3.08 | 3.77 | 2.32 | 2.13 | 70 | — | — | — | 83 | 77 | — | 81.84 | |
| Multiple dec. stageExit layer=8, Look-ahead=0, Parameters=2.9M, MACs (G/s)=25.41, Algo. latency (ms)=40, Comp. latency (ms)=18.312026.06 | 3.07 | 3.62 | 2.31 | 2.15 | 70 | — | — | — | 84 | 77 | — | 83.1 | |
| Parallel conv. (MoE)Exit layer=12, Look-ahead=0, Parameters=3.7M, MACs (G/s)=31.67, Algo. latency (ms)=40, Comp. latency (ms)=25.052026.06 | 3.07 | 3.78 | 2.31 | 2.15 | 70 | — | — | — | 83 | 77 | — | 82.93 | |
| Early reflectedTraining target=Early reflected, GAN correction=false2026.03 | 3.06 | 3.23 | 2.26 | 2.81 | 85 | 12.28 | 2.87 | 2.66 | 90 | 84 | 82 | 87.62 | |
| Specialized (upper bound)Exit layer=4, Look-ahead=0, Parameters=2.0M, MACs (G/s)=19.15, Algo. latency (ms)=40, Comp. latency (ms)=10.092026.06 | 3.05 | 3.65 | 2.26 | 2.06 | 68 | — | — | — | 82 | 76 | — | 82.75 | |
| Early reflected + GAN correctionTraining target=Early reflected, GAN correction=true2026.03 | 3.04 | 3.53 | 2.3 | 2.78 | 84 | 12.25 | 2.97 | 2.75 | 90 | 84 | 85 | 88.13 | |
| Parallel conv. (MoE)Exit layer=8, Look-ahead=0, Parameters=2.9M, MACs (G/s)=25.41, Algo. latency (ms)=40, Comp. latency (ms)=18.312026.06 | 3.04 | 3.73 | 2.28 | 2.13 | 70 | — | — | — | 83 | 77 | — | 82.72 | |
| Early-exitExit layer=4, Look-ahead=0, Parameters=2.0M, MACs (G/s)=19.15, Algo. latency (ms)=40, Comp. latency (ms)=10.092026.06 | 3.03 | 3.57 | 2.19 | 2.02 | 67 | — | — | — | 82 | 75 | — | 81.13 | |
| Rank 22026.03 | 3.01 | 3.21 | 2.3 | 2.79 | 85 | 13.11 | 2.93 | 2.94 | 90 | 85 | 84 | 88.05 | |
| Rank 12026.03 | 3.01 | 3.41 | 2.4 | 2.95 | 86 | 14.33 | 3.01 | 2.83 | 91 | 86 | 85 | 88.92 | |
| Rank 32026.03 | 3 | 3.45 | 2.31 | 2.74 | 84 | 13.06 | 3.3 | 3.08 | 89 | 84 | 83 | 87.94 | |
| Multiple dec. stageExit layer=4, Look-ahead=0, Parameters=2.0M, MACs (G/s)=19.15, Algo. latency (ms)=40, Comp. latency (ms)=10.092026.06 | 2.98 | 3.41 | 2.19 | 2.02 | 67 | — | — | — | 82 | 75 | — | 81.86 | |
| Parallel conv. (MoE)Exit layer=4, Look-ahead=0, Parameters=2.0M, MACs (G/s)=19.15, Algo. latency (ms)=40, Comp. latency (ms)=10.092026.06 | 2.96 | 3.52 | 2.16 | 2 | 67 | — | — | — | 82 | 75 | — | 81.69 | |
| Baseline2026.03 | 2.94 | 2.89 | 2.11 | 2.43 | 80 | 11.29 | 3.32 | 2.85 | 86 | 79 | 80 | 84.96 | |
| TF-GridNetCausal=non-causal2026.06 | 2.94 | 2.89 | 2.11 | — | — | — | — | — | — | — | — | 84.96 | |
| Noisy2026.03 | 1.84 | 1.69 | 1.56 | 1.37 | 61 | 2.53 | 7.92 | 5.51 | 75 | 62 | 63 | 81.29 | |
| Noisy2026.06 | 1.84 | 1.69 | 1.56 | 1.34 | 50 | — | — | — | 74 | 61 | — | 81.29 |