Table 6 The effect of different feature aggregation methods on recognition performance
From: Anchor voiceprint recognition in live streaming via RawNet-SA and gated recurrent unit
Aggregation Method | VoxCeleb-E | VoxCeleb-H | CN-Celeb | ||||||
|---|---|---|---|---|---|---|---|---|---|
EER | DCF08 | DCF10 | EER | DCF08 | DCF10 | EER | DCF08 | DCF10 | |
Average Pooling | 3.20% | 0.16 | 0.52 | 5.35% | 0.25 | 0.65 | 21.85% | 0.73 | 0.93 |
Max Pooling | 6.66% | 0.34 | 0.80 | 10.56% | 0.47 | 0.86 | 23.66% | 0.80 | 0.96 |
ASP | 5.85% | 0.31 | 0.77 | 9.41% | 0.44 | 0.85 | 23.50% | 0.79 | 0.96 |
SAP | 4.24% | 0.22 | 0.63 | 6.97% | 0.33 | 0.78 | 22.40% | 0.76 | 0.94 |
Ghost VLAD | 3.01% | 0.16 | 0.52 | 5.01% | 0.24 | 0.63 | 21.32% | 0.73 | 0.93 |
Bi-GRU | 2.80% | 0.15 | 0.52 | 5.04% | 0.25 | 0.67 | 22.31% | 0.77 | 0.95 |
GRU | 2.54% | 0.14 | 0.47 | 4.52% | 0.22 | 0.65 | 22.24% | 0.76 | 0.94 |