Table 6 Comparison of model performance on V1 and V2 datasets.
From: Dynamic convolution models for cross-frontend keyword spotting
Model | (Par., Flops.) | V1 | V2 | ||
|---|---|---|---|---|---|
Acc | Best | Acc | Best | ||
TCNet829 | (145K, 4.40M) | – | 96.2 | – | – |
TCNet1429 | (305K, 8.26M) | – | 96.6 | 96.53 | 96.8 |
KWT-131 | (607K, –) | 97.05 | 97.28 | 97.72 | 97.73 |
LeTR-12832 | (617K, –) | 97.61 | – | 97.82 | – |
LightConv41 | (105K, 7.40M) | 96.88 | 97.0 | 97.24 | 97.3 |
DyConv41 | (107K, 7.69M) | 96.89 | 97.1 | 96.26 | 97.4 |
TENet642 | (54K, 3.95M) | – | 96.4 | 96.83 | 97.0 |
TENet1242 | (100K, 6.42M) | – | 96.6 | 97.10 | 97.3 |
DARTS43 | (93K, 4.9M) | 96.63 | 96.9 | 96.92 | 97.1 |
F-DARTS43 | (188K, 10.6M) | 96.70 | 96.9 | 97.11 | 97.4 |
N-DARTS43 | (109K, 6.3M) | 96.79 | 97.2 | 97.18 | 97.4 |
DynTempConvNet | (62K, 6.11M) | 97 | 97.17 | 97.48 | 97.56 |
