Table 8 Performance comparison of models on noise datasets across different SNR levels on v2 dataset.
From: Dynamic convolution models for cross-frontend keyword spotting
Noise | SNR (dB) | Model | |||
|---|---|---|---|---|---|
DynTempConvNet + DML | DynTempConvNet | TENet6 | TCNet14 | ||
Urban | 20 | \(95.98 \pm 0.12\) | \(95.83 \pm 0.24\) | \(95.77 \pm 0.15\) | \(95.31 \pm 0.22\) |
15 | \(95.13 \pm 0.21\) | \(94.69 \pm 0.21\) | \(94.58 \pm 0.09\) | \(94.08 \pm 0.26\) | |
10 | \(93.63 \pm 0.14\) | \(93.19 \pm 0.19\) | \(92.71 \pm 0.23\) | \(91.87 \pm 0.11\) | |
5 | \(90.29 \pm 0.18\) | \(89.42 \pm 0.16\) | \(88.33 \pm 0.24\) | \(87.77 \pm 0.13\) | |
0 | \(84.17 \pm 0.10\) | \(82.63 \pm 0.27\) | \(81.13 \pm 0.12\) | \(80.17 \pm 0.25\) | |
WHAM | 20 | \(96.06 \pm 0.15\) | \(96.02 \pm 0.14\) | \(94.96 \pm 0.11\) | \(95.17 \pm 0.22\) |
15 | \(94.81 \pm 0.18\) | \(94.71 \pm 0.25\) | \(93.65 \pm 0.09\) | \(93.60 \pm 0.27\) | |
10 | \(93.35 \pm 0.13\) | \(92.54 \pm 0.21\) | \(91.15 \pm 0.24\) | \(90.96 \pm 0.12\) | |
5 | \(87.75 \pm 0.21\) | \(87.25 \pm 0.15\) | \(85.96 \pm 0.23\) | \(85.19 \pm 0.10\) | |
0 | \(78.67 \pm 0.08\) | \(78.12 \pm 0.19\) | \(76.29 \pm 0.14\) | \(74.96 \pm 0.29\) | |
