Table 3.
Classification accuracies of various combinations of source separation and sound activity detection methods for one of the microphone configurations of the noisy SDL database, across four different noise configurations. Despite this, source separation methods do provide relative reductions in error rate of up to 7%, 48%, and 2% compared with CSLE alone for the radio normal, rain, and radio loud conditions, respectively.
| Method | Radio normal | Rain | Radio loud | Talking |
|---|---|---|---|---|
| (Mic 1 and Mic 2) | (Mic 1 and Mic 2) | (Mic 1 and Mic 2) | (Mic 1 and Mic 2) | |
| Baseline (channel with best SNR + Baseline SAD) | 44.04% | 61.39% | 28.58% | 39.54% |
| SAD alone: baseline SAD | 56.33% | 77.78% | 32.92% | 46.96% |
| SAD alone: CSLE | 71.96% | 97.18% | 40.25% | 69.45% |
| ICASDL_Apriori | 44.54% | 61.06% | 29.71% | 48.75% |
| ICASDL_Apriori + CSLE | 71.63% | 98.01% | 41.04% | 66.95% |
| ICABestMic | 43.83% | 61.30% | 29.71% | 41.54% |
| ICABestMic + CSLE | 72.13% | 97.08% | 41.38% | 67.12% |
| ICALeastNoise | 43.04% | 60.88% | 30.46% | 44.83% |
| ICALeastNoise + CSLE | 72.00% | 98.10% | 41.50% | 68.25% |
| ICASDL_BestSNR | 44.46% | 60.88% | 29.50% | 41.63% |
| ICASDL_BestSNR + CSLE | 73.04% | 97.50% | 41.08% | 67.75% |
| ICASDL_LeastNoise | 43.83% | 60.56% | 30.42% | 45.75% |
| ICASDL_LeastNoise + CSLE | 73.79% | 96.81% | 40.88% | 68.75% |
| ICASSA | 44.08% | 61.90% | 29.83% | 40.92% |
| ICASSA + CSLE | 71.04% | 97.04% | 41.58% | 68.21% |