표 3 / Table 3 REVERB challenge 2014 데이터에 대한 실험 결과(word error rate, %) / Experimental results on REVERB challenge 2014 evaluation set
Experimental condition | SimData | RealData |
Room1 | Room2 | Room3 | Average | Room1 | Average |
Acoustic model | Near | Far | Near | Far | Near | Far | Near | Far |
CNN | 5.74 | 6.74 | 7.28 | 12.42 | 8.16 | 15.00 | 9.22 | 24.66 | 26.06 | 25.36 |
VDCNN | 4.88 | 5.54 | 5.81 | 8.80 | 6.34 | 10.26 | 6.94 | 19.35 | 20.26 | 19.81 |
DenseNet | 5.00 | 5.20 | 5.29 | 8.15 | 6.18 | 8.90 | 6.45 | 18.62 | 19.51 | 19.07 |
MR-DenseNet | 4.47 | 5.12 | 5.22 | 7.43 | 5.71 | 8.73 | 6.11 | 17.66 | 18.50 | 18.08 |
reverberant voice enhancement and recognition benchmark: REVERB, convolutional neural network: CNN; very deep convolutional neural network: VDCNN, multi-resolution DenseNet: MR-DenseNet