Skip to main content

Advertisement

Table 18 Experimental results on the real reverberant data by using NN training data ā€˜5s.20uā€™ (combined)

From: Single-channel dereverberation by feature mapping using cascade neural networks for robust distant speaker identification and speech recognition

Method Frame selection Acoustic model Digit accuracy (%)
InCar Lounge MeetR Office Average
LMCN + Proposed (24 NNs) Linear 8-1-0 Clean 69.4 38.5 78.9 75.3 65.6
Clean-rt 85.9 60.9 90.3 93.7 82.7
Multi-rt 82.4 85.1 90.9 86.0 86.1
Clean+multi-rt 89.6 91.0 94.4 93.5 92.1
Linear 16-1-0 Clean 67.8 35.1 79.4 78.0 65.1
Clean-rt 85.7 58.8 90.7 94.7 82.5
Multi-rt 82.6 84.2 90.2 89.3 86.6
Clean+multi-rt 89.2 90.9 94.4 93.5 92.0
Linear 32-1-0 Clean 67.0 30.0 74.6 77.6 62.3
Clean-rt 84.7 53.0 89.2 94.1 80.3
Multi-rt 78.4 82.0 88.4 87.9 84.2
Clean+multi-rt 86.5 89.2 93.1 92.0 90.2
Skip1 8-1-0 Clean 69.1 38.0 78.4 75.6 65.3
Clean-rt 85.3 61.3 90.9 93.5 82.8
Multi-rt 81.0 82.5 91.2 84.2 84.7
Clean+multi-rt 89.2 90.7 95.1 93.4 92.1
Skip1 16-1-0 Clean 67.2 34.7 75.8 77.6 63.8
Clean-rt 85.2 58.5 89.6 93.6 81.7
Multi-rt 83.2 84.8 91.5 87.1 86.6
Clean+multi-rt 90.0 92.9 95.5 94.3 93.2
  1. The bold text represents the best average performance.