Skip to main content

Advertisement

Table 5 Experimental results on the use of modified multiple NNs configuration for known positions (matched conditions)

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

Method Dataset Error rate reduction (%)
P01 P02 P05 Avg.
Proposed (1 NN) + CMN 1s.1u 30.8 19.4 21.7 24.0
1s.5u 34.7 27.8 23.5 28.7
3s.15u 35.4 33.3 23.5 30.8
Proposed (6 NNs) + CMN 1s.1u 40.8 44.0 19.7 34.8
1s.5u 45.5 62.7 16.3 41.5
3s.15u 51.0 71.4 23.5 48.7
Proposed (12 NNs) + CMN 1s.1u 33.0 33.7 22.9 29.8
1s.5u 40.6 54.0 32.0 42.2
3s.15u 55.2 71.4 27.5 51.4
Proposed (24 NNs) + CMN 1s.1u 17.9 2.5 10.5 10.3
1s.5u 30.2 31.0 22.2 27.8
3s.15u 46.9 42.9 13.7 34.5
  1. The experiments were done by using the first testing scheme and skip1 3-1-0 frame selection. The results are shown in term of ERR relative to the ‘CMN’. The bold text represents the best average performance for each training data number.