EURASIP Journal on Audio, Speech, and Music Processing

Table 19 Experimental results on the real reverberant data by using NN training dataset ‘1s.20u’ (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 (%)
Method	Frame selection	Acoustic model	InCar	Lounge	MeetR	Office	Average
LMCN + Proposed (24 NNs)	Skip1 8-1-0	Clean-rt	84.9	58.5	90.7	93.6	81.9
		Multi-rt	80.9	82.7	90.4	83.6	84.4
		Clean+multi-rt	89.1	90.7	94.7	93.0	91.8
	Skip1 16-1-0	Clean-rt	84.5	55.4	89.5	93.7	80.7
		Multi-rt	78.2	80.5	88.5	84.5	82.9
		Clean+multi-rt	88.5	90.6	94.0	93.9	91.7

The bold text represents the best average performance.

Back to article page