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

EURASIP Journal on Audio, Speech, and Music Processing

Table 7 Experimental results by using combined training data (three known positions) on the first testing scheme

Method	Dataset		Speaker identification rate (%)
Method	Dataset		P01	P03	P05	P02	P04	Avg. known	Avg. unknown	Avg. all
Prop. (24 NNs) + CMN	combd	1s.15u	88.9	90.4	93.9	93.8	90.2	91.1	92.0	91.4
Prop. (24 NNs) + CMN	combd	3s.45u	90.5	93.7	94.5	96.2	92.0	92.9	94.1	93.4
Prop. (12 NNs) + CMN	combd	1s.15u	90.9	92.5	93.0	94.2	92.2	92.1	93.2	92.6
Prop. (12 NNs) + CMN	combd	3s.45u	93.5	94.8	94.2	96.2	93.2	94.2	94.7	94.4
Prop. (6 NNs) + CMN	combd	1s.15u	90.8	91.2	91.6	94.4	93.8	91.2	94.1	92.4
Prop. (6 NNs) + CMN	combd	3s.45u	92.7	92.2	92.5	96.0	94.0	92.4	95.0	93.5

The known environments include P01, P03, and P05, while the unknown environments include P02 and P04. The experiments were done by using skip1 7-1-0 frame selection. The bold text represents the best average performance for each training data number.