Table 6 WER on the evaluation data using DNN based acoustic models. “MAP” refers to the DNN based MFCC feature mapping
MAP | Simulated rooms | Real | Avg. | ||||||
|---|---|---|---|---|---|---|---|---|---|
Room1A | Room2A | Room3A | Room1 | ||||||
near | far | near | far | near | far | near | far | ||
Single channel + Clean condition training | |||||||||
No | 10.6 | 19.3 | 23.2 | 69.3 | 30.2 | 74.6 | 68.0 | 66.2 | 45.2 |
Yes | 9.3 | 10.6 | 12.7 | 21.4 | 16.5 | 25.1 | 40.2 | 39.0 | 21.9 |
Single channel + Multi condition training | |||||||||
No | 8.7 | 9.4 | 10.5 | 16.5 | 13.4 | 20.0 | 35.4 | 34.3 | 18.5 |
Yes | 8.9 | 8.8 | 8.8 | 13.9 | 11.4 | 15.5 | 32.2 | 32.7 | 16.5 |
MVDR(2ch) + Multi condition training | |||||||||
No | 8.6 | 9.6 | 9.1 | 14.9 | 11.6 | 18.3 | 33.3 | 30.7 | 17.0 |
Yes | 8.5 | 8.6 | 7.9 | 12.4 | 10.1 | 14.8 | 29.1 | 29.1 | 15.1 |
MVDR(8ch) + Multi condition training | |||||||||
No | 7.8 | 8.3 | 8.3 | 10.8 | 9.8 | 13.3 | 24.8 | 25.1 | 13.5 |
Yes | 7.5 | 8.2 | 7.4 | 9.7 | 8.9 | 11.3 | 22.7 | 24.4 | 12.5 |