Table 4 WER when using acoustic models trained with (‘ ✓’) and without (‘-’) processing the training data with the SE front-end used during testing. The results are averaged over the acoustic conditions of SimData and RealData, for the development set
From: Strategies for distant speech recognitionin reverberant environments
Processing | Train w/ | SimData | RealData |
|---|---|---|---|
SE front-end | |||
Distant | - | 8.3 % | 24.1 % |
WPE(1ch) | - | 7.5 % | 22.8 % |
✓ | 7.0 % | 24.4 % | |
WPE(8ch) + MVDR | - | 6.2 % | 16.1 % |
✓ | 5.0 % | 20.5 % |