Skip to main content

Advertisement

Table 2 WER (%) comparisons on SimData: Clean-Model and Reverb-Model stand for baseline systems of clean-condition training and multi-condition training, respectively

From: Joint training of DNNs by incorporating an explicit dereverberation structure for distant speech recognition

System T60 = 0.25 s T60 = 0.61 s T60 = 1.10 s
Clean-condition training    
Clean-Model 21.49 50.21 91.69
DNN-PP 14.64 17.04 40.99
Multi-condition training    
Reverb-Model 6.75 7.96 21.10
DNN-PP 7.08 8.49 17.61
DNN-FM1 6.77 7.96 19.79
DNN-JT1 6.06 6.93 16.88
  1. DNN-PP stands for pre-processing, DNN-JT1 for conventional joint training structure, and DNN-FM1 for intermediate result of DNN-JT1