Skip to main content

Table 2 ASR WERs (%) of various types of features as DNN input with DNN-HMM for Dev

From: Front-end technologies for robust ASR in reverberant environments—spectral enhancement-based dereverberation and auditory modulation filterbank features

Dev. 1ch ASR task SimData RealData
  Utterance-based batch processing mode Room1 Room2 Room3 Avg. Room1 Avg.
  with multi-condition training Near Far Near Far Near Far   Near Far  
DNN-HMM MFCC(L=R=5,#143) 6.61 8.82 8.31 17.03 9.97 20.03 11.79 28.70 28.91 28.80
  AMFB(#117) 6.22 7.74 7.49 14.52 9.79 16.64 10.39 25.39 27.00 26.19
  FBANK(L=R=5,#440) 5.80 6.91 7.17 13.90 8.28 15.01 9.50 26.08 25.91 25.99
  FBANK(L=R=5,#440)+SE 5.83 7.23 7.44 14.84 8.68 14.89 9.81 25.60 25.76 25.67
  AMFB-FBANK(#360) 5.19 6.32 7.69 12.50 7.72 13.95 8.89 22.27 26.66 24.45
  AMFB-FBANK(#360)+SE 5.70 6.86 7.44 12.10 7.79 13.33 8.86 22.14 25.56 23.84