Table 6 The average WERs (%) of Eval. dataset for the single channel dataset comparing with other teams. Multi-condition training dataset and the trigram language model were used for all teams
From: Environment-dependent denoising autoencoder for distant-talking speech recognition
Team | Acoustic | Feature | Dereverberation | SimData | RealData | Ave. | ||
|---|---|---|---|---|---|---|---|---|
model | method | |||||||
REVERB-challenge | GMM | MFCC | CMVN | 25.27 | 47.48 | 36.38 | ||
baseline | ||||||||
J. Alam et al. [45] | DNN | MFCC | maximum likelihood inverse | 11.1 | 32.4 | 21.8 | ||
filtering-based dereverberation | ||||||||
Y. Tachioka et al. [46] | MMI-SGMM | MFCC and PLP | Single-channel dereverberation | 10.05 | 28.06 | 19.01 | ||
with estimation of | ||||||||
reverberation time | ||||||||
This paper | MMI-SGMM | MFCC | One-step environment- | 7.04 | 28.66 | 17.85 | ||
dependent | ||||||||
DAE | ||||||||
This paper | DNN | MFCC | One-step environment- | 7.46 | 28.11 | 17.79 | ||
dependent | ||||||||
DAE | ||||||||
This paper | MMI-SGMM +DNN | MFCC | One-step environment- | 6.41 | 26.83 | 16.62 | ||
dependent | ||||||||
DAE | ||||||||