This repo holds the codes for the paper
Recurrent Convolutional Neural Network For Speech Processing, Yue Zhao, Xingyu Jin, and Xiaolin Hu, to appear in the 42nd IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2017), New Orleans, USA.
Clone Kaldi. Follow the documentation to build the toolkit. (Remember to install IRSTLM via ./tools/extras/install_irstlm.sh since the default setting does not include it.)
Install CNTK.
Put the scripts in timit-scripts/ under egs/timit/s5/ in the Kaldi repository.
Run ./run.sh to produce a HMM model by tri3 system.
Run ./train_rcnn_tri00.sh to generate the RCNN model described in our paper.