This repository contains code to reproduce the results from the paper A Prioritized Grid Long Short-Term Memory RNN for Speech Recognition.

To cite this work, please use

@inproceedings{hsu2016prioritized,
  title={A prioritized grid long short-term memory RNN for speech recognition},
  author={Hsu, Wei-Ning and Zhang, Yu and Glass, James},
  booktitle={Spoken Language Technology Workshop (SLT), 2016 IEEE},
  pages={467--473},
  year={2016},
  organization={IEEE}
}

This project uses Kaldi for feature extraction, inital HMM-GMM model training, forced alignment, and decoding. Neural network-based acoustic model training was done using CNTK.

Place files in Kaldi example script directories (e.g. kaldi/egs/hkust/s5) and run:

cntk_scripts/run_cntk_pglstm_5l.sh \
    --expdir <exp_dir> \
    --ali_src <ali_src> \
    --train_src <train_src> \
    --test_src <test_src> \
    --cn_gpu <cntk_bin> \

• exp_dir: directory to dump experiment results
• ali_src: directory containing forced alignment results ali.*.gz, kaldi GMM model final.mdl, and kaldi senone counts file final.occs.
• train_src: directory containing training set features feats.scp
• test_src: directory containing test set features feats.scp
• cntk_bin: path to CNTK binary