This repository contains the original code of Gated Recurrent Context (paper links: arXiv, IEEE Transactions on ASLP).
This repository includes a pipeline scripts for training & inference of AED (Attention-based Encoder-Decoder) models for speech recognition task on LibriSpeech dataset.
All the included scripts are largely based on the RETURNN toolkit.
Installation process includes data download & preparation.
Basically follow the same process as installation guide of original setup by Albert Zeyer, only with a little modification as follows.
As this repository already contains RETURNN toolkit scripts in returnn/ directory, the '03_git_clone_returnn.sh' script do not have to be executed.
If you want to use up-to-date version of RETURNN scripts rather than the RETURNN originally included in this repository, please check the instructions in '03_git_clone_returnn.sh' and follow them.
Just execute scripts '01_pip_install_requirements.sh' to '21_prepare_train.sh', in index-order of pipeline bash-scripts.
CUDA_VISIBLE_DEVICES=[n] bash 22_train.sh [expname][expname] denotes the name of each experiment (e.g., 'E1.GSA_BiLSTM').
[n] denotes index of a GPU (e.g., '0')
(Only single-GPU experiments had been conducted.
Multi-GPU training is possible by using Horovod within RETURNN training, but not tested yet.)
CUDA_VISIBLE_DEVICES=[n] bash 23_recog.sh [expname] 0CUDA_VISIBLE_DEVICES=[n] bash 24_report_best_recog.sh [expname]