End-to-end Automatic Speech Recognition Systems - PyTorch Implementation

For complete introdution and usage, please see the original repository Alexander-H-Liu/End-to-end-ASR-Pytorch.

New features

1. SpecAugment
2. Label Smoothing
3. VGG encoder with Layer Normalization
4. Learning rate scheduler
5. Stabilizing joint acoustic and language model beam decoding by threshold

Instructions

Dataset

Download librispeech datset first in OpenSLR website

Training

Modify script/train.sh, script/train_lm.sh, config/librispeech_asr.yaml, and config/librispeech_lm.yaml first. GPU is required.

bash script/train.sh <asr name> <cuda id>
bash script/train_lm.sh <lm name> <cuda id>

Testing

Modify script/test.sh and config/librispeech_test.sh first. Increase the number of --njobs can speed up decoding process, but might cause OOM.

bash script/test.sh <asr name> <cuda id>

LibriSpeech 100hr Result

This baseline is composed of a character-based joint CTC-attention ASR model and an RNNLM which were trained on the LibriSpeech train-clean-100. The perplexity of the LM on the dev-clean set is 2.79.

Decoding	DEV WER(%)	TEST WER(%)
Greedy	14.74	14.80
B=2 + LM	12.89	12.93
B=4 + LM	11.67	11.74
B=8 + LM	11.35	11.42