Introduction

This is ASR source code for Online Compressive Transformer from National Chiao Tung University, Taiwan. The code also includes Synchronous transformer, it refer to the original paper[1].

Requirements

• ESPnet
• Python 3.6.1+
• gcc 4.9+ for PyTorch1.0.0+

Optionally, GPU environment requires the following libraries:

• Cuda 8.0, 9.0, 9.1, 10.0 depending on each DNN library
• Cudnn 6+, 7+
• NCCL 2.0+ (for the use of multi-GPUs)

Installation

The installation is the same with ESPnet verison 1. If you want to get more detail with installation, you can refer to ESPnet tutorial.

Please using this espnet version to install, the latest espnet maybe have something wrong to run our source code.

cd tools
make

Train

Online compressive transformer doesn't use the language model, so we skip the stage 3.

cd egs/aishell/com_asr
# if you need prepare the dataset
./run_com.sh --stage -1  

cd egs/aishell/com_asr
# if you just train the model
./run_com.sh --stage 4 

Inference

tag is the model name in egs/aishell/com_asr/exp/ folder.

cd egs/aishell/com_asr
./run_com.sh --stage 5 --recog_set "dev test" --tag "compressive_256GLU+CTC_.2_.3_.2_all_grad_ws9_ver2"

model link

the trained models was put to gdrive, you can download in the following links and put in egs/aishell/com_asr/exp/.

Aishell

• Online compressive transformer version 1
• Online compressive transformer version 2
• Synchronous transformer[1]
• Transformer

Reference

[1] Tian, Z., Yi, J., Bai, Y., Tao, J., Zhang, S., & Wen, Z. (2020). Synchronous Transformers for end-to-end Speech Recognition. ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 7884-7888.

Acknowledge

Online compressive Transformer uses ESPNET1 framework.