tencent-ailab/3m-asr

3M-ASR for End-to-End Speech Recognition

This project is used to build an End-to-End Speech Recognition system based on Mixture-of-Experts(MoE) model. MoE is an efficient way to train a large scale model and we have proved its efficiency on public dataset. More details about the algorithm can be found in "3M: Multi-loss, Multi-path and Multi-level Neural Networks for speech recognition".

Installation

• Clone this repo

git clone https://github.com/tencent-ailab/3m-asr.git

• Install Conda: please see https://docs.conda.io/en/latest/miniconda.html
• Create Conda env:

conda create -n moe python=3.8
conda activate moe
pip install -r requirements.txt
conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.9.0 cudatoolkit=11.1 -c pytorch -c conda-forge

• Follow the instruction under directory fastmoe to install fastmoe

Performance Benchmark

We evaluate our system on the public WenetSpeech dataset and the recipe of Conformer-MoE is provided(trained on 24 V100). CER results are listed below and the first three lines are provided by WenetSpeech

Acknowledge

• We used FastMoE to support Mixture-of-Experts model training in Pytorch
• We borrowed a lot of code from WeNet for the implementation of Conformer and data processing

Reference

[1] SpeechMoE: Scaling to Large Acoustic Models with Dynamic Routing Mixture of Experts(InterSpeech 2021)

[2] 3M: Multi-loss, Multi-path and Multi-level Neural Networks for speech recognition(Submitted to InterSpeech 2022)

Citation

@inproceedings{you21_interspeech,
  author={Zhao You and Shulin Feng and Dan Su and Dong Yu},
  title={{SpeechMoE: Scaling to Large Acoustic Models with Dynamic Routing Mixture of Experts}},
  year=2021,
  booktitle={Proc. Interspeech 2021},
  pages={2077--2081},
  doi={10.21437/Interspeech.2021-478}
}

@article{you20223m,
  title={3M: Multi-loss, Multi-path and Multi-level Neural Networks for speech recognition},
  author={You, Zhao and Feng, Shulin and Su, Dan and Yu, Dong},
  journal={arXiv preprint arXiv:2204.03178},
  year={2022}
}

Contact

If you have any questions about this project, please feel free to contact [email protected] or [email protected]

Disclaimer

This is not an officially supported Tencent product