tuanio / asr-toolkit

E2E Speech Recognition Toolkit with Hydra and Pytorch Lightning

ASR E2E Toolkit

Encoder output

• Input: inputs, input_lengths
• Output: outputs, output_lengths

Decoder:

• Input: targets, target_lengths, encoder_outputs (optional)

• Output: decoder_outputs

• Nếu có encoder_outputs thì mới sử dụng attention

Với mỗi lớp encoder và decoder thì đều phải có tham số output_dim

References

• https://github.com/sooftware/conformer

Variable

set HYDRA_FULL_ERROR=1

Run

• python main.py -cp conf -cn configs

Training Template

• Conformer-Transformer-AED Vivos: https://colab.research.google.com/drive/10dpTTy5huj7SjLst2Jj_Iff2aXLYJoBS?usp=sharing
• Conformer-Transformer-AED LibriSpeech: https://www.kaggle.com/code/tuannguyenvananh/conformer-transformer-aed-librispeech

Install package

conda create -n train_env -c pytorch-nightly -c nvidia -c pytorch -c conda-forge python=3.7 pytorch torchaudio cudatoolkit pandas numpy