End-to-End Audiovisual Speech Recognition

Introduction

This is the respository of End-to-End Audiovisual Speech Recognition. Our paper can be found here.

The video-only stream is based on T. Stafylakis and G. Tzimiropoulos's implementation. The paper can be found here.

This implementation includes 2-layer BGRU which consists of 1024 cells in each layer while Themos's implementation uses 2-layer BLSTM with 512 cells.

Dependencies

python 2.7
pytorch 0.3.1
opencv-python 3.4.0

Dataset

The results obtained with the proposed model on the LRW dataset. The coordinates for cropping mouth ROI are suggested as (x1, y1, x2, y2) = (80, 116, 175, 211) in Matlab. Please note that the fixed cropping mouth ROI (FxHxW) = [:, 115:211, 79:175] in python.

Training

This is the suggested order to train models including video-only model, audio-only model and audiovisual models:

i) Start by training with temporal convolutional backend, you can run the script:

CUDA_VISIBLE_DEVICES='' python main.py --path '' --dataset <dataset_path> \
                                       --mode 'temporalConv' \
                                       --batch_size 36 --lr 3e-4 \
                                       --epochs 30

ii)Throw away the temporal convolutional backend, freeze the parameters of the frontend and the ResNet and train the LSTM backend, then run the script:

CUDA_VISIBLE_DEVICES='' python main.py --path './temporalConv/temporalConv_x.pt' --dataset <dataset_path> \
                                       --mode 'backendGRU' --every-frame \
                                       --batch_size 36 --lr 3e-4 \
                                       --epochs 5

iii)Train the whole network end-to-end. You can run the script:

CUDA_VISIBLE_DEVICES='' python main.py --path './backendGRU/backendGRU_x.pt' --dataset <dataset_path> \
                                       --mode 'finetuneGRU' --every-frame \
                                       --batch_size 36 --lr 3e-4 \
                                       --epochs 30

Notes

every-frame is activated when the backend module is recurrent neural network.

dataset need be correctly specified before running. Code has strong assumptions on the dataset organisation.

temporalConv_x.pt or backendGRU_x.pt are the models with best validation performance on step ii) or step iii).

Models&Accuracy

Stream	Accuracy
video-only	83.39
audio-only	97.72
audiovisual	98.38

The results are slightly better than ones reported in the ICASSP paper due to further fine-tuning of the models. Please send email at pingchuan.ma16 <AT> imperial.ac.uk with name and affiliation for the pre-trained models.

Reference

If the code of this repository was useful for your research, please cite our work:

@article{petridis2018end,
  title={End-to-end audiovisual speech recognition},
  author={Petridis, Stavros and Stafylakis, Themos and Ma, Pingchuan and Cai, Feipeng and Tzimiropoulos, Georgios and Pantic, Maja},
  booktitle={ICASSP},
  pages={6548--6552},
  year={2018},
  organization={IEEE}
}