In this repository, we provide two example models for the 2023 GENEA challenge:

1. the first (monadic) model is only trained on the main agent's own speech; while
2. the second (dyadic) model also considers the interlocutor's speech and movements as its conditioning input.

The models are based on the paper "The IVI Lab entry to the GENEA Challenge 2022 – A Tacotron2 Based Method for Co-Speech Gesture Generation With Locality-Constraint Attention Mechanism".

The implementation is kindly provided by Che-jui Chang from Rutgers University. If you use this repository in your work, please cite their paper using the following .bib entry:

@inproceedings{chang2022ivi,
  title={The IVI Lab entry to the GENEA Challenge 2022--A Tacotron2 based method for co-speech gesture generation with locality-constraint attention mechanism},
  author={Chang, Che-Jui and Zhang, Sen and Kapadia, Mubbasir},
  booktitle={Proceedings of the 2022 International Conference on Multimodal Interaction},
  pages={784--789},
  year={2022}
}

To install the required libraries, activate your python environment and use the following command:

pip install -r requirements.txt

Note: This repository was tested under Ubuntu 20.04 and Python 3.9 using the above libraries, but it may still work under different configurations.

Get the dataset from the official GENEA 2023 repository and unzip all files to a folder of your choosing.

Download and unzip the word embedding from FastText. Put the file "crawl-300d-2M.vec" under the project directory.

wget https://dl.fbaipublicfiles.com/fasttext/vectors-english/crawl-300d-2M.vec.zip
unzip crawl-300d-2M.vec.zip
rm crawl-300d-2M.vec.zip

Preprocess the data and save them as h5 files. (This might take a while.)

python process_data.py -d <path_to_dataset> 

The above script generates the following four h5 files under the project directory:

• trn_interloctr_v0.h5
• trn_main-agent_v0.h5
• val_interloctr_v0.h5
• val_main-agent_v0.h5

Calculate audio statistics based on the training set:

python calculate_audio_statistics.py

We will use the following command to generate pipelines (*.sav) for converting our motion representations to euler rotations.

python create_pipeline.py

We provide two baseline configurations: a monadic system (generating gestures from the agent's own speech only), and a dyadic system (which is also conditioned on the interlocutor's behaviour).

The training configurations are defined in Tacotron2/common/hparams_monadic.py and Tacotron2/common/hparams_dyadic.py. You might want to edit the following options:

1. Enter the output_directory and device of your own choice.
2. Add the checkpoint_path if you would like to resume training.
3. Make sure to edit n_acoustic_feat_dims, 78 for full-body training, and 57 for upper-body training only.

Train the model using the following commands:

cd Tacotron2
python train_monadic.py # (Option 1)
python train_dyadic.py  # (Option 2)

The weights and logs can be found under the output_directory. It takes roughly 20k~30k iterations to produce decent results.

The below codes generate all gestures given the processed validation data (val_main-agent_v0.h5). You can use the same scripts to generate the motion for all test sequences once the test inputs are released near the end of the challenge.

python generate_all_gestures.py -ch <checkpoint_path> -t full

The bvh files should be generated under "Tacotron2/outputs/" folder. By defaut, the cuda device "0" is used. If you prefer to use a different cuda device for inference, please edit line 23 in the Tacotron2/common/hparams_monadic.py or Tacotron2/common/hparams_dyadic.py files.

Note: You can import the bvh files to Blender to visualize the gesture motions. If you would like to render the motions, please reach out to the repository provided by GENEA Challenge 2023.