This repository contains the code for the ZeroEGGS project from this article. It also contains our stylized speech and gesture dataset

Create and activate a virtual environment to work in, e.g. using Conda:

conda create -n zeggs python=3.8
conda activate zeggs

Install CUDA and PyTorch 1.12.x For CUDA 11.3, this would look like:

conda install pytorch torchvision cudatoolkit=11.3 -c pytorch

Install the remaining requirements with pip:

pip install -r requirements.txt

You may need to install sox on your system

The following styles are present in the ZEGGS dataset:

This repository contains large files. In order to clone this repository including the the large zip files, you need to use git lfs. If you still get errors, directly download zip files.

The speech and gesture data are contained in the ./data/Zeggs_data.zip, ./data/Zeggs_data.z01, and ./data/Zeggs_data.z02 files. You must put all of these parts to the same folder, and extract .zip file by WinRAR or Winzip.

When you extract the zip file, there are two folders:

• original folder contains the original data where the animation and audio files are in their raw version and not processed.

• clean contains aligned animation and audio data and without unwanted audio of other speaker. For more details on how these files have been processed check data_pipeline.py

All the animation sequences are in the BVH file format and all the audio data are in WAV format.

Extract the data from the Zeggs_data.zip file and place it in the data folder. Next run:

python data_pipeline.py

This processes data and creates the necessary files for training and evaluation in the "processed" folder. You can customize the data pipeline by changing data_pipeline_conf.json config file. Two suggested configurations are provided in the configs folder. You should change the configuration file name in the script.

You can use pre-trained models stored in ./data/outputs/saved_models. However, if you want to train the model from scratch, run:

python ./main.py -o <configs> -n <run_name>

For example, to train the model with the default configuration, run:

python ./main.py -o "../configs/configs_v1.json" -n "zeggs_v1"

After training is finished or using provided pretrained models (provided in ./data/outputs), you can generate gestures given speech and style as input using generate.py. The output will be save in bvh format. For full functionality (blending, transitions, using pre-extracted style encodings, etc. ) you need to directly use generate_gesture function. Otherwise, you can use CLI as explained below.

You can run the inference using the CLI in two ways:

The CLI command looks like this:

python ./generate.py -o <options file> -s <style file> -a <audio file>

where options file is similar to the training config file but contains the path to the saved pretrained models and other required data. For example, you can run:

python ./generate.py -o "../data/outputs/v1/options.json" -s "../data/clean/067_Speech_2_x_1_0.bvh" -a "../data/clean/067_Speech_2_x_1_0.wav"

To get more help on how to set other parameters such as seed, temperature, output file name, etc., run the command below:

python ./generate.py -h.

You can generate a batch of animated gestures from a csv file containing audio and style file paths along with other parameters by running:

python ./generate.py -o <options file> -c <CSV file>

For example, you can run:

python ./generate.py -o "../data/outputs/v1/options.json" -c "../data/test/evaluation_example_based.csv"

The compatible FBX files (male and female characters) along with a MotionBuilder bvh2fbx script is provided in ./bvh2fbx folder. You need to modify the path to your python environment and also the FBX character in bvh2fbx.py and the paths to the MotionBuilder plugins and executable in bvh2fbx.bat scripts. Put all the pairs of bvh and wav files (with the same name) in the ./bvh2fbx/Rendered folder and run bvh2fbx.bat script.

If you use this code and dataset, please cite our paper

@article{ghorbani2022zeroeggs,
  author = {Ghorbani, Saeed and Ferstl, Ylva and Holden, Daniel and Troje, Nikolaus F. and Carbonneau, Marc-André},
  title = {ZeroEGGS: Zero-shot Example-based Gesture Generation from Speech},
  journal = {Computer Graphics Forum},
  volume = {42},
  number = {1},
  pages = {206-216},
  keywords = {animation, gestures, character control, motion capture},
  doi = {https://doi.org/10.1111/cgf.14734},
  url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.14734},
  eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1111/cgf.14734},
  year = {2023}
}

© [2022] Ubisoft Entertainment. All Rights Reserved