search

GestureGeneration / Speech_driven_gesture_generation_with_autoencoder

This is the official implementation for IVA '19 paper "Analyzing Input and Output Representations for Speech-Driven Gesture Generation".
https://svito-zar.github.io/audio2gestures/
Apache License 2.0
107 stars 27 forks source link
autoencoder deep-learning generation gesture gesture-controller gesture-generation hri keras neural-network tensorflow

readme

Aud2Repr2Pose: Analyzing input and output representations for speech-driven gesture generation

Taras Kucherenko, Dai Hasegawa, Gustav Eje Henter, Naoshi Kaneko, Hedvig Kjellström

ImageOfIdea

This repository contains Keras and Tensorflow based implementation of the speech-driven gesture generation by a neural network which was published at International Conference on Intelligent Virtual Agents (IVA'19) and the extention was published in International Journal of Human-Computer Interaction in 2021.

The project website contains all the information about this project, including video explanation of the method and the paper.

Demo on another dataset

This model has been applied to English dataset.

The demo video as well as the code to run the pre-trained model are online.

Requirements

Initial setup

install packages


# if you have GPU
pip install tensorflow-gpu==1.15.2

# if you don't have GPU
pip install tensorflow==1.15.2

pip install -r requirements.txt

install ffmpeg

# macos
brew install ffmpeg
# ubuntu
sudo add-apt-repository ppa:jonathonf/ffmpeg-4
sudo apt-get update
sudo apt-get install ffmpeg

 

 

How to use this repository?

0. Notation

We write all the parameters which needs to be specified by a user in the capslock.

1. Download raw data

2. Split dataset

python data_processing/prepare_data.py DATA_DIR
# DATA_DIR = directory to save data such as 'data/'

Note: DATA_DIR is not a directory where the raw data is stored (the folder with data, "dataset" , has to be stored in the root folder of this repo). DATA_DIR is the directory where the postprocessed data should be saved. After this step you don't need to have "dataset" in the root folder any more. You should use the same DATA_DIR in all the following scripts.

After this command:

3. Convert the dataset into vectors

python data_processing/create_vector.py DATA_DIR N_CONTEXT
# N_CONTEXT = number of context, in our experiments was set to '60'
# (this means 30 steps backwards and forwards)

Note: if you change the N_CONTEXT value - you need to update it in the train.py script.

(You are likely to get a warning like this "WARNING:root:frame length (5513) is greater than FFT size (512), frame will be truncated. Increase NFFT to avoid." )

As a result of running this script

If you don't want to customize anything - you can skip reading about steps 4-7 and just use already prepared scripts at the folder example_scripts

 

4. Learn motion representation by AutoEncoder

Create a directory to save training checkpoints such as chkpt/ and use it as CHKPT_DIR parameter.

Learn dataset encoding

python motion_repr_learning/ae/learn_dataset_encoding.py DATA_DIR -chkpt_dir=CHKPT_DIR -layer1_width=DIM

The optimal dimensionality (DIM) in our experiment was 325

Encode dataset

Create DATA_DIR/DIM directory

python motion_repr_learning/ae/encode_dataset.py DATA_DIR -chkpt_dir=CHKPT_DIR -restore=True -pretrain=False -layer1_width=DIM

More information can be found in the folder motion_repr_learning

5. Learn speech-driven gesture generation model

python train.py MODEL_NAME EPOCHS DATA_DIR N_INPUT ENCODE DIM
# MODEL_NAME = hdf5 file name such as 'model_500ep_posvel_60.hdf5'
# EPOCHS = how many epochs do we want to train the model (recommended - 100)
# DATA_DIR = directory with the data (should be same as above)
# N_INPUT = how many dimension does speech data have (default - 26)
# ENCODE = weather we train on the encoded gestures (using proposed model) or on just on the gestures as their are (using baseline model)
# DIM = how many dimension does encoding have (ignored if you don't encode)

6. Predict gesture

python predict.py MODEL_NAME INPUT_SPEECH_FILE OUTPUT_GESTURE_FILE
# Usage example
python predict.py model.hdf5 data/test_inputs/X_test_audio1168.npy data/test_inputs/predict_1168_20fps.txt
# You need to decode the gestures
python motion_repr_learning/ae/decode.py DATA_DIR ENCODED_PREDICTION_FILE DECODED_GESTURE_FILE -restore=True -pretrain=False -layer1_width=DIM -chkpt_dir=CHKPT_DIR -batch_size=8

Note: This can be used in a for loop over all the test sequences. Examples are provided in the example_scripts folder of this directory

# The network produces both coordinates and velocity
# So we need to remove velocities
python helpers/remove_velocity.py -g PATH_TO_GESTURES

7. Quantitative evaluation

Use scripts in the evaluation folder of this directory.

Examples are provided in the example_scripts folder of this repository

8. Qualitative evaluation

Use animation server

 

Citation

If you use this code in your research please cite the paper:

@article{kucherenko2021moving,
  title={Moving fast and slow: Analysis of representations and post-processing in speech-driven automatic gesture generation},
  author={Kucherenko, Taras and Hasegawa, Dai and Kaneko, Naoshi and Henter, Gustav Eje and Kjellstr{\"o}m, Hedvig},
  journal={International Journal of Human–Computer Interaction},
  doi={10.1080/10447318.2021.1883883},
  year={2021}
}

Contact

If you encounter any problems/bugs/issues please contact me on Github or by emailing me at tarask@kth.se for any bug reports/questions/suggestions. I prefer questions and bug reports on Github as that provides visibility to others who might be encountering same issues or who have the same questions.