• Requirements
• Data Extraction
• Data Preprocessing
• Training Different Models
• Pretrained Model
• How to run an example
• Citations
• FAQs

You may install the requirements by running the following command

sudo pip3 install -r requirements.txt

The project is built for python 3.5 and above. The other libraries are listed below

• OpenCV (sudo pip3 install opencv-contrib-python)
• Dlib (sudo pip3 install dlib) with this file unzipped in the data folder
• Python Speech Features (sudo pip3 install python-speech-features)

For a complete list refer to requirements.txt file.

I used the tools below to extract and manipulate the data:

• ffmpeg (sudo apt-get install ffmpeg)
• YouTube-dl

I extracted the data from youtube using youtube-dl. It's perhaps the best downloader for youtube on linux. Commands for extracting particular streams are given below.

• Subtitle Extraction

youtube-dl --sub-lang en --skip-download --write-sub --output '~/obamanet/data/captions/%(autonumber)s.%(ext)s' --batch-file ~/obamanet/data/obama_addresses.txt --ignore-config

• Video Extraction

youtube-dl --batch-file ~/obamanet/data/obama_addresses.txt -o '~/obamanet/data/videos/%(autonumber)s.%(ext)s' -f "best[height=720]" --autonumber-start 1

(Videos not available in 720p: 165)

• Video to Audio Conversion

python3 vid2wav.py

• Video to Images

ffmpeg -i 00001.mp4 -r 1/5 -vf scale=-1:720 images/00001-$filename%05d.bmp

To convert from BMP format to JPG format, use the following in the directory

mogrify -format jpg *.bmp
rm -rf *.bmp

Copy the patched images into folder a and the cropped images to folder b

python3 tools/process.py --input_dir a --b_dir b --operation combine --output_dir c
python3 tools/split.py --dir c

You may use this pretrained model or train pix2pix from scratch using this dataset. Unzip the dataset into the pix2pix main directory.

python3 pix2pix.py --mode train --output_dir output --max_epochs 200 --input_dir c/train/ --which_direction AtoB

To run the pix2pix trained model

python3 pix2pix.py --mode test --output_dir test_out/ --input_dir c_test/ --checkpoint output/

To convert images to video

ffmpeg -r 30 -f image2 -s 256x256 -i %d-targets.png -vcodec libx264 -crf 25 ../targets.mp4

Link to the pretrained model and a subset of the data is here - Link

Download and extract the checkpoints and the data folders into the repository. The file structure should look as shown below.

obamanet
|
└─ data
|   | audios
|   | a2key_data
|   ...
|
└─ checkpoints
|   | output
|   | model.h5
|   ...
└─ train.py
└─ run.py
└─ run.sh
...

Run the following commands

bash run.sh <relative_path_to_audio_wav_file>

Example:

bash run.sh data/audios/karan.wav

Feel free to experiment with different voices. However, the result will depend on how close your voice is to the subject we trained on.

If you use this code for your research, please cite the paper this code is based on: ObamaNet: Photo-realistic lip-sync from text and also the amazing repository of pix2pix by affinelayer.

Cite as arXiv:1801.01442v1 [cs.CV]

• What is target delay for RNN/LSTM?
• Keras implementation of time delayed LSTM?
• Another link for the above