voca

Why we included this project

For anyone building digital humans, VOCA does something concrete: it takes a plain audio recording and a static character mesh and outputs a frame-by-frame sequence of animated face meshes that follow the spoken words, with lip motion and expressions that vary by speaker identity. Because the output is built on the FLAME head model, you are not locked into one rigid result; the animation can be re-edited afterward, for example to adjust head pose or add eye blinks. As the official implementation of the CVPR 2019 paper, complete with a working demo script, it is a solid reference for teams experimenting with speech-driven facial animation or studying how such systems are architected. Just be ready for the setup: it targets Python 3.6 and TensorFlow 1.14, and you will need to download the trained models from external sources.