Our research showcases a new, 3-D, photo-real talking head with freely controlled head motions and facial expressions. It extends our prior, high-quality, 2-D, photo-real talking head to 3-D. First, we apply a 2-D-to-3-D reconstruction algorithm frame by frame on a 2-D video to construct a 3-D training database. In training, super-feature vectors consisting of 3-D geometry, texture, and speech are formed to train a statistical, multistreamed, Hidden Markov Model (HMM). The HMM then is used to synthesize both the trajectories of geometric animation and dynamic texture. The 3-D talking head can be animated by the geometric trajectory, while the facial expressions and articulator movements are rendered with dynamic texture sequences. Head motions and facial expression also can be separately controlled by manipulating corresponding parameters. The new 3-D talking head has many useful applications, such as voice agents, telepresence, gaming, and speech-to-speech translation.
We propose an HMM trajectory-guided, real image sample concatenation approach to photo-real talking head synthesis. It renders a smooth and natural video of articulators in sync with given speech signals. With an audio/video footage as short as 20 minutes from a speaker, the proposed system can synthesize a highly photo-real video in sync with the given speech signals.
No comments:
Post a Comment