Disclosed herein includes a system, a method, and a non-transitory computer readable medium for rendering a three-dimensional (3D) model of an avatar according to an audio stream including a vocal output of a person and image data capturing a face of the person. In one aspect, phonemes of the vocal output are predicted according to the audio stream, and the predicted phonemes of the vocal output are translated into visemes. In one aspect, a plurality of blendshapes and corresponding weights are determined, according to the corresponding image data of the face, to form the 3D model of the avatar of the person. The visemes may be combined with the 3D model of the avatar to form a 3D representation of the avatar, by synchronizing the visemes with the 3D model of the avatar in time.