| 摘要 |
<p>PROBLEM TO BE SOLVED: To directly generate the animation of the face from voices by converting a face sequence to an optimum sequence based on the control parameters of the face and moving the animation of the face based on the control parameters. SOLUTION: Features in the video of the corners of the mouth are traced and this information is used for training a hidden Markovian model 14 of face, which is composed of a state machine 16 for face and mapping to a constitution 18 of face, while using an entropy. The hidden Markovian model of face is learnt and mapping to the features of audio is evaluated from the state of the face. By this mapping, a hidden Markovian model 22 of voice to be moved by a constitution 20 to audio can be prevented from the audio having a moving state 16 of the face. This model 22 is utilized for predicting the state of the face from a new audio 28. A state sequence 30 of face optimum for supplying an optimum orbit 36 through the constitutive space of the face is supplied by communicating with the mapping 18 of the face.</p> |
