Research On Voice-driven 3D Lip Animation Algorithm

Voice-driven 3D lip animation algorithm belongs to the interdisciplinary of voice signal processing and 3D animation. It can be applied in various of 3D animation areas which need to synchronize the voice and lip. For example,3D animation,3D game, virtual anchor, teaching video, etc. At present, researches about voice-driven lip animation algorithm at home an aboard are not so many, lip animations making is usually manual, which wastes lots of time and manpower. Therefore, researching to the voice-driven 3D lip animation algorithm has a certain economic significance and application value.In the voice-driven 3D lip animation algorithm, the reality of lip animation is directly affected by the performance of the mapping from voice to lip. At present, the available voice-driven 3D lip animation algorithms have several difficulties and problems:(1) The pronouncing regularities of phonemes in different languages are different, which make the mapping relationships from voices to lip animations be hard to generate;(2) When using back-propagation neural network to map the voice characteristic coefficients to lip, the speed and precision are usually limited by the quantity of training samples and the network structures;(3) The file formats of 3D face models are varied but there is no united standard to lip animation, so the commonality is not enough.In this paper, based on the available voice-driven 3D lip animation algorithms, the following improvements have been done:Firstly, we analyze the pronouncing regularities of Mandarin Chinese and English and try to generalize them by the International Phonetic Alphabet, then record a phoneme library.Secondly, we try to replace the BPNN to GMM and DAG-SVM to classify the phoneme, and modify the DAG-SVM to improve its performance.Finally, we use the 3D morphing animation technique in DirectX to achieve a general and virtual 3D lip animation and program a GUI which combined with the improved classifying algorithms.The experimental results show that the performance of method in this paper is better and meets the expected requirement.