Lip synchronization using linear predictive analysis

Linear predictive analysis is a widely used technique for speech analysis and encoding. The authors discuss the issues involved in its application to phoneme extraction and lip synchronization. The LP analysis results in a set of reflection coefficients that are closely related to the vocal tract shape. Since the vocal tract shape can be correlated with the phoneme being spoken, LP analysis can be directly applied to phoneme extraction. We use neural networks to train and classify the reflection coefficients into a set of vowels. In addition, average energy is used to take care of vowel-vowel and vowel-consonant transitions, whereas the zero crossing information is used to detect the presence of fricatives. We directly apply the extracted phoneme information to our synthetic 3D face model. The proposed method is fast, easy to implement, and adequate for real time speech animation. As the method does not rely on language structure or speech recognition, it is language independent. Moreover, the method is speaker independent. It can be applied to lip synchronization for entertainment applications and avatar animation in virtual environments