动态视位模型及其参数估计

A Dynamic Viseme Model and Parameter Estimation

视觉信息可以加强人们对语音的理解,但如何在可视语音合成中生成逼真自然的口形是个复杂的问题.在深入地研究了人们说话过程中口形变化的规律后,提出了一个基于控制函数混合的动态语音视位模型.并针对汉语发音的特点给出了一种系统的从训练数据学习模型参数的方法,这比依靠主观经验人为指定模型参数更为可靠.实验结果表明,视位模型和通过训练数据学习得到的模型参数可以有效地描述汉语发音过程中口形的变化过程.

Visual information can improve speech perception. But how to synthesis the realistic mouth shape is a complex problem. After studying the rule of lip movement in speaking, a dominance blending dynamic viseme model for visual speech synthesis is proposed in this paper. Furthermore, considering the characteristic of Chinese speech, a systemic learning method is given to learn the model parameters from training data, which is more reliable than desire parameters according to subjective experience. Experimental results show that the dynamic viseme model and learning method are effective.