This paper proposes a continuous speech driven photo realistic visual speech synthesis approach based on an articulatory dynamic Bayesian network model (AF_AVDBN) with constrained asynchrony. In the training of the AF_AVDBN model, the perceptual linear prediction (PLP) features and YUV features are extracted as acoustic and visual features respectively. Given an input speech and the trained AF_AVDBN parameters, an EM-based algorithm is deduced to learn the optimal YUV features, which are then used, together with the compensated high frequency components, to synthesize the mouth animation corresponding to the input speech. In the experiments, mouth animations are synthesized for 80 connected digit speech sentences. Both qualitative and quantitative evaluation results show that the proposed method is capable of synthesizing more natural, clear and accurate mouth animations than those from the state asynchronous DBN model (S_A_DBN).
展开▼
