Lijuan Wang, Yi-Jian Wu, Xiaodan Zhuang, and Frank Soong
22 May 2011
In this paper, we propose a minimum generation error (MGE) training method to refine the audio-visual HMM to improve visual speech trajectory synthesis. Compared with the traditional maximum likelihood (ML) estimation, the proposed MGE training explicitly optimizes the quality of generated visual speech trajectory, where the audio-visual HMM modeling is jointly refined by using a heuristic method to find the optimal state alignment and a probabilistic descent algorithm to optimize the model parameters under the MGE criterion. In objective evaluation, compared with the ML-based method, the proposed MGE-based method achieves consistent improvement in the mean square error reduction, correlation increase, and recovery of global variance. It also improves the naturalness and audio-visual consistency perceptually in the subjective test.
In ICASSP 2011