Czech Audio-Visual Speech Synthesis with an HMM-trained Speech Database and Enhanced Coarticulation

The task of visual speech synthesis is usually solved by concatenation of basic speech units selected from a visual speech database. Acoustical part is carried out separately using similar method. There are two main problems in this process. The first problem is a design of a database, that means estimation of the database parameters for all basic speech units. Second problem is a way how to concatenate selected basic phonetic units so as to eliminate the coarticulation effect. Both problems are aimed in our work, resulting in the Czech audio-visual speech synthesizer. We use HMM training process instead of some form of averaging for obtaining statistically best parameters for all basic phonetic units. For solution of a coarticulation effect we use the method of dominance functions. This paper presents the new Czech audio-visual synthesis. The designed talking head provides now intelligible speech and is ready for further work on the adopting real head look, which will incorporate model adaptation to a real head shapes and applying a texture. Key-Words: audio-visual speech synthesis, hidden Markov models, coarticulation.