Photorealistic 2D audiovisual text-to-speech synthesis using active appearance models
Wesley Mattheyses, Werner Verhelst · 2010
Audiovisual text-to-speech (AVTTS) synthesizers are capable of generating a synthetic audiovisual speech signal based on an input text. A possible approach to achieve this is model-based synthesis, where the talking head consists of a 3D model of which the polygons are varied in accordance with the target speech. In contrast with these rule-based systems, data-driven synthesizers create the target speech by reusing pre-recorded natural speech samples. The system we developed at the Vrije Universiteit Brussel is a data-based 2D photorealistic synthesizer that is able to create a synthetic visual speech signal that is similar to standard 'newsreader-style' television recordings.