Patent US 4,841,575

Visual images of the face of a speaker are processed to extract during a learning sequence a still frame of the image and a set of typical mouth shapes. Encoding of a sequence to be transmitted, recorded etc. is then achieved by matching the changing mouth shapes to those of the set and generating codewords identifying them. Alternatively, the codewords may be generated to accompany real or synthetic speech using a look- up table relating speech parameters to codewords. In a receiver, the still frames and set of mouth shapes are stored and received codewords used to select successive mouth shapes to be incorporated in the still frame.