Specifying affect and emotion for expressive speech synthesis

Speech synthesis is not necessarily synonymous with text-to-speech. This paper describes a prototype talking machine that produces synthesised speech from a combination of speaker, language, speakingstyle, and content information, using icon-based input. The paper addresses the problems of specifying the text-content and output realisation of a conversational utterance from a combination of conceptual icons, in conjunction with language and speaker information. It concludes that in order to specify the speech content (i.e., both text details and speaking-style) adequately, selection options for speaker-commitment and speaker-listener relations will be required. The paper closes with a description of a constraint-based method for selection of affect-marked speech samples for concatenative speech synthesis.