eCIRCUS: building voices for autonomous speaking agents.

This paper describes our work integrating automatic speech generation into a virtual environment where autonomous agents are enabled to interact by natural spoken language. The application intents to address bullying problems for children aged 9-12 in the UK and Germany by presenting improvised dramas and by asking the user to act as an “invisible friend” of the victimised character. As we are addressing an elementary school environment one specification of the resulting voice was building agecorresponding young school kids voices. The second specification addresses building a low-resource speech generation system which is capable to run on older school computers but is still fast enough in response time to guaranty a fluent conversation between the agents. Third requirement was integrating the speech-module with the agents. We focus on the speech generation system itself, pointing out possible implementation issues in building non-controlled speech interaction in virtual environments Furthermore we describe the problems arising in building unit-selection based child’s' voice TTS and shows alternative methods to child’s voice recording by deploying voice transformation methods.