Method and system for selection of acoustic elements at run-time for speech synthesis