This work uses combined speech input and output to converse with a user seeking to control a computer, or to access information from it. Conversational techniques allow the computer gracefully to limit the vocabulary that is likely to be spoken, facilitating speech recognition. Discourse techniques aid error detection and correction. Domain knowledge and learning about user preferences allow for determining a user�s needs more efficiently. Listening for various forms of back-channel response from the listener permits a talking computer to gauge user interest better, and determine at what level of detail to describe the requested information.