Um and uh are generally considered to be indicative of dysfluency and uncertainty in speech production. However, analysis of the academic seminar indicates that the distribution of um and uh is not random. In specific well-defined environments um is used to indicate the underlying structure of the talk. Although Swerts (1998) has already suggested that fillers such as um and uh could be treated as discourse markers in Dutch, the notion that such tokens are functioning as discourse markers has not been developed in detail. This paper analyses the role played by um in a series of computer science seminars. Using traditional conversation analysis techniques, the paper focuses on the way in which um indicates structure in the academic seminar by maintaining coherence across bits of talk. It thus argues that in specific well-defined environments um functions as a discourse marker. This paper therefore addresses such issues as the role and function of um in seminar talk, the environments in which it occurs, and its use in indicating the structure of the talk to the listening audience.