Accurate detection and segmentation of spontaneous multi-party speech is crucial for a variety of applications, including speech acquisition and recognition, as well as higher-level event recognition. However, the highly sporadic nature of spontaneous speech makes this task difficult. Moreover, multi-party speech contains many overlaps. We propose to attack this problem as a tracking task, using location cues only. In order to best deal with high sporadicity, we propose a novel, generic, short-term clustering algorithm that can track multiple objects for a low computational cost. The proposed approach is online, fully deterministic and can run in real-time. In an application to real meeting data, the algorithm produces high precision speech segmentation.