The development of motion capturing devices like Microsoft Kinect poses new challenges in the exploitation of human-motion data for various application fields, such as computer animation, visual surveillance, sports or physical medicine. In such applications, motion segmentation is recognized as one of the most fundamental steps. Existing methods usually segment motions at the level of logical actions, like walking or jumping, to annotate the motion segments by textual descriptions. Although the action-level segmentation is convenient for motion summarization and action retrieval, it does not suit for general action-independent motion retrieval. In this paper, we introduce a novel semantically consistent algorithm for partitioning motions into short and further non-divisible segments. The property of semantic consistency ensures that the start and end of each segment are detected at semantically equivalent phases of movement to support general motion retrieval. The proposed segmentation algorithm first extracts relative distances between particular body parts as motion features. Based on these features, segments are consequently identified by constructing and analyzing a one-dimensional energy curve representing local motion changes. Experiments conducted on real-life motions demonstrate that the algorithm outperforms other relevant approaches in terms of recall and precision with respect to a user-defined ground truth. Moreover, it identifies segments at semantically equivalent phases with the highest accuracy.