This report presents the EvoMove musical companion that generates music according to performer moves. It uses wireless sensors (accelerometers, gyroscopes and magnetometers) to acquire a continuous stream of information about the motions of the performer. This stream is analyzed on-the-fly to exhibit categories of similar moves and to identify the category of new incoming moves. The main feature is that these categories are not predefined but are determined in a dynamic way by a new subspace clustering algorithm. This algorithm, called SubCMedians, stems from our previous work on the Chameleoclust+ algorithm (Deliver-able 5.1) and is targeted towards more efficient processing. SubCMedians is a median-based subspace clustering algorithm using a weight-based hill climbing strategy and a stochastic local exploration step. It is shown to exhibit satisfactory quality clusters when compared to well-established clustering paradigms, while allowing for fast on-the-fly handling of the sensor data. The music generation itself relies on a tiling over time of audio samples, where each sample is triggered according to the moves detected in the data stream coming from the sensors.