We present a machine-learning-based adaptive digital music interface for stroke rehabilitation. The interface uses smartphone inertial sensors to capture upper limb movement, and for each user, it trains a model for classifying the user’s movements. The resulting classification model is then used to map different movements to sounds. The interface provides a simple user interface that allows the participants to train the classification model and choose the associated real-time feedback from a variety of percussion sounds. High accuracy results suggest that the proposed adaptive digital music interface can be used as an engaging and motivating tool in music-based interventions with stroke patients.