The role of an interactive music system (IMS) is to accom-pany musicians during live performances, like a real musi-cian. It reacts in realtime to audio signals from musicians, according to a timed specification called mixed score, writ-ten in a domain specific language. Such goals imply strong requirements of temporal reliability and robustness to un-foreseen errors in input, yet not so much studied in the computer music community. We present the application of model-based testing tech-niques and tools to a state-of-the-art IMS, including the fol-lowing tasks: generation of relevant input data for testing (including timing values) following coverage criteria, com-putation of the corresponding expected output, according to the semantics of a given mixed score, black-box execution of the test data and verdict. Our method is based on formal models compiled directly from mixed scores, and passed, after conversion to timed automata, to the model-checker Uppaal. This fully automatic approach has been applied to real mixed scores used in concerts and the results obtained have permitted to identify bugs in the target IMS.