Combining Score and Filter Based Models to Predict Tempo Fluctuations in Expressive Music Performances

Tempo variations in classical music are an important means of artistic expression. Fluctuations in tempo can be large and sudden, making applications like automated score following a challenging task. Some of the fluctuations may be predicted from (tempo annotations in) the score, but prediction based only on the score is unlikely to capture the internal coherence of a performance. On the other hand, filtering approaches to tempo prediction (like the Kalman filter) are suited to track gradual changes in tempo, but do not anticipate sudden changes. To combine the advantages of both approaches, we propose a method that incorporates score based tempo predictions into a Kalman filter model of performance tempo. We show that the combined model performs better than the filter model alone.