Robust Polyphonic Midi Score Following with Hidden Markov Models

Although modern audio score following systems work very well with low polyphony performances, they are still too imprecise with highly polyphonic instruments such as the piano, or the guitar. On the other hand, these instruments can easily output Midi information which shows that our work on robust Midi score following is still needed. We propose an adaptation to Midi input of our HMM-based stochastic audio score follower, focusing the attention on the piano as our test instrument. The acoustic salience of the Midi notes is modeled by an amplitude envelope, taking into account the sustain pedal, from which note match and attack probabilities are derived. Tests with a complex piano piece played with many errors showed a very high robustness.