Proceedings of the 6th Sound and Music Computing Conference (SMC 2009)
We present a performance rendering system that uses a probabilistic
network to model dependencies between score and
performance. The score context of a note is used to predict
the corresponding performance characteristics. Two extensions
to the system are presented, which aim at incorporating
the current performance context into the prediction,
which should result in more stable and consistent predictions.
In particular we generalise the Viterbi-algorithm,
which works on discrete-state Hidden Markov Models, to
continuous distributions and use it to calculate the overall
most probable sequence of performance predictions. The
algorithms are evaluated and compared on two very large
data-sets of human piano performances: 13 complete Mozart
Sonatas and the complete works for solo piano by Chopin.