Prediction of notes from vocal time series produced by singing voice

Aiming at optimal prediction of the correct note corresponding to a vocal time series we trained a classification algorithm on the basis of parts of interpretations of Tochter Zion (Händel) and tested the algorithm on the remaining parts. As classification algorithm we use a radial basis function support vector machine together with a "Hidden Markov" method as a dynamisation mechanism and some smoothing for categorical data. With this we were able to obtain a minimum of 5% average classification error and a maximum of 26% on data from an experiment with 16 singers.