Score Following Using Spectral Analysis and Hidden Markov Models

This paper presents an approach to score following. The real-time alignment of a performance with a score is obtained through the use of a hidden Markov model. The model works on two levels. The lower level compares the features of the incoming signal with the expected ones. Groups of states of the lower level are embedded in states at the higher level, which are used to model the performance by taking into account the possible errors a performer may make. The performer's position on the score is computed through a decoding technique alternative to classic Viterbi decoding. A novel technique for the training of hidden Markov models is proposed.