Probabilistic sequence alignment methods for on-line score following of music performances.

Abstract

Existing systems for automated accompaniment of musical performers assume every performance note is specified in advance in a written score. Musical styles such as Jazz and Blues often use scores that specify only main harmonies, overall structure, and important melodies. The exact realization of the score varies greatly from performance to performance. For such music, a system must be able to handle performances that start anywhere in the form of the score, repeat or skip sections any number of times, play at unexpected tempi, and contain improvisations over the harmonic structure of the piece, without necessarily performing the written melody (such as during a Jazz solo). This dissertation describes the techniques I developed for following semi-improvised music: extensions of existing string matching and HMM score following techniques; improved score representation; a new method of integrating beat tracking and score following; a new chord transcription system that infers chords from varied surface instantiations of notes; and new training methods that let the system train its error models one set of scores and performances and thereafter follow performances of pieces on which the system has not been trained. Experimental results in the paper show that the system is able to correctly label chords in improvised performances roughly 75% of the time (the exact value depends on the test corpus). Training the system to a performer's individual style lets the system correctly align the chord sequence inferred from the performance to the score, with a resolution of one or two beats (depending on the test set). This result can be achieved by training the system with as few as five example score-performance pairs. Once trained, the system can follow new performances and scores without additional training, achieving results as good as those achieved for the training set.