Aligning Semi-Improvised Music Audio with Its Lead Sheet

Existing audio-score alignment methods assume that the audio performance is faithful to a fully-notated MIDI score. For semi-improvised music (e.g. jazz), this assumption is strongly violated. In this paper, we address the problem of aligning semi-improvised music audio with a lead sheet. Ourapproachdoesnotrequirepriortrainingonperformances of the lead sheet to be aligned. We start by analyzing the problem and propose to represent the lead sheet as a MIDI file together with a structural information file. Then we propose a dynamic-programming-based system to align the chromagram representations of the audio performance and the MIDI score. Techniques are proposed to address the chromagramscaling, keytranspositionandstructuralchange (e.g. a performer unexpectedly repeats a section) problems. We test our system on 3 jazz lead sheets. For each sheet we align a set of solo piano performances and a set of fullband commercial recordings with different instrumentation and styles. Results show that our system achieves promising results on some highly improvised music.