For a system to be able to generate realtime accompaniment topreviously unknown songs, it must predict their harmonic development,i.e. the chords to be played. We claim that such a system must combinelong-term experience, to identify typical chord sequences (e.g. II–Vand II–V–I), with ‘on-the-fly’ adaptation totrack-recurrent structures (e.g. choruses and refrains) of the particularsong being played. We have implemented a prediction system using a neuralnetwork model that encompasses prior knowledge about typical chordsequences. The results achieved are very encouraging, and rather betterthan those reported in the literature. However, our predictor could notadapt its behaviour to the idiosyncrasies of each song, since onlinelearning is difficult in neural networks. In this paper, we propose anextension to our previous work by the inclusion of a rule-based sequencetracker, which detects recurrent chord sequences while the song is beingperformed. We show that this hybrid model, which combines a neural network predictor with a rule-based sequence tracker, improves thesystem's performance.