Abstract:

For a system to be able to generate real time accompaniment for previously unknown songs, it must predict their harmonic development, i.e., the chords to be played. We claim that such a system must combine long term experience, to identify typical chord sequences (e.g., II-V and II-V-I), with ``on the flight'' adaptation to track recurrent structures (e.g., choruses and refrains) of the particular song being played. We have implemented a prediction system using a neural network model that encompasses prior knowledge about typical chord sequences. The achieved results have been very encouraging, really better than those reported in the literature. However, our predictor could not adapt its behavior according the idiosyncrasies of each song, since on-line learning is nearly impossible in neural networks. In this paper, we propose an extension of our previous work by the inclusion a rule-based sequence tracker, which detects recurrent chord sequences while the song is being performed. We show that this hybrid model, combing a neural network predictor with a rule-based sequence tracker, improves the system's performance.