Keiji Hirata/Satoshi Tojo/Alan Marsden/Masatoshi Hamanaka

Music Analyzer that Can Handle Context Dependency

Appropriate handling of context dependency is crucial in music analysis. For example, each occurrence of a repeated phrase may have a different musical meaning. The musical meaning derived from a phrase can be represented by a tree, with different tree structures representing different meanings.  We propose a cognitive model of musical context dependency in which the key ideas are tension-relaxation grammar, the separation of bottom-up and top-down processes, and expectation-based parsing. A tension-relaxation grammar may work effectively in discovering distant relationships. Parsing with a tension-relaxation grammar is used to generate a global normative form, which contains the information of context dependency. The model extracts local structures in a bottom-up manner while identifying a global normative form within a piece of music corresponds to the top-down analysis. Then, by unifying the local structures with the global normative form, we obtain the whole consistent tree structures reflecting context dependency. We propose that, every time one listens to a piece, one's expectations are based on the most recent listening experiences and may elaborate and/or revise previous expectation. Hence, we consider the model with two input channels, a score and an expectation; every time a piece of music is input to the model with an expectation previously obtained in the circular manner, an expectation is to be elaborated and revised, and accordingly a more valid tree structure is generated. Through this circular process, the output structure is gradually accommodated with context dependency and converges to a valid tree structure.