Version Detection for Historical Musical Automata.

Musical automata were very popular in European homes in the pre-phonograph era, but have attracted little attention in academic research. Motivated by a specific application need, this paper proposes a first approach to the automatic detection of versions of the same piece of music played by different automata. Due to the characteristics of the instruments as well as the themes played, this task deviates considerably from cover version detection in modern pop and rock music. We therefore introduce an enhanced audio matching and comparison algorithm with two main features: (1) a new alignment cost measure – Off-Diagonal Cost – based on the Hough transform; and (2) a split-and-merge strategy that compensates for major structural differences between different versions. The system was evaluated on a test set comprising 89 recordings of historical musical automata. Results show that the new algorithm performs significantly better than the reference system based on Dynamic Time Warping and chroma features without the above-mentioned new features, and that it may work well enough to be practically useful for the intended application.