Music Information Retrieval Using Biologically Inspired Techniques

The computational modeling of our perception of music similarity is an intricate, unsolved problem with various practical applications. Many of the current approaches aim at solving it by employing heuristics, such as expert intuition or music theory, which limit their application to narrow contexts, e.g., certain types of music, certain music representations. This dissertation stems from the observation that biological sequences and music items (melodies, chords sequences) share a number of resembling concepts, from the way they are represented, the way they are transformed through time, to the problems most prominent in their respective fields of study (bioinformatics and music information retrieval). Bioinformatics however, has a long history of algorithm development that offers data-driven instead of heuristic-based solutions. Naturally, such solutions can be adapted to the general task of modeling music similarity. More specifically, this dissertation considers the tasks of melodic and chord sequence similarity, intra-family similarity (the specific similarity among related music items), outlier detection (identifying the musical items that do not belong in a group) and polyphony reconstruction (arranging temporally-corrupted polyphonic voices to their original state). At the same time, this work provides a reliable insight into the mechanics of similarity perception by performing a data-driven analysis on the essential concept of music stability. As such, we consider this dissertation to be a fine balance between the occasionally contradicting goals of music information retrieval, problem-solving and knowledge acquisition.