Grammar Induction for Musical Melodies

Universiy of Southern California, Spring 2007

in ISE575/EE675/CSCI575/PSYCH675

by Reid Swanson (2007)

Analysis

Human

Even though we all possess these remarkable abilities it is quite another thing to understand the mechanisms that govern the structure and meaning of an utterance or piece of music. There are several common approaches for analyzing the structure of music. One approach common earlier in the development of Western music were written treatises which explicitly defined the adherence of a piece to a set of rules. The information gained from a treatise is usually enough to determine if a piece belongs in a particular genre. It is also usually detailed enough so that a composer, even an automated one, can generate music in that style. However, this type of prescriptive knowledge does not capture all the intricacies that musicians actually play, nor does it have any explanatory power of why they play it the way they do. More descriptive approaches such as Psychological and task can provide a richer cognitive model for why music takes on the structures that it does. For example in Roozendal (1993) set up an experiment that examined a composer's process of writing new material. Through this work tasks that composer's typically had trouble with were highlighted, such as remembering complex themes. Although this work was primarily used to direct research in designing composition software the weaknesses and strengths demonstrated in the experiment also suggest a deeper reason why music has developed the way it has.

Statistical

Statistical analysis is another approach that has become popular especially since the high availability of computers. These approaches range from simple frequency counts to the use of complex data structures such as Hidden Markov Models and other complex statistical models. Statistical models provide a much better descriptive account of music allowing one to easily see the differences across periods or genres through changes in pitch distribution or tempo changes, for example. However, most statistical analysis lacks any substantial information about the overall structure of the music. Although the analysis can tell you that a piece is primarily composed of As, Bs and Cs it can not tell you that the beginning of the piece was primarily As the middle Bs and Cs and the ending a return to As.