Listener disagreements

  • Published as “Audio properties of perceived boundaries in music” in Music Theory Online.
  • Article link, BIB.
  • Note: the online article includes figures, sound files, video illustrations, and score examples!

For theories of music cognition, positing an ideal listener is a useful abstraction, and models for predicting analyses are usually deterministic. But when we look at the actual musical analyses of listeners, we find many disagreements. In order to predict these, a model should certainly incorporate listener-dependent factors such as “level of musical training” or “number of times a piece is heard.” These and others can be studied systematically in psychology experiments, but perhaps it is too optimistic to hope that every single listener attribute could be isolated and measured in this way in order to predict all disagreements. Instead, what can we learn from a small case study that attempts to trace exactly how (a few, very specific) listener disagreements arise?

Isaac Schankler and I conducted a case study of ourselves that revolved around three of his improvisations with Mimi. (Mimi, or Multi-Modal Interaction for Musical Improvisation, is software that listens to a performing partner and then improvises along with them by imitating what it has heard and remixing it.) Isaac recorded three performances, which we each analyzed. The figure below illustrates our analyses for one of these songs, but we also made notes for the justifications for each section and boundary. We compared our analyses, and through our conversation and introspection tried to trace back the many surface disagreements to their deeper origins.

Figure 1

We decided that the following was a plausible chain of causation: first, differences in expectation, followed by differences in information, the perception of a piece’s opening moments, and finally their attention. Working backwards:

  1. At the surface level, our disagreements seemed directly attributable to paying attention to different features. That is, we used our descriptions of musical features to justify our boundaries and labels: e.g., one might say “I hear this as a boundary because I attend to the change in register.”
  2. Some of these disagreements appeared, in turn, to be directly attributable to disagreements about opening moments: if we disagreed about how to characterize the opening theme of the piece, we would naturally disagree about how to characterize later parts of the piece that reuse this material. (This is perhaps the result of local expectations established in the opening moments.)
  3. Differences in information about the piece led us to focus on different things: Isaac, as the performer, had different knowledge about the creation of the piece than I had, and this affected the analyses.
  4. Finally, we noted differences in analytical expectations (not the local expectations mentioned earlier), which I like to think of as prejudices about what an “analysis” should look like. For example, one might believe (as I seemed to) that analyses should have roughly equal-sized parts, or at least 3 and no more than 7 parts, or that ABC and AA′A′′ are “uglier” analyses than ABA.

This study contrasted with the rest of the work that was included in my PhD thesis, since it engaged more with the field of music theory than with music psychology or music information retrieval. Still, it provided a rich set of hypotheses to follow up, and guided much of the rest of the thesis.