Music Complexity Measures predicting the Listening Experience

Madsen, S. T. and Widmer, G.

This paper presents a computational model calculating the level of
attraction a voice in a score is likely to require at a given
time. The model is based on a music information complexity measure.

The assumption is that we continuously while listening tend to focus
on the most complex (least repetitive) voice, experiencing this as
foreground. This is reflected in the model - based on the complexity
measured in all voices over a short time window, the model predicts
the most complex voice to be the most interesting in that time window.

We discuss how to measure music complexity of pitch and rhythm, and
examine which factors are the most important.

We present a computational model measuring the information of each
voice in a score during a short time window. For each window we can
plot the current complexity of each voice. Moving the window, we get
curves reflecting how active each voice is over time. A curve staying
on top for a period is predicted to attract most attention in that
period. A graphical interface presents the score with emphasized
predicted notes while playing the score.

We tend to perceive complex structures in the most structured way
possible. Thus measuring complexity is in our model measured in terms
of (in)predictability. We use entropy of pitches, intervals and
durations of notes as measures of complexity. Compression algorithms
will also be discussed.

One way to automatically test the correctness of our model is by
trying to predict the melody notes in a score with an annotated
melody. A music student has prepared Haydn and Mozart string quartets
for testing purposes. The melody in this type of music often coincides
with what we prefer to listen to.

We test which factors (pitches, intervals, durations) that have the
greatest impact on melody note prediction.

For the kind of music where a melody is quite clear we present a model
fairly competent in correctly predicting the melody notes (supports
the assumption). However if the melody is of a more static character,
the most interesting accompaniment line will be expected preferred
listening to.