Expectation networks have been proposed as a computationally simple method
for learning tonal expectations associated with individual scale degrees
(Verosky, 2019). Using principles of activation and decay, expectation
networks infer the expectation of encountering a given event type followed
in the near (but not necessarily immediate) future by any other event type.
The current work outlines how these learned expectations can be used to
predict melody continuations and tests the predictions against listener
responses to a melodic cloze task previously used to compare two other
models of melodic expectation, IDyOM and Temperley’s Gaussian model
(Morgan, Fogel, Nair, & Patel, 2019; Pearce, 2005; Temperley, 2008).
Results of multinomial logistic regression indicate that all three models
account for unique variance in listener predictions, with coefficient
estimates highest for expectation networks. Despite expectation networks’
computational simplicity relative to IDyOM, direct comparisons between
IDyOM and expectation networks similarly yielded higher coefficient
estimates for the latter. Although all three models are limited in their
ability to incorporate global, hierarchical information about pitch
structure, expectation networks seem to benefit from a tendency to predict
all three notes of the tonic triad at cadence points while ranking the
tonic as the most probable continuation. Our findings suggest that
generalized scale degree expectations as captured by expectation networks,
stereotypical pitch sequences as captured by IDyOM, and immediate
intervallic expectations as captured by Temperley’s model all factor into
real-time listener predictions to varying extents, highlighting several
possible areas for future work.