By Eric Brochu
Computers cannot, of course, appreciate the emotional qualities of music. But can they describe music with emotional adjectives that match what a human might expect? I have implemented a system, MILQ, to explore this hypothesis.
Using a large data set, a selected set of labels (including both genre and style labels like Indie Rock and tone labels like Cathartic), and proven feature extraction techniques, I constructed a set of nonlinear logistic discriminative networks using logistic discriminative techniques, which compute marginal probabilities for each label. Such techniques and other Machine Learning methods have been used before to construct genre classifiers and my model works well for those.
Estimating the probabilities of the tonal labels is much more difficult, however, as these can have a very strong cultural component, as well as an acoustical one. Therefore, I add a second Bayesian network stage. This uses a set of labels from the logistic network as the priors for the belief of each label, treating the labels as nodes in a directed, loopy Bayesian network. Using a modified version of loopy belief propagation, the posterior of each label is conditioned on its neighbours. Evaluations suggest that the model can be used with a fair degree of accuracy to assign tone-based adjectives to music.