SurpriseNet

Yi-Wei Chen, Hung-Shin Lee, Yen-Hsing Chen, Hsin-Min Wang, "SurpriseNet: Melody Harmonization Conditioning on User-controlled Surprise Contours," in Proc. ISMIR, 2021.

In this study, we proposed SurpriseNet, which is based on a conditional VAE model and combines a surprise contour from the transition probability in a Markov chain, to achieve a user-controlled melody harmonization task. From the generated samples, we observed that the model could accurately generate various harmonic chord progressions according to the given surprise contours.

The experiments were conducted on the Hooktheory Lead Sheet Dataset, which contains rich intonation data, such as Roman, symbol, secondary, and mode data.

Generated Samples

We gave out the following six types (a) - (f) of function as the surprise contours to the model. The graphs indicate the surprise level of chords varing with time. Here we selected some generated samples and demonstrated below: