Ground truth
The human-arranged chord progression.
Loads when you press play.
Precomputed listening demo
Hear surprise take shape.
Compare precomputed SurpriseNet outputs across six user-controlled contours and three melody excerpts.
Harmony, guided by expectation.
SurpriseNet conditions a variational autoencoder with a contour derived from chord-transition probabilities.
- Melody excerpts
- 3
- Surprise contours
- 6
- Audio-visual samples
- 39
One melody, six surprise shapes.
Pick an excerpt and contour, then compare the source progression with two model variants. Everything here is served as static media; this page does not run model inference.
Choose an excerpt
The selection stays shareable in the page URL.
Shape the surprise contour
The graph controls how chord surprise changes over time.
SurpriseNet
Generated with the selected contour.
Loads when you press play.
Weighted SurpriseNet
Generated with additional surprise weighting.
Loads when you press play.
What to listen for.
The contour represents harmonic unexpectedness over time. It guides chord selection without changing the input melody.
- Start with the source. Learn the original harmonic rhythm and phrase boundary.
- Compare the same contour. Listen for where either model delays or releases tension.
- Switch the shape. Notice how the harmonic path changes while the melody stays fixed.
Paper and source.
The model, training code, evaluation utilities, and this explorer now live in one repository history.
“SurpriseNet: Melody Harmonization Conditioning on User-controlled Surprise Contours”
Yi-Wei Chen, Hung-Shin Lee, Yen-Hsing Chen, and Hsin-Min Wang. ISMIR 2021.