Statistical version of Foote's "checkerboard" novelty and structure detector
Foote's novelty detector is a clever little piece of engineering. A high novelty detection function indicates a significant change.
It's so clever because it actually doesn't just measure differences between to adjacent stretches of audio (as my Structural Change feature in its simplest form), but it also takes into account the homogeneity of the parts to the left and the right.
However, as far as I know, it does so using a simple DSP convolutive approach.
A person with a good understanding of statistics could revisit the checkerboard function, and
- try to understand the statistical patterns behind it
- develop a new version based on mathematical statistics that measures how "significantly" two adjacent parts of audio (or other timeseries differ from each other.
If more time and enthusiasm remains:
- Vamp implementation (Vamp plugins run in Sonic Visualiser, which is awesome, and they run fast)
- interpretation of the significant difference between two adjacent parts as complexity -- "does that make sense?" would be a question to answer.