Analysing sound features (frequency, tempo, gesture) as inputs to probabilistic rules generates visuals that respond to music without one-to-one mapping

Mathieu Le Sourd’s masterclass treats music as a ‘total audiovisual phenomenon’ by extracting sound features — frequency bands, tempo, and other musical gestures — and feeding them not into direct one-to-one parameter mappings but into randomness/probability rules that ‘give life to visual content.’ The distinction matters: a direct mapping (bass amplitude to circle size) is predictable and quickly reads as mechanical; using the analysed feature to weight a probabilistic choice among visual actions produces ‘unexpected outcomes’ that still feel musically driven. This sits between rigid audio-reactivity and pure noise: the audio biases the distribution, chance supplies the surprise. The technique generalises to any audio-visual system where you want reactivity without literalism.

Examples

Tempo drives the probability that a new shape spawns each beat, while a frequency-band threshold biases which of several colour palettes is chosen — the visuals track the music but never repeat identically. Contrast: a spectrum bar directly scaled by FFT magnitude, which is reactive but not generative.

Assessment

Design a mapping where one extracted audio feature (e.g. spectral centroid) weights a probabilistic choice among three visual behaviours. Specify the feature range, the probability weights it produces, and why this reads as musical but non-literal.