Modular synthesizer · Evolutionary search · Runs in your browser

Pick the sound
you like better.

Auracle plays you two patches, learns which one you preferred, and starts making more of it. The panel here is real. Press play and choose.

Play the instrument ▸ Read the guide

Free, no account, nothing to install. Your patches and your taste model stay in your browser.

A —

B —

What it thinks you like

Press 1 and 2, then pick the one you prefer.

A four-knob miniature. The real instrument measures forty coordinates and fits a posterior over up to five taste lenses.

The problem

Sound design makes you choose between fast and deep.

Explore

Presets and randomizers. Fast, and shallow. You audition until something works, and nothing accumulates. Tomorrow you start over.

Construct

Patch from scratch. Deep, and slow. You get exactly what you can already imagine, at the speed you can build it.

Infer

Auracle treats it as inference. It generates, you judge, it fits a model of your taste and searches with it. The work you put in on Tuesday is still there on Friday.

The instrument

Three views, one loop.

The PLAY view: a patch bank on the left, an eight-module rack wired with green audio cables and amber modulation cables, the node bank catalogue on the right, and a keyboard docked along the bottom.

PLAY. The whole rack, with knobs at true positions in musical units. Edit it and lock parts of it while it runs live under your hands.

How it learns

A model that can be wrong in public.

Your picks condition one latent quantity: how much you would like a patch. Taste is a maximum over lenses rather than a single direction, so you are allowed to love dark drones and bright plucks without the model averaging them into a preference for neither.

u(x) = max_k θ_k · φ(x)

Search then proposes toward you: the fitted taste reshapes the grammar's own proposal weights, not just the score.

π_β(x) ∝ p_grammar(x) · e^βu(x)

And every duel is forecast before you answer it, then scored against a proper scoring rule. So the instrument can tell you when it does not know yet. Early on, it does.

Read how it works ▸

A coefficient plot: named perceptual coordinates down the left, horizontal amber bars extending left and right of a centre line, each with a thinner whisker showing the credible interval. — Every coefficient with its credible interval. A long bar whose whisker crosses the centre line is a guess, and the display says so.

How it fits together

Two loops, running at different speeds.

The machine evaluates thousands of candidates against what it has learned about you, silently, and surfaces a curated few. Interactive-evolution tools usually do the opposite: they ask you to rate a whole population, every generation, until you quit.

Green flows continuously with no human in it. Amber moves at the pace you answer questions, and carries your taste back into how the search proposes, not only how it scores.

The module palette

Forty-one modules, from source to output.

sourcesA wavetable, a physically-modelled pluck, a formant oscillator that speaks vowels, supersaw, VCO, noise
shapeWavefolder, distortion in three drive modes, bitcrush
filterSVF, diode ladder, parametric EQ, vocoder
spaceDelay, reverb, granular, pitch shift
motionChorus, phaser, flanger, tremolo, vibrato
dynamicsCompressor, ducker, noise gate, each with a control input you can sidechain into
combineCrossfade and ring modulation, merging two chains into one
modulationLFO, envelope, sample-and-hold, slew, follower, Euclidean. These chain: s&h rand → quantize → slew before it ever reaches a cutoff

Every knob is an address in the genome, so turning one edits the patch the search is working on. Lock a knob and refinement leaves it alone.

Start

Pick the sound
you like better.

What it thinks you like