From measurement
to music.

A photograph, an audio clip, or a short video, depending on the modality — an iris close-up, a voice recording, a pulse video, a face photo, a handwriting sample, a fingertip photo, a palm photo. That's the only input. Nothing else about you factors into the composition unless you separately opt in to DNA fusion (more on that below).

We run well-established, published computer-vision techniques — not a black-box model — against your capture to compute a specific, named set of measurements. For an iris, that includes things like the texture pattern across the iris ring and the ratio between pupil size and iris size. For voice, things like pitch and the shape of the frequency spectrum. For handwriting, stroke geometry. Every measurement has a name, a definition, and a reason it was chosen.

These techniques are deterministic: run the same file through the same version of our extractor twice, and you get identical numbers both times. There's no sampling, no randomness, no model weights that might answer differently on a re-run.

A documented equation turns each measurement into a musical parameter — a note, a rhythmic value, a choice of instrument, an octave. We don't publish the exact coefficients (that's the part of the system that's actually ours), but the category of mapping is public: a measurement that describes texture complexity tends to influence melodic density; a measurement that describes a ratio or proportion tends to influence pitch or interval choice; a measurement that describes a signal's rate (like a heartbeat) tends to influence tempo. The result is a MIDI skeleton — a set of notes and timings, not audio yet — that a derivation table on your composition page can show you, feature by feature.

Because Layer A is what actually decides the music, we can make a strong claim about it: the same capture, run through the same model version, always produces the same MIDI. We test this directly — every modality has an automated check that a fixed reference capture reproduces byte-identical output. If you ever see two runs of the same file disagree, that's a bug, not expected variance.

Several modalities also get run through a pretrained foundation model — the kind of large vision/audio model that produces a dense numeric vector (an “embedding”) rather than a named, interpretable measurement. These are powerful at capturing subtle similarity — “this iris pattern resembles that one” — but they arrive at that judgment in a way that resists a clean explanation.

We use these embeddings for exactly one purpose today: showing you nearest-neighbor comparisons (which of our sample captures your capture most resembles, in the model's own geometry) as an exploratory panel next to your composition. An embedding is never allowed to change a note, a tempo, or an instrument choice. If that ever changes for a specific feature, we'll say so plainly, right next to that feature — not bury it in a settings page.

The interpretable MIDI skeleton from Layer A, plus a text description of its mood and instrumentation, goes to a third-party music-generation service that renders it into audio. That service receives musical parameters — key, tempo, instrument choices, a short description — never your raw biometric file and never the measurement values themselves.

If you also have a DNA profile through our sibling product, LifeSong, you can opt in to have it apply a deterministic adjustment on top of your biometric composition — same idea as Layer A, in that the same DNA and the same biometric capture always produce the same result, but we don't publish the specifics of how the genetic features are derived or how the delta is computed. This is off by default and requires its own separate consent. Full detail on the privacy side is in our Privacy Policy.