Neural Audio Synthesis as Compositional Medium for EEG Data Sonification
One night of sleep EEG, turned into sound: bandpower from each electrode navigates the latent spaces of three RAVE neural audio models, and the scored sleep stages decide which of those voices you hear at any moment.
Three full-length pieces, each a whole night of sleep time-compressed 30:1. They run about fourteen and a half minutes; the excerpts below are the way in if you are standing at the poster.
The voice and modular models, gated against each other by the scored sleep stages: the stage decides which electrode-and-model pairings are audible, so the texture turns over as the sleeper moves between stages. The organ model is not part of this piece — it stands alone below.
Coming soon. The full composition will be posted here before the conference.
The same night carried entirely by the voice model — the whole recording through one latent space, without the stage-driven switching between models.
Coming soon. This piece will be posted here before the conference.
The night through the organ model — sustained, pad-like textures, generated with the encoder deactivated.
Coming soon. This piece will be posted here before the conference.
Seven passages from the composition, each cut around a sustained change of sleep stage. Stage boundaries come from the three-scorer human consensus scoring of the night. Times in brackets are positions in the full composition.
The first descent out of wakefulness [0:33]. Scoring here is unstable — the sleeper crosses into N1 at 0:41 and N2 at 0:48, then surfaces again — and that instability is audible as the texture keeps resettling.
Coming soon. This excerpt will be posted here before the conference.
Download: composition · voices · organ
The first descent into slow-wave sleep [2:44], crossing into N3 at 2:54 as delta power rises and takes over the spectrum.
Coming soon. This excerpt will be posted here before the conference.
Download: composition · voices · organ
The night's first rapid eye movement period [3:35], entered at 3:42. Faster rhythms return and the palette of models opens up again.
Coming soon. This excerpt will be posted here before the conference.
Download: composition · voices · organ
One of the night's longest unbroken REM stretches [6:12], beginning at 6:19 — long enough to hear the material develop rather than just change.
Coming soon. This excerpt will be posted here before the conference.
Download: composition · voices · organ
The night's longest stretch of slow-wave sleep [12:30], entered at 12:38 and broken by an arousal to wake at 13:10 — the sharpest stage change in the piece.
Coming soon. This excerpt will be posted here before the conference.
Download: composition · voices · organ
How the piece closes [13:52]: a last crossing into REM at 13:58, running to the end of the night.
Coming soon. This excerpt will be posted here before the conference.
Download: composition · voices · organ
One complete sleep cycle end to end [4:46] — down through N2 into slow-wave sleep at 5:22, back up to N2 at 5:43, into REM at 6:19, and out again at 7:23. The longest excerpt, and the one that shows the architecture of the night rather than a single boundary.
Coming soon. This excerpt will be posted here before the conference.
Download: composition · voices · organ
The source recording is one night from the Bitbrain Open Access Sleep dataset. Every sound you hear is generated unconditionally by the RAVE models — no audio is fed into them; only the EEG-derived control signals move through their latent spaces. Nothing is sample playback.
Full details are in the ICAD 2026 paper: read the paper (PDF, 4 pages). Code and data processing: github.com/stefvasi/Navigating_Sleep