Lab · Music
Resonance Sessions
Three guided listening journeys, each shaped to a designed arc that settles, rises, peaks and resolves. You can open each one with a song that moves you, and a short spoken intention starts you off. Every track is made by AI and labelled that way.
AI-made. The music is generated by ACE-Step 1.5 and the voice by Piper TTS. A person designed the arcs, chose the takes and mixed them.
No brainwave claims. Nothing here entrains, tunes or "shifts" your brain. It is music, shaped with care. What it does for mood or relaxation is yours to notice.
Not therapy. These sessions are not treatment and not for anyone in crisis. Music can amplify difficult feelings as well as good ones.
Safety. Don't listen while driving or doing anything that needs your attention. Headphones are optional. A Stop button is on every screen, and Esc works too.
Choose a session
Three arcs
The line on each card is the designed intensity arc: how much energy, density and pulse the music has at each point. Every session has a silence control of the same length with the same spoken opening and return cue, just no music, so you can compare the two in your Experience Ledger.
How these were made
Arc first, then music
1. The arc is designed before any music exists. Each session is a list of segments of about two and a half minutes, each with an intensity between 0 and 1: still (no pulse), pulse (a soft, even beat enters), build, peak, and ease. Intensity sets two things: the words we give the music model (how dense, how rhythmic, what enters or drops out) and a loudness target in the mix, from −34 LUFS at 0 to −17 LUFS at 1. Segments overlap by 20 seconds with equal-power crossfades. Every session ends with a long quiet resolve and a 45-second fade, because the way back matters as much as the peak (Kaelen et al. found the music has to carry people back down, not just up).
2. Rhythm aims for the middle of the groove curve. Witek et al. found that the urge to move and the pleasure people feel both follow an inverted U against syncopation: medium syncopation scores highest. So the peak segments ask for “moderate syncopation, not busy, not complex”, lower segments ask for an even, simple pulse, and all the rhythmic segments in a session share one moderate tempo (84, 92 or 96 BPM), so crossfades never collide two tempi. The “tempo” dimension of the arc is carried by pulse density (none, then even, then syncopated) rather than by speeding up.
3. Seeding and voice. Blood & Zatorre’s chills came from music the listeners chose; Kaelen et al. found that how people related to the music predicted outcome, not how intense it was. We cannot generate your favourite song, so we let you play it yourself as the threshold. It plays from your own device and is never uploaded. Huels et al. found rhythm changed everyone’s brain activity but the altered state came with trained intent, so each session opens with a spoken intention rather than relying on sound alone.
4. Models. Music: ACE-Step 1.5 (the acestep-v15-sft diffusion model with its 1.7B planning language model), run locally on one NVIDIA T4 in float32, instrumental only, 50 diffusion steps, fixed seeds. Voice: Piper TTS, voice en_US-amy-medium, slowed slightly. Mixing, loudness placement and encoding: our own script with ffmpeg. No human musicians played on these tracks. Lecamwasam & Chaudhuri found listeners often can’t tell calm AI music from human music, and that labels change what people report, which is one more reason to label it plainly.
Listen-check: designed arc vs measured loudness
Each finished master was measured with ffmpeg’s EBU R128 meter (short-term loudness, median of each 10-second window) and set beside the loudness the arc asked for. The first 30 seconds and last 60 are left out of the statistics, because the spoken voice and the final fade sit there on purpose.
Honest limits
- The loudness arc is partly the model and partly our gain placement. We normalised each segment to its target, so a match on loudness shows the mix follows the design. It does not show the model followed the words about density and groove. That is checked by ear.
- “Moderate syncopation” is a prompt, not a measured quantity. We did not compute a syncopation index on the output. The model decides what the words mean.
- We make no claim that these sessions relax you, deepen anything or change your brain. The silence control and the Ledger questions are there so you can see for yourself whether music does anything for you beyond sitting quietly with the same words.
- The voice is synthetic, and some people find synthetic voices off-putting. If it jars, that is useful data. Say so in the Ledger notes.
- The peak segments are generated, not composed. Some transitions are rougher than a human producer would allow.
Built on
- Kaelen et al. (2018), The hidden therapist: the arc that rises, peaks and resolves; liking the music matters more than intensity; music can also amplify distress, hence the stop control.
- Witek et al. (2014), Syncopation, body-movement and pleasure in groove music: moderate syncopation at the peak, not maximum complexity.
- Blood & Zatorre (2001), Intensely pleasurable responses to music: peaks are personal, so you seed the session with your own song.
- Huels et al. (2021), Neural correlates of the shamanic state of consciousness: rhythm plus intent, not rhythm alone, so every session opens with a spoken intention.
- Ingendoh et al. (2023), Binaural beats to entrain the brain?: weak, inconsistent evidence (5 of 14 studies), so no binaural beats and no entrainment claims.
- Lecamwasam & Chaudhuri (2025–26), Listener perceptions of AI and human-composed music: labels change what people report, so every track is labelled AI-made.
Summaries of all six are in the reading list.
Resonance Session
Open with a song that moves you (optional)
Your song becomes the threshold you cross before the session begins. Pick a file on this device: it plays here in your browser and is never uploaded. Or play it yourself, from any app you like, then come back.
Not therapy. Not for anyone in crisis. Don't use while driving. Stop is always top right, and Esc works too.
The threshold
Your song
Settling
0:00 / 0:00
Close your eyes if you like. The voice will tell you when the music is coming to rest.
Returned
Welcome back
Take your time. Feel your feet, look around the room, have a sip of water.