- Recognisable listened music reconstructed from invasive cortical recordings.
- Known imagined songs and individual pitches classified above chance from scalp EEG.
- Imagined melodies decoded and mentally manipulated in MEG studies.
- Auditory attention and imagined commands measured using wearable around-ear EEG.
- Deliberate silent articulation recognised from face and neck muscle signals.
OPEN INVENTION PROGRAMME · 001
The instrument
between imagination
and sound.
Yaan has composed music continuously in his head since he was a teenager. Thought Music is the serious attempt to let that music cross the last gap—privately, deliberately and in real time.
INPUTdeliberate musical intent
OUTPUTeditable MIDI + stems
RULEthe musician remains the author
SCROLL TO ENTER THE LAB ↓
00 / ORIGIN
“I make music in real time as I go along in my head. I have done since I was an early teenager. I always do it most when I’m walking.”
That is the starting material—not a trend and not a gimmick. As Kainer, Yaan already knows arrangement, rhythm, tension, release and sound design at professional level. The missing instrument is the one that can follow the composition before hands, voice or a DAW interrupt it.
If nobody builds it in time, we will.
01 / WHAT IS TRUE NOW
The frontier is real.
The finished instrument is not.
Current science can identify pieces of imagined music. It cannot yet read an arbitrary unheard song from a consumer headset. Our plan begins exactly where the evidence ends.
- Continuous high-fidelity reconstruction of spontaneous imagined music from wearable EEG.
- A universal decoder that works without personal calibration.
- Reliable neural audio capture during natural walking with no motion compensation.
- A system that can infer missing detail without occasionally inventing the wrong music.
- Anything honestly described as effortless consumer “mind reading.”
2023 · PLOS BIOLOGY
Recognisable Pink Floyd reconstructed from direct auditory-cortex recordings.
Listened music · invasive ECoG · 29 clinical participants ↗
2023 · IEEE TNSRE
Seven imagined pitches decoded from scalp EEG above chance.
35.68% mean; 14.3% chance · non-invasive ↗
2024 · PLOS BIOLOGY
Imagined melodies and their deliberate mental manipulation decoded with MEG.
71 participants · short controlled melodies ↗
2021 · J NEURAL ENG
Imagined speech commands classified using wearable ear-centred EEG.
Proof of wearable neural intent, not free-form speech ↗
2018–2025 · MIT ALTEREGO
Deliberate internal articulation captured through a private neuromuscular wearable.
Silent intent · user-controlled · limited vocabulary ↗
2024 · INTERSPEECH
Dry-electrode neckband points toward discreet EMG-to-speech control.
Promising 11-word pilot · only two subjects ↗
Research rule: strong results in constrained tasks do not automatically become a general thought decoder. Every claim here states the sensing method and limitation.
02 / THE INVENTION
Do not decode a waveform first.
Decode a musician.
Full audio contains far more information than musical intent. The first useful instrument learns Yaan’s compact language of pulse, contour, tension, density, timbre and form—then renders it through his own palette.
Sense
Around-ear EEG, dry neck/jaw EMG and walking IMU capture complementary evidence.
Align
Listen → imagine → reproduce sessions give the model timing and musical ground truth.
Decode
A personal confidence-gated model emits musical intent tokens—not fake certainty.
Render
A local engine creates editable MIDI and stems using Kainer’s instruments and decisions.
THE DIFFERENTIATING ENGINE
Never invent more than the evidence can carry.
The decoder keeps several plausible musical continuations alive at once. Weak evidence changes their probability; deliberate corrections teach the personal instrument. If confidence falls, it holds, repeats or simplifies instead of composing over Yaan.
SHADOW capture silently
ECHO hear it one beat or bar later
DRONE output only pulse and tonal centre
INBOX choose between phrase hypotheses later
Why delayed echo? Immediate feedback can overwrite the internal phrase and make the musician follow the machine. A one-beat or one-bar delay leaves the imagination in front.
Captured
Supported by an explicit muscular, gestural, vocal or future neural signal.
Inferred
The best reversible interpolation between captured anchors.
Completed
Detail added under musical rules Yaan explicitly selected.
Generated
Optional autonomous texture, always separable and removable.
PROPOSED WEARABLE · NON-CLINICAL CONCEPT
Three weak signals become one useful instrument.
- EAR / AROUND-EAR EEG
- Auditory attention, imagery timing and coarse melodic information.
- NECK + JAW EMG
- Deliberate silent commands, contour and rhythmic articulation.
- MOTION / GAIT
- Separates walking artefacts from intent and supplies contextual pulse.
- BONE CONDUCTION
- Private feedback without sealing the user away from the path around him.
03 / LIVE CONTROL-LAYER PROTOTYPE
Build the instrument’s nervous system now.
This browser lab does not read thoughts. It exercises the layer a future neural decoder will control: tempo, contour, harmonic tension, density and section intent, with walking-motion input and exportable event data.
ENGINE OFFLINENO BIOSIGNAL SENSOR CONNECTED
BAR 0011.1
DELIBERATE INTENT TOKENS
WAITING FOR DELIBERATE INPUT…
Prototype boundary: audio is generated locally in your browser from the visible controls. “Simulate decoder” sends synthetic confidence-gated intent tokens so the proposed interface can be tested without pretending a neural sensor is attached.
04 / EXPERIMENT 001
FIRST FALSIFIABLE QUESTION
Can Yaan’s imagined pulse and phrase identity be decoded repeatedly while walking?
No vague “brainwave mood” demos. The first experiment has an intended phrase, a label, a held-out day and a result that can fail.
- 01Prepare eight original Kainer phrases across two tempos and four rhythmic identities.
- 02Listen, imagine for eight seconds, then immediately hum, tap or play the phrase as ground truth.
- 03Repeat stationary and walking while recording EEG, EMG, IMU and exact event markers.
- 04Compare EEG-only, EMG-only, movement-only and fused models on entirely held-out days.
- 05Proceed only if fusion beats every single modality and false activation remains low.
05 / BUILD BY EVIDENCE GATES
No fantasy dates.
Each stage must earn the next.
The destination remains direct imagined music. The route is deliberately staged so useful instruments emerge before perfect neural reconstruction exists.
Walking capture companion
Define Yaan’s personal musical intent grammar, collect timestamped sessions and harden the real-time output engine.
Silent intent band
Dry EMG + IMU for deliberate commands, rhythm and contour. Target high precision, low false activation and sub-250 ms command latency.
Ear-neural fusion
Add around-ear EEG and prove it contributes information beyond muscles and walking motion.
Personal music latent
Decode beat grid, pitch probabilities, harmonic function, density and stem activity into editable DAW material.
Direct thought-to-sound
Use future higher-bandwidth sensing to reconstruct continuous imagined music while preserving intent, privacy and authorship.
06 / NON-NEGOTIABLES
Instrument, not sonifier.
Random EEG changes mapped to random filters are not the music Yaan hears.
Deliberate on-state.
The system follows only music intentionally handed to it. A physical off-state is mandatory.
Local by default.
Raw neural and myoelectric data stays on the musician’s device unless explicitly exported.
Uncertainty is visible.
The decoder says “I’m not sure” instead of manufacturing a convincing false phrase.
Authorship survives AI.
Generative models render constrained intent. They do not quietly become the composer.
Walking is the lab.
Movement is a primary operating condition, not noise hidden by a stationary demo.
THOUGHT MUSIC · OPEN NOTEBOOK 001
You hear it.
The instrument learns to follow.
The science is early enough to require honesty and mature enough to justify starting. So we start.