Φ
Φ
PURE C · 6,679 EMOTION NORMS · 1,000× REAL-TIME

The Voice That Feels.

Every word carries its own emotion — arousal shapes tempo, valence bends pitch, dominance drives amplitude, layer distance paints stereo space. Two voices coupled as oscillators converge, diverge, or lock. The synthesis of three years of language research, fused into a voice that sounds different because it’s saying something different.

4
Engines
solo → duet → chorus
6,679
Words
emotion-mapped
~1,000×
Real-Time
compute speed
11
EEG Tests
10 null, 1 real

Three Parts of the LLM Problem

Can a machine find meaning in language without being trained on meaning? We investigated three paths. The voice is their synthesis.

01

The Language Disc

19,888 words. Bigram and trigram chains harvested from the NUMEN codebase. No training, no gradients — direct counting. The cloze engine predicts at 31.6% top-1 from pure statistics.

logos.qlib · 12 MB
02

The Layer Trigger

5,056 words with measured skip-gram distance. Each word’s natural resonance depth in the language — how far its meaning radiates. Built from 92 billion token-pairs.

layers.bin · 20 KB
03

The Emotion Binding

6,679 words matched to Warriner 2013 norms: valence (positive/negative), arousal (calm/excited), dominance (submissive/dominant). The one real EEG finding wires arousal to tempo.

warriner.bin · 107 KB

How Emotion Becomes Sound

Four measured parameters, four acoustic dimensions. Nothing arbitrary.

ArousalTempo & Attack

Pupil contracts with arousal (p=0.005, 11/12 brains). High arousal = faster, sharper.

ValencePitch Register

Positive words get brighter harmonics, higher overtones. Negative words go darker.

DominanceAmplitude

Dominant words hit harder. Submissive words recede. The word’s authority shapes volume.

Layer DistanceStereo Pan & Reverb

Deep-layer words spread wide via golden-angle rotation. Surface words stay centred.

PRESS PLAY

Four Engines, One Voice

From a single chain to a five-voice chorus. Each engine is pure C, zero dependencies, ~1,000× real-time.

SINGLE CHAIN

THE VOICE

One language chain, emotionally modulated. Arousal shapes tempo and attack. Valence bends pitch register. Dominance controls amplitude. Layer distance paints stereo space.

5,029 words/s1,069× real-time60 words
seed: the light inside the machine
0:00 / 0:00
MULTI-TURN MOOD

THE CONVERSATION

Takes your words, chains a response, speaks it. Running mood tracks valence and arousal across the whole conversation with exponential decay. The conversation has inertia.

4,542 words/s999× real-time198 words
seed: "all words are tied to an emotion"
0:00 / 0:00
COUPLED OSCILLATORS

THE DUET

Two independent chains with coupled emotional states. Resonance pulls them together. Counterpoint pushes them apart. A real dynamical system in emotional space.

4,381 words/s948× real-time568 words
seed: light hope joy ↔ dark fear pain
0:00 / 0:00
RING TOPOLOGY

THE CHORUS

Five voices — Warden, Librarian, Healer, Dragon, Bridge — coupled in a ring. Each voice pulls (or pushes) its two neighbours. Five pitch registers spread across stereo.

4,464 words/s940× real-time600 words
seed: justice · knowledge · love · fire · light
0:00 / 0:00

The Coupling Matrix

Two voices, seeded on opposite emotions (light vs dark). 30 turns. Coupling strength k controls how hard they pull (or push). Measured gap in Valence×Arousal space.

RESONANCEk=0.10
CONVERGED
start gap0.714
end gap0.535
Δ = -0.179 · final V gap = 0.53
COUNTERPOINTk=0.10
DIVERGED
start gap1.071
end gap0.801
Δ = -0.270 · final V gap = 0.80
RESONANCEk=0.40
LOCKED
start gap0.178
end gap0.137
Δ = -0.041 · final V gap = 0.14
COUNTERPOINTk=0.40
FIGHTING
start gap1.606
end gap1.252
Δ = -0.354 · final V gap = 1.24

At k=0.40 resonance, the voices fuse — gap drops to 0.14 immediately. At k=0.40 counterpoint, they fight — swinging ±1.0 in valence each turn. A real coupled oscillator system in emotional space.

The EEG Evidence

ZuCo 1.0, 12 subjects, 83,712 word tokens, 54,790 EEG epochs. 11 emotion tests. 10 null. 1 real. Published in full.

Pupil size vs word arousal11/12 brains negative. The one replicated effect.
r = −0.015, p = 0.005REAL
Alpha band power vs emotion
p = 0.73NULL
Theta band power vs emotion
p = 0.51NULL
EPN 200–300ms vs emotionGrazed. Published as a miss.
p = 0.0503NULL
LPP 400–700ms vs emotion
p = 0.82NULL
Full ERP waveform (10 bands)
all p > 0.05NULL
Layer distance vs emotion
r = −0.049, p = 0.30NULL

HONEST CEILING

The one real finding (pupil × arousal, r=−0.015, p=0.005) is ~1,000× too small to predict anything. It’s real. It replicates across 11 of 12 brains. It’s useless for prediction. The voice uses it as a design principle — high arousal compresses — not as a measured acoustic effect. We publish every null because the nulls are the finding.

Pure C. Zero Dependencies.

numen_voice.c
413
lines · Single chain
numen_converse.c
243
lines · Conversation
numen_duet.c
~350
lines · Coupled pair
numen_chorus.c
~380
lines · Ring of five

All four engines compile with gcc -O2, link against logos_lib.c and -lm. No external libraries. No GPU. No training. No gradients.