Audio Codec · Early Access 2026

Sound, decomposed.
f  =  ∇φ  +  δψ  +  h

Three orthogonal layers. Exact reconstruction.
The mathematics of Hodge theory applied to audio.

41 dB
SNR vs MP3 25 dB
NEON
Hardware accelerated
—
Decode time, this device
Live demo

Hear it split apart.

Real decomposition running in WebAssembly — powered by strong bisimulation-driven bitrate allocation and topological noise cancellation.

Hodge Decomposition — Live —
Loading sample…
Drop audio file here
∇φ TONE
δψ ATTACK
h SOUL
0:000:00
ξ — — decode — — compression — 4.1× H¹ phase lock — active
Δ = dδ+δd  ·  ∇φ tone   δψ attack   h soul  ·  β(T²) = (1,2,1)
The concept

Three layers. One file.

The split is mathematically unique — every sound decomposes into exactly one set of these three layers, and they sum back to the original exactly.

∇φ

TONE

The melodic tendency — where the sound wants to go. Pitch and structure, stripped of hits and noise.

= gradient component  exact 1-form
δψ

ATTACK

The rotational energy. Drum hits, string plucks, percussive transients — isolated cleanly.

= coexact component  co-closed 1-form
h

SOUL

The invariant core that's neither tone nor attack — a topological fingerprint unique to every recording.

= harmonic component  Δh = 0

f = ∇φ + δψ + h  ·  ⟨∇φ, δψ⟩ = 0  ·  Δh = 0  ·  decomposition is unique

Fidelity

Numbers don't lie.

Prophet-6 synthesizer, 48 kHz. Same file-size class as MP3 — 60% higher fidelity. Runs on S23+ (Snapdragon 8 Gen 2) and Apple M1–M3 with ARM NEON SIMD.

MP3 128k
25.5 dB
AAC 128k
29.2 dB
Ogg 128k
31.2 dB
HodgeCodec
41.1 dB
WAV (ref)
∞ lossless
4.1× compression at exact harmonic reconstruction · SI-SDR 28.4 dB · 91 automated tests  ·  preprint ↗
Reserve your spot

Join the waitlist

Beta opens after arXiv publication. Founding Supporters get lifetime API access and attribution in the paper.

Music Producer
Developer / Engineer
Researcher
Audiophile