AIAPS

AIAPS (the AI Audio Protection Standard) is a personal project I built to give audio creators a way to formally declare their recordings off-limits for AI model training. It pairs a custom-built audio plugin with a Next.js and Supabase web platform. Protection rests on three components, a perceptual fingerprint, a timestamped public registry record, and an AIAPS-PROTECTED notice written to file metadata, and the audio itself is never altered: exports are bit-perfect.

The Plugin

The core of the project is a custom audio plugin built on JUCE and distributed as a macOS standalone app, AU, and VST3, so creators can protect tracks directly inside their DAW. The plugin batch-processes files, derives a perceptual fingerprint from the audio, writes the AIAPS-PROTECTED notice to the file's metadata tags, and registers the track in the AIAPS registry. All processing runs locally on the user's machine, so the audio never leaves their computer, and nothing is embedded in the sound itself. Releases ship as signed, notarized installers through a scripted macOS build pipeline.

The plugin authorizes against the platform with an OAuth-style device flow: it requests a short code, the creator enters it on the site, and the plugin polls for a 30-day bearer token, so no password is ever typed into the plugin itself.

Protection & Verification

The fingerprint is a Philips-style robust hash: spectral energy comparisons across Bark-spaced bands are majority-voted over the opening minutes of the track and hashed into a 256-bit key, so the result survives lossy encoding, format conversion, and delivery edits like trims and fades. Because the samples are copied in their native representation with no float round-trip, a protected WAV or AIFF nulls bit-for-bit against its source, a claim exercised by dedicated test harnesses covering both fingerprint robustness across AAC bitrates and bit-perfect output.

The platform exposes a verification flow so anyone can confirm a track's protected status through fingerprint lookup, artist search, or its AIAPS ID, making the declaration independently checkable rather than self-asserted. Verification also still recognizes files protected by earlier plugin versions, so no creator's protection is orphaned by the standard evolving.

Every registered track receives a sequential AIAPS ID and a permanent certificate permalink, and the public registry is searchable by ID, fingerprint, or artist, with opt-in artist pages grouping protected tracks by album.

Web Platform

The web side is built on Next.js and Supabase, the same stack I used for Stagelet. Supabase handles authentication and stores the registry of protected works, with account login gating the dashboard where creators manage their registered tracks. The site also publishes the specification, compliance and governance documentation, verification tools, and an FAQ, currently in early access for macOS. Page copy is driven by a JSON content model with a shared section renderer, and a single source of truth holds the standard's canonical definitions so wording never drifts between the spec, FAQ, and site. An integrated AI chat assistant, powered by the OpenAI API, helps visitors navigate the standard and get answers about protecting their work.

Why I Built It

As generative audio models advanced, I wanted a concrete, standards-style answer to the question of how creators can opt out of training. AIAPS was my attempt to design that end-to-end: the DSP and plugin tooling, the perceptual fingerprinting layer, the registry and verification infrastructure, and the public-facing spec, built as a self-directed project spanning native audio code and modern web development.