ObraVera Audio Authentication

Provenance and integrity for human-created audio. Embed signed credentials at production time, verify them after compression and editing, and recover them from a watermark when metadata is stripped.

Status: Beta. Currently distributed under a proprietary beta-testing license tied to a pilot program. This README is structured open-source-style for clarity; see the License section below for the actual terms.

Modern audio platforms strip metadata, re-encode aggressively, and normalize loudness to a few dB. None of those operations are malicious, but all of them destroy ordinary metadata-based provenance.

Separately, neural voice cloning has reached the point where same-speaker verification by acoustic similarity (cosine on a speaker embedding) cannot reliably distinguish a real recording of someone from a high-quality clone of them.

Provenance and similarity are different problems. Voiceprint matching asks "does this sound like X?" — a probabilistic similarity claim. ObraVera asks "did X authorize this specific file?" — a verifiable provenance claim. Only the second is robust against cloning. ObraVera's job is to make that second claim cheap to make and hard to break.

• Custom metadata — ID3v2 (MP3), RIFF INFO (WAV), Vorbis comments (FLAC, OGG)
• C2PA manifest — for ecosystems requiring industry-standard provenance (Adobe Content Credentials, BBC/Reuters/AP workflows)
• Watermark — choose AudioSeal (neural, 16-bit + ECC, robust to loudness normalization), audiowmark (patchwork, 128-bit, MP3-resilient), or both for redundancy
• Chromaprint fingerprint — perceptual hash for tamper detection
• Ed25519 signature — cryptographic proof of credential origin
• Local registry — SQLite-backed recovery: a watermark payload alone is enough to retrieve the full credential

A SHA-256 binding hash ties watermark payload(s) to the credential's Ed25519 signature. This catches the swap-watermark attack: lifting a real watermark from one file and pasting it into another is detected because the binding no longer matches.

A separate web verifier at verify.obravera.com lets anyone hold their phone up to a speaker playing a watermarked recording and confirm the watermark over the air — no file upload, no special equipment, just a microphone and a browser. This proves the watermark survives the speaker → air → microphone chain, which is the weakest link in any "unknown recording" verification scenario.

• Desktop GUI (CustomTkinter) — Embed, Verify, Settings tabs
• Command-line interface — embed, verify, extract subcommands
• DAW Quick Apply — single-file headless mode for integration into Pro Tools / Logic / TwistedWave workflows
• Batch pipelines — embed across many files with progress tracking and per-file server registration
• PWA verifier at verify.obravera.com

                  ┌──────────────────────────────────────────┐
                  │              Audio file (in)             │
                  └──────────────────────────────────────────┘
                                       │
                                       ▼
        ┌─────────────── embed_credential ────────-─────────┐
        │                                                   │
        │   Metadata     (ID3v2 / RIFF INFO / Vorbis)       │
        │   C2PA         (optional manifest)                │
        │   Watermark    (AudioSeal | audiowmark | both)    │
        │   Fingerprint  (Chromaprint perceptual hash)      │
        │   Signature    (Ed25519 over binding hash)        │
        │   Registry     (local SQLite, payload-indexed)    │
        │                                                   │
        └───────────────────────────────────────────────────┘
                                       │
                                       ▼
                  ┌──────────────────────────────────────────┐
                  │       Authenticated audio file (out)     │
                  └──────────────────────────────────────────┘

Verification reads whichever layers survived and falls down the chain: metadata → C2PA → watermark → registry recovery. A green verdict requires exactly one path to succeed end-to-end and a valid signature binding. The fingerprint flags content tampering even when the credential itself remains valid.

Distribution-ready macOS builds are on the Releases page.

The first time you open ObraVera, macOS will display a Gatekeeper prompt: "ObraVera is an app downloaded from the internet. Are you sure you want to open it?" This is the standard prompt for any signed app downloaded from outside the App Store. Click Open. Subsequent launches are silent.

The app is signed with an Apple Developer ID and notarized through Apple's Notary Service.

Open ObraVera from Applications. The Embed tab walks you through:

1. Pick an audio file (or a folder, for batch mode)
2. Fill in the credential information (work title, AI usage declaration)
3. Choose a preset (Standard, Quick Audition, Maximum) or set parameters manually
4. Click Embed

The Verify tab takes any audio file and reports which authentication layers passed and which failed, with diagnostic detail when something has gone wrong.

# Embed (default: AudioSeal watermark + fingerprint + signature)
obravera embed input.wav --credential cred.json -o output.wav

# Embed with audiowmark instead (better MP3 survival, 128-bit payload)
obravera embed input.wav --credential cred.json --watermark audiowmark -o output.wav

# Embed with both engines for redundancy
obravera embed input.wav --credential cred.json --watermark both -o output.wav

# Verify
obravera verify output.wav

# Verify with diagnostic output (bit-level corruption analysis)
obravera verify output.wav --diagnostic

obravera-apply --file recording.wav --preset standard --headless

Useful for keyboard-shortcut integration into Pro Tools / Logic / TwistedWave: embed the credential without breaking out of the DAW.

ObraVera stands on these projects' work:

• AudioSeal (Meta) — neural audio watermarking
• audiowmark (Stefan Westerfeld) — patchwork watermarking with BCH error correction
• Chromaprint (Lukáš Lalinský) — acoustic fingerprinting
• SpeechBrain — speaker embedding models (voiceprint feature, in development)
• c2pa-python — C2PA manifest tooling
• ISCC — International Standard Content Code

Without these, ObraVera would be a much narrower project.

Version: 0.5.0 — beta, pilot distribution.

In active development:

• Speaker voiceprint matching — investigative 1:1 same-speaker comparison using SpeechBrain ECAPA-TDNN. Calibration phase in progress; a careful look at what voiceprint matching can and cannot promise before shipping.
• Windows installer — planned.

Proprietary beta-testing license. Distribution is limited to authorized pilot participants.

For licensing inquiries outside the pilot program, contact:

Brady Ridgway — brady@obravera.com