RightsRelay

Revoke once, regenerate every affected synthetic-voice derivative.

RightsRelay is a small production-minded demo for the 2026 Backblaze Generative Media Hackathon. A media team creates a voiceover from a source script, records which synthetic voice profile was used, and keeps a verifiable Genblaze manifest. If a voice profile becomes unavailable or its operational use window changes, RightsRelay finds affected campaigns and regenerates them with an approved replacement while preserving the old revision and hash chain.

This is operational rights tooling, not a legal-compliance engine. A status in RightsRelay records a team's workflow decision; it does not determine whether a license, consent, or contract is legally valid.

Why it exists

Synthetic media is usually stored as disconnected output files. A producer may know that a voice can no longer be used, but still lack a reliable answer to:

• Which current campaign files depend on that voice?
• What replaced each affected derivative?
• Can a reviewer verify that the replacement bytes match the recorded run?
• Are source scripts and regenerated media durably stored rather than tied to a provider's expiring URL?

RightsRelay makes those relationships explicit and mechanically verifiable.

Real pipeline

1. The source script is hashed and ingested with Pipeline.ingest.
2. A custom PiperTTSProvider(SyncProvider) runs a local neural TTS model.
3. Genblaze binds the script input, voice profile, operational metadata, WAV hash, and replacement lineage into a canonical manifest.
4. In B2 mode, ObjectStorageSink uploads source and derivative assets with content-addressable keys, then stores the manifest.
5. Revoking a profile queries the reverse index and reruns every affected campaign with its designated replacement profile.

The demo uses Piper en_US-libritts_r-medium, a 904-speaker, 22.05 kHz model fine-tuned from English lessac medium on LibriTTS-R train-clean-360. The official model card identifies the dataset license as CC BY 4.0. Piper TTS is used locally; no text or media is sent to a hosted model API.

Run locally

Python 3.11 is recommended.

uv venv --python 3.11 .venv
uv pip install --python .venv\Scripts\python.exe -e ".[dev]"
.venv\Scripts\python.exe scripts\download_voice.py
$env:RIGHTSRELAY_VOICE_MODEL_PATH="$PWD\.models\en_US-libritts_r-medium.onnx"
$env:RIGHTSRELAY_VOICE_CONFIG_PATH="$PWD\.models\en_US-libritts_r-medium.onnx.json"
.venv\Scripts\python.exe -m uvicorn rightsrelay.app:app --app-dir src --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000.

Backblaze B2 mode

Create a private bucket with a bucket-scoped application key and set the variables from .env.example. Never use a master key. Secrets stay server-side and are excluded from responses and logs. Raw source scripts and canonical Genblaze manifests remain private; the public /proofs/{campaign_id} endpoint returns hashes and lineage fields without the prompt or filesystem paths.

B2_KEY_ID=...
B2_APP_KEY=...
B2_BUCKET=rightsrelay-demo
B2_REGION=us-west-004
B2_PUBLIC_URL_BASE=

The app deliberately disables automatic bucket lifecycle changes. Audio is read from private B2 by the service and returned through the same-origin media route; no signed credential-bearing URL is persisted or exposed to the browser.

The private B2 bucket also stores a stable SQLite reverse-index snapshot. A free Render instance restores it after an ephemeral-filesystem restart. If the bucket is empty, RIGHTSRELAY_SEED_DEMO creates an idempotent before/after example: Atlas revision one followed by a Nova replacement revision. Scripts, audio, full manifests, and the index survive process restarts, while content-addressed assets are deduplicated.

Anonymous mutations are serialized, capped at 24 campaigns, and rate-limited per client in the public demo. A production deployment should add real user authentication and tenant isolation.

Tests

$env:RIGHTSRELAY_TEST_MODEL_DIR="$PWD\.models"
pytest

Tests exercise real offline neural synthesis, canonical manifest verification, audio and provenance tamper detection, redacted public proofs, reverse lookup, idempotent demo seeding, and revocation-triggered regeneration. No B2 credentials or network calls are required by the default suite.

Disclosure and licenses

• Application code: Apache-2.0.
• Genblaze: MIT.
• Piper TTS package: see the upstream package license.
• en_US-libritts_r-medium dataset/model attribution: CC BY 4.0, model card linked above.
• AI coding assistance was used during implementation. All generated changes were locally tested and are disclosed rather than presented as unaided work.