Aria Voice runs on the InSync Peacock H100 — our hardware, our rack, our power bill. We are actively building sovereign XTTS with streaming as the cutover path away from ElevenLabs. The doctrine is plain: BYOK passthru pricing today, sovereign cutover the day XTTS quality matches ElevenLabs. Not before. Quality first, sovereignty second.
The wrapper-first doctrine: keep external until sovereign matches on quality, then sever. Updated continuously as the cutover proceeds.
Our rack. Our power. Our cooling. Hosts the inference stack and all encrypted transcripts at rest. Customer data never leaves the box uninvited.
AES-256-GCM transcript encryption deployed 2026-05-12. Off-server key copies stored separately. Privacy gate doctrine: no commercial-caller transcripts without consent.
Custom text-to-speech running on our own GPU, streaming token-by-token. Replaces the current third-party voice tether. Currently in shadow-measurement against ElevenLabs.
We are writing our own inference, tooling, and platform — not vendoring an SDK that pretends to be ours. The repo is the moat.
Today's TTS. Marked-up passthrough billing for paid customers. Replacement (sovereign XTTS) in active construction. Cutover when quality parity proven, not before.
Today's brain. BYOK passthru — the customer pays Anthropic directly through their own key. Sovereign replacement target: an H100-hosted model that matches Claude Haiku 4.5 quality. We cut over the day the bench says we beat it, not the day we wish we did.
Multi-party compute privacy layer. Phase 1 deployed 2026-05-12 22:45 — AES-256-GCM transcript encryption live. Built for partners who need structural privacy, not policy promises.
Every state-change emits a scored splat. Hash-chained. Replayable. Compliance-grade audit without the compliance theater.
Transcripts written through an AES-256-GCM envelope. Key material isn't in the database; the database can't decrypt itself if compromised.
Twilio in → Peacock H100 inference → response back. No third-party brain in the middle once the sovereign cutover lands.
For partners who need verifiable-compute (regulated industries, sensitive transcripts), Arcium MPC privacy compute is the substrate. Phase 1 deployed and live.
Commercial caller transcripts are not saved without consent. Allowlist-only until we ship the consent-capture flow before any voice-product cutover.
The CertusOrdo splat ledger records every voice-call action. Hash-chained. Verifiable. If you need to prove what happened on a call, the receipt is signed.
Privacy deletion requests propagate through the splat chain. Erasure events themselves are logged — proof of deletion is part of the audit trail.
Voice systems where "trust us" is not an acceptable answer. The audit trail and the sovereign hardware story are designed for the room where the lawyer asks where the bits live.
The same call-handling pattern Aria Receptionist uses, plus encrypted-at-rest, plus the MPC layer, plus a written DPA. We will sign the paperwork that vendor APIs cannot.
If you care about voice synthesis as a research substrate, the sovereign XTTS work is shadow-measured publicly. Talk to us about co-publishing the eval.
This is the part of the InSync stack that takes the longest to build and is the hardest to copy. If you want to see where it is today — what's owned, what's still a tether, what the eval bench looks like — book 30 minutes.
book a sovereignty review