ORI Utter

Utter is an audio orchestration framework that turns any text agent into a realtime voice agent. It handles the parts of a voice loop that are hard to get right — semantic turn-taking, streaming synthesis, interruption, voice identity, mid-call handoffs — so your agent only has to answer the next question. Utter is the same framework that runs live Waterr meetings. It ships with a curated voice roster, a swappable reasoning layer, and a per-scenario toggle inside Waterr — and is available as a standalone framework for teams building on their own agents. Status: Research Preview. Enable per-scenario from the scenario editor’s pipeline picker, or request access for the standalone framework.

​

Get started

​

Audio orchestration

Utter is one system with four cooperating layers. Kyutai’s open streaming STT and TTS foundations provide the low-latency audio substrate; Utter contributes the turn engine, the reasoning contract, the voice curation, and the meeting-level primitives that make the loop feel like a conversation instead of a transaction. The four layers: The Kyutai substrate gives us a low-latency streaming loop. What makes Utter Utter is what sits around it: the semantic turn engine that arbitrates the loop, the reasoning contract that lets any LLM slot in without breaking timing, the voice roster we’ve selected for meeting durability, and the meeting-level primitives — multi-agent handoff, participant memory, custom functions — the loop is designed to serve.

​

What you can build

​

BYOA (Bring Your Own Agent)

Utter’s reasoning layer is a contract, not a fixed model. Point it at:

• Waterr’s default backbone — no configuration; use the framework as shipped.
• Your own OpenAI-compatible endpoint — configure a custom endpoint against Waterr’s MCS proxy and Utter routes reasoning to your model. Your agent stays on your infra; only the audio loop runs on Utter.
• A fine-tuned Utter checkpoint — see Fine-tuning.

BYOA is what lets Utter be the voice layer over an agent you already trust — a domain-tuned coding interviewer, a clinical reasoning agent, a compliance-trained sales rep — without giving up the turn engine, the voice head, or the meeting primitives.

​

Fine-tuning

Utter supports supervised fine-tuning on the reasoning layer for enterprise deployments. Typical use cases:

• A support agent tuned on your product’s actual conversations
• A recruiter agent that follows your interview rubric to the letter
• A medical intake agent that speaks in the terminology of your specialty
• A voice register or refusal pattern specific to your brand

Fine-tuning runs on a customer-scoped checkpoint — your data never enters a shared model. Contact the team to scope a run.

​

Dedicated & on-prem deployment

For customers with data-residency, latency, or compliance requirements, Utter is available as a dedicated deployment:

• Dedicated cloud — Utter running in a single-tenant environment in your region of choice. No noisy-neighbor variance on latency; audit-friendly logging.
• On-premises — the full Utter stack (turn engine, reasoning layer, voice head) deployed inside your network. Voice data never leaves your infrastructure. Suitable for healthcare, finance, defense, and regulated public-sector workloads.

Both configurations support BYOA and fine-tuning against your own checkpoint.

​

Numbers (Preview, honest)

​

Limitations

Being honest about the Preview cut:

• English-first. Other languages work but quality varies by voice.
• 30-minute session cap. Longer sessions are on the roadmap; today they need to be re-established.
• Multi-agent handoff coverage varies by pipeline. See the Multi-Agent matrix for what’s supported today.

​

Related

• ORI-Realtime 1.5 — the developer API surface exposing Utter over WebSocket.
• Multi-Agent — mid-call persona handoffs.
• Custom Functions — tools Utter can call mid-turn.