Introducing System-One Voice & SARAH Switch
Voice agents have sounded human for years. Production automation has not followed.
That gap is the question we have been engineering against since the first on-premise SIP voice stack shipped on US3: if the model can speak, why do most enterprises still route every decision through a human reviewer, a ticket queue, or a brittle IVR tree?
Today we are publishing the pattern we use in production: System-One voice (fast, structured, software-shaped decisions on the phone) and SARAH Switch (the routing layer that turns utterances into typed actions your CRM, calendar, and compliance rules can trust).
This is not a chatbot press release. It is how Omega Revenue Engine and SARAH AI Suite behave when latency, audit, and revenue are on the line.
Frontiers, old and new
| Rented cloud chat voice | System-One + SARAH Switch | |
|---|---|---|
| Optimized for | Human preference: natural dialogue, long answers, open-ended help. | Calibrated operations: classify, route, book, escalate, and record with explicit outcomes. |
| Inputs | Sequential audio streamed to a remote API; state lives in vendor memory. | Streaming STT on your lane; session state in your PostgreSQL + portal config; optional sovereign GPU path. |
| Outputs | Generated speech strings. Tool use is optional, often hallucinated, and parsed after the fact. | Structured actions: extension transfer, calendar hold, CRM note, compliance script, or bounded natural-language reply when policy allows. |
| Sampling | Token-by-token generation; long tail latency on complex turns. | Parallel harness: STT frame, retrieval, policy check, and TTS chunk pipeline tuned for sub-second perceived response on socket lanes. |
| Confidence | Models rarely say when they are guessing; enterprises add humans in the loop. | Switch returns explicit branches: speak, transfer, book, refuse, or ask one clarifying question (never two in a row on sales lanes). |
| Use cases | Demos, copilots, and pilots where failure is cheap. | PSTN revenue lines, regulated intake, multi-language reception, and fleet extensions (3000, 3011, 3031, 3334) with booking and KB retrieval wired in brain.py. |
What SARAH Switch is
Think of SARAH Switch as a frontier-intelligence router: unstructured audio and caller context in; a small set of allowed actions out. The large model still reasons, but the harness constrains freedom so software downstream does not parse JSON from a rambling paragraph.
On extension 3011 (Chris office PSTN) and 3000 (Voice Link browser), the same prompt family and KB spine apply. On 3031, the lane sells a single offer. On 3334, conference identity stays separate from human accept console 3333. The Switch pattern is what keeps those lanes from bleeding into each other.
Evidence from production
Latency shape
We publish no synthetic benchmark league tables. What we measure on US3 is operational: time from end-of-utterance to first TTS byte on golden socket lanes, with WhisperLive streaming STT only (batch STT is forbidden in our stack). System-One shaping targets the band where callers still feel listened to, not lectured at.
Nuance
Your mileage depends on carrier, codec, and GPU load. PSTN on Crazytel into 3011 carries extra listen-window tuning versus WebRTC on 3000.
Structured booking
Voice booking flows run through the same capture verifier used before every gateway recycle: intents map to calendar holds and CRM-safe fields, not free-text promises the business cannot fulfill.
Knowledge without gibberish
Read.ai C-suite corpus, CXO teachings pages, and offer-specific landings (for example the $4M software engagement) ingest into extension_knowledge with pgvector retrieval. Outbound chat and voice paths pass through guards so raw tool JSON never reaches a human channel.
What this enables next
- Smart if-branches in revenue code. Score a lead, route by territory, or refuse out-of-policy pricing without hand-writing every edge case.
- Map-reduce over conversations. Thousands of executive sessions become teachings bullets (see CXO teachings) rather than one unusable transcript dump.
- Real-time voice UX. When the harness owns turn shape, product teams can ship phone surfaces that feel like a senior operator, not a FAQ reader.
We give a FAQ
- Where does the name System-One voice come from?
- From the Kahneman fast-and-slow distinction, adapted for engineering: System-One here means decisions that must be right often enough, fast enough, and typed enough for software to compose. Slow reasoning still exists on socket 0 for build and research lanes; voice sales and accept paths are shaped for speed.
- Is SARAH Switch a smaller model?
- No. Production voice uses the 235B Brain on sovereign GPUs. Switch is policy and orchestration around the model, not a separate tiny weights file.
- How is this different from an AI Chat on a phone?
- Chat optimizes for open-ended strings. We optimize for governed actions on your extensions, your calendar, and your audit log, on hardware you control.
- Can I try it?
- Voice Link and ORE extensions are provisioned per engagement. Start with extension 3000 in the browser or schedule a call with the team.
See the lane live. Browser Voice Link on extension 3000, or speak with SARAH on your provisioned ORE extension.
Open Voice Link 3000 Schedule a call