Voice receptionist — Path A is live, Path B is proven, the phone leg is yours

What you do next — 3 steps, about 10 minutes

1. LiveKit Cloud dashboard: Telephony → Phone numbers → +1 760 965 5049 → Assign dispatch rule → SDR_p6seEBH2cQ4q.

   Why you and not the system: the API call that does this is broken on LiveKit’s side — every field-name and flag variant returns invalid_argument, and the one workaround that did write cleanly turned out not to actually bind. The dashboard is the only lever left.

2. Call (760) 965-5049 from your cell and talk to it.

   Why it matters: this is the only thing never yet tested. Listen for six beats — greeting, it asks your name, it reads your number back digit by digit, it asks your reason, it asks permission to pass it on, then one confirmation and it hangs up. If any beat is missing or it loops, stop and tell me which one.

3. Only when you are satisfied: GHL → Settings → Phone Numbers → (760) 642-1892 → forwarding → +1 760 965 5049.

   Why last: this is the cutover. Until you do it, Path A keeps answering exactly as it does now. Leave the Voice AI backup box checked — it still fires on ring-timeout no matter where the forward points, so you keep a net under the net. Revert is the same screen, old value ends 1622.

How it was built

Path A — 19 minutes on 09-11, get something answering

1. Pulled the existing Nano research before building anything.22:57 A Nano session had already written the cost research, a call-screening audit and a staged LiveKit worker with 35 passing tests. Read those instead of redoing them.
2. Proved the local speech stack.23:00 Whisper on M2 healthy, Kokoro healthy, Ollama 12 models, the OpenAI key on file invalid — which is what made the whole thing local-only by default.
3. Wired the screener to LiveKit.23:01 Inbound trunk locked to carrier ranges (LiveKit refuses an open trunk), dispatch rule to agent business-screener, worker installed as a LaunchAgent, explicit dispatch delivered a job to it.
4. Found which GHL sub-account owns the 760 by API, not by asking you.23:03 Three sub-account tokens exist; two are dead or 401. EROS answered 200 and its number list contains 1892, friendly name “My Legacy Management”, forwarding to the cell ending 1622.
5. Discovered GHL’s Voice AI API was reachable on that sub-account.23:04 Which meant a native agent could be attached to the 760 with no new number, no carrier account and no forwarding change.
6. Schema discovery, with one honest miss.23:05 A minimal probe POST intended to surface validation errors instead created an agent. Recorded as a block so it is not repeated.
7. Reconfigured that agent in place.23:06 Name, business, voice, welcome line, CMS-safe message-taking prompt (no advice, no quotes, not a “specialist”, no government affiliation), notifications, backup mode, number attached.
8. Verified by re-reading the record, not by the 200.23:06 One agent on the location, correct number, backup mode confirmed.
9. Recorded state everywhere it has to live.23:07–23:12 Lane on both boxes, packet, control script, A0 card, session log, memory, blocks, identifiers into the operator env on M2 and Nano.

Path B — 09-12 into 09-13, own the whole stack

1. Built our own speech-to-text on Nano. A streaming zipformer running int8 under LaunchAgent com.aios.stt, websocket on :8992 with a health endpoint. Real-time factor 0.06, first partial about 60 ms.
2. Benched it instead of trusting it — and it failed. 40 receptionist utterances in clean and telephone-band audio: 23% word error, and only half to six-in-ten phone numbers transcribed digit-perfect. A receptionist that reads your callback number back wrong is worse than no receptionist.
3. Tried the cheap fix first: insurance hotwords. Medicare, Medigap, annuity, My Legacy Management and the rest. The library needed a vocabulary file nobody documents, which cost a block to find. Result: 23% → 19% word error, digits unchanged. Honest verdict: not enough.
4. So the design changed rather than the tuning. Two-pass: the streaming model still does partials and decides when you stopped talking, but the final transcript is produced by Whisper large-v3-turbo running in-process on Nano’s GPU, with a 3-second timeout that falls back to the streaming text. 0.8% and 0.4% word error, 100% of phone numbers digit-exact, and the final arrives in 0.30 s — faster than the reference Whisper server at 1.10 s.
5. Picked the language model by measurement, not by size. Given a realistic transcript and asked to pull out name, digits and consent: the 3B model got the digits wrong both times (fast, useless), the 27B model was right but took 14.6 s (correct, unusable on a live call), and a mid-size model was right both times in 4.1 s warm. That one won.
6. Pinned it warm. A cold model is a 20-second silence on a live call. A LaunchAgent keeps it resident and writes a health line every 4 minutes across all five services.
7. Abandoned the softphone route on evidence, not on frustration. Every SIP attempt from a softphone into LiveKit Cloud answered “404 No trunk found” — with allowlists, with auth, with numbers, with every trunk variant. LiveKit Cloud expects carrier-style origination. That path is dead and, as it turns out, not needed.
8. Proved the whole call without a phone line. Dispatched the agent into a room and published a recorded caller into it, with an env-gated test mode in the worker. Room call-test-223536 saved a correct record: name, number, consent, category and a full reason string, with both sides of the transcript and per-turn timings in the log.
9. Fixed the one real bug that test exposed. The model tried to save the record before it had the reason, inferring your consent from “yes that’s right.” Now the save is refused unless name and reason are present, a deterministic normalizer owns the digits rather than the model, and the call ends after exactly one confirmation.
10. Rented the number and hit the last wall. +1 (760) 965-5049, the first number free on every plan. The dispatch rule exists and points at our agent. Binding the two together fails server-side on every API variant — which is why step 1 above is yours.

Architecture — the two flows

Everything inside Path B after the carrier hop runs on hardware you own, at zero per-minute inference cost. The only rented pieces are the number and LiveKit’s media edge. Path A stays attached as the backstop either way — the GHL agent fires on ring-timeout regardless of where the forward points.

Proof — claim, method, evidence

Speech-to-text bench — why the design changed twice

Deepgram-API compatible (added 06:30 PT 09-13). aios-stt answers Deepgram's own wire protocol on /v1/listen (Results, UtteranceEnd, Metadata; KeepAlive, Finalize, CloseStream). Proof: LiveKit's official Deepgram plugin pointed at base_url=http://127.0.0.1:8992/v1/listen returned 15 interim and 1 correct final transcript with END_OF_SPEECH. So "own Deepgram" is literal: a drop-in, self-hosted replacement, $0 per minute.

40 scripted receptionist utterances — names, phone numbers, reasons, confirmations, times — each rendered clean and through a telephone-band filter. Paired cells are clean / phone. Word error is measured after a shared normalizer, so no engine is rewarded for punctuation. v1’s final is near zero because the streaming decoder has already emitted its text; v2 spends 0.30 s on the second pass and buys a 22-point error drop and every digit. First partial is measured from the first audio frame and is bound by the harness’s chunk size, which is why it is identical across v1 and v2 — the model’s own first partial is about 60 ms at a real-time factor of 0.06. Raw runs: stt-bench/BENCH-run1-greedy.md, BENCH-run2-hotwords.md, BENCH-run3-v2.md, results-run*.json.

Latency chain — measured on the end-to-end test

So a typical turn lands around 2.5–4.5 s and a bad one near 8 s. That is the honest number and it is the thing your test call should be judged on as much as the wording. If the pauses feel long, the lever is the model tier, and that is a measured trade against digit accuracy — the fast model got the phone number wrong every time.

Controls

Path A — run from M2

Path B — on Nano

Open decisions — 4

Also done, not asked for

• A reusable speech benchmark kit — 40 receptionist utterances, clean and telephone-band, shared normalizer, digit-exactness scoring. Any future speech engine gets judged against the same ruler instead of an impression. stt-bench/
• Six blocks recorded so none of these walls costs an hour twice: the GHL probe that creates, the LiveKit command wrappers, the softphone dead end, the undocumented vocabulary file, the broken number binding, and a private-directory requirement in the record store.
• Two research documents that close questions rather than open them: the English/Spanish options table, and the self-hosting decision table with a numeric trigger.
• Exact click-paths written down for both UI steps, including the revert value, because neither has an API.

Files: comms/outputs/2026-09-11-voice-agent-lane/ — this page, README.md, agent-config.json, ghl-voice-agent.sh, GHL-FORWARDING.md, NUMBER-BINDING.md, EN-ES-PLAN.md, SELF-HOST-EDGE-PLAN.md, stt-bench/. Worker, models and LaunchAgents live on Nano. Path B artifacts named in the proof table were produced and read on Nano during the 09-12/09-13 session; this page was assembled on M2 and does not re-verify them.