Every call answered. Even the thousandth.
Voicewire puts real-time AI voice agents on disaster hotlines. When a flood, earthquake or storm sends call volume past what any team can staff, every caller reaches a voice that listens, their location and needs are logged for responders, and life-threatening calls go straight to a human.
Early stage: the real-time voice engine is built and running. Telephony and field pilots are next.
- Incoming calls / min
- 90
- Answered by an agent
- 0
- Urgent, routed to dispatch
- 0
- Callers on hold
- 0
Illustrative simulation, not measured data. Assumes a 24-line human team at four minutes per call.
Hotlines are staffed for an ordinary day
Emergency and helpline numbers are sized for normal demand. A disaster brings the whole region to the phone at once, in the hour when answers matter most.
The event
A quake, a flood crest, a storm landfall. Thousands of people reach for their phones at the same moment.
Lines saturate
Every operator is on a call. New callers hear hold music, busy tones, or nothing at all.
The queue hides the urgent
Someone trapped by rising water waits behind a hundred callers asking where the nearest shelter is.
Every line picks up
Agents answer every call in parallel, handle routine questions, and move urgent callers to the front of the human queue.
One call, start to finish
The agent talks like a calm operator, reads back what matters, and fills a structured incident record while the caller speaks. Callers can interrupt at any point, and the agent stops and listens.
- Agent
Flood line. Are you safe right now?
- Caller
Water's coming into the house. My mother can't walk.
- Agent
I'm logging that now. What's your address, or the nearest landmark?
tool: log_incident(people=2, mobility="1 cannot walk", hazard="rising water") - Caller
Behind the school on River Road.
- Agent
Behind the school on River Road. Is that right?
read-back before anything is saved - Caller
Yes.
- Agent
Someone who can't move with water rising is urgent. I'm connecting you to a dispatcher now. Stay on the line.
escalate: warm transfer to human dispatch, record attached
Illustrative call. Names and places are invented.
A voice engine built for the surge
Voicewire runs its own streaming speech pipeline, so a reply starts playing while it is still being written, and one server holds many calls at once.
Streaming end to end
Speech is transcribed as the caller finishes, the reply streams token by token, and audio starts on the first clause.
Interruptible
Voice activity detection hears the caller over the agent. Playback stops, and the agent remembers only what the caller actually heard.
Agentic, with guardrails
Tools run in parallel with timeouts. Actions that change anything need the caller's spoken confirmation, enforced in code.
Self-hosted or API
Speech models run on our own GPUs. The language model swaps between self-hosted vLLM and hosted APIs, with automatic failover.
One server, many calls
Stateless model adapters, shared batched inference, and an event loop that never blocks, built for concurrent sessions.
Measured every turn
Each turn records speech-to-text time, time to first token, tool time and time to first audio.
Where we are
Honest status. The engine works today; the pieces that put it on real phone lines come next.
| Capability | Status | Notes |
|---|---|---|
| Real-time voice pipeline | Built | Streaming speech-to-text, language model and text-to-speech with barge-in. |
| Agent core | Built | Conversation memory, tool calling, confirmation for actions, model failover. |
| Phone lines | Next | Inbound and outbound calls over the public phone network and SIP. |
| Surge scaling on AWS | Next | GPU inference on EC2, admission control, hold-and-callback, multi-region failover. |
| Responder dashboard | Planned | Live incident queue, map and priority, exportable to existing dispatch tools. |
| Local languages | Planned | Speech models evaluated on noisy, narrowband phone audio in the languages of pilot regions. |
| Field pilot | Planned | A disaster drill with a partner agency, measured on answer rate and escalation accuracy. |
Run a drill with us
We are looking for disaster management agencies, NGOs and helpline operators to shape the first pilot. If your lines go quiet on a normal day and overflow on the worst one, we would like to talk.