Skip to content

use case

verified July 2026

TTS for insurance claims calls.

A claims line has two modes: the ordinary Tuesday, and the morning after the storm, when every policyholder in a county calls at once. The workload is defined by its worst day, and pricing capacity for that day is the whole problem.

01

Hear the workload

A first-notice-of-loss exchange, both sides rendered on this API. Notice the agent’s second turn: a long, structured read-back — tow, shop, rental — delivered as one utterance, well inside the 2,000-character cap.

Audio — First notice of loss — both voices are this engine

rendered 2026-07-16

alex

“Someone rear-ended me on the 405 about an hour ago. Everyone's fine, but the bumper's hanging off and I've never filed a claim before.”

agent

“I'm glad everyone's okay — that's the part that matters. Your claim is open as of right now, and I've texted a link for photos of the bumper.”

alex

“Do I need a police report? A tow?”

agent

“No report for this one. The tow is covered and twelve minutes out, the shop is booked for Tuesday, and your rental is waiting at the same lot.”

both voices synthesized — lewis and mia, stock

02

The surge problem

Catastrophe traffic multiplies call volume overnight and holds it there for days. Sizing committed lines for the storm means paying for idle capacity all year; sizing for the ordinary Tuesday means busy signals on exactly the day policyholders remember. Burst lines price the storm itself: traffic past your committed lines spills over at $10 a line-day, only on the days the surge is real, with a cap you set in the dashboard.

03

What a claims call demands of synthesis

  • Patience with corrections — claimants fix dates and details mid-sentence, so barge-in has to stop the read-back and pick it up again.
  • The policyholder’s language: one cloned voice carries the same identity across 23 languages.
  • A record that answers auditors: watermark at generation, per-key accounting, and reference audio that never enters a training set.

See also

Related sheets.

Thirty minutes, your production script, the live latency readout — measured in front of you.

Bring this use case to the call