Skip to content

use case

verified July 2026

TTS for healthcare scheduling calls.

A clinic’s voice traffic is two workloads wearing one coat: the front desk answering live, and the scheduling engine calling out — recalls, confirmations, waitlist backfills. The economic case is that one line can wear both.

01

Hear the workload

A scheduling exchange, both sides rendered on this API — the caller moving a dialysis-adjacent appointment for the third time, the agent resolving it in two turns.

Audio — A scheduling line — both voices are this engine

rendered 2026-07-16

maya

“I need to move my mom's appointment again. Third time — her dialysis keeps shifting and I can never get through to a person.”

agent

“You're through now, and I have her chart open. Doctor Okafor has Thursday at ten or Friday at eight fifteen — which fits better around the dialysis?”

maya

“Thursday at ten.”

agent

“Done — moved, no fee, and the intake forms carried over. I've texted the confirmation, and I'll remind you both the night before.”

both voices synthesized — jenny and ava, stock

02

One line, two workloads

The desk is busiest mid-morning; the recall list runs at lunch and the confirmations at dusk. Because a line prices concurrency rather than characters, the outbound jobs ride the same line the desk uses at zero marginal cost — $150 covers the desk, the recalls, and the confirmations together. On a character meter each workload bills separately, and the recall list gets shortened the first time finance reads the invoice.

03

What a practice demands of synthesis

  • Speak as the practice, not a stock voice: ten seconds of whoever runs the desk becomes the clinic’s voice on every call.
  • Match the transport to the job: the live desk holds a WebSocket per call; the confirmation job uses one-shot WAV synthesis and needs no socket at all.
  • Keep the jobs auditable: the scheduler and the desk run on separate API keys, each readable at GET /v1/usage.
  • Cover the patient roster’s languages — one cloned identity across 23 languages.

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