TTS for AI customer-support agents.
8,000 min
one busy stream, every month
One socket per call, an utterance per turn, and playback that stops mid-word when the caller talks over it. Synthesis sits at the floor of the turn budget so the headroom goes to the stage you cannot buy down.
The latency budget
Humans hand a conversation back and forth in about 200 milliseconds, and callers start interrupting an agent that takes much past a second. A working budget for one turn, drawn to scale:
fig.
One turn of a support call, working budget per stage
Fig
Speech-to-text final result
150, 300 ms
LLM first token
200, 400 ms
Synthesis first audio
at the floor
Network + turn detection
100, 200 ms
callers interrupt past here
The whole turn stacks those budgets. Synthesis at the floor of its row hands the headroom to your LLM, the slowest stage you cannot buy down.
The cost shape
A desk doing 400 four-minute calls a day speaks about 48,000 minutes a month, six busy streams: $240 to $800 each on published meters, $150 here. Once characters cost money, someone asks the agent to say less.
What the desk demands of synthesis
- Stop mid-sentence and resynthesize when the caller barges in, persistent WebSocket, one socket per call.
- Speak the brand’s voice on every ticket: instant cloning from a ten-second reference, no training job.
- Survive your current vendor’s bad night: automatic failover.
Notes
How do I size streams for a support desk?
Count busy-hour simultaneous conversations, not monthly minutes. A desk stacking six concurrent calls at peak needs six streams; overflow past them spills to burst streams at $10 a stream-day rather than ringing busy.
What happens to the agent voice during an interruption?
Barge-in is one message on the open socket: playback stops mid-word, the answer is resynthesized inside the same turn budget, and the interrupted characters cost nothing because nothing is metered.
A key and one stream to build this on, the same production API this page measures.
Get a key for this use case