Estimate a Voice Agent's Per-Minute Cost Before You Commit

You have a voice-agent quote in front of you, and the line that matters is the one that scales with every spoken word. A vendor can make the voice sound smooth while the stack spends most of its money on speech synthesis. Treat the minute as a unit and force each component to show its share.

TTS is the largest line in the minute

Speech synthesis scales with every word the agent says, so the TTS bill is the first line in the worksheet.

Inworld announced the TTS-2 family, consisting of Realtime TTS-2 and Realtime TTS-2 Flash, on September 2, 2026. Realtime TTS-2 Flash recorded a 25 ms time to first byte in Coval benchmarking. The Flash number gives you a speed comparison point when you are choosing a stack.

Inworld's on-demand TTS price starts at $25 per million characters, and its monthly Growth plan lowers the listed rate to $12.50 per million characters. The spread is the case for committed-use pricing: predictable volume can move the per-character rate.

A per-minute worksheet beats a vendor quote

Start with characters per minute from your own script, not a vendor sample. Write the agent's likely reply length for a normal turn, then multiply by the number of turns you expect in a minute. You are done when you have a low and high range that your team can defend.

Convert the TTS line by multiplying characters per minute by the per-million-character rate and dividing by one million. Split the minute into STT, TTS, and LLM. STT captures speech, TTS produces speech, and the LLM chooses the reply. The TTS share changes with wordiness: a wordy agent grows it, while a terse agent may let the LLM share dominate. Keep the split visible instead of a blended rate.

Apply committed-use pricing only after you can defend the volume. A monthly plan prices predictable flow; it does not replace a volume estimate. A small pilot may make the on-demand rate safer. Steady traffic can make the committed rate the real cost. Finish when the plan choice follows from usage, not from a sales call.

Keep the TTS rate as the dominant line, then add the other lines as checks. The worksheet should show TTS per minute, LLM per minute, and STT per minute. A small TTS line suggests an unusually terse script or model. A dominant TTS line points to the main cost driver.

The hidden specs need a live test

Run a first-chunk latency test with the same prompt, network, and endpoint. Request the first audio chunk, not the full sentence. A low first-chunk number can change the feel of the conversation even when total generation time is similar. Realtime TTS-2 supports sub-100 ms first-chunk latency, and the TTS-2 models support cross-language voice coverage of more than 200 languages and more than 500 dialects while maintaining the same voice identity.

For agents that switch languages, test cross-language identity. Check whether the same character remains recognizable as the language changes, rather than whether the voice sounds natural in each language. The coverage claim above is the one to verify in your own test.

Test inline steering with a short instruction at the start of the text. Have the model deliver a line as calm, urgent, or playful, then listen for whether the delivery changes without changing the words. Realtime TTS-2 allows developers to pass a natural-language delivery instruction inline at the start of the text. When the instruction leaves the delivery unchanged, the steering feature is not doing the work you need.

Do not let a benchmark number replace a product test. A latency result can be true in a lab and useless in production. Put the same test through your network, your payload, and your user-facing timeout. The number you keep is the one your users hear.

After Talkpal added Inworld Realtime TTS, a four-week A/B test showed TTS costs down 40%, feature usage up 7%, and retention up 4%. That result belongs in a business case, not in a demo script: cost, usage, and retention can move together when TTS is cheaper and faster.

Bring the estimate to the decision meeting with the TTS share, the latency result, the identity result, and the steering result. Request the missing test for any vendor gap in your worksheet. Probe the volume assumption behind a committed rate that seems cheap.