# What Voice AI Costs: STT, LLM, TTS and Telephony

> What a phone voice AI costs across STT, LLM, TTS and telephony, what each part scales with, which conversation patterns raise cost, and how to measure it.

Published: 2026-11-14
Source: https://voicast.jp/en/blog/voice-ai-cost-breakdown

## Key points

- Voice AI cost splits into four parts (speech-to-text, LLM, text-to-speech, telephony), and nearly all of it scales with call length and the number of turns.
- Conversation design moves the bill more than unit prices do. Long prompts, long answers and repeated questions make calls longer and LLM requests bigger.
- Don't trust spreadsheet estimates. Place test calls per scenario, record cost and duration per call, and compare by outcome.

A phone voice AI costs money in four places: speech-to-text (turning the caller's voice into text), the LLM (deciding what to say), text-to-speech (turning the reply into audio) and telephony (the number and the call itself). All four grow with call length and the number of back-and-forth turns, so the fastest way to cut cost usually isn't a cheaper vendor; it's a conversation that finishes sooner. Here's what each part scales with and how to measure your real per-call cost.

## The four cost components

Behind every call, four things are running:

| Component | What it does | What drives its cost |
|---|---|---|
| Speech-to-text (STT) | Transcribes the caller | Call duration (audio streams even while nobody is talking) |
| LLM | Reads the transcript and writes the reply | Turns × tokens sent per turn (prompt, history, tool results) |
| Text-to-speech (TTS) | Speaks the reply | Amount of text the AI says |
| Telephony | Phone numbers and call minutes | Numbers held and minutes connected |

Most APIs bill STT by audio duration, LLMs by input and output tokens, and TTS by characters, but the exact units vary by provider, so check the pricing page of whatever you use. Telephony is typically a monthly fee per number plus per-minute charges; see [Twilio's voice pricing](https://www.twilio.com/en-us/voice/pricing) for one example.

On top of that sit operational costs: the servers that relay audio, monitoring, and storage for recordings and transcripts. If you build the stack yourself, engineering and on-call time belong on the bill too; see [build vs buy voice AI](https://voicast.jp/en/blog/build-vs-buy-voice-ai).

## Why LLM cost grows during a call

Of the four, the LLM is where conversation design makes the biggest difference. Chat-style LLM APIs are stateless: each time the AI replies, you send the system prompt plus the conversation so far.

- Every extra turn makes the history you resend longer.
- A long system prompt is paid for on every single turn.
- A bulky tool response stays in the history and rides along on every turn after it.

Both the number of turns and the size of each request compound as the call goes on. Trimming the prompt to what matters, and returning only the fields the AI needs from tools, helps cost and latency at the same time (see [function calling in live calls](https://voicast.jp/en/blog/voice-ai-function-calling)).

## Conversation patterns that drive cost up

Each of these either lengthens the call or inflates what you send per turn:

1. **Long answers.** Reading a paragraph-long FAQ answer costs TTS characters and call time. On the phone, answer in a sentence or two and offer details another way.
2. **Repeated questions.** Every "sorry, could you spell that?" is another turn. Vocabulary hints and read-back confirmation help capture names and numbers the first time.
3. **No clear purpose up front.** If the greeting doesn't ask why they're calling, you get small talk.
4. **No defined ending.** If the AI doesn't wrap up once it has what it needs, the call runs until the caller hangs up.
5. **Late handoffs.** Dragging out a call the AI can't resolve means you pay for the AI and then for a human anyway.

The flip side: better conversation design is cost reduction.

## Measuring real cost

Estimates built from unit prices miss how real callers behave. Measure instead:

1. With a setup close to production, place several test calls per scenario: booking, rescheduling, a simple question, a handoff.
2. Record duration and cost for each call.
3. Average them by outcome (details collected, question answered, handed off, caller hung up).
4. Listen to the longest calls in each group and look for turns you could remove.
5. Change the setup and measure the same scenarios again.

Keep watching outcome mix and conversation duration month by month once you're live, so you can see what each setup change did. [Metrics for running a phone AI](https://voicast.jp/en/blog/phone-ai-metrics) covers what to track.

## Doing this with voicast

voicast charges ¥10 per minute of call time (excl. tax), covering the speech recognition, LLM and speech synthesis parts together. Phone numbers and lines stay on your own Twilio (or similar) account, so telephony is billed there separately.

Each call's cost is on the result, in `billed_yen` alongside `duration_ms`:

```sh
curl -s https://api.voicast.jp/v1/calls/by-provider/CA1234567890abcdef \
  -H "authorization: Bearer $VOICAST_API_KEY"
# { "outcome": "completed", "duration_ms": ..., "billed_yen": ..., ... }
```

For totals over a period, use `GET /v1/usage`. With `group_by` set to `tenant`, `day` or `channel`, it returns calls, seconds, minutes and an estimated amount (`amount_yen`, before tax) for each group. It counts calls that ended within the period, and minutes are the group's total seconds divided by 60, rounded up. Calls in the test environment are counted at the same price as live ones.

```sh
curl -s "https://api.voicast.jp/v1/usage?from=2026-12-01&to=2026-12-31&group_by=tenant" \
  -H "authorization: Bearer $VOICAST_API_KEY"
```

If you resell voice AI to your own customers, [billing your customers for voice AI usage](https://voicast.jp/en/blog/voice-ai-usage-billing-saas) shows how to use this breakdown.

## FAQ

### Can I know what a call will cost before going live?

Run test calls for each scenario and measure cost and duration per call. The same scenario varies with how people talk and how often they repeat themselves, so look at an average over several calls rather than a single one.

### Am I paying while the caller is silent?

Telephony bills while the call is connected, and speech recognition runs while audio is streaming. That's one more reason to handle silence deliberately rather than letting a call sit idle. voicast itself bills by call time.

### Do test calls cost money?

In voicast, yes: test environment calls are counted at the same price as live calls, and there are no free minutes. Plan your test scenarios and place only the calls you need.
