# What is an AI voice agent? How one works and where it is used

An AI voice agent listens to a caller, decides what to say with a language model, and answers in a synthetic voice. See how the parts fit and what a call costs.

Published 2026-11-06. Paxa Labs.

Calling a business used to mean a menu of numbered options or a wait for a person. An AI voice agent answers the call itself. It listens to what the caller says, works out a reply, and says it out loud, turn after turn.

## What is an AI voice agent?

An AI voice agent is software that holds a spoken conversation. It is built from three parts that run in turn. Transcription writes down what the person said, a language model decides the reply, and text to speech says it.

People reach one by calling a phone number, tapping a microphone button in an app, or speaking to a kiosk or a device.

## How does an AI voice agent work?

1. It listens: Streaming transcription with automatic turn detection writes down each turn and reports when the caller stops. Paxa STT Lite Realtime does this at 12.5 credits per minute of audio received.
2. It decides: A language model reads the turn, the conversation so far, and your instructions, then writes the reply. Through your own systems it can also look up an order or book a slot.
3. It speaks: Text to speech turns the reply into audio and streams it back as it is generated. Paxa TTS Flash v2 speaks English, Thai, and Mandarin, with 50 voices to choose from.
4. It handles interruptions: When the caller cuts in, the agent stops talking and listens. The transcription session is told while the agent's audio plays, and it does not take the agent's own voice for a turn.

## Why does turn detection matter?

A spoken conversation has no send button. The agent has to work out from the audio when the caller has finished, and answer then. Answer too early, and it talks over the caller. Wait too long, and the silence sounds like a dropped line. [Realtime Thai transcription turns](https://paxalabs.com/guides/realtime-thai-transcription-turns) looks at where turns fall in Thai speech.

## Why does a voice agent need streaming?

Each part starts before the part ahead of it has finished. Transcription sends each turn as it ends, the language model streams its reply, and text to speech starts speaking before the whole reply is written. [Realtime text to speech](https://paxalabs.com/realtime-tts) and [realtime speech to text](https://paxalabs.com/realtime-speech-to-text) cover the two voice ends.

## Where are AI voice agents used?

- Customer service lines that answer common questions and hand harder calls to a person. See [the call centers page](https://paxalabs.com/use-cases/call-centers).
- Bookings and reminders for clinics, restaurants, and deliveries.
- Companions and characters that talk back. See [the companions page](https://paxalabs.com/use-cases/companions) and [the AI companion page](https://paxalabs.com/ai-companion).
- Voice control inside apps, cars, and devices.

## What should you decide before you build one?

- The languages your callers speak. Speech covers English, Thai, and Mandarin, and transcription is multilingual.
- The voice. Listen to every voice on [the voices page](https://paxalabs.com/voices) and pick one that fits the brand.
- What the agent may do on its own, and when it hands the call to a person.
- How you tell callers that they are speaking to an AI.
- Which language model writes the replies. Paxa supplies the listening and the voice, and the language model is your choice.

## What does an AI voice agent cost to run?

A call costs the transcription of the caller's audio plus the speech of the agent's replies. The language model is billed by whoever provides it.

| Part | Model | Rate |
| --- | --- | --- |
| Listening | Paxa STT Lite Realtime | 12.5 credits per minute of audio received |
| Speaking | Paxa TTS Flash v2 | 10 credits per 1,000 characters |

Take a 5-minute call in which the agent says 1,500 characters. Listening to the whole call costs 62.5 credits, pauses included, and speaking costs 15 credits. That is 77.5 credits, or $0.08, before the language model.

A new account's 100 free credits cover both parts while you build, with no card. They are for personal use and evaluation. Buy any plan or credit pack, and the account's audio and transcripts are yours to use commercially, including those made before the purchase. [The Terms](https://paxalabs.com/terms#content-and-output) set this out.

## How do you start building one?

1. Hear the voices: Type a reply your agent would give into [the Speech tool](https://paxalabs.com/app/speech) and listen to it in several voices.
2. Test the listening: Open Transcribe live in [the Transcribe tool](https://paxalabs.com/app/transcribe), speak as a caller would, and watch where each turn ends.
3. See a working agent: [The voice agents page](https://paxalabs.com/use-cases/voice-agents) shows how the parts connect in an agent on a phone line.

> Paxa STT Lite Realtime is a research preview. Its output and options may change between releases.

## Common questions

### How is a voice agent different from a chatbot?

A chatbot reads and writes text. A voice agent adds transcription before the language model and speech after it, and it has to manage turns as they happen. [Realtime vs recorded transcription](https://paxalabs.com/guides/realtime-vs-recorded-transcription) explains the listening side.

### Can an AI voice agent speak Thai?

Yes. Paxa TTS Flash v2 speaks Thai, and English names inside a Thai sentence are read in the same voice. Transcription is multilingual and keeps each word in the script it was spoken in.

### Which language model does it use?

The one you choose. Paxa provides the listening and the voice, and your application calls its language model between them.

### What happens when the caller interrupts?

Your application stops the agent's audio, and the transcription session takes the caller's new turn. Telling the session when the agent is speaking keeps the agent's own voice from counting as a turn.