# LiveKit text to speech for voice agents

A LiveKit voice agent whose users speak more than one language needs a voice that reads mixed sentences. Point LiveKit's OpenAI plugin at Paxa with a base URL and an API key, and the agent speaks English, Thai, and Mandarin in any Paxa TTS Flash v2 voice. The same plugin adds Paxa transcription.

## One plugin, pointed at Paxa

The speech and transcription objects come from LiveKit's OpenAI plugin, set with a Paxa base URL, key, model, and voice. The notes on each argument are the setup guide's own.

## From a key to a speaking agent

The agent keeps its LLM and its session code. Paxa takes the speech side, and the transcription side when you add it.

1. **Install the plugin**: Install livekit-agents with its openai extra, and set PAXA_API_KEY to a key from the dashboard.
2. **Point the TTS at Paxa**: Create openai.TTS with base_url set to https://api.paxalabs.com/v1, model set to paxa-tts-flash-v2, a Paxa voice id, and response_format set to pcm, the format the plugin plays without decoding.
3. **Add transcription**: Create openai.STT with the same base URL and model paxa-stt-lite-v1-preview, and pass both to AgentSession with a VAD to find where each turn ends.

Paxa TTS Flash v2 lists at $10 per 1M characters. Paxa STT Lite lists at $0.50 per hour of audio, billed by the millisecond.

## Questions

### What is LiveKit Agents?

LiveKit describes Agents as a realtime framework for voice, video, and physical AI agents. A voice agent streams audio through speech to text, an LLM, and text to speech, and Paxa can serve the speech to text and text to speech stages.

### Which voices can the agent use?

The agent can use any of the 50 Paxa TTS Flash v2 voices, and each one speaks English, Thai, and Mandarin. [GET /v1/voices](https://paxalabs.com/docs/api/voices) lists their ids, and [the voice library](https://paxalabs.com/voices) plays them. OpenAI's voice names also work and map to Paxa voices.

### When does a long Thai reply start speaking?

The agent speaks a reply one sentence at a time. Thai written without punctuation counts as one sentence, and a long Thai reply then starts speaking only after the LLM finishes it.

### How does transcription work in LiveKit?

`openai.STT` sends each finished turn to Paxa as one recording, and the session's VAD finds where each turn ends. Pass `language` as a hint, such as `th`, or set `detect_language` to `True` to let the model detect the language. Paxa STT Lite is a Research preview.

## Start with free credits

Sign in with Google, GitHub, or Hugging Face and spend 100 one-time free credits on your own text.