Create each key on its own page, paste it below, then copy the finished lines into your project's .env file. What you paste stays in this tab. It is never sent to Wirestack or saved.
Responses API through LiveKit Inference. Use the Responses API with your LiveKit Inference credentials, with structured output and web search. REST is the default. You can turn on WebSocket transport to reuse the connection across turns.
Live translation. OpenAI gpt-realtime-translate and Google gemini-3.5-live-translate-preview, both as beta translation models.
New speech plugins. Zoom AI STT, Kalpa kalpa-tts-1 TTS (English and Hindi), Rime STT and Prism speech-to-speech, Fish Audio transcribe-stream-1 realtime STT, and new public Respeecher TTS models in Swedish, Brazilian Portuguese, German and Arabic.
Security. The Respeecher plugin now sends WebSocket API keys in handshake headers, not in the URL.
Prerequisites
A LiveKit Cloud project with LIVEKIT_URL, LIVEKIT_API_KEY and LIVEKIT_API_SECRET set.
Python with livekit-agents 1.8.11 or later.
Rime STT and Prism need Python 3.11 or later, and Rime STT needs the stt extra.
Steps
Upgrade livekit-agents to 1.8.11.
Create the LLM with the new Responses class. Official example:
Python
from livekit.agents import inferencellm = inference.responses.LLM("openai/gpt-4.1", use_websocket=True)
Leave out use_websocket=True to use REST.
Check the behavior changes below before you deploy.
Behavior changes in this release:
Inference Responses now uses each known model's lowest reasoning effort by default. Set reasoning=None to keep the provider's default.
A late STT final that arrives after the user starts speaking again now joins the resumed turn.
During speech that can't be interrupted, user transcripts are now kept, without stopping playback or generating a reply.
Sarvam TTS model bulbul:v2 is sunset and bulbul:v3-beta is removed.
Test it
Pass the new llm to your agent session and start a call. The agent should answer as before. If answers seem less thorough, set reasoning=None and compare.
Not confirmed
The release notes do not show code for the translation models or the new plugins.
The release notes do not show how web search or structured output are switched on for inference.responses.LLM.
Wirestack