LiveKit is a realtime platform that enables developers to build video, voice, and data capabilities into their applications. Building on WebRTC, it supports a broad range of frontend and backend platforms.
LiveKit is powering high-volume voice applications for enterprises like OpenAI, Oracle, ebay, Spotify, and more.
With LiveKit's Rime integration and the Agents framework, you can build AI voice applications that are responsive and sound realistic.
What the integration includes
The integration adds Rime as a text-to-speech plugin inside LiveKit Agents, LiveKit’s framework for building realtime AI agents. Your agent pipeline handles speech recognition and language model reasoning however you choose, and Rime handles the voice, streaming synthesized audio into the LiveKit session as text is generated so callers hear a response without a long pause.
Configuration takes a few lines. Pick a Rime voice and model, pass your API key, and the plugin manages the streaming connection for you. From there the same setup works across LiveKit’s supported platforms, whether your agent lives in a web app, a mobile app, or a phone line.
Why latency decides voice agent quality
In a phone call or a live web session, the gap between a caller finishing a sentence and hearing a reply determines whether the agent feels conversational. Rime’s models are engineered for low time-to-first-audio, which keeps that gap short even before network and LLM time are added. Paired with LiveKit’s WebRTC infrastructure, the result is natural turn-taking at production call volumes.
Choosing a Rime model for your agent
Every Rime model works through the same LiveKit plugin, so the choice comes down to what your agent needs. Coda is the default for new applications: 184 voices across 8 languages, sub-100ms model latency, and the highest voice-quality scores in our human evaluations. Mist v3 is the pick when time to first audio matters most, reaching roughly 37ms time to first audio, and it adds support for custom pauses. Mist v2 offers inline pronunciation control for cases where exact phonetics matter, like medication names on a healthcare line. Switching models is a one-parameter change in the plugin configuration, which makes it easy to test candidates against real calls.
Handling interruptions and phone-grade audio
Production callers interrupt, and a good agent stops talking when they do. Rime streams audio with word-level timestamps, which lets your LiveKit agent cut synthesis mid-sentence and keep its transcript aligned with what the caller actually heard. Output formats cover telephony as well as web and mobile, so the same agent logic can answer a phone line in the morning and a browser session in the afternoon.
Scaling to production
When a prototype becomes a product, the constraints change. Rime’s enterprise plans add unlimited concurrency, SLAs, and dedicated support for high call volumes. Teams with strict compliance requirements can route traffic through regional cloud endpoints or run Rime in a dedicated VPC or fully on-premises, keeping caller audio inside their own network while LiveKit handles the realtime transport.
Getting started
More information is available in LiveKit's documentation, and the Rime docs include a complete tutorial for building a voice agent on LiveKit. A good first project is a simple agent that answers a question out loud. You can have one running locally in an afternoon, then swap voices from the Rime catalog until you find the right fit for your brand.
Plus check out this fun example from the LiveKit team, which uses Rime + DeepSeek + LiveKit.
Beyond LiveKit
LiveKit is one of several ways to bring Rime into a realtime application. Rime also connects to Pipecat, Vapi, Daily, SignalWire, and VideoSDK, and every model is available directly over HTTP and WebSocket APIs if you’re building your own pipeline. One API token works across all of them.
More integrations coming soon!
More from the lab
Writing for the ear: Prompting your TTS to sound human
Rime was founded by linguists. We spent years studying how people actually talk before we trained models on it, so our prompting guide teaches something different: how to transcribe speech the way it’s spoken. Below is the short version of that guide and gives some context behind why we do it this way.
.png)
Text to Speech for Business: How to Choose a TTS App That Scales
This guide shows you how to pick a TTS app that holds up in production.
What Is Back-Channeling in Conversational AI? A Plain-English Guide
This guide explains back-channeling in conversational AI in plain English. You'll learn what it is and how a voice agent decides when to use it. You'll also see how it produces a natural cue and why timing makes or breaks the effect.