Try Rime in LiveKit free through September.
ElevenLabs models stopped working on LiveKit Inference on August 31. Switching to Rime takes one line of code, and is free through September. See why callers stay on the line longer with Rime.
Built into LiveKit Inference.
No extra API key to manage, just choose the Rime voice that matches your use case and start testing on real conversations. Usage, rate limits, and billing stay inside LiveKit Cloud. Change one string and redeploy.
One string, and you're live
Voice IDs do not transfer between providers, so pick a new voice from the Rime catalog. Every Rime API voice is supported in Inference.
What you get
- Coda and Mist v3 voice models are available today
- All Rime API voices, browsable in the Rime voices catalog
- Works in LiveKit's no-code Agent Builder
- Free usage through September 30 for teams switching from ElevenLabs
Built for speed and quality
Two model families, one API. Both are trained by people who study how humans speak.
Coda rime/coda
Our default for new agents. Coda pairs an LLM backbone with a dedicated speech inference engine and is trained on real, full-duplex conversations rather than voice actors reading scripts. It sounds like someone who is actually listening.
- Word-level timestamps for aligned transcripts
- Highest voice-quality scores in Rime's human evals
- English, Spanish, French, German, Hindi, Japanese, Portuguese, Arabic
Mist v3 rime/mistv3
Built for speed. Mist v3 is our fastest model and the one to reach for when every millisecond of time-to-first-audio counts, like telephony and IVR at scale.
- Custom pauses for natural pacing control
- Need inline pronunciation control? Use Mist v2
- English, Spanish, French, German
Hear it before you switch.
Same script, two voices. Listen for what a caller hears in the first three seconds.
A number readback, digit by digit. The place robotic voices give themselves away first.
The same arena prompt, two voices. Judge it the way your callers will.
A live support exchange, where tone decides whether the caller stays.
More people finish once they stay.
Completion efficiency asks a narrower question: given that a call reached a certain point, how likely was the respondent to finish? Past the thirty second mark, Rime voices pulled away. The advantage was statistically significant against both providers, and it held up after multiple-comparison correction.
Rime's advantage was statistically significant against both providers.
Read the study results.
Callers vote with the hang-up button.
Voice quality is table stakes. What separates providers is what happens to real call outcomes. Two independent research firms measured it.
"Mist v3 is incredibly fast and deterministic pronunciation is a game-changer. When you're powering live conversations at scale, you can't have a model guessing brand names or proper nouns."Tom Shapland · Product Manager, LiveKit
Find your voice, free
Switch a LiveKit Inference agent from ElevenLabs to Rime and your usage is free through September.
Swap the string
Point your session at rime/coda in LiveKit Inference. No Rime account and no new API key. You're live in minutes.
Run it for real
Put Rime on production traffic through September. Watch your own completion and hang-up numbers.
Decide in October
Keep what works. Usage continues at standard LiveKit Inference rates, and high volume teams can talk to us about enterprise options.
Switch before September ends.
One line of code between your agents and a voice that keeps callers on the phone.