Tasks and inputs
Prism is listed for Audio, Audio-to-audio. It accepts audio, text and produces audio. Use the publisher’s input format and preprocessing for your chosen deployment.
Rime Prism is a speech-to-speech model for enterprise voice agents that settles each reply as inspectable text first, with parallel tool calls.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: No published price on the launch page; free trial available. Through LiveKit Inference, usage is billed to your LiveKit account. Confirm rates on your own account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Compliance teams dislike speech-to-speech models for one reason: nobody can read what the model is about to say. Rime Prism, launched on 28 October 2026, is built around that complaint. It settles each reply as text before turning it into audio, so a rule, a rewrite or a guardrail can step in at the moment of decision.
It is still one end-to-end model, not a text model glued to a TTS engine. The text is an intermediate checkpoint, and the model keeps listening while it talks.
For a bank or a telco, the inspectable layer is the feature. You can block a promised refund, swap a product name, or tag content before a customer hears it. Phonemic conditioning handles the words generic voices mangle, such as surnames, account terms and branded plans.
Tool calls run in parallel with speech. The agent can say something useful while a lookup is in flight, and filler phrases can be configured per tool.
| Item | Detail |
|---|---|
| Model ID | rime/prism on LiveKit Inference |
| Alternative | Rime's own plugin, livekit-plugins-rime[realtime], Python 3.11+ |
| Output | Audio only |
| Turn detection | Server-side; cannot be disabled |
tool_choice |
auto or none only |
| Reported reply time | ~300 ms |
| Endpoint region | Global (no EU endpoint listed) |
| Access | Free trial; LiveKit bills usage to your LiveKit account |
The constraints are real. Audio-only output means you cannot pair it with a separate TTS voice. Imported context accepts text and completed tool calls, not images. There is no session resumption, so a dropped session needs a fallback adapter. Create one model instance per session. Node.js is not supported by the plugin.
On a tau-Voice run without reasoning before speaking, Rime reports an overall 56.6 and 0.57 s to first audio. It says two models score higher overall but take 1.24 to 1.35 s to start talking. In the telecom slice it reports 82.5 against 45.6 for a GPT-Live model. Treat all of this as publisher-reported. For context on the alternatives, see GPT-Realtime-2.1 and Gemini 3.8 Live.
Rime does not label Prism GA or preview; we list it as generally available because it has a public trial and a LiveKit listing. Some outlets reported it a week after the 28 October launch. Per-minute pricing figures circulate, but none appeared on a Rime page we read, so we print none. Check the price on your own account before you budget. Supported languages and the full voice list are also unstated; only a voice called astra appears in examples.
Replay a set of recorded support calls through your agent with Prism, and log three things: how often the text checkpoint caught something you would have blocked, time to first audio from your users' regions, and whether names come out right with and without phonemic hints. Then try barge-in. Server-side turn detection you cannot switch off will suit some call flows and frustrate others.
More options live under audio-to-audio models.
Compare the model on the work you need to complete.
Prism is listed for Audio, Audio-to-audio. It accepts audio, text and produces audio. Use the publisher’s input format and preprocessing for your chosen deployment.
Compare pronunciation, voice consistency and time to the first playable audio. Check the service’s billing unit for the actual script.
Use the model identifier and account requirements for the selected API route. Confirm its request limits and output format with one complete task before connecting a larger workflow.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Liquid AI
LFM3-Audio-2B is Liquid AI's open 2B speech model: ASR, TTS, voice chat and interleaved text-and-speech output from one local checkpoint.
Alibaba
Qwen-Audio-3.1-Realtime Flash is Alibaba's lighter full-duplex speech model with audio and text in and out, function calling, web search and a 262K context.
ElevenLabs
Eleven Realtime v1 Max is ElevenLabs' reasoning speech-to-speech agent for support and multi-step tool workflows, OpenAI Realtime-compatible.
ElevenLabs
Eleven Realtime v1 Mini is ElevenLabs' fast, low-cost speech-to-speech voice agent stack, with an OpenAI Realtime-compatible API and 90+ languages.
xAI
Grok Voice 3.0 is xAI's speech-to-speech model for real-time voice agents, replacing Grok Voice 2.0 at the same $0.08-per-minute price.
Amazon
Nova 2.5 Sonic is Amazon's speech-to-speech model on Bedrock for real-time voice agents, with async tool calling, polyglot voices and a 256K context.