Introduction

Rime, the voice company behind the Mist and Coda text-to-speech models, launched Prism on October 28. It's a speech-to-speech model, so it takes a caller's audio in and gives spoken audio back, with no separate transcription and TTS steps stitched together.

The twist is in the middle. Prism listens to the actual audio, writes out what it's going to say as text, and only then speaks. That text layer is something you can read, check and rewrite before anyone hears it. It's a closed model, and you can try it today on Rime's free tier.

What Shipped

From Rime's launch post, the Prism page, the pricing page and the WebSocket docs:

  • Modality: speech in, speech out, with a text reply generated before the audio.
  • Access: closed weights. You use it through a WebSocket API or Rime's SDKs for Python, Node.js, Go and Rust, plus LiveKit and Pipecat integrations.
  • Audio format: you send 16 kHz mono PCM16 and get back 24 kHz audio plus the text. Rime says 8 kHz phone audio works in and out too.
  • Speed: 0.57 seconds to first audio on the Artificial Analysis harness, by Rime's own run.
  • Tools: tool calls run in parallel with speech, so the agent can look something up while it talks.
  • Turn-taking: half duplex at launch, with interruption handling.
  • Fine-tuning: Rime calls it the first speech-to-speech model you can fine-tune. Self-serve RL and supervised fine-tuning APIs are being built with beta customers, and fine-tuning and custom voices are Enterprise-only for now.
  • Deployment: Rime cloud on every plan. VPC and on-prem are Enterprise options.

Why "Text Before Speech" Matters

Most voice agents today are either a chain (speech-to-text, then a language model, then TTS) or a fused speech-to-speech model. The chain is easy to control but throws away tone, hesitation and laughter once the audio becomes plain text. Fused models hear all of that but give you nothing to inspect before they talk.

Prism tries to keep both. It decides what to say from the audio, so it can still react to a frustrated or laughing caller. But every word passes through a text checkpoint first, where your guardrails, required disclosures or pronunciation fixes for names and account numbers can step in.

The Numbers, With Caveats

These are Rime's own figures on Artificial Analysis's τ-Voice benchmark, which scores tool use and instruction following on simulated customer-service calls. Treat them as vendor-reported.

  • Prism: 56.6 on τ-Voice, 0.57 s to first audio, no reasoning step.
  • GPT-Live-1 (reasoning on): 59.3, 1.24 s.
  • GPT-Realtime-2.1 High (reasoning on): 45.7, 1.21 s.
  • Gemini 3.8 Live (reasoning off): 30.1, 1.18 s.
  • Telecom slice: Prism 82.5 against 45.6 for GPT-Live-1.

So Prism isn't the top scorer overall. Its pitch is that it gets close to the reasoning models at about half the wait. Rime also shared a customer test with Revmo, a food-ordering agent: after training on 10 hours of real calls, Prism placed 94% of orders only after the caller confirmed, against 57% for the OpenAI model Revmo runs today.

One note: the CEO's launch post on X said about 300 ms response time, while the blog and pricing page say 0.57 s to first audio. I'm going with the published 0.57 s.

How to Try It Today

Sign up at app.rime.ai. Every account gets $50 of usage a month with no card, and that covers Prism. Past that:

  • Flex: $0.080 per minute of call audio, no monthly fee.
  • Scale: $0.064 per minute, $100 a month in credit that rolls over up to $200.
  • Enterprise: volume rates, fine-tuning on your calls, custom voices, VPC or on-prem.

At $0.08 a minute, the free $50 works out to roughly 625 minutes of Prism a month.

From a backend, connect to wss://prism.api.rime.ai/v1/realtime with your API key as a Bearer token. Each connection is one conversation.

Pro Tip

Because Prism writes its reply first, you can test how it handles tricky details before you build anything around it. Use a first test call script like this one, packed with the numbers and names voice agents usually fumble:

Hi, I'm calling about order 4417-B. It's under Siobhan Nguyen, and the callback number is 555-0182. Can you move pickup to Friday at 6:15?

Then read the text Prism generated against what you heard. If the order code or name is wrong in the text, that's where a pronunciation rule or guardrail belongs.

Why It Matters for Audio Creators

Prism is built for call centers, not music or narration. Still, it's a sign of where voice is heading: models that hear emotion in audio and still give you a script you can edit. If you build interactive characters, voice games or talking companions, that mix of expressiveness and control is exactly what's been missing.

The limits are real for creative work, though. There's no custom voice outside Enterprise, turn-taking is half duplex, and Rime hasn't listed supported languages for Prism yet.

Conclusion

Rime Prism is a closed speech-to-speech model that listens to audio, writes a reply you can inspect, then speaks it in about 0.57 seconds. It costs $0.08 a minute after a free $50 monthly allowance, and fine-tuning is on the way. If you're building a voice agent and the "fast but uncontrollable" trade-off has bothered you, it's worth a free-tier call.

—Anabel ♡