Three voice and video launches in one week all try to skip the old “turn it into text first” step. Prism skips it halfway, on purpose.
The short version: Prism is a new voice AI from Rime, launched October 28, 2026, for businesses that answer phone calls with AI. It hears the caller’s real voice, writes its reply as text so the company’s rules can check every word, then speaks, in about half a second. It’s English only and sold to businesses, so you won’t download it. You may hear it when you call a clinic or a bank. Block’s Dessa, by contrast, never uses text and isn’t public yet.
What is Rime’s Prism?
Rime is a San Francisco company, founded in 2022, that makes AI voices for businesses: the voice you hear when an automated agent answers a call. It raised $24 million from investors in July and said it would focus on speech-to-speech models. Prism is the result.
Rime calls Prism a speech-to-speech model: sound goes in and sound comes out, from one AI model instead of three separate ones stuck together. Rime’s CEO, Lily Clifford, admits she didn’t see it coming: “Eighteen months ago I said, very publicly and very confidently, that speech-to-speech models were a decade out from real use.”
Five terms, in plain English
- Speech-to-speech
- An AI that takes in audio and answers in audio, rather than working only with written words.
- Pipeline (or cascade)
- The older setup: one tool writes down what you said, a chatbot writes a reply, a third tool reads it aloud.
- Half-duplex
- Taking turns, like a walkie-talkie. Prism works this way, but you can still cut in.
- Full-duplex
- Listening and talking at the same time, the way people do on a phone call. Dessa works this way.
- Tool call
- When the AI looks something up or makes a change, such as finding your appointment or freezing a card.
How does Prism work?
Most voice bots you’ve talked to run a three-step pipeline. They turn your speech into text, a chatbot writes an answer, and a voice reads that answer out. Your tone gets lost in step one: a stressed “fine” and a happy “fine” look the same on paper.
Prism is trained as one system with three jobs: “it listens, it writes what it’s going to say, and when it speaks, it is still hearing the user.” The written step is still there, but it comes after Prism has heard your voice, and it gives a business a place to check the reply against its rules before you hear it.
Is that really different from the old three-step setup? Rime says yes: in its words, “The text is the output checkpoint, not the input to the reasoning.” We can’t look inside the model to confirm that, so treat it as Rime’s description.
- It takes turns, but listens throughout. Rime’s pricing page lists “Half duplex at launch, with interruption handling.” If you talk over it, it stops and answers what you just said.
- No dead air. While it looks something up, it says something like “Let me check that,” according to Rime’s documentation.
- English only, four voices, called aurora, wawona, brando and astra.
- Cloud by default. Big customers can run it on their own servers instead.
Listen to a Prism call
You can’t talk to Prism in a browser, but Rime’s homepage plays a recorded call, about a minute and a half long: a patient named John rescheduling a doctor’s appointment. We played it through and followed the transcript and the “under the hood” panel beside it.
The moments worth noticing, with the tool calls and rules the panel logged:
- It spells your name back. After “John Smith” it says “J-O-H-N S-M-I-T-H. Is that right?” The side panel shows the rule behind it: “Spell back the name before using it.”
- No records until you’re verified. It asks for a date of birth first, under the rule “No record access until identity is verified.”
- Stress gets noticed. The caller says “Sorry, I’m like super stressed out.” The panel notes “Caller sounds stressed,” and Prism answers “I understand. Let’s get this sorted for you” before booking. He said it in words, so a text-based bot could catch this one too; Rime says Prism also picks up stress from tone of voice.
- Contact details are optional. It offers a text or email confirmation; he says no, and the panel logs “Caller declined. No phone or email requested.”
From our listenThe caller asked for “middle of November.” Prism searched November 1 to 20 and offered the 5th and the 6th, which is early November. He took the 6th anyway. It’s a small slip, but it’s in the company’s own showcase demo, so it’s worth checking dates when an AI agent books something for you.
Rime’s launch video for Prism. Video: Rime.
How is Prism different from Dessa?
Five days before Prism, Block announced Dessa, another speech-to-speech model. Both hear your voice directly instead of a transcript, and both are pitched as sounding human. They make opposite bets on text:
Old-style phone bot
- How it hears you
- Turns your speech into text first
- Text in the middle?
- Yes, before it thinks
- Talking over it
- Usually waits for you to finish
- Can you use it?
- You already do, on many support lines
Cheap and common, but tone gets lost.
Rime Prism
- How it hears you
- Works with your voice directly
- Text in the middle?
- Yes, after it thinks, so a business can check it
- Talking over it
- Takes turns, stops when you cut in
- Can you use it?
- Launched for businesses Oct 28, 2026
Built for companies that need to control every word.
Block Dessa
- How it hears you
- Works with your voice directly
- Text in the middle?
- None at all
- Talking over it
- Listens and talks at the same time
- Can you use it?
- Not yet; safety researchers first
The most human-sounding demos, still a research project.
Rime’s argument for keeping text is control. A bank or a clinic has to be able to read what its AI is about to say, block anything it shouldn’t say, and add required disclosures. In Rime’s words, with fully audio-only models “you get a model that hears, but you lose the ability to read what it’s about to say.” Block says giving up that checkpoint buys Dessa natural timing: in its demos it laughs in the right place and says “mm-hm” while you talk.
The same week brought a third version of this idea, in video. With Simulon’s Reality-Aware Rendering, you don’t type a prompt at all: you act out a scene with an AI character on your iPhone, and the app reads your movement, your voice and the room. Dessa skips text entirely, Prism skips it on the way in, and Simulon skips the prompt box.
How fast is Prism, and how good?
Rime says Prism starts speaking about 0.57 seconds after you finish talking, roughly half the wait of the realtime voice models from OpenAI and Google it tested.
Two cautions. Rime ran these tests itself. And its CEO’s launch post claims a 300-millisecond response, “4x faster” than OpenAI, which doesn’t match the 0.57 seconds in Rime’s own documentation. We’ve used the documented number.
Want help putting this into practice?
Join the free Beginners in AI community on Skool. You get a 14-day starter path with one small build a day, plus a place to ask questions when you get stuck. No experience needed.
Join the free community →Free to joinWill you hear Prism when you call a company?
Possibly, but you won’t be told the brand. Rime sells to businesses with high call volumes. Its Prism page names the phone software company Dialpad as a design partner and quotes Twilio, which runs phone systems for other businesses. Rime’s earlier voice customers include Asurion, Mayo Clinic and Upstart.
Rime’s terms of service ban customers from “representing that a voice is human when it is not.” But Rime doesn’t add a disclosure to calls itself; the business has to. That sits awkwardly beside the CEO’s launch pitch of a model “you can’t tell from a human in conversation.”
In the US, the FCC ruled in 2024 that AI-generated voices in robocalls count as “artificial” under the robocall law, which covers calls a company makes to you. When you call a company yourself, rules vary. The simplest test still works: ask “Am I talking to a person?”
Can you try Prism yourself?
Only if you’re comfortable with a little code. Rime’s free tier gives $50 of use a month with no card, and Prism is included. The quickstart has you run a small Python or Node program and talk to Prism through your computer’s microphone. After the free credit, Prism costs $0.08 a minute, or $0.064 on the $100-a-month Scale plan.
For everyone else, the homepage demo above is the way to hear it. If you’re curious about voice AI you can use today, ChatGPT, Gemini and Claude all have voice modes, and our guide to ChatGPT Voice with plugins shows one.
Why does a half-second matter?
On a phone call, a one-second pause feels long. It’s the gap where you start to wonder if the line dropped. Fast voice AI could matter even more on smart glasses, which Google and Samsung pitch around “help through voice interaction”. That’s our reading: Prism itself is a phone-call product with no announced glasses plans. Our best AI glasses guide covers what’s on sale.
Common questions
Is Prism the same as Siri or Alexa?
No. Siri and Alexa are assistants you use on your own devices. Prism is a voice engine that businesses build into their phone lines to answer customers.
Can I interrupt it?
Yes. It takes turns, but it keeps listening while it talks, and if you speak over it, it stops and answers you.
Does it speak other languages?
Not yet. Prism is English only at launch, with four voices.
Could a company copy my voice with it?
Only with written permission. Custom voices are for enterprise customers only, and Rime’s terms require “express written consent” from anyone whose voice is recorded for that purpose.
How is it different from ElevenLabs?
ElevenLabs is best known for reading text aloud, such as narration and audiobooks. Prism is built for live, two-way phone conversations. Our ElevenLabs guide covers that side.
Is Dessa better?
They aim at different jobs. Dessa’s demos sound more natural, but you can’t use it yet. Prism is available now and lets a business control what it says.
Made for regular people, by a real person
A free daily newsletter that explains new voice AI like Prism and Dessa before you meet it on the phone. Curated and edited by a human, in plain English.
Free forever. Unsubscribe anytime.
Sources
- Rime: Introducing Prism (October 28, 2026)
- Rime: Prism product page
- Rime docs: Prism overview
- Rime docs: Prism quickstart
- Rime pricing
- Rime homepage demo call
- Rime terms of service
- Rime: company page
- TechCrunch: Rime picks up $24M Series A (July 2026)
- FCC: AI-generated voices in robocalls ruling (2024)
- Block: Dessa
- 9to5Google: Android XR glasses (May 2026)