Blog

The State of Voice AI Agents: What Deepgram's Pipecat Deal and Rime's Prism Mean for Your Phone Line

In two days, Deepgram agreed to buy Daily, the company behind the open-source Pipecat framework, and Rime launched a speech-to-speech model you can check before it speaks. If you're thinking about an AI receptionist, here's what changed and what to ask before you buy one.

Caleb Dorris 29 October 2026, 9:36 AM 10 min

Update, October 29, 2026

Phonic-1 just topped a voice AI benchmark. Here's how I'd read that

On October 29, Phonic launched Phonic-1, a speech-to-speech model built for voice agents. It's in first place on Artificial Analysis's Speech to Speech Index with a score of 87.1, just ahead of Grok Voice 3.0 at 86.9. That index weights four things equally: reasoning on spoken questions, an agentic task test, human preference in live conversations and task success.

The feature that matters most for a receptionist is one Phonic says only its model has. Phonic-1 can call tools, like looking up an appointment, while the caller is still talking, instead of saying "let me check on that" and going quiet. Phonic's docs claim under 600 milliseconds from the caller's speech to the model's reply, and Artificial Analysis measured 0.61 seconds to first audio. Phonic says pricing starts at $0.07 a minute, and new sign-ups get $250 in free credits through November 30.

Point 4 below still applies. A 0.2-point lead is a tie in practice, and none of these tests is your phone line, your callers or your background noise. The talk-while-looking-it-up claim is the one I'd test hardest, because that's where a receptionist either sounds sharp or leaves a caller sitting in silence.

If you want a reference point, call my live AI receptionist demo and listen for how it handles follow-up questions. Then put the same questions to any model you're considering before you trust its numbers.

If you run a small business and you've been wondering whether an AI receptionist is ready to answer your phone, this was a useful week to watch. Two announcements landed back to back, and between them they show where voice AI is heading: more open plumbing underneath, and more control over what the AI actually says to your customers.

You don't need to buy anything from either company. But the news changes what you should expect from whoever sets up your phone line.

What happened

Deepgram is acquiring Daily, the maker of Pipecat (October 29). Deepgram sells speech-to-text, text-to-speech and voice agent infrastructure. Daily runs real-time audio and video networks, and it built Pipecat, an open-source framework developers use to wire a voice agent together: the phone or web connection, the speech recognition, the language model, the voice, and the hand-offs between them. Deepgram calls Pipecat the world's most popular open-source voice agent framework. That's Deepgram's claim, but Pipecat is widely used: Daily says it has more than 200 integrations and almost 400 contributors.

The companies didn't disclose the price.

The part that matters most for anyone building on Pipecat is what happens to it:

It stays free and open source under its existing license. Deepgram's announcement says Pipecat "will remain fully open source and free to all developers," and the license on Pipecat's GitHub repository is BSD 2-Clause, a permissive license that lets anyone use and modify the code commercially.
An independent Pipecat Foundation is being set up. According to Daily, a steering committee "that represents the entire realtime AI ecosystem, including companies that compete with Deepgram" will run it. Neither company has named members or a date yet, so for now this is a commitment, not a working organization.
Pipecat stays vendor-neutral. Developers can keep mixing models and providers from different companies. Daily says that for its existing customers, "nothing will change."

Deepgram also introduced the Real-Time Context Standard (RCS). This is an open protocol, co-developed with the Pipecat team, for carrying context through a voice agent. The problem it addresses is real: in most voice agents, your caller's words get turned into text, and everything else gets lost along the way. Their tone, a whisper, an unclear word or a noisy background all disappear. An annoyed "oh, great" reaches the AI looking exactly like a happy one.

RCS adds three kinds of information: context from the audio, instructions for how the language model should use it, and guidance on how the reply should sound. In practice, that means an agent can ask the caller to repeat themselves when it didn't hear clearly, or sound apologetic when the caller is clearly frustrated. Pipecat already has early, English-only tools for it.

Deepgram also previewed a Voice Agent Platform built on Pipecat. It's described as "forthcoming," so it isn't something you can buy today.

Rime launched Prism (October 28). Rime makes voice models for businesses that run phone agents. Prism is a speech-to-speech model: instead of passing your caller's audio through separate speech-to-text, AI and voice steps, one model listens and replies. That usually makes the conversation feel more natural, but it has a catch. With most speech-to-speech models, you can't see what the model is about to say until it's already said it.

Prism's pitch is that it fixes that catch. Here's what Rime claims:

It writes its reply as text before speaking it, so your rules can read and even rewrite every word before the caller hears it. Rime calls this "an inspectable text layer to run guardrails over before the model speaks."
It's fast. Rime says Prism responds in about 300 milliseconds. Its own benchmark write-up gives a different figure, 0.57 seconds to first audio on an independent testing setup, so treat "300 ms" as a best case rather than what every call will get.
It can run on a company's own servers or private cloud, not just Rime's. At launch, Rime's developer docs list a cloud API, English only and four voices. Rime pitches Prism at enterprises, and the self-hosted options are likely to matter most to them.
Rime says big names already use its voice models, including Mayo Clinic, Asurion, Upstart and Dialpad. That refers to Rime's models in general, not to Prism. Dialpad is the company Rime names as a Prism design partner.

Rime lists Prism at $0.08 a minute on its pay-as-you-go plan. The company raised a $24 million Series A in July.

All of those performance claims come from the vendors themselves. None of them has been tested independently on calls like yours yet.

What this means if you're considering an AI receptionist

Put the two announcements together, and the direction is clear. The hard parts of a phone agent aren't the voice anymore. They're control, context and not getting locked in. Here's how that plays out for a small business.

1. Open plumbing means less lock-in, if your setup uses it

Most AI receptionist products you can buy are closed subscriptions: their phone number, their AI, their monthly fee, raised whenever they like. Pipecat staying open, under a permissive license and a foundation that includes Deepgram's competitors, is good news for the other approach. You own the setup, and you can swap the AI model or voice provider later without starting over.

What to ask: "If a better or cheaper voice model comes out next year, can we switch without rebuilding?"

2. Tone is becoming something an AI receptionist can act on

The difference between a good front desk and a bad one is often tone. They notice when someone's upset, slow down for an elderly caller, or repeat back a confusing address. Standards like RCS and models like Prism are trying to keep that information from getting thrown away. These tools are early, but expect a good AI receptionist to react to how people sound, not just what they say.

What to ask: "What does it do when the caller is frustrated or it didn't catch something?" The right answer is that it asks again or hands off to a person. It shouldn't guess.

3. You should be able to see and limit what it says

Prism's selling point, text you can check before it's spoken, points at something every business should insist on. You need rules about what the AI is allowed to say: no quoting prices you haven't approved, no promising appointment times it can't book, and always disclosing that the caller is talking to an AI. Whatever model is underneath, you should be able to read every call afterward as a transcript.

What to ask: "Where are the rules about what it can and can't say, and can I read a transcript of every call?"

4. Treat speed and accuracy numbers as marketing until you hear it yourself

Every voice AI company leads with latency figures and benchmark scores. Rime's own materials give two different response times for the same model. The only test that matters is a call on your own phone, with your own questions, in your own accent, with your own background noise.

What to ask: "Can I call a live demo right now?"

Where I fit in

I build AI receptionists for small businesses, and I design them around exactly these questions. The one I set up for you answers on the first ring. It asks the follow-up questions a good front desk would, captures the caller's name, number, request and urgency, and sends you a summary, transcript and recording. It runs on accounts in your name, with your card on file, so you pay the phone and AI providers directly at their list prices. You pay me once for setup. There's no monthly fee to me, and no platform you have to keep renting.

Because it's built on parts you own, the news this week is the kind of change you benefit from instead of getting stuck behind. When a better model or a better open standard shows up, your line can move to it.

Call my live AI receptionist demo at (629) 213-4695 and watch your own call appear on the page as you talk. That page also shows what it would cost to run at your call volume. If it sounds right for your business, get a free quote.

If you're wondering about the other direction, AI agents calling your business, I wrote about that in AI Agents Are Starting to Call Small Businesses. Is Your Phone Line Ready?

FAQ

Is Pipecat still free after the Deepgram acquisition?

Yes. Deepgram says Pipecat will stay fully open source and free under its existing license, which is BSD 2-Clause on GitHub. An independent Pipecat Foundation is being set up to steward it.

How much did Deepgram pay for Daily?

The companies didn't disclose the price.

What is the Real-Time Context Standard?

It's an open protocol from Deepgram and the Pipecat team for carrying information like tone, whispers and unclear words through a voice agent, along with guidance on how the reply should sound. It was introduced on October 29, 2026, and Deepgram is inviting others to contribute to it.

What is Rime Prism?

Prism is Rime's speech-to-speech model for voice agents, launched October 28, 2026. It writes its reply as text before speaking, so a business's rules can check it first. Rime says it responds in about 300 milliseconds and can be deployed on-premises or in a private cloud.

Do I need Deepgram or Rime for an AI receptionist?

No. They're two of several companies that make the underlying models and tools. A small business needs a receptionist that answers reliably, follows your rules, hands off to a person when it should, and doesn't lock you in. The vendor underneath matters less than how it's set up.

Sources

All posts