Some AI problems show up in research. Others only appear when you start thinking about how to put the tech in front of thousands of customers. At Deepgram, we endeavor to push the frontier AND productize fast, all while balancing speed with safety. More with @jason, @mishindan1, @gsivulka, and me on @ThisWeeknAI. piped.video/watch?v=FnVyEYKc…

I spent a few years working on dark matter detectors, including the LUX-ZEPLIN (LZ) experiment. At the time, I definitely didn’t imagine that experience would eventually lead to a company building voice AI. It’s exciting to see the latest result from LZ getting a new signal to chase, but the bigger thing is what work like this creates around it. When you try to answer a really hard scientific question, you have to build new instruments, new simulations, and new ways to work with massive amounts of complex data. That work doesn’t just stay inside the original experiment. It creates capabilities that go in directions nobody expected. That’s what happened with @DeepgramAI. We were working with waveform data in particle physics and trying to understand it. That led us to think differently about how machines could process complex signals at scale. We weren’t trying to build voice AI, but eventually that’s where the work took us. That’s what I find cool about moments like this. The discovery matters, obviously, but so do all the things that become possible because people decided to go after a hard problem in the first place. Congratulations to the LZ team, excited to see what comes next…

Voice AI is about to have a big moment. Excited to get the people building it together at Speak ‘26 in SF.

The weird thing about speech is that the meaning of a sentence rarely lives in the sentence itself. “Sure.” That can mean absolutely, I’m annoyed, I don’t believe you, or please stop talking. The words are identical. The context is what gives them meaning - and that context carries through the conversation. If someone has been frustrated for five minutes, the next response shouldn’t suddenly sound upbeat. If you’ve just solved their problem, the voice should feel different. Expressiveness is not a style setting. Expressiveness should be a product of conversational context.

Deepgram CEO @deepgramscott joined @furrier from @theCUBE for #NYSEWired to talk Flux TTS, Deepgram crossing $100M ARR, and voice as a primary interface. Catch the whole conversation: piped.video/watch?v=D3xDOYbE…

Introducing @DeepgramAI Flux TTS, a text-to-speech model built for live conversation. Voice AI is moving from demos to real customer calls, and the bar there is different. Text-to-speech needs to be able to handle the nuances of how we speak. Can it handle interruptions? A changed order? A string of numbers read off fast? Can it respond quickly enough, and sound expressive enough, to feel like a real conversation? Flux TTS was built for exactly that. It tracks context across turns, adapts mid-call, carries natural expressiveness, and responds in as low as 80ms. These are voices you'll want to have a conversation with. Build with Flux TTS for free until September 12th: deepgram.com/product/text-to… Also, I asked Flux about Flux. Hear what it had to say.

The same sentence can mean five different things. Take this sentence: I never said he stole the money → Someone else said it. I never 𝑺𝑨𝑰𝑫 he stole the money → I might have implied it, but I didn't say it. I never said 𝑯𝑬 stole the money → Someone else may have stolen it. I never said he 𝑺𝑻𝑶𝑳𝑬 the money → Maybe he borrowed it. I never said he stole the 𝑴𝑶𝑵𝑬𝒀 → Maybe he stole something else. Same words. Same grammar. Completely different information. Humans do this without thinking. We use emphasis, timing and intonation to communicate information that isn't explicitly encoded in the text. That's a fascinating problem for TTS. A text-to-speech model isn't just being asked to pronounce a sequence of words correctly. It has to make decisions about information the text never explicitly gave it. Which word matters? What is being contrasted? What does the speaker actually mean? Making a voice sound human is one problem, but knowing how a human would choose to say it is a much harder one!. Very cool stuff.

Experience matters. But borrowing decades of someone else's experience through a conversation is one of the biggest unfair advantages you can have. That's something I've come to appreciate more every year.

It's interesting how quickly expectations change. 2024: AI that talks. 2026: AI worth talking to.

Here's what caught my attention over the last few weeks. OpenAI, Anthropic, and Amazon are all investing heavily in voice. Despite competing on almost everything, they seem to agree on where AI is heading next. That tells me we're entering a new phase. We've spent the last few years making models smarter. The next few will be about making them better to interact with - more natural, more responsive, and better at understanding people in real time. The companies that win won't just build the most capable AI. They'll build the AI people actually want to have conversations with. What do you think will matter more over the next few years: better models, or better interactions? ⚖️

What separates the great AI founders from everyone else? There are two kinds of AI founders. The ones building the foundations: models, infrastructure, chips, tooling. And the ones building with AI to solve real customer problems. Both can build exceptional companies. But they're playing completely different games, with different moats, different economics, and different paths to success.

Two of the biggest AI labs shipped voice-led products this week. The obvious takeaway is that voice is becoming a primary interface to AI. The more important takeaway is that voice puts the entire AI stack on the clock. A voice model can hear or speak. A real-time voice system has to listen, understand, reason, use tools, handle interruptions, respond, and recover without dropping the conversation. Think about a robot. The planner decides where to go. A faster system keeps it upright while it gets there. It doesn’t matter how intelligent the planner is if the robot falls over. Voice agents work the same way. The LLM is the planner. The real-time voice layer keeps the conversation upright. That means knowing whether someone is finished speaking. Handling “yeah” and “uh-huh” correctly. Starting useful work before the entire utterance has arrived. Letting a person interrupt. Maintaining context while another model or tool works in the background. Recovering when something takes longer than expected. This is why the voice AI market is moving beyond individual speech models. The next phase is about real-time systems. And it won’t be won by the company with the most impressive 30-second demo. It will be won in production: ● tail latency, not average latency ● task completion, not just naturalness ● noisy rooms, accents, names, and addresses ● thousands of concurrent conversations ● observability, deployment control, and predictable economics OpenAI and Anthropic are validating what voice builders have known for years: when AI can act, its interface has to be as dynamic as human conversation. @DeepgramAI has spent 11 years building for that constraint. Voice is already becoming the real-time operating layer for AI. Now the race is to make it durable at scale.

The best CEO lessons I've learned didn't come from business books. They came from race cars, skate parks, golf, and watching people who are simply better than me. Where do you go to learn when it isn't from work?

Here's a thought: the best voice AI probably won't impress anyone. It will simply... work. Nobody celebrates electricity when they turn on a light, or Wi-Fi when a webpage loads. The technologies that have the biggest impact eventually become invisible because they fit so naturally into our lives. I think voice AI will follow the same path. The moment we stop noticing the technology is probably the moment it's finally ready. 😎

Is bigger always better in AI?

More often than anything, people ask me: "Will people really want to talk to AI?" I think they're asking the wrong question. 😆 We already talk to machines every day. We speak to our cars, ask smart speakers questions, navigate with voice, and spend hours on automated customer support systems. The behavior already exists! What's changing isn't people's willingness to speak. It's the quality of the conversation. Voice AI is finally reaching a point where talking to software can feel natural instead of frustrating. That's an exciting shift to watch.

Everyone talks about latency, hallucinations, or word error rate. Those things matter, but I think there's a much simpler benchmark for voice AI. Would you interrupt it? Humans interrupt each other all the time. We change our minds halfway through a sentence, pause to think, correct ourselves, and accidentally talk over one another. That's just how conversation works. A voice agent that can handle all of that naturally is much closer to feeling human than one with a perfect voice or the lowest latency. What do you think?