ElevenLabs Hit $300 Million ARR by Building Voice Before Anyone Believed It Was the Interface

Sarah Guo introduces Mati Staniszewski, co-founder and CEO of ElevenLabs, a company founded to change how people interact with each other and with computers through voice. In three years it has reached more than $300 million in run rate. The products sit on foundational audio models: speech that sounds human, speech understanding, and the orchestration that makes those pieces interactive. On top sit a creative platform for narration, ads, movies, and dubs, and an agents platform for customer experience, personal education, and immersive media.

The company has grown to 350 people. It started in Europe as a remote-first outfit and still is, with hubs in London, New York, Warsaw, San Francisco, Tokyo, and Brazil. ARR is roughly 50/50 between self-serve subscriptions on the creative side and sales-led enterprise work on the agents side. More than five million monthly actives use the creative product. A few thousand enterprise customers range from Fortune 500s to fast-growing AI startups.

Guo groups ElevenLabs with Midjourney, Suno, and HeyGen: 2022-era creation tools that investors found unintuitive. Who really wants to make voices? Dubbing, on its face, is not a huge market. Staniszewski's answer starts at home. He and his co-founder are both from Poland. Watch a foreign movie there and every voice, male or female, is narrated by a single speaker with a flat delivery. It is a terrible experience. As soon as people learn English, they switch off that version, and yet it still covers the majority of content.

Staniszewski had worked at Palantir. His co-founder had worked at Google. Both believed that information would eventually be available globally, in every language, at high quality. The sharper question was whether you could carry the original voice, emotion, and intonation across, not just the words. Imagine switching this podcast to Spanish and still hearing Sarah and Mati. That is what they later did for Lex Fridman's interview with Narendra Modi: listeners could stay inside the same voices while the language changed.

Digging further, they saw three layers. Creation was still expensive and unintuitive: hire talent, book a studio, wrestle with tools. The technology to recreate a specific voice at high quality did not really exist. And the interactive layer looked broken. After decades of computing, most people still spend their time on a keyboard looking at a screen, even though speech is the interface humanity started with. They wanted to solve that.

Fast-forward from 2022 and many people now share the belief that voice is the interface of the future, on smartphones, computers, and robots. In 2022 it was not a consensus. Doing product and research at the same time is still hard. Staniszewski says he is lucky: he has known his co-founder for 15 years, calls him the smartest person he knows, and leans on him to create the research foundation that the products stand on.

Guo notes the old startup rule: do one thing well. ElevenLabs is treating itself as an interaction company, with research plus products for at least two markets. Sequencing started with a blunt finding. They tried existing models for narration and dubbing and got speech so robotic that people did not want to listen. That is where his co-founder's research mattered. He assembled a team, did much of the work himself, and created a new way to generate the audio.

Internally the unit of work is a lab: researchers, engineers, and operators pointed at one problem. The first was voice, a group of roughly five people whose mission was to narrate in a more human way. Research came first, then a simple layer so people could use it, then a broader suite for a full audiobook and a full movie dub. Once speech sounded human, the next problem was obvious. To interact with technology you have to bring knowledge on demand. That spawned an agent lab, combining text-to-speech with language models and speech-to-text, then wiring the stack into legacy systems, functions, deployment, testing, monitoring, and evaluation.

Other labs were pulled by users. A music lab appeared because creators wanted to add licensed music while they made speech. Sound, image, and video partner models are now coming into the same suite. Dubbing itself did not feel like a market with a huge push, but they believed content should travel naturally across languages. Staniszewski still thinks that market will be immense, not only for movies. Real-time conversation across languages is the Babel fish from *The Hitchhiker's Guide to the Galaxy*: language barriers to communication and creation coming down. On the agent side, some demands are predictable integrations. Others are subtler, like understanding not only what was said but how it was said, so the reply can carry that expression.

Guo's buyers are usually not machine-learning scientists, and audio still lacks a full eval suite. How does a company pick a good voice? Staniszewski splits the problem in two: how you benchmark the field, and how you select a voice for a use case. On selection, ElevenLabs deploys what he calls a voice sommelier, a voice coach with an exceptional voice of their own, plus a team that helps enterprises find brand-right audio. There is now a celebrity marketplace too, so a company can license iconic talent such as Sir Michael Caine.

Briefs vary wildly. One of the biggest European companies asked for a voice as robotic as possible, a request so counterintuitive the team had to work backwards. In Japan and Korea, callers are routed by age: younger customers get an excitable, famous market voice, older customers a calm, slow one. Partners recently asked for a New York voice with a Long Island accent, which Staniszewski had not known was a distinct thing. He expects voices to become dynamic. Evening and tired, morning news, different energy for different moments, eventually a voice created from inferred preference.

Benchmarks remain unsolved. Speech quality scores exist, but whether you like the speech depends so much on the voice that comparing model A to model B with different voices can reverse the result even when quality differs. He has seen that clearly on Artificial Analysis leaderboards. A third gap sits under the labels. Early on they asked traditional vendors to annotate not only what was said but how: emotion, accent, delivery. Most could not do it well, because you have to hear the clip and have the skill to describe it. ElevenLabs had to build that data itself.

Customer support is the enterprise use case ramping fastest. Cisco, Twilio, and TELUS Digital are all elevating voice agents there. The more interesting shift inside that domain is from reactive to proactive. Meesho, a leading e-commerce company in India, started with refunds and package tracking. The agent now sits on the website as part of the front experience. Shoppers can speak, ask for help navigating to an item, or ask what to give as a gift. The system uses the offer and the questions to show items, move through the site, and potentially all the way to checkout. Square is on the same path, from voice ordering into discovery, so connected businesses can give customers that full-funnel experience.

A second category is the move from static to immersive media. Stories and IP that today have one delivery path can become something you talk to. With Epic Games, ElevenLabs brought Darth Vader's voice into Fortnite so millions of players could speak with Vader live in the game. Staniszewski sees that as a theme: talk to a book, talk to a character, shift the whole space. Education is the application he is most excited about, a personal tutor in your headphones. On chess.com you can learn from Hikaru Nakamura or Magnus Carlsen. MasterClass, working with FBI negotiator Chris Voss, lets you finish the lesson and then call him for a practice negotiation.

The most ambitious case combines those threads. Staniszewski recently went to Ukraine to work with the Ministry of Digital Transformation on what he calls a first agentic government. The plan is to change how ministries run. Government customer support, benefits, employment, the process of leaving the country, would run through a digital app. There would be proactive alerts for citizens, and an education system delivered as personal tutoring. The operating model impressed him as much as the product: engineering leaders inside each ministry, feeding a central platform, despite everything else happening in the country.

Guo asks how a Fortune 500 buyer should choose among Palantir or a large consultancy, a platform such as ElevenLabs or OpenAI, and a use-case specialist such as Sierra. Staniszewski, who started at Palantir and still blends forward-deployed engineering into ElevenLabs, is blunt. If you want one pointed solution, they are probably not the best choice. If you want to deploy across support, internal training, and sales experiences that raise the top line, the platform plus their engineers is a strong fit. Fortune 500s often already have platform investments and will build some pieces themselves. Palantir and consultancies have more resources for a full digital-transformation journey. Sierra is strong on a specific use case. International coverage, languages, and integrations are another filter. Summarized: many use cases, plus help deploying conversational agents as a new customer interface, is where he wants the bake-off.

No Priors is named for the habit of watching market assumptions collapse. In 2022 and 2023, Staniszewski heard the same question: Google and OpenAI can do this, so why do you get to persist? His first answer is product. Capturing value in agents and creation requires integrations, deployment help, and a layer that meets customers where they are. He does not think the big labs will focus on that product work. The second answer is research itself. Biased as he is, he says ElevenLabs is beating those labs on text-to-speech, speech-to-text, and orchestration benchmarks. Credit goes to his co-founder and a small research group.

Audio, in his view, is not primarily a scale game. You need architectural and model breakthroughs more than you need more people. He estimates there are maybe 50 to 100 researchers worldwide who can do that work at a high level, and ElevenLabs has about 10 of them. The difference is obsession plus a short path to production: those people see how users actually interact, then go back into the model. He also admits the other side of the mystery. Those labs have extraordinary talent too. Why they have not matched the quality is, in part, unknown.

Open source is the next pressure. Internally the narrative is that models will commoditize. For some use cases differences will still matter; for most they will not, in two, three, or four years, nobody knows. A fine-tuning layer on top will be valuable. Base models will get pretty good. That is why product matters. On asynchronous narration, open source is already great and out-of-the-box quality gaps are shrinking. What most models have not figured out, and where he thinks ElevenLabs still wins, is controllability: making the voice say it the way you intend. Interaction is a different race. Cascaded speech-to-text, language model, and text-to-speech is good enough for customer experience, but still away from a conversation that would pass a Turing test. He puts that at about a year. Real-time translation, the live Babel-fish version, is maybe two years.

Guo offers an uncomfortable belief: most technology advantages last a year or ten, not forever. They buy momentum, not an infinite defense. Staniszewski does not argue. Research is a head start, six to 12 months of getting the better experience to the customer earlier, often built in parallel so the product is waiting when the model lands. Long-term value is the ecosystem around it: brand, distribution, the collection of voices, integrations, and workflows. Inside the company a rough rule of thumb governs the wait-or-build tension. Product teams can see the research roadmap and still ship something else if it helps the customer. If it will take longer than three months, they will probably build it. If it is less than that, they probably will not wait, or they will not bother building a stopgap. Research and product stay parallel so that when a result publishes, the corresponding product capability is already there.

Current research splits along the same two businesses. On the creative side they have been stacking controllable text-to-speech, a speech-to-text model that covers almost 100 languages including low-resource ones, and a fully licensed music model, with more work on how audio meets existing video. On the agent side the focus is real-time speech-to-text and text-to-speech. They just released Scribe v2, under 150 milliseconds, 93.5 percent accuracy across the top 30 languages on FLEURS, a cut they use because they serve so many more languages than most vendors. A new orchestration mechanism, due in a couple of months, is meant to cut end-to-end latency and carry the emotional context of the conversation into the reply. They are also investing in a more fused speech-to-speech approach. For reliable enterprise work over the next year, he still prefers the cascaded stack: more structure, visibility into each step, tool calling. Speech-to-speech may be the choice when you want expressiveness and can tolerate hallucination.

Asked about AI companions, he thinks they will be a big business. He is not personally excited and the company does not spend much time there. Assistants, companions, and characters will blur. What he wants is the Jarvis version: a super-assistant that understands him, opens the blinds, talks about the weather, and puts music on. Education will mix a large share of time with AI tutors and an explicit share of time with no technology at all, human to human, so people still learn the social half. Not everything should be personified. Talking to an oven can stay static. Code does not need much extra emotion. He does hope devices recede into a pocket or an attached element so people stay more immersed in real life. Andrej Karpathy's line, a decade of agents and then a decade of robots, matters here. When you interact with robots, voice will be a key input and output. Robots, he agrees, will be personified.

The application he thinks still has not kicked in is education with voice: a teacher on your headphone or speaker who understands you, stays with you, and delivers the right content. Commercial partners exist. Schools and universities, in a safeguarded way that still supports the social part of education, have not really absorbed it. The version he is excited about is a little science-fictional and a little practical: Richard Feynman or Albert Einstein, or another teacher you love, delivering the lecture. That, he says, will be sick.