ElevenLabs, founded in 2022 with just over 400 people, reached over $400m in revenue by building a frontier audio AI business without hundreds of millions in upfront capital.
Mati Staniszewski says voice will become the primary interface for robots and other technology, and that authenticated human voices will need to be marked as AI voices spread.
ElevenLabs trains some models that will not pay off for a year or two, as it optimises for long-term customer impact and has assembled more than 1,000 annotators, including voice coaches and musicians.
How ElevenLabs Built Frontier Voice AI Without Billions in Funding
ElevenLabs bet early that voice would become the interface for technology while rivals ignored audio. Because audio models needed far less compute, a small, revenue-funded team could build frontier models without billions. Mati Staniszewski argues the lasting advantage comes from the craft of natural speech and the product layer around the models.
Takeaways
Audio was an overlooked niche in 2022. While the industry chased text and image models, audio models were smaller and needed less compute. That let ElevenLabs build frontier models and reach over $400m in revenue with just over 400 people, without the hundreds of millions usually assumed necessary.
Early revenue funded the model training. The team started remotely between London and Warsaw and recruited top researchers by scraping GitHub and contacting them based on their work. ElevenLabs monetized quickly to pay for training and raised outside money only later, as its ambitions grew.
Dubbing set the product roadmap. The idea grew from Polish films voiced by one monotone narrator. Solving dubbing needs transcription, understanding, translation and speech, so ElevenLabs built context-aware text-to-speech first, then speech-to-text, then real-time models for voice agents, and finally music generation.
Voice agents now reach well beyond support. Deliveroo agents call restaurants to capture opening times and update riders and customers. Deutsche Telekom and ElevenLabs itself run inbound sales on agents. Ukraine's government gives citizens near the front line a voice line for information on safety and education.
Emotional intelligence is the next frontier. Agents are reliable for support but still too slow and miss emotion in sensitive conversations. ElevenLabs is moving emotional understanding toward production, since in a negotiation delivery and pauses matter as much as content. AI music also cannot yet match top-chart songs.
The moat is artistry plus the product layer. Speech-to-text is technology, but text-to-speech is artistry that needs focused user feedback. ElevenLabs employs over 1,000 annotators, including voice coaches and musicians, trains some models that pay off only later, and builds workflows, templates and a library of over 20,000 user voices.
A flat organization with no titles. Most managers have about ten direct reports and core teams stay under ten people. Engineers sit inside non-technical teams such as legal and people to build automation, and a scoring system governs sales concessions on indemnity and liability clauses.
Where voice goes next
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Voice will become the primary interface for robots and other technology.
Agents will talk to agents. At a hackathon, two voice agents recognized each other and switched to a more efficient language. Startups already have agents call venues for prices, then call back with a budget.
Agents that only take orders today could eventually interrupt humans in negotiations.
Authenticated human voices will be watermarked, and everything else will be assumed fake, while human-to-human interaction gains value.