ElevenLabs CEO Mati Staniszewski says voice AI is technically ready in many cases, and the bottleneck is deployment by automotive firms and big enterprises.
ElevenLabs reports $350m in ARR at the end of 2025 and added $100m of net new ARR in a single quarter, with over half its enterprise business now sales-led.
Staniszewski says voice agents already pass the Turing test in customer support but not yet in interactive gaming, and expects them to pass in all cases within about a year.
Voice AI is ready, and deployment is now the bottleneck
ElevenLabs is betting that voice AI is already good enough for most uses and that the bottleneck has moved to deployment. Its lead rests on proprietary speech data and a reliable cascaded architecture, offered as a horizontal platform that others build domain apps on. Enterprise demand, led by customer support, is turning that into fast revenue growth.
Takeaways
Voice technology is ready, but deployment lags. John Collison notes his phone cannot read a PDF aloud and no car has good voice input. Staniszewski says automakers and big companies adopt too slowly, though linking real-time voice to context only recently became possible and drew big enterprise adoption.
ElevenLabs' edge comes from better data. Cofounder Piotr brought transformer and diffusion methods into speech, but in-house labelers annotating how speech is delivered made the difference. Models deduce accents and emotion themselves instead of using hard-coded parameters. Speech to text began as an internal annotation tool.
Cascaded systems beat speech to speech for now. Speech to speech models are faster but must stay small, so they lose capability, reliability and visibility into each step. ElevenLabs optimises the cascaded chain of speech to text, a language model and text to speech.
Voice passes the Turing test only in support. Collison argues text LLMs passed long ago while voice is far behind. Staniszewski says agents already pass in customer support, the most ready use case, but not in interactive gaming. He hopes they pass everywhere within about a year.
Revenue is compounding quickly through enterprise sales. ARR was $350m at the end of 2025, with $100m of net new ARR added in one quarter. ElevenLabs raised $500m at an $11bn valuation. Over half of enterprise business is sales-led, with customers including Deutsche Telekom, Klarna, Meta and IBM.
Voice models are small, but capital spending is high. Models run from a few billion to low tens of billions of parameters, versus hundreds of billions for leading LLMs. Pricing is per minute for agents and transcription, per token otherwise. Fused LLM-voice models could reach hundreds of billions of parameters.
Small flat teams, with culture building the company. ElevenLabs keeps product and research teams under 10 people and puts a technical lead in each non-technical team. Staniszewski says whether the flat structure works will be clear in five to ten years.
Uses beyond customer support
Transcription tuned to a specific voice handles accents and crowded rooms, and already separates speakers extremely well.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Universal translation, which Staniszewski links to The Hitchhiker's Guide to the Galaxy, lets the company's own sign-up agent gather leads in every language.
Restored voices for people with ALS and throat cancer, and for a Neuralink patient.
Ukraine's Diia app now offers government services such as booking health checkups by voice. Staniszewski calls Ukraine the most tech-forward government he has seen.