ElevenLabs, founded in 2022, reports more than $300m in annual recurring revenue and 350 employees across remote hubs worldwide.
Mati Staniszewski says the company will build any feature that research would take longer than three months to deliver, since research gives only six to twelve months of head start.
Staniszewski predicts open-source base audio models will commoditise within two to four years, leaving fine-tuning and controllability as the real differentiators, and argues AI labs will not build the product layer.
Why ElevenLabs thinks voice AI's moat sits above the model
ElevenLabs, now at 300m in ARR with 350 staff, is betting that leadership in voice AI comes from pairing its own audio research with a broad product ecosystem. Staniszewski expects base models to commoditise and the big labs to skip the product layer. The durable edge therefore lies in the voices, integrations and workflows built on top of the models.
Takeaways
Research buys months, the ecosystem buys years. Staniszewski says model research gives customers a six to twelve month head start. Lasting value comes from brand, distribution, voices, integrations and workflows. If a research fix would take more than three months, product teams build the feature themselves.
Audio needs new architectures more than scale. Only about 50 to 100 researchers can do this audio work, and ElevenLabs employs about 10. Staniszewski expects open-source base models to commoditise within two to four years, leaving fine-tuning and controllability as differentiators. He doubts big labs will build the product layer.
Customer support leads, proactive agents come next. Support is the fastest-growing agent use, with customers such as Cisco and Twilio. The next step is proactive assistance, seen with Meesho in India and Square's voice ordering, alongside character work such as Darth Vader's voice in Fortnite.
Customers struggle to judge voice quality. Benchmarks are weak because the voice itself shifts how listeners rate speech, so ElevenLabs runs a voice sommelier and a celebrity marketplace. One European firm wanted the most robotic voice possible, while a Japanese and Korean company wanted calm voices for older callers.
Cascaded pipelines still win for enterprise reliability. Fused speech-to-speech models suit expressive uses that tolerate hallucination, but cascaded systems will serve enterprises for the next year. The new Scribe v2 speech-to-text model runs under 150 milliseconds with 93.5% accuracy across the top 30 languages on FLEURS.
ElevenLabs fits companies wanting one voice platform. Asked whether enterprises should buy from Palantir, consultancies, OpenAI or Sierra, Staniszewski says ElevenLabs suits firms running several voice experiences on one open platform, especially with global users. Palantir and large consultancies have more resources for broad transformation.
What Staniszewski expects next
Conversation that passes a Turing-style test is at least a year away. Real-time dubbing and translation should arrive in about two years and open an immense market.
AI tutors will take a large share of learning time, alongside some human-only time. Voice tutoring could become one of the biggest uses of voice, perhaps with lecture notes read by Richard Feynman or Albert Einstein.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
He is sceptical of AI companions as a social product, though they could help with loneliness, and is more excited about a Jarvis-style assistant.