Mati Staniszewski on Voice AI and a Simulated Aristotle

ElevenLabs co-founder Mati Staniszewski introduces a live voice AI demonstration. His stated goals are making information accessible as audio, representing diverse voices and languages, and making voice a way to interact with technology.

The first demo reads a welcome message aloud with a British accent and expressive intonation. Staniszewski highlights the system’s use of context to shape delivery, including its pronunciation of the Swedish word fika.

He describes an ElevenLabs voice marketplace where people can create and share synthetic voices and receive compensation when they are used. He plays a sample modeled on his own voice.

A Swedish voice then reads a welcome in Swedish and English. Staniszewski points to the ability to switch languages while retaining recognizable vocal characteristics, and imagines speakers being heard in a listener’s chosen language.

For conversational AI, he creates an agent prompted with information about Aristotle’s life and ideas. He emphasizes that this is a constructed AI character, then asks it what Steve Jobs might have asked Aristotle.

The simulated philosopher suggests questions about technology, virtue and balance. Asked for a Greek maxim, it offers the idea of knowing oneself and explains its relevance to science and art. These answers are generated by the agent rather than historical statements.

Summit host Lauren Crichton asks about enterprise uses. Staniszewski mentions education, healthcare and assistants, including an example of nurses using an agentic workflow to help contact patients and gather updates.

He also describes an interactive article experience in which a reader can listen and ask questions. He says reliability, scaling, taking the intended action and avoiding hallucinations remain difficult, particularly in healthcare, finance and logistics.

Crichton asks whether speech models actually understand what they say. Staniszewski says timbre, emotion and intonation can provide context beyond text and that models increasingly use surrounding material to select a delivery, while the deeper question of understanding remains open.

He gives irony as an example: a sentence may need a different delivery once the rest of the story is known. He imagines audiobooks with multiple voices and sound, requiring a system to track who speaks throughout a book.

Crichton asks what happens to language instincts if people outsource writing, translation and speech. She compares writing to physical strength after mechanization: still valuable, but no longer something everyone must exercise in the same way.

Staniszewski suggests language learning may become more of a chosen cultural and creative pursuit. He also hopes voice models can help preserve recordings of languages, accents and voices, and says some governments are discussing educational uses.

Looking ahead, he imagines voice interfaces putting technology in the background so people can remain present in their surroundings. His second hope is a real-time translation experience like the Babel fish in The Hitchhiker’s Guide to the Galaxy.