Synthesia built an interactive AI avatar of a TechCrunch reporter

Synthesia, a video-generation startup known for building AI avatars for corporate use, made an interactive digital twin of a TechCrunch reporter that is trained to answer questions about only one of her stories. It is the first time the company has built an interactive avatar for a journalist, or for anyone besides Alexandru Voica, its head of corporate affairs, whose own avatar answers common press questions about the company.

Originally based in the U.K., Synthesia competes with rivals like D-ID, HeyGen, and Colossyan in building AI avatars for enterprise training videos. The company hit a $4 billion valuation earlier this year and said last year it had crossed $100 million in annual recurring revenue. Synthesia builds three types of products: a video-creation and distribution platform where a user types a script and a classic avatar reads it back, an agentic platform called Sessions where people interact with avatars through surveys or roleplay, and an API platform that lets customers combine Synthesia’s video and voice models with other services. It recently launched a Sessions product called Roleplay Sessions, which lets employees practice tasks such as sales pitches with an interactive AI avatar that responds to and scores their answers.

Building the reporter’s avatar required a visit to a mini film studio inside Synthesia’s New York office, where staff took numerous photos of her and recorded a two-minute sample of her voice. It took Synthesia’s team a couple of days to produce two interactive avatars — one with glasses, one without — plus two personal avatars that simply read back whatever script they are given rather than converse.

The interactive avatar runs on a chain of models: a voice-to-text model transcribes what a person says, an agentic language model interprets the text and can act on it, a text-to-voice model generates the spoken reply, and a video model that Synthesia built animates the avatar as it talks. Synthesia’s own video and voice models power the pipeline by default, but customers can swap in alternatives from Cartesia, ElevenLabs, Google, or OpenAI, and can host their avatars on their own cloud or pay Synthesia to host them.

Because the avatar is deterministic — limited to the single story it was trained on — it redirects any unrelated question back to that topic instead of improvising. That limitation showed up when the reporter’s mother, who had already called the demo amazing, tried to stump the avatar with questions only she would know; it just kept redirecting her back to the story, prompting her to joke, “I don’t remember giving birth to two of you.”

Why it matters

Synthesia’s push into deterministic, single-purpose interactive avatars — for its own PR and now for outside test cases — points to a future where AI stand-ins could front some public-facing communication. Whether that extends to journalism, where trust is the core asset, remains an open question the company’s own experiment doesn’t resolve.