Synthesia Creates First Interactive AI Avatar for a Journalist, Trained on One Article

Voica had already sent the reporter an interactive virtual avatar of himself this summer, trained to answer common press questions about Synthesia, such as what it does and how it works. The day before, the reporter had been on a panel where public relations professionals asked whether reporters minded pitches that used AI-generated text. TechCrunch described Voica's avatar as a step beyond that, calling it the final boss of using AI in public relations.

In September, Synthesia invited the reporter to its new office space in New York. The company, originally based in the U.K., is among a group of digital avatar startups that includes D-ID, HeyGen, and Colossyan. It reached a $4 billion valuation earlier this year and said last year that it had crossed $100 million in annual recurring revenue. Synthesia lets enterprises build interactive training videos with AI avatars. It recently launched a product called Roleplay Sessions, which lets employees practice tasks such as sales pitches with an interactive AI avatar that responds and scores their responses.

At the New York office opening, Synthesia asked the reporter if they would like their own AI avatar. The reporter agreed. To build it, the reporter entered a mini film studio inside Synthesia's office, where the company took numerous photos and captured a two-minute recording of the reporter's voice. The reporter had to consent to the avatars being made. Synthesia created a personal avatar that reads whatever script it is given, with and without glasses, and two interactive avatars that can talk back and listen, also with and without glasses.

The interactive avatar was trained on the reporter's article about why venture-backed startups commit more fraud than non-venture-backed startups. TechCrunch said this was the first time Synthesia had made a digital avatar for a journalist, or for anyone outside Voica. The avatar's technology stack includes voice-to-text, video, language, and text-to-voice models. Synthesia's own video and voice models are part of the stack, though the company also allows customers to choose alternatives from other labs such as Cartesia, ElevenLabs, Google, or OpenAI. Enterprises can host their avatars on the cloud of their choice or pay Synthesia to host them.

According to TechCrunch, the voice-to-text model turns speech into text, an agentic language model interprets the text and can take actions based on it, a text-to-voice model turns a response into audio, and a video model built by Synthesia animates the avatar as it talks. Synthesia builds three types of products: a video-creation and distribution platform with classic avatars, where someone types in a script and the avatar repeats it; an agentic platform called Sessions, where people can interact with avatars in surveys or roleplay; and an API platform that lets customers combine Synthesia video and voice models with other technology services to build interactive avatars or other products. The team took a couple of days to make the reporter's avatars.

The reporter first played with the personal avatars and typed a fairly generic script about fall arriving in New York, the reporter's favorite time of year. The voice was fairly accurate, and the reporter was glad it did not pick up any hoarseness from the recording. Some non-tech friends found the personal avatar both interesting and creepy. The interactive avatar was deterministic, meaning it would only say what it was trained to respond to. When the reporter asked where they had worked before TechCrunch or what part of New York they lived in, the avatar redirected each question back to the venture fraud story. Friends did not think the interactive voice sounded much like the reporter and thought the likeness was not as good as the personal avatar, but they still found it close enough to be somewhat creepy. The reporter's mother called it 'amazing.' The reporter's parents tried to ask questions 'only they would know' about the reporter, but the model did not answer and redirected them each time to the venture fraud story.