Synthesia moves digital twins from reading texts to dialogue simulation and sales team training

Audio-visual digital embodiment technologies are shifting from mere tools for reading pre-written texts to interactive agents capable of listening and responding instantly, a step that reveals the direction of the AI-driven video production sector toward corporate training models and live simulation. Synthesia, a company founded in the United Kingdom and recently expanded with a new office in New York, highlights this shift by building interactive digital twins that rely on a composite technical architecture integrating speech-to-text models, agent-based language models, and generative audio-visual models.

The company, whose market value exceeded four billion dollars earlier this year and that recorded recurring annual revenue of more than one hundred million dollars last year, competes in a market that includes firms such as D-ID, HeyGen and Colossian. Its new technical focus, however, is on an interactive platform that enables organisations to run interactive simulation sessions for employees, for example training sales teams to deliver pitches and receive immediate feedback and evaluations from the digital twin, alongside its traditional video-generation platform and API platform.

These interactive models rely on a distributed computing architecture that separates perception functions, response generation and visual rendering.The technical pipeline starts with a model that converts the user’s speech into written text, after which an agent-based language model interprets the meaning and determines the response path, then a speech-generation model turns the text into audio, and finally Synthesia’s video model animates facial features and synchronises lip movement. The company gives its institutional clients flexibility to choose voice and language providers from labs such as OpenAI, Google, Eleven Labs and Cartesia, as well as the option to host these twins on the client’s private cloud or on the company’s own infrastructure.

The creation of these characters undergoes a verification process that includes on-site recording in a dedicated studio, capturing multiple images and a voice recording of no more than two minutes, with the requirement of explicit consent from the character’s owner, and the technical processing takes a few days. Current interactive models use a deterministic approach that adheres to the facts and data they were trained on, such as answering a specific press kit or a particular company’s policies, automatically redirecting the user to the relevant domain if questions fall outside the training scope, which limits the risk of linguistic hallucination.

This technical pathway changes the equation for professional training and customer service in the Gulf markets, Egypt and the broader Arab region.Companies in the financial and telecommunications sectors, which rely on large sales and outsourcing teams in Cairo, Riyadh and Dubai, face the option of reducing recurring training costs by building interactive training agents that assess an employee’s skills before they go into the field. The ability to host on private clouds also gives regional organisations the capacity to adopt the technology while meeting data-sovereignty requirements and strict controls over the flow of sensitive information.