ElevenLabs Launches Eleven v3, Text-to-Speech Model with Emotional Control

This experimental version of their speech synthesis technology includes new expressiveness features. The model allows generating voices with different emotions through specific audio tags and supports conversations between multiple speakers, characteristics developed after detecting demands from the audiovisual sector.

The system incorporates support for over 70 languages and uses tags inserted in text to modify tone and vocal expressions. Users can apply commands like [whispers], [sighs] or [excited] directly in their scripts to generate specific effects. The technology also allows combining multiple tags in the same phrase to create more complex expressions.

The multi-speaker dialogue functionality operates through an API that processes JSON structures, where each object represents a different speaker's intervention. The system automatically manages transitions between voices, tone changes and conversational interruptions, generating a cohesive audio file that simulates natural conversations.

Development of this version has been oriented toward sectors requiring greater vocal expressiveness, such as film production, video game development, education and accessibility tools. Developers indicate that technical audio quality was no longer the main limitation, but rather the ability to generate nuanced emotions and believable dialogues.

The v3 model requires greater precision in prompt formulation compared to previous versions. For applications needing real-time response or conversational use, maintaining v2.5 Turbo or Flash models is recommended, while a real-time version of v3 is being developed.

This update is part of the evolution since the launch of Multilingual v2, which had already found adoption in professional productions across various sectors. The new model seeks to cover expressive needs that previous versions did not fully satisfy in advanced multimedia content applications.