ElevenLabs Launches Eleven v3 Alpha with expressive speech, multi-language & dialogue API

ElevenLabs has released Eleven v3 in alpha, introducing a text to speech model focused on expressiveness and realism. The new version supports over 70 languages, making it suitable for a broader set of global applications. Inline audio tags now enable control of tone, emotion, and non-verbal cues, letting users produce nuanced and realistic voices by annotating prompts with tags like excited, whispers, and sighs. Example combinations, such as 'happily``shouts We did it! laughs,' allow for richer emotional inflection. You can hear some of these tonal and expressive changes in this video.

The new Text to Dialogue API enables multi-speaker audio generation using JSON input to define speaker turns, transitions, and emotional shifts. The model also improves stress, cadence, and naturalness in speech output.

Currently, Eleven v3 is available through the ElevenLabs app and includes an 80% usage discount until the end of June. While a real-time version is in development, live use is recommended with Eleven v2.5 Turbo or Flash. API access for v3 is not public yet but can be requested, and enterprise or self-serve users may activate the model via the UI. Full documentation and prompting guide are available for reference.