ElevenLabs released version 3 of its voice model this summer, and the leap is from narration to performance. Earlier text-to-speech, including ElevenLabs’ own, produced clean, competent, but emotionally flat readings; v3 adds fine-grained expressive control, native multi-speaker dialogue, and inline audio tags that direct delivery — a whispered aside, a laugh, a pause, a shift in emotional register mid-sentence. It’s the difference between a voice that reads your words and one that can act them, and it moves synthetic speech into territory previously reserved for voice actors.
Watch: Introducing Eleven v3 (alpha) — Our Most Expressive Text to Speech Model (YouTube)
Direction, not just synthesis
The interesting design move is that v3 lets you direct the voice rather than only feed it text. Audio tags embedded in the script cue specific deliveries — emphasis, emotion, non-verbal sounds, timing — so a creator shapes a performance the way a director gives notes, rather than settling for whatever neutral reading the model defaults to. Combined with genuine multi-speaker dialogue, that makes conversational scenes, character work, and dynamic narration possible from text alone. For the audio side of this site’s coverage, it’s the voice counterpart to the sound-design and music tools we’ve tracked — the missing piece for anyone assembling a whole soundscape from generated parts.
Where it lands, and the questions it raises
The obvious applications are dubbing, audiobooks, game and animation dialogue, accessibility narration, and prototyping voice for interactive work — anywhere convincing, controllable speech is expensive to record. The same capability sharpens the field’s hard questions: consent and likeness for voice cloning, disclosure, and the real economic pressure on working voice actors. ElevenLabs continues to pair the technology with voice-verification and rights tooling, but expressive, directable synthetic performance makes those concerns more concrete, not less — the more human the output, the higher the stakes around whose voice it’s built from.