ElevenLabs: Eleven v3

Eleven v3 is ElevenLabs’ expressive text-to-speech model for emotional dialogue, character discussion and audiobook narration. It converts a written script into spoken audio and uses context to shape emotional range and delivery. The producer supports both single-voice speech and a dialogue workflow with multiple speakers. This card concerns the named v3 edition, kept separate from v3 Conversational and the later v4 family.

Inline audio tags can direct emotion, vocal delivery and non-verbal reactions. The manufacturer gives examples such as [sad], [angry], [happily], [whispers], [shouts], [laughs], [clears throat] and [sighs]. Place the cue near the words or reaction it should affect. For a narrative scene, write the emotional transition into the actual dialogue and use the tag to make the intended sound explicit, rather than relying on a detached description of the scene.

V3 does not support SSML <break> tags. Its prompting guide uses audio tags, punctuation, ellipses and text structure for pauses and pacing. That approach supplies a performance direction rather than an exact timing contract. For a hesitant line, an ellipsis and the relevant vocal cue can make the intended hesitation clear. If a recording must fit a precise timeline, use the selected take and editing workflow to place it rather than treating a pause tag as a fixed-duration unit.

Voice selection remains important. The documented guidance says a delivery that is already present in the voice’s training material is easier to reproduce, while unusual requests may be less reliable. Match the voice to the character or narrator and review whether the chosen tags produce the intended delivery. The current v3 prompting section says Professional Voice Clones are not fully optimised for this edition; v4’s documented clone improvements and native pronunciation controls are separate edition features. The product guide also marks Speed, Similarity and Speaker Boost unavailable for ordinary v3. Prepare pacing and delivery through the voice, tags, punctuation and text structure rather than inheriting another edition’s sliders.

Eleven v3 supports more than 70 languages, including English, French, German, Japanese, Korean, Spanish, Portuguese and Russian. Choose a voice appropriate to the language and regional accent and prepare the actual spoken wording before synthesis. Language support does not translate a script automatically. For a localised character scene, preserve the speaker’s intent and reaction while adapting the wording naturally for the audience.

A text-to-speech request supports up to 5,000 characters. Divide longer scripts at complete scene or narration boundaries so each segment has a clear beginning and ending. The producer provides create and stream modes for speech as well as create and stream modes for dialogue. For dialogue, associate each turn with its intended speaker and retain their sequence. Keep an emotional change with enough surrounding text to make the transition understandable.

The shared producer audio workflow offers MP3, PCM and telephony encodings such as μ-law, with additional formats listed in its TTS guide. Pronunciation dictionaries provide a way to keep names and technical terms consistent. The dictionary guide explicitly lists v3 for phoneme rules and limits its multilingual phonetic extension to v4; that does not make v4’s slash-wrapped native IPA syntax a v3 feature. Prepare important spellings or spoken aliases and check the result with the selected voice.

For a practical performance brief, provide a finished episode rather than an outline: a narrator’s introduction, the character’s actual line and the reaction that closes the scene. Specify which content is spoken, where a vocal reaction belongs and the emotional direction of each beat. Review pronunciation, voice identity, unintended reading of directions and the transition between speakers. The result is an expressive recording of the supplied script, with a chosen take that can be reused or revised in the audio project.