ElevenLabs Introduces Eleven v4 for Directable Emotional Text-to-Speech

ElevenLabs has introduced Eleven v4, its most emotive text-to-speech model to date, designed to transform scripts into directed performances rather than monotone readings. The model allows users to control tone, pacing, and non-verbal sounds like sighs and laughs through in-line instructions.

Directing Performance with Audio Tags

The core feature of Eleven v4 is the use of Audio Tags embedded directly within the script to guide the AI's delivery. Users can insert tags such as [furious], [sarcastic], [whispers], or [crying] to dictate the emotional state of a specific line. The model interprets these cues to adjust pitch, pace, and intensity, effectively acting as a director for the voice actor.

Beyond explicit tags, Eleven v4 leverages punctuation and emotional continuity to shape the output. Ellipses slow down delivery, dashes indicate interruptions, and exclamation marks add intensity. If a sentence begins with an emotional tag, the model maintains that tone across the line, allowing for iterative scene building.

Practical Application and Limits

The platform demonstrates the model's range by rendering the same sentence—"I didn’t think you’d actually come back"—with five different emotional tags: [furious], [excited], [sarcastic], [crying], and [worried]. Each tag produces distinct acoustic changes, such as hardened consonants for anger or stretched vowels for sarcasm.

For optimal results, ElevenLabs recommends:

• Iterating on direction: Swapping tags (e.g., [worried] instead of [sad]) rather than rewriting text if the delivery misses the mark.
• Generating full scenes: Using paragraphs or passages up to 10,000 characters per generation to provide sufficient context for the model.
• One emotion per phrase: Applying a single tag per sentence segment to avoid conflicting instructions that degrade output quality.

Availability

Eleven v4 is available in the ElevenLabs Text to Speech app. To encourage adoption, Creator+ subscribers are granted 3x credits for the new model until October 12.