How to Generate Expressive Speech with ElevenLabs v3 Text to Speech
ElevenLabs v3 is the third generation of the company’s text-to-speech model. The headline change is emotional control: you can tag a line as whispered, excited, or sarcastic and the voice actually shifts.
If you have used earlier ElevenLabs versions, v3 feels less like a narrator reading and more like someone performing.
What is ElevenLabs v3?
It is a TTS model with inline audio tags — short bracketed cues like [whispers] or [laughs] that sit inside your script and steer delivery. It supports multiple languages and lets you clone a voice from a sample, or use the stock voice library.
Access runs through the ElevenLabs web Studio and API. Credits are consumed per character, and the v3 model costs more per character than the older fast models.
How do you generate a clip step by step?
- Create a project in the ElevenLabs Studio.
- Pick a voice, or clone one from at least a minute of clean audio.
- Paste your script and set the model to v3.
- Add tags inline: [excited], [whispers], [sighs], or a pace cue.
- Generate a short section first, listen, then render the full piece.
- Download as MP3 or WAV.
Render a paragraph before the whole script. Tags that sound great in your head sometimes land as overacting on the first try.
Which tags are worth using?
The reliable ones are pacing and volume rather than emotion: [whispers], [pause], [slower]. Emotional tags work but are strong — [angry] on a calm sentence sounds like a different take entirely. Use one tag per few sentences, not one per clause, or the voice wobbles between modes.
What are the costs and limits?
Characters bill on every generation, including ones you throw away, so rough drafts get expensive. Voice cloning needs editorially clean audio; a noisy sample clones the noise. Long scripts can lose consistency, and a voice’s energy in minute one may not match minute ten.
Check your plan’s character quota before committing to a 20-minute audiobook chapter.
Who is it for?
Podcasters adding a second voice without a second person, YouTube creators who want narration with real inflection, and game teams generating placeholder dialogue. It is a poor fit for legal disclaimers where a flat, neutral read is actually required, or for anything needing perfect pronunciation of technical jargon without review passes.
Key Takeaways
- ElevenLabs v3 adds inline audio tags for emotion and pacing control.
- Pacing tags like [whispers] and [pause] are the most reliable.
- Credits bill per character including discarded drafts.
- Voice cloning needs clean audio; noise gets cloned too.
- Strong for narration and dialogue; weak for flat, neutral legal reads.
