How to Actually Direct ElevenLabs v3 Like a Voice Actor (Not a TTS Engine)

1. Write stage directions in brackets, not adjectives in your script

This is the whole shift. If you remember nothing else, remember this one.

ElevenLabs Audio Tags are words wrapped in square brackets that the new Eleven v3 model can interpret and use to direct the audible action. They can be anything from [excited], [whispers], and [sighs] through to [gunshot], [clapping] and [explosion]. Treat them exactly like the script directions you’d hand a voice actor in a booth. Short, clear, sitting right before the line they apply to. Audio tags are simple text instructions you place inside [brackets] directly in your prompt. Think of them as short, clear direction notes, the same type of cues you’d give a voice actor in the studio.

The old workflow was to type a paragraph, hit generate, hate it, regenerate, hate it again, fiddle with sliders, regenerate. That’s done. The Eleven v3 model changed this entirely. The v3 architecture natively understands bracketed “stage directions.” Instead of complex coding syntax, you act as the audio director using simple tags in square brackets. Stop describing how you wish the line sounded in your head, and start writing the direction into the script.

Practical rule: every emotional beat that matters gets a tag. Every non-verbal sound you want, a sigh, a laugh, a sharp inhale, gets a tag. Pauses you actually want? Tag them. If you’re not bracketing, you’re not directing. You’re hoping.

2. Pick the right stability mode (and stop leaving it on default)

The stability slider is the most misunderstood control in v3, and most people leave it wherever it landed. Don’t. The stability slider is the most important setting in v3, controlling how closely the generated voice adheres to the original reference audio.

Here’s the real map. Creative: More emotional and expressive, but prone to hallucinations. Natural: Closest to the original voice recording, balanced and neutral. Robust: Highly stable, but less responsive to directional prompts but consistent, similar to v2.

If you took anything from step 1, here’s the corollary: For maximum expressiveness with audio tags, use Creative or Natural settings. Robust reduces responsiveness to directional prompts. In plain English: if you spent 20 minutes salting your script with [whispers] and [laughs] and you’re on Robust, you wasted 20 minutes. The model will politely ignore most of them.

My rule of thumb: Natural for narration, audiobooks, anything where you want most of your tags to land but you don’t want the model going off on its own. Creative for character work, ads, anything where you want the model to swing. Robust only when you need v2-style consistency for long-form, low-direction reads.

3. Write longer prompts than feels necessary

This one’s unintuitive, and it’s why so many short demos sound flat.

With V3’s alpha release, short prompts can cause inconsistent outputs. To mitigate this, aim for prompts exceeding 250 characters. This allows the model to build context and maintain flow, resulting in smoother, more natural-sounding narration. 250 is the floor, not the goal. The model uses the surrounding text to figure out the emotional shape of the line you actually care about.

So if you need a single dramatic sentence, give it a setup sentence and a follow-through. Testing a tag on one line? Write three lines and put the tag on the middle one. The model treats your script like a scene, and a scene with one line of dialogue and no context will sound like exactly that.

This also means: stop generating one-word vibe checks (“[excited] Wow!”) and concluding v3 is broken. It isn’t. You’re starving it.

4. Pick a voice that can actually do the thing you’re asking for

You cannot tag your way out of the wrong voice. This burns people constantly.

Match tags to your voice’s character and training data. A serious, professional voice may not respond well to playful tags like [giggles] or [mischievously]. And the inverse: Don’t expect a whispering voice to suddenly shout with a [shout] tag.

The model’s emotional range is bounded by what’s in the voice’s training audio. If the voice was cloned from a calm, measured corporate read, no tag is going to produce a manic laughing fit. There’s nothing in the source for the model to draw on. Users should select voices with emotional range in their training data to maximize the impact of audio tags.

Practical move: before you commit to a voice for a project, run your three hardest tags against it on a test paragraph. Whisper, laugh, shout, whatever your script actually demands. If two of three land convincingly, you’ve got a workable voice. If none of them do, switch voices before you write another line. You’ll save yourself days.

One more thing if you’re a returning ElevenLabs user with cloned voices: Professional Voice Clones (PVCs) are currently not fully optimized for Eleven v3, resulting in potentially lower clone quality compared to earlier models. During this research preview stage it would be best to find an Instant Voice Clone (IVC) or designed voice for your project if you need to use v3 features. Translation: your prized PVC isn’t the move on v3 right now. Use an IVC or a library voice.

5. Stack tags for compound emotions, not single ones

Single tags are fine. Stacked tags are where v3 starts to sound like an actor.

You can even stack them (e.g., [whispers][nervous] I think someone is out there.). The two cues fuse. You don’t get a whisper, then nervousness. You get a nervous whisper, which is what you actually wanted.

The same trick works for an emotional arc inside a single line. Straight from the ElevenLabs docs: Here are some frequently used tags to direct emotional performance: Emotional states: [excited], [nervous], [frustrated], [sorrowful], [calm] Reactions: [sigh], [laughs], [gulps], [gasps], [whispers] Cognitive beats: [pauses], [hesitates], [stammers], [resigned tone] Tone cues: [cheerfully], [flatly], [deadpan], [playfully]

Mix categories. Emotional state plus reaction plus tone cue, [tired][sigh][resigned tone], gives the model a fully specified beat to play. One tag tells it the temperature. Three tags tell it the temperature, the physical action, and the attitude.

A grief example that ships, lifted from a real prompt pattern: [sobbing] I don’t know why I’m [sniff] [sniff] crying this hard… [crying] it just feels like [sigh] a lot right now. [clear throat] I know I’ll be okay, I just need a minute to [sigh] let it out. Notice how the tags sit inside the sentence, not just at the front. That’s the move.

6. Use punctuation and capitalization as a second set of controls

Tags are the loud lever. Punctuation is the subtle one, and v3 actually listens to it.

Punctuation and capitalization – Ellipses create pauses; capitalization adds emphasis; proper punctuation improves natural rhythm. Read that twice. The model treats your script formatting as direction.

So: write the way you’d write a screenplay, not a press release. Em dashes for interruptions. Ellipses for trailing-off hesitation. A SHOUTED WORD in caps for actual emphasis (sparingly, caps everywhere just makes the model push too hard). A comma where you want a beat. A period where you want a stop.

And on pauses specifically, there’s a subtle gotcha that catches people coming from v2. Eleven v3 does not support SSML break tags. Use the techniques described in the Prompting Eleven v3 section for controlling pauses with v3. If you’ve got old SSML <break> tags in your scripts, strip them. Use ellipses, em dashes, or natural sentence breaks instead. Alternatives to include dashes (- or, ) for short pauses or ellipses (…) for hesitant tones. However, these are less consistent.

Less consistent than what v2 gave you, but more reliable than ignoring it.

7. Use Text to Dialogue for multi-character work, don’t fake it with one voice

If your script has two people talking, you’ve been doing it the hard way: generate each side separately, splice them together in your DAW, hope the energy matches. Stop.

The official page also introduces a Text to Dialogue flow where structured speaker turns can generate a multi-speaker audio output with pacing, turn changes, and interruptions. This feature alone justifies the v3 upgrade if you do any narrative work. The model handles the handoffs, the natural pacing between turns, the way a real conversation lurches and overlaps, in a single generation. You can’t fake that by stitching solo takes.

You can now also create multi-speaker dialogues that feel spontaneous, handling interruptions, shifting moods, and conversational nuance with minimal prompting. Pair it with tags inside each speaker’s lines and you’re directing a scene, not assembling a podcast in post.

One important caveat, because real-time use is a common ask: ElevenLabs says teams should stay on v2.5 Turbo or Flash for real-time and conversational use cases for now. v3 is the model for produced content. Audiobooks, ads, characters, trailers, narration. It’s not the model for a live chatbot voice. Use the right tool for the job.

A bonus, because it’ll save you generations: hit the Enhance button before you regenerate

When you’ve written a line and you can hear in your head exactly how it should sound but you don’t know which tags to reach for, let the UI help.

In the ElevenLabs UI, you can automatically generate relevant audio tags for your input text by clicking the “Enhance” button. Behind the scenes this uses an LLM to enhance your input text with the following prompt: You can combine multiple audio tags for complex emotional delivery. It’s not always right, but it’s a fast way to see which tags the model itself thinks fit your line, and you can hand-edit from there instead of starting from a blank script.

The one habit that ties it all together: stop thinking of v3 as text-to-speech and start thinking of it as a one-person voice cast you’re directing. Every tag is a note to the actor. Every punctuation mark is a timing instruction. Every stability setting is how much rope you’re giving them. The people getting cinematic output aren’t getting lucky, they’re directing, and the model is following. Start treating the script as a stage direction document and your hit rate goes up overnight.