AI voice design creates a new voice from a written description. For a persona, start with a voice brief (language, gender, age, quality, timbre, pacing, delivery), write the prompt in ElevenLabs' recommended format, compare the three previews on real captions, and save the winner as the persona's permanent voice.
AI voice design creates a brand-new synthetic voice from a written description, so you can give an AI persona a voice that fits its character without copying a real person. You describe the language, gender, age, tone and pacing, the tool generates preview voices, and you save the best one. The skill is in the description: a clear voice brief turns vague results into a voice that sounds intentional.
Checked October 2026 against ElevenLabs' Voice Design documentation and prompting guide and its Voice Design API quickstart. Other tools work differently.
Voice design sits alongside cloning in a creator's audio toolkit. For how the two compare and which tools offer them, start with our complete guide to voice cloning and AI voices.
AI voice design: the workflow end to end
- 01Voice brief
Seven attributes on one page
- 02Prompt
In the recommended format
- 03Previews
Three options per generation
- 04Test on real captions
Hooks, lists, questions
- 05Lock it
Save and use everywhere
The order matters because the brief stops you from generating voices at random. Without one, it is easy to keep regenerating until something sounds nice in isolation, then discover it does not fit the persona's look, niche or audience.
Step 1: Write the AI persona voice brief
Before opening any tool, write down seven attributes. They map directly onto what ElevenLabs' prompting guide says the model listens for: age, gender, tone, accent, pacing, emotion and style, plus audio quality.
| Attribute | What it decides | Example for a lifestyle persona |
|---|---|---|
| Language and variant | Who the audience is | Native English, General American |
| Gender | How the persona presents | Female |
| Age range | Maturity and energy | Late 20s |
| Quality | Clean or deliberately lo-fi | Studio-quality recording |
| Timbre | The texture of the voice | Warm, slightly husky |
| Pacing | Speed and rhythm | Upbeat, quick but clear |
| Delivery | Emotion and style | Friendly, conversational, a little playful |
Match the voice to the persona you already have. A persona's name, niche and visual style set expectations; a voice that contradicts them feels off even when viewers cannot say why. Our guide to AI influencer names and brand positioning covers the persona side.
Step 2: How to create a custom AI voice prompt
ElevenLabs recommends this format for Voice Design prompts:
Its prompting guide makes several practical points. More descriptive prompts usually give more accurate, nuanced voices, although a simple prompt like "a calm male narrator" can be enough for a neutral voice. For clean output, add a phrase such as "perfect audio quality" or "studio-quality recording". For deliberately rough audio, such as a phone call effect, leave quality descriptors out or say so explicitly.
- Writing "accent" when you mean intonation
- Leaving language and dialect implicit
- Only adjectives, no pacing or delivery
- Contradictory traits in one line
- No quality descriptor for clean audio
- Describe intonation and emphasis directly
- Name language and regional variant first
- Add pacing and delivery sentences
- Pick one dominant mood
- Add "studio-quality recording"
The first pitfall is the one ElevenLabs calls out: using "accent" when you mean intonation can trigger unwanted dialect shifts. Its guidance is to be explicit about language and regional variant in the first sentence to prevent drift.
Step 3: Design a voice for your character with previews
When you generate, ElevenLabs creates three voice options from your prompt. You pay credits only for the preview text, once, even though three samples are generated, so it is worth writing preview text that sounds like your real content rather than a generic sentence.
- 1Write preview text from your niche
A hook, a list item and a question in the persona's style.
- 2Generate three previews
Listen once without judging, then again with the brief beside you.
- 3Score against the brief
Age, timbre, pacing and delivery: does each match?
- 4Refine one attribute at a time
If pacing is wrong, change only the pacing sentence and regenerate.
- 5Shortlist two, test on longer scripts
A 30-second caption reveals issues a single line hides.
Changing one attribute at a time is the quickest way to learn how the model responds. If you rewrite the whole prompt every round, you cannot tell which change helped.
Step 4: Lock the voice as brand audio
Once you have a winner, save it to your voice library. From there it works like any other voice in ElevenLabs, including through the API. Then treat it as fixed: same voice, same settings, every video. Viewers recognise a persona by its voice as much as its face, and switching voices mid-series undoes that recognition.
- The exact prompt that produced the voice
- The saved voice name and ID
- Generation settings you use for every video
- Three reference clips: hook, list, question
- Words to spell phonetically for correct pronunciation
- Where and how you disclose AI-generated audio
A style sheet also protects you if you change tools later. With the prompt and reference clips saved, you can recreate a close match elsewhere instead of starting from scratch.
Disclosure and platform rules
A designed voice is still AI-generated audio, and platforms have rules about that. Meta's misinformation policy says it requires people to disclose, using its AI-disclosure tool, whenever they post organic content with photorealistic video or realistic-sounding audio that was digitally created or altered, and that it may apply penalties if they do not. Meta also adds "AI info" labels to content when it detects industry-standard signals or when people self-disclose.
For a persona channel, the simplest approach is to be open about it everywhere: say in the bio that the account is an AI persona, use the platform's AI label on posts with synthetic voice, and never use a designed voice to imply that a real person said something. That keeps you inside platform rules and avoids the trust problems that come when audiences discover a voice was synthetic.
- Bio: a short line such as "AI persona, created by" followed by the brand or studio name.
- Posts: turn on the platform's AI label when the voice or video is realistic.
- Sponsored content: follow the platform's branded content rules as well as AI disclosure.
Designed voice or cloned voice?
A designed voice suits a fictional persona: it belongs to the character and copies no one. A cloned voice suits creators who want a persona to sound like them. If you go the cloning route, our guide to cloning your own voice covers recording setup and consent rules, and our ElevenLabs tutorial covers the interface.
ElevenLabs' billing docs say the free plan includes three custom voice slots for Voice Design voices, while cloning needs the Starter plan or above, so designing a persona voice is a low-cost place to start. For the full persona workflow, from look to voice to content plan, our AI Influencers program covers each step.
AI voice design: FAQ
What is AI voice design?
AI voice design creates a new synthetic voice from a text description instead of a recording. You describe the voice, such as language, gender, age, tone and pacing, and the tool generates preview voices that match. Because no real person is recorded, it suits AI personas that should not imitate anyone. ElevenLabs calls its version Voice Design.
How do I write a good voice design prompt?
ElevenLabs recommends a format: native language, gender, age range and quality level first, then one or two sentences about timbre, pacing and delivery. More detail usually gives more accurate results, though simple prompts like "a calm male narrator" can work for neutral voices. Adding "studio-quality recording" helps when you want clean audio. Checked October 2026.
How many voices does Voice Design generate?
ElevenLabs says that when you press generate, it creates three voice options. You are charged credits only for the preview text, once, even though three samples are generated. You can then save the one you like to your voice library and use it anywhere you use other voices.
Can I design a voice on the free plan?
ElevenLabs' billing docs say the free plan offers three custom voice slots you can use for voices made with Voice Design, while cloning needs the Starter plan or above. Plans change, so check the current pricing page before relying on that for a project.
Why does my designed voice keep changing accent?
ElevenLabs warns against writing "accent" when you mean intonation, because it can trigger unwanted dialect shifts. Name the language and regional variant in the first sentence of your prompt, and describe intonation, emphasis and delivery patterns separately. That keeps the voice anchored to the dialect you intended.
Should my AI persona use a designed voice or a cloned voice?
A designed voice is usually the safer choice for a fictional persona, because it does not copy a real person. A cloned voice suits creators who want the persona to sound like themselves. Whichever you choose, pick one and keep it for every video, so the voice becomes part of the persona's identity.
Give your persona a voice people recognise.
The AI Influencers program, included in All Access, covers personas, voice, image and video workflows, with the other three programs, live coaching and the private community in one subscription.
Free AI creator updates
Join the free Telegram channel for voice, image and video tool changes as they happen.