To make a voice with ElevenLabs Voice Design, describe it in a short prompt, in this order: language, gender and age range, audio quality, persona, emotions, and timbre, pace and delivery. ElevenLabs gives this exact order as its recommended format.1
Use a few words for the persona, two or three emotions, and one or two sentences on timbre, pace and delivery. Then press generate, listen to the three voices you get, and save the one you like.1
- The format: language, gender and age, quality, persona, emotions, then timbre, pace and delivery.
- Be specific: ElevenLabs says more descriptive prompts tend to be more accurate, but a short one such as “A calm male narrator” can work for a plain voice.
- Two traps: do not write “accent” when you mean intonation, and avoid effect words such as “reverb” or “phone”.
- Preview text matters: it should match the voice, and a full sentence beats a short phrase.
- Experimental: ElevenLabs says Voice Design is still experimental and that a professional voice clone is more consistent.
Voice Design makes a brand-new voice from words alone, like describing an actor to a casting director. It is not a clone: no real person’s recording goes in. Timbre means the physical sound of a voice, such as deep, airy or raspy. Guidance scale is a dial for how strictly the tool follows your description.
What is Voice Design, and when should I use it?
The plain point: use it when the voice you need is not in the voice library and you want to try ideas quickly.
ElevenLabs says Voice Design “helps creators fill the gaps when the exact voice they are looking for isn’t available in the Voice Library,” and that it is “best for quick exploration and iteration, and output quality can vary depending on your prompt and use case.”1 You find it under Voices, My Voices, Add a new voice, Voice Design, or through the API. If you need the most consistent, production-ready quality, ElevenLabs points you to a Professional Voice Clone instead, if one in its library fits.1
Each time you press generate you get three options. ElevenLabs says the only charge is the credits for the preview text, once, even though three samples come back, and that saving a voice “will take up one of your voice slots.” This site lists no prices; ElevenLabs’ plans page has the current numbers.1
What is the best format for a Voice Design prompt?
The plain point: follow ElevenLabs’ fixed order. Put the language and dialect first, so the voice does not drift.
ElevenLabs’ recommended format is three lines. The first gives the language, the gender and age range, and the quality level. The second gives a persona of two to five words and two or three emotion adjectives. The third gives one or two sentences about timbre, pacing and delivery.1 It also says: “Be explicit about language and dialect — always specify the language and regional variant in the first sentence to prevent drift.”1
Here is that format as a fill-in template, followed by a worked example written for this page (not ElevenLabs’ own).
ElevenLabs says shorter prompts can work too: “sometimes short and simple prompts can also work, especially when you’re aiming for a more neutral or broadly usable voice.” Its own example is “A calm male narrator.” But “more descriptive and granular prompts tend to yield more accurate and nuanced results.”1
Start from the three-line format. Put the language and dialect in the first sentence. Add detail only for the things that matter to your project.
How do I describe age, tone and gender?
The plain point: describe how the voice sounds, with specific words, and keep sound separate from mood.
| What | ElevenLabs’ advice | Words it lists |
|---|---|---|
| Age | Says: use specific terms; age shapes maturity, vocal texture and energy. | Examples: “in their 20s”, “woman in her 40s”, “man in his 80s”. |
| Tone and timbre | Says: the physical quality of the voice, separate from emotion or attitude. | Examples: “gravelly”, “airy”, “warm”, “tinny”. |
| Gender | Says: it tends to influence pitch and weight, so describe the sound, not only the label. | Example: “A lower-pitched, husky female voice”. |
| Pacing | Says: state the speed and rhythm, and say where and why it changes. | Examples: “Deliberate and measured pacing”, “Staccato delivery”. |
Where each line comes from.1
How do I get the accent right?
The plain point: name the accent clearly, say how strong, and never use “accent” for anything but a regional way of speaking.
ElevenLabs says “thick” often gives better results than “strong” when you say how marked an accent should be, and that you should avoid vague words such as “foreign” or “exotic.”1 The warning that matters most: “Don’t use ‘accent’ when you mean intonation — this can trigger unwanted dialect shifts.” If you mean a rising, questioning or flat way of speaking, say that instead.1 It also suggests combining the accent with other traits for control, with its own example of “A sarcastic old woman with a thick New York accent, speaking slowly.”1
What about audio quality and sound effects?
The plain point: ask for the clarity you want, and leave out words that name effects.
ElevenLabs says the tool aims for “clean and natural-sounding audio” by default, and lists quality levels you can name: “Ok quality”, “Good quality”, “Very good quality”, “Excellent quality”, “Studio quality” and “Broadcast quality.” But it warns that these phrases “can sometimes reduce the accuracy of the prompt in general if the voice is very specific or niche.”1 It also says to avoid “FX words — terms like ‘reverb,’ ‘echo,’ ‘phone,’ or ‘tape’ can negatively affect output quality.” For a deliberately rough sound, such as a phone call, it suggests leaving the quality words out or writing phrases like “Low-fidelity audio”.1 If you want the voice to sound as if it is in a room, ElevenLabs’ Text to Speech page has separate audio effects that you add after the voice exists, which is covered in how to make ElevenLabs voices sound natural.
Why does the preview text matter?
The plain point: the preview text is the script the voice performs, so it should fit the voice you described.
ElevenLabs says the preview text “acts like a performance script — it sets the tone, pacing, and emotional delivery that the voice will attempt to match,” so it should “complement the voice description, not contradict it.” Its example of a clash is a calm, reflective voice asked to shout an angry line, and it adds that “The model will try to reconcile that mismatch, often leading to unnatural or inconsistent results.” It also says “longer preview texts tend to produce more stable and expressive results.”1 In this site’s use, that means writing two or three sentences in the voice’s own register, for example:
What does the guidance scale do?
The plain point: it sets how strictly the tool follows your words. Higher is stricter but can hurt sound quality on unusual voices.
ElevenLabs says it “dictates how closely the Prompt is adhered to. Higher values will stick to the prompt more strictly, but could result in poorer audio quality if the prompt is very niche, while lower values will allow the model to be more creative at the cost of prompt accuracy.” Use high values “when accent or tone accuracy is of paramount importance.”1 The example prompts on ElevenLabs’ page use guidance scales between 20% and 40%, so those are sensible places to start. Loudness is a separate control: it sets the volume of the preview and of the saved voice.1
If a voice sounds wrong, change one thing at a time: first the accent wording, then the preview text, then the guidance scale. That way you learn which change helped.
What if the voice comes out wrong?
The plain point: most problems trace back to one of five wording choices, and ElevenLabs names each.
| Problem | Likely cause | Try |
|---|---|---|
| Wrong dialect or odd accent | Cause: “accent” used for intonation, or no dialect in the first sentence. | Try: name the language and region first; describe intonation as intonation. |
| Muffled or odd sound | Cause: an effect word such as “reverb” or “phone”. | Try: remove it and add a quality level. |
| Does not match your niche idea | Cause: a very specific voice plus quality words, or a low guidance scale. | Try: drop the quality phrase; raise guidance a little. |
| Abrupt, uneven preview | Cause: a very short preview phrase. | Try: a full sentence or short paragraph. |
| Flat, not performing | Cause: Voice Design voices on v4 “may be less performative.” | Try: audio tags in the script, or another model. |
On v4, ElevenLabs says “Voices created with Voice Design work with Eleven v4, our latest model, and with earlier models including Eleven v3. They support audio tags. On Eleven v4 they may be less performative than on earlier models.”2 The tags are explained on the ElevenLabs cheat sheet.
What should you not use Voice Design for?
The plain point: do not use it to imitate a real person, and do not expect the same voice every time from a short prompt.
ElevenLabs’ own advice is to use a Professional Voice Clone when you need the highest consistency, as noted above. On real people, this site’s advice is plain: describe a type of voice, not a named person. A description built to sound like a particular known speaker is the same legal question as cloning one, covered on what the law says about cloning a voice or a face. For a voice that belongs to nobody, see also prompting music and voice. That paragraph is this site’s reading, not an ElevenLabs statement.
Questions people ask
Is Voice Design the same as voice cloning?
No. Voice Design makes a voice from a text description; cloning copies a voice from recordings. ElevenLabs describes Voice Design as letting you “create a unique voice from a text prompt.”2 The site’s page on voice cloning covers the other route.
How many voices does Voice Design make at once?
Three. You listen to all three, then save one. Saving uses one of your voice slots (source above).
Can I use a Voice Design voice with Eleven v4?
Yes. ElevenLabs says these voices work with Eleven v4 and earlier models, and that they may be less performative on v4 (source above).
Will the voice sound the same every time?
The saved voice is the same voice, but each generated take still varies, as the natural-sounding guide explains. For the most consistent results, ElevenLabs recommends a Professional Voice Clone over Voice Design (source above).
Where to go next
- “How do I make it sound natural?” See voice, model, settings and script.
- “How do I add emotion and pauses?” See the ElevenLabs cheat sheet.
- “Are there other voice tools?” See ElevenLabs vs the alternatives.
- “What is text-to-speech?” Read the plain answer, then what voice cloning is.
- “Can I use Claude to drive ElevenLabs?” See what ElevenLabs skills are.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does ElevenLabs say to put in the first sentence of a Voice Design prompt?
- The persona and three emotions
- The language and regional variant
- The guidance scale
Show the answer
Answer: The language and regional variant. Its guide says to specify the language and regional variant in the first sentence to prevent drift.
2. What can go wrong if you write “accent” when you mean intonation?
- It can trigger unwanted dialect shifts
- The voice comes out silent
- Nothing, the two words mean the same
Show the answer
Answer: It can trigger unwanted dialect shifts. ElevenLabs warns that using “accent” for intonation can trigger unwanted dialect shifts, so describe intonation or delivery instead.
3. What should the preview text do?
- Be as short as possible
- Contradict the description so the model can balance it
- Match the voice you described, and a full sentence beats a short phrase
Show the answer
Answer: Match the voice you described, and a full sentence beats a short phrase. The preview text acts like a performance script, so it should fit the voice description, and longer text tends to give steadier, more expressive results.
- 1ElevenLabs, Voice Design, read at source 27 Oct 2026“Native <Language>. <Gender>, <Age range>. <Quality level>.” then “Persona: <2–5 words>. Emotion: <2–3 adjectives>.” then “<1–2 sentences about timbre, pacing, delivery>”; “When you hit generate, we’ll generate three voice options for you.” ↑a“Persona: <2–5 words>. Emotion: <2–3 adjectives>.” then “<1–2 sentences about timbre, pacing, delivery>”; “When you hit generate, we’ll generate three voice options for you.” ↑b“Voice Design helps creators fill the gaps when the exact voice they are looking for isn’t available in the Voice Library.” and “Voice Design is best for quick exploration and iteration, and output quality can vary depending on your prompt and use case.” ↑c“Professional Voice Clones (PVC) are currently the highest quality voices on our platform”; “Voice Design is still experimental.” ↑d“The only charge for using voice design is the number of credits to generate your preview text, which you are only charged once even though we are generating three samples for you.” and “After generating, you’ll have the option to select and save one of the generations, which will take up one of your voice slots.” ↑e“For consistent results, structure your prompt using this format:” followed by the three lines above. ↑f“Be explicit about language and dialect — always specify the language and regional variant in the first sentence to prevent drift.” ↑g“In general, more descriptive and granular prompts tend to yield more accurate and nuanced results.” and “sometimes short and simple prompts can also work, especially when you’re aiming for a more neutral or broadly usable voice. For example, “A calm male narrator” might give you exactly what you need”. ↑hAge “helps define its maturity, vocal texture, and energy. Use specific terms”, with “Young adult” / “in their 20s”, “woman in her 40s”, “man in his 80s”; Tone/Timbre “Refers to the physical quality of the voice, shaped by pitch, resonance, and vocal texture. It’s distinct from emotional delivery or attitude.” with “Gravelly” / “raspy”, “Airy” / “breathy”, “Warm” / “mellow”, “Tinny” / “metallic”; Gender: “you can push beyond simple categories by describing the sound instead of the identity”, “A lower-pitched, husky female voice”; Pacing: “Being explicit about pacing is essential”, “If the pacing shifts, be sure to indicate where and why”, “Deliberate and measured pacing”, “Staccato delivery”. ↑i“‘thick’ often yields better results than ‘strong’ when describing how prominent an accent should be.” and “Avoid overly vague descriptors like ‘foreign’ or ‘exotic’ — they’re imprecise and can produce inconsistent results.” ↑j“Don’t use “accent” when you mean intonation — this can trigger unwanted dialect shifts. Instead, describe intonation, emphasis, or delivery patterns.” ↑k“Combine accent with other traits like tone, age, or pacing for better control. E.g., ‘A sarcastic old woman with a thick New York accent, speaking slowly.’” ↑l“By default, ElevenLabs aims to produce clean and natural-sounding audio”; the list “Ok quality” to “Broadcast quality”; “Including these types of phrases can sometimes reduce the accuracy of the prompt in general if the voice is very specific or niche.” ↑m“Avoid FX words — terms like “reverb,” “echo,” “phone,” or “tape” can negatively affect output quality.”; for intentionally low quality, “either leave out quality descriptors entirely or explicitly include phrases like” “Low-fidelity audio”. ↑n“The preview text acts like a performance script — it sets the tone, pacing, and emotional delivery that the voice will attempt to match.” “your preview text should complement the voice description, not contradict it.” “The model will try to reconcile that mismatch, often leading to unnatural or inconsistent results.” “longer preview texts tend to produce more stable and expressive results.” ↑o“Dictates how closely the Prompt is adhered to. Higher values will stick to the prompt more strictly, but could result in poorer audio quality if the prompt is very niche, while lower values will allow the model to be more creative at the cost of prompt accuracy.” and “High values are recommended when accent or tone accuracy is of paramount importance.” ↑pthe “Example Prompts and Text Previews” table lists guidance scales from 20% (“Movie Trailer Voice”, “Squeaky Mouse”) to 40% (“Relatable British Entrepreneur”, “New Yorker”); “Controls the volume of the Text to Preview generation, and ultimately the voice once saved.” ↑q
- 2ElevenLabs, Voice Design, FAQ, read at source 27 Oct 2026“Voices created with Voice Design work with Eleven v4, our latest model, and with earlier models including Eleven v3. They support audio tags. On Eleven v4 they may be less performative than on earlier models.” ↑a“Voice Design lets you create a unique voice from a text prompt.” ↑b
Read at source 27 Oct 2026: ElevenLabs, Voice Design (prompting guide, parameters, example prompts and FAQ). ElevenLabs marks Voice Design as experimental, so the guide may change.
The worked example, the preview text, the problem-and-fix table and the advice not to describe a named person are this site’s own, built from the guide, not claims from ElevenLabs. No voice was generated for this page.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK