The short version
- Describe age and gender first, then pitch, texture, pace and mood, plus an optional role.
- Generate two or three takes of the same short test sentence; every take is a new speaker.
- Save the take you like as a voice in My Voices.
- Use the saved voice for all your texts: it keeps the same speaker across long texts and languages.
Open Voice Design. Stuck for ideas? The 🎲 Surprise me button writes a random description that follows the rules below.
How Voice Design works
The voice designer reads your description and invents a speaker that fits it. It doesn't look the voice up anywhere; it imagines one. That has two consequences:
- The same description gives a different voice each time. Think of it as casting: you describe the role, and every take brings a new actor who fits the brief.
- Saving a take locks it. A saved voice is a fixed speaker. It sounds the same in every future generation, in any language, however long the text.
The recipe: five qualities and a role
| Quality | Examples that work |
|---|---|
| Age and gender | young adult woman, middle-aged man, older woman |
| Pitch | deep, low, warm mid-range, bright, light, husky |
| Texture | smooth, slightly raspy, breathy, crisp, velvety, gravelly, polished |
| Pace | slow and deliberate, relaxed, steady, brisk, lively |
| Mood | calm, confident, warm, playful, serious, mysterious, reassuring |
| Role (optional) | like a documentary narrator, a late-night radio host, a patient teacher |
Put them together with this template:
A [age] [man/woman] with a [pitch], [texture] voice, speaking at a [pace] pace with a [mood] tone, like a [role].
Leading with age and gender keeps the most important facts unmissable. Keep the description under about 40 words: one or two sentences beat a paragraph.
Examples by use case
Audiobook narration
A mature woman with a warm, smooth voice, speaking at a steady pace with a gentle tone, like an audiobook narrator.
YouTube explainer
A thirty-something man with a bright, crisp voice, speaking at a brisk pace with a friendly, confident tone, like a tech reviewer talking to camera.
Meditation and sleep content
A young adult woman with a low, breathy voice, speaking at a slow, deliberate pace with a calm, reassuring tone, like a meditation guide.
Game or audio-drama villain
An older man with a deep, gravelly voice, speaking at a slow pace with a cold, mysterious tone, like the villain in an audio drama.
E-learning and training
A middle-aged woman with a warm mid-range, clear voice, speaking at a relaxed pace with a patient, encouraging tone, like a teacher explaining something.
What doesn't work, and what to do instead
Celebrities and real people
Voice Design doesn't imitate real people, and requests naming them are declined. The model couldn't reproduce them anyway; in our tests, “like [famous actor]” produced a generic voice. Describe the qualities you associate with that voice instead: deep, slow, gravelly, warm, authoritative.
Ethnicity, nationality and accents
Descriptions such as a nationality or an accent are unreliable: the designer often falls back to a neutral, General American-sounding voice. For speech in another language, set the Language selector and write the text in that language. The description itself can stay in English.
Contradictions and audio effects
“Calm but very excited” or “whispering and shouting” pull the voice two ways at once; pick one. Studio terms like reverb, EQ or “on the phone” don't apply: the designer creates a clean, dry voice.
Let casting do the picking
The designer doesn't always follow every word: a “deep” or “fast” voice comes out as asked most of the time, not always. Cast 3 voices generates three short takes of your description (about 10 seconds), measures them, and marks how each one matches what VoiceAct can check: man or woman, child, deep or high, fast or slow. The best match is labelled; play them, then click Use this voice to generate your text with it, or Save to keep it in My Voices. Descriptions in any of the ten languages are understood. Casting costs one short sample's characters.
A workflow that gets results fast
- Write a short test sentence that resembles your real script (same language, similar mood).
- Generate two or three takes. If none is right, change one or two words of the description rather than rewriting it.
- When a take sounds right, click Save Voice. Saving needs at least five seconds of audio, so test with a sentence or two, not a single word.
- Generate your real script with the saved voice from My Voices.
Long texts and consistency
Long scripts are generated in segments of about a minute each and joined with natural pauses. A saved voice is identical in every segment. If you generate a long text directly in Voice Design without saving, VoiceAct designs the voice from your first sentences and then locks it for the rest, so the speaker doesn't change halfway through.
Other languages
- Voice Design works for all ten supported languages; set the Language selector to match your text.
- English-looking words inside other languages (brand names, URLs) can pick up an English accent. Spelling them the way they're pronounced in your language helps.
- Russian is in beta: expect the occasional mispronunciation.
Troubleshooting
| Problem | Try this |
|---|---|
| Wrong age or gender | Move age and gender to the very start of the description. |
| Every take sounds similar | Change the pitch and texture words; they shape the timbre most. |
| Flat or robotic delivery | Add a mood and a role, and use punctuation in your text (commas, question marks, exclamation marks). |
| A name or term is mispronounced | Spell it phonetically in the text. |
| Great take, but you moved on | History keeps every generation to replay and download. To reuse a voice, save it right after generating. |
Using designed voices responsibly
Designed voices are synthetic and you may use the audio commercially. Every file carries an inaudible AI watermark and, in downloads, signed Content Credentials that mark it as AI-generated. See the Terms for the full rules.