Voice Design
Write the voice you can already hear
You know how the voice in your head sounds. Voice Design turns that into a sentence and the sentence into a voice id, in seconds. Here is what a prompt can say, what each part changes, and how to get from close to right.
A prompt that shipped
A British female voice, 20 to 30, glossy and confident, with girly chatter, high pitch, fast pacing, high energy and bright sparkling resonance. Ideal for a friendly receptionist or assistant.
This one produced Freya, a voice in the library today.
What it produced
- FreyaEnglish · 20 to 30 · Female0:00
What a prompt
can say
Eleven things, and no prompt needs all of them. Accent, age range and the three dials carry most of the result; the rest is tuning once the voice is close.
- Accent and locale
- The clause that does the most work, and the one worth being specific about. Gradium maps 57 named accents across five languages, so "a Rioplatense Spanish voice" and "a Bavarian German voice" both land, where "a Spanish voice" or "a German voice" leaves the choice open. Each accent has a page here with the wording that produces it.
- Gender
- Male, female, or left out. Leaving it out is a real option rather than an omission: the rest of the prompt will usually settle it, and a voice that sits between the two is often what a brand actually wants.
- Age range
- Use two numbers. "20 to 30" and "60 to 75" both work, where "young" and "old" underperform because they leave the boundary open. Age and pitch move together, so asking for gravity without moving the age range tends to return a young voice performing gravity.
- Purpose
- What the voice is for, in your words: a support line, an audiobook, a game character. This steers delivery as much as sound, because it says what a good read of the line would be. It is the clause people skip and then miss.
- Character
- Two or three adjectives for the personality rather than the acoustics. "Glossy and confident", "warm and unhurried", "clipped and dry". Past about four they start to cancel out.
- Pitch, pace and energy
- The three dials that move a voice most, and the three to change one at a time. Low to high, slow to fast, calm to boisterous. Anything you can hear a difference in belongs here rather than in the character clause.
- Resonance and timbre
- The texture: bright and sparkling, warm and rounded, thin, gravelly, breathy. Two voices can share a pitch and a pace and still not sound alike, and this is usually why.
- Modifiers and applications
- Anything left: a speech habit, a hint of another language underneath, or the formats the voice has to hold up in. Most useful last, once the voice is close and you are tuning it.
One prompt,
clause by clause
Freya's prompt with each clause named, so you can see which one to change. Every clause below is doing something you can hear.
| Clause | This prompt |
|---|---|
| Accent / locale | British |
| Gender | female |
| Age range | 20 to 30 |
| Character | glossy and confident |
| Pitch | high pitch |
| Pace | fast pacing |
| Energy | high energy |
| Resonance / timbre | bright sparkling resonance |
| Modifiers | girly chatter |
| Applications | a friendly receptionist or assistant |
Read it as a sentence and it is one line of description. Read it as clauses and it is a set of dials. That is the whole method: the sentence is how you write it, the clauses are how you change it.
Getting from close
to right
The first voice is rarely the one you ship, and that is the normal way round. Four habits that make the second one better than the first.
- Ask for three, not one
- A prompt describes a space rather than a point, so generate a small batch and pick from it. Three voices from one prompt also show you which clauses are doing the work and which are being ignored.
- Move one clause at a time
- Pitch, or pace, or the stated purpose. On its own. Two changes at once and you cannot tell which one did it; three and you have a different voice rather than a variation.
- Audition on a line that exposes the accent
- Neutral copy hides the thing you are listening for. Every accent page here lists the words and vowels that give that accent away, and a line built from two or three of them settles it in five seconds.
- Name the purpose before adding adjectives
- "For a support line" changes the read more than a fifth adjective will, because it says what a good delivery of the line would be. Adjectives past the fourth tend to pull against each other.
The wording
for every accent
Each accent has its own page: what makes it sound like itself, the words that expose it, and prompts that produce it. Start from the one you need rather than from a blank box.
- Design an Irish AI voice
- Design an American AI voice
- Design a British AI voice
- Design a Canadian AI voice
- Design a Scottish AI voice
- Design an Australian AI voice
- Design a New Zealand AI voice
- Design an Indian English AI voice
- Design a South African AI voice
- Design a West African English AI voice
- Design a Standard French AI voice
- Design a Parisian French AI voice
- Design a Southern French AI voice
- Design a Belgian French AI voice
- Design a Quebecois French AI voice
- Design a Standard German AI voice
- Design a Bavarian AI voice
- Design an Austrian German AI voice
- Design a Swiss German AI voice
- Design a Castilian Spanish AI voice
- Design a Mexican Spanish AI voice
- Design a Caribbean Spanish AI voice
- Design a Colombian Spanish AI voice
- Design a Chilean Spanish AI voice
- Design a Rioplatense Spanish AI voice
- Design a European Portuguese AI voice
- Design a Standard Brazilian AI voice
- Design an African Portuguese AI voice
Prompts by
what the voice is for
Questions
about prompting
One or two sentences. The limit is 500 characters and most good prompts use half of it. Length is not what makes a prompt work: naming the accent, an age range and the three dials in one sentence beats a paragraph of adjectives, because adjectives past the fourth or fifth start pulling against each other.
Yes, and give it as a numeric range rather than a word. "20 to 30" and "60 to 75" both work, where "young" and "old" underperform because they leave the boundary open. Age and pitch tend to move together, so a prompt asking for gravity without moving the age range usually returns a young voice performing gravity.
Yes. A designed voice is a normal voice id with the same endpoints, the same SDK and the same latency as any library voice, so nothing about your pipeline changes once you have it. If a library voice runs in your agent today, a designed one will too.
Almost always because more than one clause changed at once. Move pitch, or pace, or the stated purpose, on its own and listen again. Change three together and the result is a new voice rather than a version of the one you had, which is useful when you are exploring and frustrating when you were close.
No. Write the prompt in English and name the language and accent inside it. "A Carioca Brazilian Portuguese female voice, 30 to 40" is an English prompt that produces a Portuguese voice.
Between one and five, and more than one is usually worth it. The same prompt describes a space rather than a point, so a batch of three gives you something to choose between and shows you which parts of your prompt are doing the work.
One that exposes the accent rather than a neutral sentence any accent could read. Every accent page here lists the words and vowels that give that accent away, and a line built from two or three of them tells you in five seconds what a paragraph of general copy will not.
Yes. There is no actor to license and no recording session behind it, which is the practical reason teams design a voice rather than cast one: the voice is yours to ship, and you can make a second one that matches it next week.
Now write
yours
One sentence, a few seconds, and a voice id you call from the API or the SDK. No recording session, no actor licensing, and a second voice that matches it whenever you need one.