Introducing Phonon ・ Our first open, on-device TTS model

How do you describe a voice? What we learned from designing 16,000 of them

A ribbon of yellow, blue and pink particles twisting across a dark, grainy background like a sound wave.
8 min read

To build the Gradium voice catalog, we designed more than 16,000 voices from written descriptions and kept about 1,400 of them. Each of those voices started as a few lines of text. Much of what we know about describing a voice comes from writing those lines and listening to what came back.

Describing a voice turns out to be harder than it sounds. When you design a voice and start describing it in a prompt, the accent and the gender usually come to mind first. Other qualities shape a voice just as much, but they're easy to forget or hard to put into words. Age is one of them. Pitch, texture and tone are others, and it isn't obvious whether words like "calm" or "lovely" mean anything to a model at all.

It's also a question we hear often from teams building with Voice Design: how do you write a prompt that gets the voice you have in mind? Many of them share their prompts and feedback with us, and each one helps us understand how people describe voices and keep improving the model.

The gap between the voice you hear in your head and the one that comes back usually comes down to a few words. Below are the patterns our voice team relies on to close it, each with samples you can play and compare.

A voice from a few lines of text

Voice Design creates a voice from a written description. Our catalog has more than 1,400 ready-made voices, each one curated and checked by our voice team. When none of them fits a use case, or when a company wants a brand voice of its own, Voice Design lets you create one that is yours alone.

Every voice Voice Design produces is newly created and unique. The model never outputs the voice of a real person.

The model learned from a large set of voices labeled with attributes like age, accent and persona. When you write a prompt, it maps your words onto what it heard during training, and it fills in anything you leave out on its own. The description is the whole interface.

Take a short prompt like "An old Irish man." reading the line "Ah, sit down there. I'll tell you how it was when I was young." The prompt gives the model a gender, an age and an accent. Everything else, such as the mood, the texture of the voice and the pace, the model chooses on its own, and it can choose differently on the next draft.

Old storyteller
Transcript

“Ah, sit down there. I'll tell you how it was when I was young.”

Short prompt

An old Irish man.

0:00/0:00

f0 106 Hz · 4.4 words/s

Left to chance:rolemoodsound

Where the voice will live matters first

Every voice is built for somewhere. Our catalog sorts voices into use cases like customer service, narration, ads and social media, and characters, because each of those asks for something different from a voice.

Whether a voice is good depends on where it will be used. A bubbly cartoon voice can be perfect for a children's game. Put the same voice on a bank's support line, though, and it would sound completely out of place.

This is why the use case is often the most informative part of a description. "For a professional support line" or "narrating a children's bedtime story" tells the model more than another adjective would, because it implies a whole set of choices about pace and energy.

The dimensions of a voice

The filters in our voice catalog give a good map of what a description can contain, because the model learned from voices labeled along the same lines. For each dimension, the table lists words that work well and how to use them.

DimensionWhat it describesWords to tryHow to use it
LanguageThe language the voice speaksEnglish, French, Spanish, German, PortugueseName it directly
Accent and regionWhere the speaker sounds like they're fromSouthern American, Quebecois, RioplatenseUse a well-known accent name. A broad one beats a rare regional one
GenderMasculine, feminine or gender neutralmasculine, feminine, gender neutralUse one of these three words
AgeAge groupyoung adult, middle-aged, elderlyUse an age group, not a number
PitchHow high or low the voice sitsdeep, low, mid-range, highPlain words like low or high work best
PaceSpeaking speedslow and deliberate, briskPair a speed with a manner: slow and deliberate
ToneMood and attitudecalm and reassuring, cheerful, dryTwo or three words, not a list
TextureThe physical quality of the soundbreathy, raspy, smooth, resonantCombine it with pitch, not on its own
Type or personaA character or rolenews anchor, tour guide, narratorAdd the other parts, not just the role
Use caseWhere the voice will be usedsupport line, audiobook, social adAdd it as context at the end

Pitch is one of the easiest things to get right. "Low" and "high-pitched" mean the same thing to almost everyone, and the model hears them the same way.

Texture sits at the other end. Even people on our voice team admit they can't reliably tell an airy voice from a breathy one. If listeners struggle to hear a difference, the model has a hard time producing it on purpose, so texture words work better as a nudge than as an exact instruction.

Accents land best when they're named precisely. "Mexican Spanish" gives the model more to go on than "Spanish with an accent." For accents our own team can't judge by ear, such as Peruvian or Bolivian Spanish, we ask native speakers from the region to confirm the result before a voice goes into the catalog. The surest picks are widely known accents, like the 28 in our catalog's filters. For a very local accent, start from the broader region it belongs to.

Looking for an accent you can't find? Tell us. We're actively adding more.

Roles help, extremes help more

A role tells the model a lot. "Documentary narrator" or "tour guide" implies a pace and an energy that would otherwise take several adjectives to describe. But a role on its own doesn't replace the rest of the prompt. When we asked for just "A tour guide.", the drafts didn't even agree on whether the guide was a man or a woman: across three drafts, measured pitch ranged from about 110 Hz, where a typical man's voice sits, to 260 Hz, where a high woman's voice sits.

Tour guide
Transcript

“And on your left, a castle that has been standing here for over six hundred years.”

Short prompt

A tour guide.

0:00/0:00

f0 111 Hz · 3.6 words/s

Five parts

Feminine, middle-aged.1 Tour guide.2 Scottish accent.3 Enthusiastic, amused.4 A bright, clear voice with a lively pace.5

0:00/0:00

f0 196 Hz · 3.5 words/s

Adds: a woman's voice (this short-prompt draft came back as a man), a Scottish accent, enthusiasm, a bright voice

Character voices like pirates or robots usually take a few more rounds of rewording. The same goes for niche roles: a news anchor is a safer bet than an obscure figure from a 19th-century novel. What works better is pushing the descriptive words to extremes: adding "very", stacking texture words, or naming a cartoon. For a client who needed playful characters for children, we asked for very high-pitched, squeaky voices, and they came back sounding like cartoon characters. Here is the same mascot, described mildly and then pushed.

Cartoon mascot
Transcript

“Hooray! You found the secret level. Let's go, let's go, let's go!”

Mild

Gender Neutral. Mascot. Friendly, cheerful. A bright, clear voice.

0:00/0:00

f0 159 Hz · 2.3 words/s

Pushed

Gender Neutral. Cartoon mascot. Bubbly, giggly and sing-song. A very high, thin, squeaky voice, bouncy and melodic, very expressive.

0:00/0:00

f0 353 Hz · 2.1 words/s

Pushes: cartoon, bubbly, giggly, sing-song, very high, squeaky

What a good description tends to look like

When we look back at the prompts that produced our best voices, they share a shape. They cover five parts, in this order:

  1. 01WhoGender and age group: masculine, feminine or gender neutral; young adult, middle-aged, older adult or elderly.
  2. 02RoleWho they are: narrator, tour guide, customer-service agent.
  3. 03AccentA well-known accent, in its own sentence: Standard British, Scottish, Australian.
  4. 04MoodTwo or three words: calm, sincere, playful.
  5. 05SoundPitch, texture and pace: a soft, breathy voice with a slow pace.

Every part is optional, but a part you leave out is decided by chance, and it can change from one draft to the next. Short, common words work best. Exact ages, towns, numbers and film references don't help. Voice Design also tidies your prompt into this shape automatically. That fixes the order, but it can't add a part you didn't give, and it keeps vague words. If you know where the voice will be used, add it at the end as context.

Those details fall into two groups. Some describe who the speaker is and stay fixed. Others describe how the speaker delivers a line, and those can change with context. We wrote more about that split in our post on context-aware voices.

Here's what that shape looks like for a support agent:

Feminine, young adult.1 Customer-service agent.2 General American accent.3 Calm, reassuring.4 A warm, clear voice with an even pace.5

And for a storyteller, the complete version of our old Irish man:

Old storyteller
Transcript

“Ah, sit down there. I'll tell you how it was when I was young.”

Five parts

Masculine, elderly.1 Grandfather.2 Irish accent.3 Playful, amused.4 A soft, slightly raspy voice with a gentle, easy delivery.5

0:00/0:00

f0 155 Hz · 4.0 words/s

Neither prompt is long. Each covers all five parts in common words, one short sentence per part.

Small changes in wording, big changes in sound

Some of the biggest differences we hear come from changing a few words. Filling in the parts a short prompt left out, or pushing a mood or sound word further, has more than once turned a voice that sounded like an ordinary narrator into one with a clear personality.

The pairs below show what that looks like. Most start from a short prompt, the kind people usually write first, and then fill in all five parts. The last one starts complete and pushes the mood and sound further. Both prompts in a pair read the same line, so the wording is the only thing that changes. Play them side by side and compare.

Superhero
Transcript

“Everybody out. Now! The bridge is going to collapse. Move, move, move!”

Short prompt

A superhero.

0:00/0:00

f0 139 Hz · 1.9 words/s

Five parts

Masculine, young adult.1 Superhero.2 General American accent.3 Urgent, intense.4 A loud, forceful voice, shouting, with a fast, hurried delivery.5

0:00/0:00

f0 227 Hz · 2.0 words/s

Adds: shouting, urgency, a fast pace

Surf instructor
Transcript

“Alright, paddle out past the break, then wait for the next set to roll in.”

Short prompt

A surf instructor.

0:00/0:00

f0 132 Hz · 3.6 words/s

Five parts

Masculine, young adult.1 Surf instructor.2 Australian accent.3 Relaxed, amused.4 A light, smooth voice with an easy, casual delivery.5

0:00/0:00

f0 115 Hz · 3.5 words/s

Adds: an Australian accent, a relaxed, casual delivery

Documentary narrator
Transcript

“Deep beneath the ice, life has found a way to survive the long polar night.”

Short prompt

A documentary narrator.

0:00/0:00

f0 155 Hz · 3.3 words/s

Five parts

Masculine, middle-aged.1 Documentary narrator.2 Standard British accent.3 Calm, sincere.4 A deep, rich voice with a slow, measured tempo.5

0:00/0:00

f0 111 Hz · 3.5 words/s

Adds: a British accent, a deep, rich voice, a slow tempo

Meditation guide
Transcript

“Breathe in slowly, and let your shoulders drop as you breathe out.”

Short prompt

A meditation guide.

0:00/0:00

f0 157 Hz · 2.8 words/s

Five parts

Feminine, young adult.1 Meditation guide.2 Standard British accent.3 Calm, gentle.4 A soft, breathy voice with a slow, measured pace.5

0:00/0:00

f0 204 Hz · 2.9 words/s

Adds: a soft, breathy voice, a slow pace

Villain
Transcript

“You came all this way. How touching. Now, sit down.”

Mild

Masculine, middle-aged. Villain. Serious. A low voice.

0:00/0:00

f0 117 Hz · 2.3 words/s

Pushed

Masculine, older adult. Villain. Intense, dark. A very low, gravelly, rough voice with slow, weighted delivery.

0:00/0:00

f0 157 Hz · 2.2 words/s

Pushes: intense, dark, very low, gravelly, rough, slow and weighted

Let the script do the acting

The prompt decides who is speaking. Pauses, pace and punch come mostly from the text the voice reads. With the same superhero voice, "Everybody out, now!" and "Everybody out. Now!" sound different: the full stop gives the shout its own beat. For the old Irish man, splitting "Ah, sit down there and I'll tell you how it was" into two sentences adds a pause and slows the start.

SuperheroSame voice, same prompt
Comma

“Everybody out, now! The bridge is going to collapse. Move, move, move!”

0:00/0:00
Full stop

“Everybody out. Now! The bridge is going to collapse. Move, move, move!”

0:00/0:00

A full stop instead of a comma gives the shout its own beat.

Old storytellerSame voice, same prompt
Comma

“Ah, sit down there and I'll tell you how it was when I was young.”

0:00/0:00
Full stop

“Ah, sit down there. I'll tell you how it was when I was young.”

0:00/0:00

Splitting the sentence adds a pause and slows the start.

From a first prompt to the voice you want

Most of what we learned comes down to three habits. Give every prompt all five parts, so nothing is left to chance. When a character needs more, push the mood and sound words further. And let the script do the acting, since the way a line is written shapes how it's delivered.

These are the same habits our voice team used to build a catalog of more than 1,400 voices from over 16,000 designs. They work for a calm support agent as well as for a squeaky cartoon mascot, and a clear first description means fewer drafts, and fewer credits, before you find your voice.

Related posts

Frequently Asked Questions