This guide helps you get a voice that sounds right for your content, whether that is a calm narrator, an energetic sports commentator or a character in a film. It covers three things: choosing the right way to get a voice, creating it in the Voice Library, and testing it so it delivers the emotion you need.
Before you start: which option is right for you?
Option | Best when | What you need |
A. Use a library voice | You want to start today and a ready-made voice fits | Nothing. Just browse and test |
B. Design a voice from a description | No library voice fits and you have no recording | A written description of the voice |
C. Clone a voice from audio | You need a specific person's voice | A clean recording (or your microphone) and the speaker's permission |
D. Clone a speaker from a dubbing project | The voice is already in a video you are dubbing | An open dubbing project |
Tip: Always try Option A first. It needs no recording and you can compare several voices in minutes.
Option A: Find a voice in the Voice Library
Click Voice Library in the left menu. You land on the Explore tab.
Start from a Handpicked for your use case card (Conversational, Narration, Characters, Social Media, Educational, Advertisement, Entertainment), or use Search and the Gender, Language, Age and Category filters. Search for words that describe the style you need, for example sports, announcer, presenter, narrator or energetic.
Click ▶ to preview each voice, and click the bookmark icon to keep the ones you like in Saved.
Test two or three shortlisted voices with the same short script (see Test your voice below), then keep the one that sounds best.
What makes a voice hold up under emotion: a clear, slightly bright tone, crisp consonants and room to rise in pitch and volume. Soft, breathy or very deep voices tend to flatten when you ask them to sound excited.
Option B: Design a voice from a description
Step 1. In the Voice Library, click Create Voice (top right).
Step 2. Select Text-to-Voice Synthesis and click Continue.
Step 3. In Describe your voice, describe the voice in 30–50 words (up to 400 characters).
Step 4. In Voice script, type a line that sounds like your real content (up to 500 characters), so you hear the voice in context.
Step 5. Click Generate voice. Each generation uses credits.
Step 6. Under Select the voice, play the three versions and select the one you like. Click Generate New for three new versions, or Continue to keep your choice.
Step 7. Add the voice details (see Voice details below) and click Save.
Step 8. Read the Confirm consent message and click Confirm. You'll see Voice created successfully!
What to put in the description: gender, age range, accent, energy, speaking style and voice texture.
Example: live sports commentator
"An energetic male sports commentator in his late 30s with a neutral British accent. Bright, clear, slightly raspy voice that rises in pitch and volume at big moments. Fast, punchy delivery, like a live stadium broadcast."
Example: documentary narrator
"A confident female presenter in her early 30s with a neutral international English accent. Warm, clear and authoritative, with a steady, measured pace and crisp diction, like a polished documentary narrator."
Not sure where to start? Click Inspire me to fill in an example description and script, then adjust them.
Option C: Clone a voice from a recording
In the Voice Library, click Create Voice, select Upload a voice sample and click Continue.
Drag your file into the upload area, or click Use your voice to record with your microphone.
If the recording has some background noise, tick Audio contains background noise. Leave it off for clean studio audio.
Click Next, add the voice details, click Save, then Confirm.
File limits: .wav, .flac, .aac or .mp3 · 3 seconds to 10 minutes · up to 40 MB.
Recording checklist
The clone copies what it hears, so the recording decides the result.
Length: 3 seconds is the minimum the upload accepts, but aim for 1–2 minutes. Clips under about 30 seconds clone less consistently.
One speaker only: no other voices, crowd noise, music or room echo under the voice.
Clean sound: quiet room, good microphone, no distortion, steady volume.
Match the style you want: if you need an excited commentator, record excited commentary, not a flat reading. A calm recording gives a calm voice.
Varied content: different sentences, names and numbers, so the clone hears all the sounds of the language.
Format: WAV or FLAC gives the best quality; MP3 and AAC also work.
Permission: you must have the speaker's consent to clone their voice.
Option D: Clone a speaker from a dubbing project
Open your dubbing project and go to the speaker settings (top right of the editor).
Click the speaker's name and scroll to Clone and Save to Library.
Name the voice and click Clone.
The same recording checklist applies: the speaker should have enough clear, solo speech in the video.
Voice details
When you save a voice, the Add Voice Details step asks for:
Voice name: something you will recognise later, e.g. Commentator – Male – Live.
Language: the language and accent of the voice, e.g. English (United Kingdom). This shapes pronunciation.
Gender: Male, Female, Not Known or Not Applicable.
Age category: Young, Adult or Old.
Category: Conversational, Narration, Characters, Social Media, Educational, Advertisement or Entertainment.
Description: filled in from your voice description; edit it if you like.
Your new voice appears under My Voices in the Voice Library.
Test your voice
Always test a new voice in Text to Speech before you use it in a project, and test every voice the same way so you can compare them fairly.
Watch this short video
Step 1. In My Voices, click the voice's ⋮ menu and choose Use voice.
Step 2. Select Text to Speech and click Use Voice. Text to Speech opens with your voice already set as the Narrator voice.
Step 3. Click Model and choose MARS8-Instruct (Most Expressive) or MARS-X Instruct (Beta). These are the models that follow emotion tags. The default, MARS8.1-Flash-Beta, is built for speed rather than emotion.
Step 4. Set Language to match your voice (for example English (United Kingdom)). If they don't match you'll see a red warning, and the accent may drift.
Step 5. Type a short test script (up to 1,000 characters): one calm line, one build-up line and one big moment, with emotion tags (see below). Tags turn orange when they're recognised.
Step 6. Click Begin Generating Speech and listen. Use Regenerate to try again, or the download button to save the audio. Earlier attempts are under the History tab.
Add emotion with tags
Put a tag in square brackets before the words it should colour. A tag can hold one or several words.
[tense, hushed] Final lap, and it's all square between the two of them...
[building excitement] She makes her move on the outside,
[excited, shouting] she's through, she's going to do it! [ecstatic] Gold!
You don't have to type tags yourself: under Emotion Controls, click a quick tag such as [laughing], [sighing] or [whispering], or click Browse all to search every tag by sound effect, speaking style and emotion. The Subtle → Extreme scale shows how strong each emotion is.
Emotions work best when they build step by step, for example [uneasy] → [nervous] → [anxious], rather than jumping straight to the strongest one.
Tags that work well:
Moment | Tag |
Build-up |
|
Big moment |
|
Celebration |
|
Near miss |
|
Narration / intro |
|
Personal story |
|
Promo / recap |
|
For the full list and intensity levels, see Emotion Tag Gradation Guide.
Write the script the way it should be spoken
Short sentences for energetic styles: "He's clear. He's away. Nobody's catching him." Long sentences sound read, not spoken.
Punctuation sets the pace: commas and "..." slow it down, "!" adds lift, a full stop resets.
Write numbers as words: "nine point eight seven seconds", "twenty twenty-six", "three-nil".
Unusual names: if one comes out wrong, spell it the way it sounds or use IPA spelling.
No stage directions such as "(crowd roars)" in the text. They are read aloud. Use a tag instead.
One moment per generation (1–3 sentences) gives the most control.
If it doesn't sound right
Change one thing at a time: first the voice, then the tags, then the script, so you can hear what made the difference.
Flat or too calm: check the model is MARS8-Instruct or MARS-X Instruct, try a stronger tag (see Emotion Tag Gradation Guide), or use a voice that was recorded or described with more energy.
Wrong accent: check that Language in Text to Speech matches the voice, and check the recording. A strong accent in the source carries into the clone.
Too fast: add commas or "..." before the line, or split it into a separate generation.
Inconsistent clone: re-record a longer, cleaner sample (1–2 minutes) and create the voice again.
Manage your voices
The Voice Library has four tabs: Explore (all available voices), My Voices (voices you created), Workspace Voices (voices shared with your workspace) and Saved (voices you bookmarked).
Click ⋮ next to a voice in My Voices to:
Use voice in Stories, Narration or Text to Speech
Copy ID (for the API)
Share, View, Add to Collection or Add to workspace
Delete voice
The number of custom voices you can create depends on your plan. See Price Plans.
Need help choosing or tuning a voice? Contact support.