How to Make an AI Voice Whisper, Laugh or Pause

Three kinds of markup cover almost everything you can direct on the Studio model: a plain-English bracket at the start of a line, six non-verbal tags for sounds like a laugh, and a break tag for pauses of an exact length. To make an AI voice whisper, put something like [whisper in a hushed style] before the words you want whispered.

That's the short version. Below is the syntax for each, what it costs, and the list of things that don't work, because the failure cases are where most of the wasted points go. Everything here is for the Studio model, which is the default on every plan, Free included, so you can try all of it with your first 60 points.

How to make an AI voice whisper

Steering means writing the delivery you want, in your own words, inside square brackets at the start of a line. There's no list of approved commands. You describe the performance and the model follows it.

Write the instruction in English even when the spoken text is in another language (the model covers 200+ of them), keep it lowercase and free of punctuation, and put it before the words it applies to. One direction set per line is the rule. If a speaker needs two moods in one paragraph, split it into 2 lines.

[whisper in a hushed style] Don't wake the others. We leave at first light.

[say quietly in a low tone with deliberate pauses] We need to talk about what happened last night.

The same bracket handles more than whispering. Here's what the documentation lists as controllable, with an example for each.

You can combine several in a single bracket, as in [say sadly with deliberate pauses in a low voice and a hushed style]. Keep them compatible. A hush and a shout in the same line give an unpredictable result.

How to add a laugh to an AI voice

Laughs, sighs and the like aren't directions; they're sounds the model performs. They go inline, exactly where the sound should happen. There are six: [laugh], [sigh], [breathe], [cough], [yawn] and [clear throat].

They also stay out of your captions. A tag shapes the audio, but it never appears in subtitle or SRT output, so you don't need to clean them out afterwards.

Use them sparingly, 1 or 2 per page of script. One laugh per 60 seconds sounds human. Four in the same 60 seconds sound like a sound-effects reel. These six tags also work on VocalLab Studio Flash, so if you've switched to Studio Flash, which costs 0.75 of the standard points, for drafting, the sounds still come through (the bracketed directions above do not).

How to add a pause in text to speech

For a pause of a specific length, drop a break tag inline where the silence should fall. Time can be in seconds or milliseconds, so 1.5s and 1500ms mean the same thing. On long scripts, the milliseconds form is the safer choice. You can use up to 20 break tags in a single request, and like the sound tags they never appear in captions.

Let me think <break time="1500ms" /> Yes. I've thought about it.

You often don't need the tag at all. A comma is a short beat, a period is a full stop, and a blank line between paragraphs gives a longer breath. Capitalizing one word adds stress to it: we NEED a real vacation lands differently from we need a REAL vacation. Reach for a break tag only when the punctuation isn't enough.

In the audiobook editor there are also dedicated pause blocks, from 0.5 to 10 seconds in half-second steps, which are easier than tags when you're laying out a long book.

Directing a whole scene, line by line

Here's how the three pieces combine in a short scene. Take a four-line exchange of about 100 characters per line. That's 400 characters, or 24 points. Add three 28-character directions and a [laugh] and you're at roughly 490 characters, so ⌈490 ÷ 17⌉ = 29 points. The direction layer cost 5 points. That's a fair price for a scene that no longer sounds like four lines read in a row.

The method is the same every time. Generate the line plain first. Add one bracket and listen again. If the change is wrong, rewrite the bracket, not the script, and change one thing per pass so you know what caused what. Two or three passes on a 25-point scene is normal; twenty is a sign the line needs different words, not a different direction.

One more habit: direct only the lines that need it. If every line has a bracket, none of them stands out, and you've paid for 30 or 40 characters of instruction on lines the voice would have read well anyway.

What doesn't work

These 5 failures are worth knowing before you burn points on them. Each one is something the product does today, not a guess.

• A line that is only a direction

  [whisper] with nothing after it is rejected, because there's no word to speak. Retrying won't change that. Put the words after the bracket.

• Directions on the wrong model

  On VocalLab Studio Flash, Pro and Lite, a steering bracket may be read out literally instead of performed. Set Model to VocalLab Studio first.

• Conflicting instructions

  A whisper plus a very loud instruction in the same bracket gives unpredictable audio. Match the delivery to the words too: cheerful directions on grim text fight the model.

• Markdown around the bracket

  A stray ** in front of a direction has made the model read the direction aloud. Paste plain text, not formatted text.

• Pitch as a number

  [in a low tone] describes a sound to aim for. Nothing numeric sits behind it, so don't expect the same depth every time.

On pitch specifically, there's no slider, and the post on why deeper AI voices aren't a setting covers the alternatives.

What directions cost, and a voice to try them on

Brackets count as characters. In the app a generation costs the character count divided by 17, rounded up, and the charge applies to the text you send, tags included. A 28-character direction such as [whisper in a hushed style] adds about 2 points to the line. Cheap, but not free, so a 69-character stack of adjectives is 5 points before a single word is spoken.

The Delivery setting, with Stable, Balanced and Creative, is separate from directions and lives in Voice settings. If a long read keeps drifting, try Stable before adding more brackets.

The voice emotion control page is the best place to practise.

For hushed reads specifically, the whisper voice generator is built for it.

And for anger, sadness and joy there's the emotional voice generator.

• 

  Gentle Female Mindfulness Voice

  Whispered and hushed reads

  • American
  • Soft, nurturing
  Use in Studio
• 

  Calm Male Narrator Voice

  Narration with quiet asides and pauses

  • American
  • Smooth, calm, friendly
  Use in Studio

Frequently asked questions