Short answer: to make a scary or creepy AI voice, write the line with Grok speech tags (<lower-pitch>, <slow>, <whisper>, [breath] and [long-pause]) and generate it with Grok text to speech. The box above already holds a scary line; the buttons load four more, from a creepy whisper to a jump scare. Pick one of 28 voices and press Generate. Trying it needs no account.
How to make a scary voice
- Pick a scary line with the buttons above the tags, or type your own. Each example is under 200 characters, so it plays without an account.
- Edit the words. Keep the tags around the parts you want low, slow or whispered; the tags are not read out.
- Pick a voice. All 28 voices accept the same tags. The clips above use Eve.
- Press Generate and listen. Press Download MP3 to save it; without an account the file ends with a short spoken credit.
A line that builds tension, then whispers the reveal:
[breath] <lower-pitch><slow>Do not turn around.</slow></lower-pitch> [long-pause] <whisper>Whatever is standing behind you has been there since the lights went out.</whisper>
Which tags make it scary
All of these are among the 26 speech tags in xAI's text to speech docs. Which tag suits which effect is our suggestion:
| Tag | What it does | Use it for |
|---|---|---|
<lower-pitch>…</lower-pitch> | Lowers the voice for the wrapped words | A narrator or a threat |
<slow>…</slow> | Drags the words out | Dread, a line that lingers |
<whisper>…</whisper> | Whispers the wrapped words | Secrets, something right behind you |
<soft>…</soft> | Speaks quietly | The calm before a scare |
[breath], [inhale], [exhale] | A breathing sound | Fear, effort, someone close by |
[long-pause], [pause] | A silence | Tension before the next line |
<loud>…</loud> | A raised, pushed delivery | The jump scare at the end |
[chuckle] | A short laugh | A villain enjoying it |
In our two pre-recorded clips of the same words in the same voice, the tagged version ran 10.0 seconds against 5.2 for the plain read, came out about 4 dB quieter on average, and left a 1.3-second silence at [long-pause]. That is the effect to expect: slower, quieter, with room for the listener to get nervous.
Ideas to try
These are suggestions, not tests we ran for each use:
- Horror stories and creepypasta narration:
<lower-pitch><slow>for the narrator,<whisper>for what the monster says. - Halloween props and haunted houses: a short whispered line on a loop, with
[long-pause]between repeats. - Games and tabletop sessions: a villain's warning with
[chuckle]at the end. - Analog horror and YouTube videos: a flat, slow emergency-broadcast read; try several voices on the same text.
Tips for a creepier voice
- Short sentences. A pause between short lines is scarier than one long sentence.
- Whisper the reveal, not the setup. Let the first line sound almost normal, then drop into
<whisper>. - Stack tags sparingly.
<lower-pitch><slow>works; wrapping every word in three tags tends to sound uneven. - Try two or three voices. Our voice pages suggest Zagan for mystery and horror narration and Orion for dramatic narration (suggestions, not xAI descriptions); the voices page has every voice reading one sentence.
- Need effects? Grok adds no reverb or distortion. For a demon or monster sound, generate the line here and add a pitch shift in an audio editor.
More moods are on the text to speech with emotion page, a whole script in a whisper is on the whisper text to speech page, and every tag is in the speech tags guide.
Limits, price and privacy
Signed out, you get 3 previews a day of up to 200 characters, downloadable with a short spoken credit at the end; a free account removes the credit, and a Google account raises the limit to 1,000 characters per clip. The tag characters count toward the limit. Your text goes to xAI's model through our Cloudflare function to make the audio and is not stored.
If you build your own app, the same model costs $15 per million characters through xAI's API; see the pricing calculator. There is no separate price for a scary voice; xAI bills text to speech per character.
Last updated: