ElevenLabs Sound Effects: How to Make SFX with AI (Guide)

The fastest way to make sound effects with AI is ElevenLabs Sound Effects (SFX): you describe the sound in one sentence — what makes it, how, what it's made of and where — pick a duration from 5 to 22 seconds or Auto, and get an MP3 file a few seconds later. It covers footsteps, impacts, creaking doors, editing whooshes, UI sounds and steady beds like rain or city hum, and with the Loop switch it makes seamless loops for long scenes. Prompts can be up to 450 characters. The model runs in the IIshki studio in your browser. Below: a prompt formula, 15 ready prompts by job, five steps, and the mistakes that make a sound come out wrong.

What ElevenLabs Sound Effects does

It is a text-to-sound generator: a description goes in, one MP3 file (44.1 kHz, 128 kbps) comes out. There is no recording or library search — the sound is created fresh for your description. The model card has four settings.

What it won't do: intelligible speech, or music with a steady beat. For voice there is ElevenLabs Text to Speech; for a track, ElevenLabs Music.

Which model to pick for the job

If you are choosing between voice engines, see the roundup for voicing videos and the guide to dubbing video with AI.

The sound effect prompt formula

Weak prompts are one or two words: "explosion", "door". The model guesses, and you get something different every time. A strong prompt answers six questions in order:

1. Source — what makes the sound: a door, a glass bottle, sneakers.
2. Action — what happens to it: opens slowly, falls, slides.
3. Material — what it's made of: oak, metal, wet asphalt. A wooden door and a steel one sound nothing alike.
4. Space — where: an empty stairwell, a small bathroom, an open field. This decides whether there's an echo.
5. Distance — how close the mic is: right up close, a couple of metres, far away.
6. Character — last: sharp, dull, tense, soft.

An example built on the formula:

Heavy oak door creaking open slowly, empty stone stairwell, mic one metre away, hollow echo, tense

Describe what you hear, not what you see. "A beautiful sunset over the sea" is a picture; "gentle surf, a few distant gulls, light wind" is a sound.

15 ready prompts by job

Editing and transitions

Fast airy whoosh moving left to right, clean, no reverb, single

Short dull bass hit for a cut, dense, with a short tail

Film camera shutter click and a short film wind, vintage

UI and games

Bright metallic coin chime with a sparkly tail, single

Soft plastic button click, very close, dry

Short rising synth success tone, friendly

Foley for videos

Sneaker footsteps on wet asphalt, steady pace, close mic

Glass bottle shattering on a concrete floor, sharp and bright

Coffee pouring into a ceramic cup, quiet kitchen, very close

Beds and atmosphere (with Loop)

Heavy rain on a tin roof, distant rolling thunder, steady

Summer forest by day: birds, rustling leaves, a distant woodpecker

Night city outside a window: occasional cars, distant siren, hum

Horror and tension

Low drone with distant metallic scraping, dark and uneasy

Slow footsteps on a creaky wooden floor overhead

Wind whistling through a narrow gap, cold, empty

How to make a sound effect: 5 steps

1

1. Open the model

Go to ElevenLabs SFX. No input files are needed — just the description.

2

2. Write the description with the formula

One sound per generation. If you need a sequence like "footsteps, then a door, then silence", make three separate sounds and join them in an editor: each comes out clean, and you place the seams exactly where your cut needs them.

3

3. Pick the duration

• Auto For single hits, clicks and whooshes: the model trims the sound to fit.
• 5 s A short effect with a tail, a transition.
• 10–22 s Beds, atmospheres, long actions (running water, a passing car).

Match the duration to your edit up front: you can trim a finished sound, but you can't stretch it without artifacts.

4

4. Set prompt adherence and looping

Strict keeps the model close to the wording — good for a specific hit or click. Loose lets it improvise — good for atmosphere, where character matters more than each detail. Start at Medium and move one step at a time. Turn on Loop only for steady beds.

5

5. Generate 2–3 takes and place them on the video

Every run is generated from scratch, so the same prompt gives different takes. Keep the best, download the MP3 and drop it onto your video. Place a whoosh slightly before the cut, and duck the music a few decibels under a loud effect so the hit lands without the track jumping. If you sign in with Telegram, the file is sent there too.

Sound and video: which comes first

There are two workflows that work.

Video first, sound second. Make the clip — for example in Kling 3.0 with the Sound switch off — then add effects and beds in ElevenLabs SFX and music in ElevenLabs Music. This gives you the most control, because every layer is separate.

Sound first, video second. Generate the sound or track, upload it as a reference to Seedance 2.0 (up to 3 audio files) and say in the prompt that the motion follows the audio — for example, that hits land on the hero's steps. This suits music videos and rhythmic scenes. If you want scene audio straight away, with no separate layer, use Veo 3.1 or Kling 3.0 with sound on: both generate video together with sound. The differences are covered in the Veo 3.1 vs Kling 3.0 comparison.

Common mistakes

What's next

Build the full soundtrack for your video: voice with the ElevenLabs guide, music with the ElevenLabs Music walkthrough, and effects and beds here. If the video isn't made yet, start by animating a photo, and chain several steps together on the canvas.