ElevenLabs SFX v3 Builds a Whole Sound Scene From One Prompt. Write It Like a Sound Editor, Not a Shopping List
Most AI sound effect tools are good at one sound at a time. Ask for a door creak and you get a door creak. Ask for a door creak in the rain with a dog barking down the street, and you used to get a muddle: the rain swallows the door, or the dog never shows up, or all three arrive at the same volume like they were recorded in the same cupboard.
ElevenLabs released SFX v3 on October 29, and the headline claim is narrow and specific. The new model is "better at following several instructions within a single prompt." That sounds like a minor line in a changelog. For anyone who cuts video, makes podcasts or builds games, it is the difference between a sound effect and a sound scene.
It also changes how you should write the prompt. The people who get the most out of this model will be the ones who already think like a sound editor.
What changed, in plain terms
A sound effect library gives you separate files. Footsteps. Rain. Distant traffic. You stack them on your timeline, set the levels, and spend twenty minutes making them feel like they happened in the same place. That last part is the craft, and it is also the tedious part.
SFX v3 tries to do the stacking for you. You describe the scene, and one generation renders the layers together. The rain sits behind the door instead of on top of it. The dog is distant because you said distant, and it sounds distant because the model placed it there, not because you pulled a fader down.
ElevenLabs lists a few other improvements. Gongs, chimes and whistles come out with fewer unwanted artifacts, which matters if you have ever generated a bell that crackled at the top. The developer docs add that music and vocal sounds are handled better and that results sound more natural overall. The company says it beat the previous model in its own benchmark testing. That is ElevenLabs grading its own homework, so treat it as a claim and let your ears decide.
Where you can use it, and what it costs
SFX v3 lives inside ElevenCreative, the company's app for creators: the Sound Effects tab, Flows, Studio and the Music Editor. It is also in the ElevenLabs mobile app and in the API for developers.
The working limits are generous for most edits:
- Each generation runs from a tenth of a second up to 30 seconds.
- You can download MP3, or 48 kHz WAV for effects that do not loop. WAV is the one you want for a video edit.
- A loop option makes ambience that repeats without an audible start or end.
- A "prompt influence" control slides from literal to loose. High follows your words closely. Low adds variation.
Money comes down to two things. First, the licence. The ElevenLabs sound effects page says that "as long as you have a paid account, you can use our sound effects royalty-free in commercial projects." The free plan is personal use only. If the sound is going into client work, a monetised video or a game you sell, you need a paid plan. Plan prices and first-month discounts change often, so check the pricing page on the day you sign up.
Second, credits. The docs say that if you set the length yourself, it costs 40 credits per second. A full 30-second ambience bed is 1,200 credits, so compare that against your plan's monthly allowance before you batch a dozen of them. The docs do not say what a generation costs when you let the model pick the length, so check your balance after your first few tries.
One snag for developers. The blog post says SFX v3 is "available now," but the docs list the API model as limited access, and warn that it "may not be available in every region." If you call it and get an "Unknown model" error, that is why. The in-app version is the easier place to start.
The prompt is a mix, so write it like one
Here is the advice ElevenLabs gives in its own launch post, and it is the most useful sentence in the whole announcement:
"Start with the main action, add the surrounding sounds, and listen to how they work together."
That is how a sound editor works. You lay the sound that carries the story first, the thing the audience must hear. Then you build the world around it. Then you listen and adjust.
A shopping-list prompt does the opposite. "Door, rain, dog, car, wind" gives the model five equal items and no idea which one matters. You get five equal sounds.
A sound editor's prompt has a hierarchy. Compare these two:
Shopping list: "door creak, rain, dog barking, car passing"
Scene: "A heavy wooden door creaks open slowly, close to the listener. Behind it, steady rain on a tin roof. Far down the street, a dog barks. A car passes on wet asphalt in the distance."
The second prompt tells the model what is near and what is far, what is steady and what is an event. It reads like the spotting notes you would hand to a foley artist. That is the whole trick.
Two more habits help.
Describe the space. "In a tiled bathroom," "in an empty parking garage," "in a small carpeted bedroom." Room sound is half of what makes layers feel like they belong together, and it is the part that takes the longest to fake by hand.
Use the verbs a person would use. The ElevenLabs docs give a telling example: "a couple arguing in French" worked better than "a couple fighting in French." Small wording changes swing results. When a prompt almost works, change one word at a time.
Put this into practice: a 15-minute test
Do this before you trust the model with a real job. You need a paid or free ElevenLabs account and a scene from something you are working on.
- Pick one real shot. A scene from your edit that currently has no sound, or temp sound you dislike. Note its length.
- Write the main action alone. One sentence, the sound the scene cannot live without. Generate it with prompt influence set high. Listen.
- Add the world. Keep your first sentence and add two or three surrounding sounds, near to far, plus the room. Generate again.
- Generate a looser take. Same prompt, prompt influence lower. Compare. Sometimes the loose version finds a texture you would not have thought to ask for.
- Drop the best take under your picture. Export WAV. Put it on the timeline and watch the shot, not a waveform. Does the main action land where the picture needs it?
- Fix the timing in the edit. Slide the clip. Trim it. If the dog barks in the wrong place, cut that bark out and move it. This is normal, and it is fast.
- Write down what worked. Keep the prompts that gave you good layers. A small personal library of prompts that suit your projects is worth more than any list online.
If step 5 gets you something you would keep, the model has earned a place in your workflow. If it does not, you lost fifteen minutes and learned how your scenes like to be described.
Where it breaks
This is the part that decides whether SFX v3 saves you time or costs you an afternoon.
It does not count well. The docs say plainly that prompts asking for an exact number of repetitions, their example is "triangle hit 4 times," are not always followed. If your scene needs exactly three knocks, generate the knock once and copy it three times on your timeline. Do not argue with the model.
It does not sync to your picture. The model has never seen your shot. A generation is a block of audio, not a performance timed to an actor's hand on a doorknob. For anything that needs frame-accurate sync, you will still slide clips by hand, or use a tool built to read video.
Thirty seconds is the ceiling. Long ambience means loops or several generations crossfaded. The docs also describe looping in terms that do not quite line up with the 30-second cap, so test the loop on your own material before you build a long scene around it.
It is not a music tool. Better music and vocal sounds inside an effect is not the same as a score. ElevenLabs points you to its separate music product for that.
The quality claims are the vendor's. There is no independent listening test yet. Your project, your speakers and your ears are the benchmark that counts.
Commercial use needs a paid account. If you make something on the free plan and it turns into client work, regenerate it on a paid account rather than guess.
The skill it rewards
There is a tempting story about tools like this: type a sentence, get finished audio, skip the sound editor. SFX v3 does not support that story, and I think that is good news.
What it rewards is knowing what a scene should sound like before you hear it. Which sound carries the moment. What sits behind it. How big the room is. Those are the decisions a sound editor makes for a living, and the model now listens when you make them.
So take one silent shot from your current project this week and spot it on paper first: main action, surroundings, room. Then type exactly that. The gap between what you wrote and what came back will tell you whether this is a tool for your rough cuts, your finals, or neither.