Sound effects
SFX v3 Prompt Guide: Write Better Sound Descriptions
Write concrete sound-effects prompts with action, material, space and event order. Includes before-and-after examples and a Chinese-to-English workflow.
A useful SFX v3 prompt describes something a listener could recognize. Start with a source and an action, then add the material, rhythm or space that makes it specific. Keep each revision small enough that you can hear what changed.
The examples below are writing suggestions, not recorded before-and-after tests. Open them in the audio workspace, choose a compatible sound-effects model and compare your own results. New to the process? Start with the beginner’s guide.
A simple prompt structure
Use this as a checklist, not a sentence you must fill completely:
Sound source + action + material or texture + space + event order.
For example:
A ceramic cup set down once on a wooden desk, a light hollow tap, heard close up in a small quiet room.
The source is a cup, the action is one contact, the materials are ceramic and wood, and the listening position is close. Each detail gives you something concrete to judge. A single clean effect may need only the source and action.
Turn a vague request into a testable description
| Starting idea | More specific prompt | What to listen for |
|---|---|---|
| A cool interface sound | One muted electronic tick followed by a short soft bell tone, compact and dry. | A clear tick, then a restrained tail. |
| Scary footsteps | Slow uneven boot steps across loose wooden floorboards, heard from the next room. | Uneven pacing and a sense of distance. |
| A cinematic transition | A fast rush of air rising in pitch, ending with a short low impact. | A distinct rise followed by an impact. |
| Rain | Steady fine rain landing on a canvas tent, with occasional heavier drops near the listener. | A stable bed with a few nearby details. |
Words such as “cinematic” can communicate a broad intent, but they do not replace a description of the audible event. Avoid stacking incompatible instructions such as “far away, close microphone, huge hall, completely dry.”
Put multiple events in a clear order
Use “first,” “then,” “followed by” or “ending with” when order matters. Separate the background bed from foreground events:
A steady low ventilation hum. First a switch clicks nearby, then a small motor spins up, ending with a brief mechanical stop.
Before generating, write a listening checklist: hum, click, motor start, stop. If an event is missing, shorten the description or generate that event separately. If your edit needs frame-accurate placement or separate volume controls, create separate assets and assemble them in an editor. A text sequence is not a guaranteed timecode instruction.
Describe material, distance and room
Change one of these dimensions at a time:
- Material: hollow wood, thin sheet metal, ceramic, loose gravel, soft fabric.
- Action: scrape, tap, drag, crack, slide, shake, open, close.
- Pacing: one hit, three spaced taps, slow irregular steps, a steady rattle.
- Perspective: close, across the room, behind a wall, passing the listener.
- Room: a small furnished room, an open field, a tiled corridor.
For a matching set, preserve the material and room while changing the action. A wooden chest opening, closing and being tapped can share a brief describing the same wood and listening position. Consistent wording helps you compare candidates, but does not guarantee perfectly matched outputs.
Keep an isolated sound focused
Start with one source and a sparse setting rather than a whole scene. If unwanted music appears, remove terms that suggest a soundtrack or musical composition, and try a simpler event description:
A single rubber shoe squeak on a polished indoor court, brief and close, with a quiet background.
“No music” can be tried as an instruction, but do not assume a negative-prompt field or a guaranteed exclusion feature. Review the complete result, including the tail. Trim or layer in an editor if the best candidate still needs cleanup.
Translate a Chinese description
Original idea: “雨后的石板路上，一个人慢慢走近，停下后轻轻敲两下木门。”
Break it into audible events: footsteps on damp stone; approaching perspective; a stop; two gentle knocks on wood. Then write:
Slow footsteps on damp stone paving approach the listener and stop, followed by two gentle knocks on a wooden door.
Preserve audible information rather than adding visual details such as clothing color or camera movement. If you ask an AI assistant to translate, request a plain sound description and tell it to preserve order without inventing extra events. Test the output; translation alone does not validate audio quality.
Revise with a small listening log
Record the prompt, selected model, exposed settings, credit estimate and one observation for each candidate. For example: “the impact is useful, but the tail is too long.” Change the tail description next while leaving the impact wording intact.
If the model exposes a prompt-adherence control, compare a small adjustment with the original setting. Do not assume that every control in the provider’s API is available in this website’s workspace. Treat each submitted variation as a separate generation when budgeting.
Common questions
Should every prompt be long?
No. Add detail when it resolves an ambiguity you can hear. Extra adjectives make comparison harder when you do not know which one affected the result.
Can I request exact timing?
A requested duration and exact placement of individual events are different things. Use the available duration control for overall length; use separate clips and an editor when the timing of each event is critical.
Can I create a seamless loop just by writing “loop”?
Do not rely on wording alone. Check whether the selected model exposes looping and audition the seam. Current official model documentation lists looping for SFX v3, while an earlier developer article describes it as v2-only. Checked November 9, 2026.
Where can I find more starting points?
Use the SFX v3 prompt library. Each entry includes a prompt, its intended use and a suggested variation. Bring one into the generator and confirm the model and credit estimate before submitting.