Blog

Guide

ElevenLabs Sound Effects v3 Prompt Guide: Organizing Sound Effect Descriptions with Material, Action, and Distance

Nov 18, 2026
On a warm sand-colored background, stepped horizontal color bands are arranged, layered in terracotta and tan tones, serving as a blog cover image.
On this page

Why prompts should be split into material, action, and distance

The most common mistake when writing sound effect prompts is to write one sentence and submit it: "a crisp click." This sentence has adjectives but no information the model can act on—what material is making the sound? Is the action fast or slow? How far is the listener from it? The sound effect generation model needs to fill in these dimensions, and the way it fills them in is uncontrollable. So running the same prompt twice gives two wildly different results, and debugging turns into a game of chance.

A more reliable approach is to split the description into three layers of structured information:

  • Material determines timbre and texture: wood, glass, metal, fabric, water, plastic, each has different resonance modes and high-frequency components. Writing the material clearly sets a baseline for the timbre.
  • Action determines rhythm and envelope: a single click or a repeated strike, a gentle placement or a heavy slam, a slow rub or a fast scrape. Action words determine the sound's attack, decay, and duration distribution.
  • Distance determines spatial sense and reverb tendency: against the ear, within one meter, on the other side of the room, far outdoors. The farther the distance, the weaker the direct sound and the higher the proportion of reverb and low frequencies, and the model's imagination of the space changes accordingly.

These three layers are not a template that must be weighted equally; they are an analytical framework: when you get a sound effect requirement, first judge which layer it depends on most, then write the corresponding words specifically, and supplement the other two layers as needed. UI feedback sounds rely almost entirely on action and material, object foley takes material as the subject, and ambient atmosphere is supported by distance and space. The following sections expand on each.

Within the 450-character limit, this framework also has a practical benefit: it naturally helps you control information density. One sentence for each of the three layers is usually enough; if you can't fill it, the description isn't specific enough; if you blow past the limit, you're piling up synonyms.

The three layers of information—material, action, and distance—each form a step, ultimately converging into one structured prompt.

API shape and the three tunable parameters

First fix the contract so that the writing methods below have a common landing point. The current task API corresponds to the model elevenlabs/sound-effects-v3, and the input fields are grouped under input:

  • prompt: required string, describes the desired sound effect, needs to cover action, scene, atmosphere, and presentation, maximum 450 characters.
  • duration_seconds: required number, in seconds, range 0.5 to 30, default 2. Automatic duration is not supported—you must actively set the length according to the sound effect type.
  • loop: optional boolean, default false, controls whether to generate a sound effect that can be played back seamlessly in a loop.
  • prompt_influence: optional number, 0 to 1, default 0.3. Lower values are more random and varied, higher values adhere more strictly to the description.

Among the three optional/semi-optional parameters, the easiest thing to overlook is the bias introduced by the defaults. prompt_influence defaults to 0.3, meaning the model has considerable freedom to "improvise," which is good during exploration, but becomes noise when you have polished the description to be precise and need reproducibility. duration_seconds defaults to 2 seconds, which is just right for short UI sounds but far from enough for ambient sounds.

Call example (based on the task API's input structure):

{
  "input": {
    "prompt": "wooden drawer sliding open slowly on a dry track, close perspective, slight creak",
    "duration_seconds": 3,
    "loop": false,
    "prompt_influence": 0.6
  }
}

For the complete fields and values of the API, refer to the model page; you can check the current parameters and integration method at ElevenLabs Sound Effects v3. The next three sections discuss how to write prompts for the three types of scenarios, then return to how the parameters work together.

The three input elements—prompt, duration, and influence strength—are expressed by the widths and convergence of the stepped bands, converging into the same generated audio result.

UI feedback sounds: short, dry, action first

UI feedback sounds require the least imagination of the three types, and are also the easiest to write badly. The requirements are clear: short, dry, clear attack, no dragging tail. When writing this type of prompt, the action should come first in the sentence, material immediately after, and distance always written as near-field.

A usable writing skeleton:

[action] + [material/object] + [perceptual qualifier]

For example:

soft plastic button click, single press, very close, no reverb, tight transient
short glass tick, light tap, near field, clean tail
quick mechanical switch flip, small metal lever, close perspective, dry

Three details are worth emphasizing:

Write whether the action is single or repeated. "click" and "rapid double click" are completely different things in terms of envelope. Most UI sounds are single, and writing "single press" reduces the probability of the model adding its own rhythm.

Use "no reverb" and "dry" to suppress space. If a UI sound has obvious reverb, it will sound dirty when placed in the UI. Use distance words like "very close" and "near field," then add a dry qualifier, and it usually converges significantly.

Keep duration short rather than long. For this type of sound effect, duration_seconds starts from 0.5, and around 1 second is often enough; if it exceeds two seconds, ask yourself whether you've written in a tail that shouldn't be there. If the duration is too long, the model tends to fill in content—dragging a clean click into an atmosphere.

Choosing the material is actually choosing the timbre family: plastic leans mid-frequency with a bit of dullness, glass leans high-frequency and crisp, metal has sustain and overtones. For the same "click" action, changing the material word covers most UI feedback needs. This is also why in this type of scenario, material is not decoration but one of the subjects.

Object foley: let material be the subject

The difficulty of object foley is that the same action sounds completely different on different materials. The word "impact" itself carries no timbre information; it is "wooden crate," "ceramic mug," and "chain" that speak. So the syntax of this type of prompt should make the material the subject, with the action and force application as the predicate.

Recommended writing order: material → action/force → additional texture → distance.

heavy ceramic mug placed on wooden tabletop, hard contact, short resonant thud, close perspective
metal chain dropped onto concrete, links colliding, bright clatter with metallic ring, one meter away
fabric curtain pulled across a metal rod, continuous friction, soft rustle with occasional ring

A few criteria to check whether the material word is really doing its job:

  • Can it still hold if you replace the material? If after replacing the material word the description feels empty, the material really is the core of the sentence; if replacing it makes no difference, you've written a generalized action and the model can only guess.
  • Should the force application be written? "placed" and "slammed" determine the strength and transient. Friction-type actions like "colliding," "scraping," and "sliding" need to state whether they are continuous, because the envelope of continuous friction is completely different from that of a collision.
  • Should resonance be left in? Metal, glass, and ceramic have recognizable lingering ring, wood is short, and fabric is almost pure noise. If you don't want sustain, write "damped" or "dry" directly in the prompt; if you do, write "resonant" or "ring out" clearly.

Regarding duration, collision types usually need only 0.5 to 2 seconds, while friction and open/close types need to cover the complete process and may need 2 to 5 seconds. Don't expect the model to cram a three-second drawer-opening action into 1 second—duration_seconds is the only means by which you tell it "how long," and it will not adapt automatically.

The stepped bands are sectioned into different material layers: hard high-frequency micro-vibrations, soft broad lingering vibrations, and concentric ripples from a single impact, corresponding to how material words rewrite object timbre.

Ambient atmosphere: write the scene with distance and space

Ambient atmosphere sounds are fundamentally different from the previous two types: they do not describe an event, but a sustained spatial state. The envelope of an event rises and falls, while atmosphere must fill the entire duration and remain stable. In this type of prompt, distance and space type are the skeleton, and material and action take a back seat.

When writing, first answer three questions:

  1. What is the space type: a small indoor room, a hall, a street, a forest, an underground garage, an open outdoor space.
  2. Where is the sound source and how far: a single sustained source, or multiple layers stacked near and far.
  3. Is the sustained state stable: does it need undulation, or is it close to a steady state like white noise.

Corresponding writing:

quiet indoor room tone, distant traffic hum through a closed window, low level, steady ambience
office background, faint keyboard taps and air conditioning drone two meters away, continuous
outdoor night field, crickets at varying distances, soft wind over grass, spacious, wide stereo image

Note expressions like "varying distances"—they yield results with more depth than a single distance. The texture of ambient sound often comes from near-far layers rather than one prominent sound. If all elements in the prompt are "very close," the result will be like an extreme close-up with no space.

Looping is the key divergence point in this section. Ambient sounds usually need to loop, and whether the loop point is natural is directly determined by the duration. If loop is set to true, what you need to consider is: does the content remain in the same state throughout the entire duration? If the prompt contains directional descriptions like "gradually rising wind," looping back will expose itself at the seam. A safe approach is to keep the atmosphere steady and give a sufficiently long duration_seconds—if the loop segment is too short, the seam appears frequently; give it a dozen seconds or even close to the upper limit, and the loop will sound continuous.

Conversely, if this atmosphere is used in a one-off scenario (transition, opening), keep loop as false, set the duration as needed, and feel free to write in rises and changes.

How parameters work with the prompt

Prompts and parameters are not two separate things; tuning should follow a fixed order to avoid changing multiple variables at once.

Step one: set duration_seconds. It is required and not automatic, so first give a reasonable value according to the sound effect type—UI sounds 0.5 to 1 second, collision foley 1 to 2 seconds, friction and open/close 2 to 5 seconds, ambient sounds more than ten seconds depending on loop needs. Once this value is set, the subsequent prompt organizes content within this time window.

Step two: set loop. Only ambient types and a few sustained mechanical sounds need true; event-type sound effects are always false. When set to true, go back and check whether the prompt contains directional descriptions (fading in, fading out, gradually moving away); if so, change them to steady-state expressions.

Step three: tune prompt_influence. The default 0.3 leans divergent and is suitable for the exploration stage—to see what the model gives you. When you already have a satisfactory result and want to reproduce the same style, raise it to the 0.5 to 0.8 range to make the output closer to the text. Conversely, if the description is too specific but always feels stiff and lacks natural randomness, you can lower it appropriately.

Step four: only then go back and change the prompt. If you are still unsatisfied after tuning the parameters, the problem is in the description itself: the material word is not specific enough, or the action does not match the duration (for example, a three-second open/close action written as a single collision).

Summarizing the focus by scenario: for UI feedback sounds, prioritize short duration_seconds and higher prompt_influence (for stable reproduction); for object foley, prioritize writing the material accurately, with medium prompt_influence; for ambient atmosphere, prioritize sufficient duration and correct loop, and prompt_influence can be lowered moderately to trade for naturalness.

Common prompt problems and how to fix them

The following situations are the most common during debugging and can all be quickly located through self-check.

Vague description. The typical feature is only adjectives and no entities: "a cool sound," "an exciting swoosh." The criterion is simple: what remains after removing the adjectives? If nothing remains, the model can only improvise. The fix is to fill in material and action, changing "a cool click" to "a short glass click, single press, close."

Information overload. Five or six elements are crammed into one prompt: material, action, environment, emotion, reverb level, and stereo width are all written, and they contradict each other. The result is that the model distributes attention evenly and none of them come through. The fix is to cut down to two or three layers by scenario: UI sounds keep only action and material, ambient sounds keep only space and distance.

Length over limit. prompt has a maximum of 450 characters; exceeding it will be truncated or fail validation. Piling up synonyms is the main cause of exceeding the limit; for example, "loud, noisy, booming, thunderous" actually only needs one. After writing, do a trimming pass: keep one synonymous adjective, and delete optional modifiers directly.

Action and duration conflict. You wrote a sustained friction action but gave only 0.5 seconds; or you wrote a single collision but gave ten seconds. In the former, the model can only compress it into a short noise; in the latter, it will fill with tail. The fix is to align duration_seconds with the semantic duration of the action.

Looping doesn't hold. Setting an event sound with obvious start and end as loop: true will inevitably make the seam abrupt. Looping applies only to steady-state content; do not enable it for event sounds.

Mixing templates from different scenarios. Using the UI sound's "no reverb, dry" to write outdoor atmosphere will suppress the sense of space; using the ambient sound's "spacious, distant" to write a click will produce a muddy mess. Writing the three scenario types separately and iterating them separately is much more reliable than applying the same template.

Treat material, action, and distance as three checkpoints and go through them before each submission; most instability problems can be eliminated at the prompt-writing stage.

Related articles

View all articles →

Start building with persistent AI artifacts

Get your API key and call image, video and audio models through one task-based API.

Read the API docs