How to prompt in Salience
Salience turns a short text description into a sound using Stable Audio 3, the same box you fill in during Make a sound. Here's what that model actually listens for, so your samples come out the way you want. For the full model-level reference, see the official guide linked at the bottom.
How a prompt is built
A Stable Audio prompt is not a sentence you write once and hope for. Stability's own guide is upfront that there is no single perfect prompt and results shift a little each time you generate, so plain, concrete wording that reads like real recording and production language tends to work best. For a one-shot, build it in this order:
- The source. What is actually making the sound: an 808, a rimshot, a mallet instrument, a vocal chop.
- The character. Two or three texture words: punchy, glassy, dusty, metallic, warm, airy.
- The action, if it matters. How it's triggered or played: plucked, struck, muted, brushed.
- The production detail, if it matters. Room, mic distance, decay length, effects: dry, close-mic, fast decay, tape saturation, vinyl crackle.
Skip the parts that don't apply. "Punchy sub 808, fast pluck, tight decay" already has all four; you rarely need more than one short clause per part.
The words that matter
- Source and instrument words: 808, rimshot, mallet, clap, hi-hat, kick, snare, vocal chop, guitar pluck, synth stab.
- Character and material words: punchy, glassy, dusty, metallic, warm, airy, gritty, bright, deep, soft.
- Action and performance words: plucked, struck, muted, brushed, tight, finger-picked.
- Production words: dry, close-mic, room, reverb tail, fast decay, zero ring, tape saturation, vinyl crackle, distortion.
- For a full-mix Loop (not a one-shot): add genre, mood, and BPM too, the way you'd describe a track: "Detroit-influenced techno, stripped-back, metallic percussion, 130 BPM."
Key and chord have their own fields in the composer, so you don't need to cram them into the text.
One-shot specific tips
Salience's default output is a single hit, not a song, so borrow from how the official guide treats sound effects: name the source, the action or duration, and the production character, in that order, and keep it to one sentence.
If you don't say how the sound ends, you'll sometimes get a longer tail than you want. Add "fast decay" or "zero ring" when you need a clean, short one-shot.
Not sure where to start? The starter prompts are quick examples of this same idea, drop one in and tweak it.
Simple vs. detailed
A short prompt is not a lesser prompt. Both of these work; the detailed one just narrows what you get.
Start simple, then add detail only for the part that isn't coming out right.
Common mistakes
- Writing a whole scene. "A sad rainy night in a cyberpunk city" describes a mood, not a sound. Name the actual source and texture instead.
- Stacking contradictory character words. "Bright and muffled" fights itself; pick the one you mean.
- Expecting lyrics. Salience (via Stable Audio 3) is built for instrumental work. Vocal-ish textures can appear, but not intelligible words.
- Forgetting the ending. No decay or ring cue can leave a one-shot with an unwanted tail.
- Genre buzzwords with no sound description. "Trap" or "lo-fi" alone tells the model less than naming the actual source and character.
16 example one-shot prompts
Written in the official guide's style (source, character, action, production) but sized for a single pad hit:
Editing an existing sound
The AI edit field on a loaded pad's Shape editor takes a change, not a full description. Its placeholder: "warmer tail with more air and a touch of vinyl dust...". Describe what should change, not the whole sound again.