Skip to content

Stable Audio

aivrar edited this page Nov 18, 2026 · 1 revision

Stable Audio: Music & Sound Effects

Stable Audio Open 1.0 (Stability AI) turns a text description into sound effects, ambiences and short music clips up to 47 seconds. It can render several variants at once, and CLAP scores each one against your prompt so the best match is marked BEST.

Stable Audio panel

Use Stable Audio for foley, ambiences, impacts, risers, jingles and loops. Use ACE-Step for full songs with vocals and lyrics.

🔊 Listen: Stable Audio sample: "cinematic thunderstorm, heavy rain on a tin roof"

Step by step

  1. Load the model on a GPU (the first load downloads ~16 GB: the model plus the CLAP scorer).
  2. Describe the sound in the prompt. Optionally add a negative prompt for things to avoid (music, voice, low quality).
  3. Set length, steps and the number of variants.
  4. Click Generate. The bar shows each diffusion step with an ETA:

Stable Audio generating: diffusion steps

  1. Listen to the variants. The one CLAP rates closest to your prompt is marked BEST:

Stable Audio results with CLAP scores

Writing prompts

Describe what, where and how it's recorded:

Goal Example prompt
Ambience rainforest at dawn, birds, distant waterfall, gentle breeze, binaural field recording
Foley footsteps on gravel, slow walk, close mic, dry
Impact cinematic boom hit, deep sub bass, long reverb tail
Sci-fi spaceship engine hum, low drone, subtle mechanical clicks
Music loop lo-fi hip hop beat, 85 bpm, dusty vinyl, mellow piano, loopable

Tips: name the genre or source, add tempo/key for music, and words like high quality, field recording, studio help. Use the negative prompt to remove unwanted voices or music from SFX.

Settings that matter

Setting Default Guidance
Steps 100 (10–200) More steps means more detail and slower renders. 100–150 is a good range.
Length 30 s (10–47) Shorter clips render faster.
Guidance scale 7 How strictly the prompt is followed; 6–9 is typical.
Variants 3 (1–4) More variants give CLAP more to choose from.
Seed random Set a number to reproduce a result.

API

curl -X POST http://127.0.0.1:5006/stable_infer -H "Content-Type: application/json" -d "{\"prompt\": \"wooden door creaks open slowly\", \"steps\": 100, \"length\": 8, \"num_waveforms_per_prompt\": 2}"

Load first with POST /stable_load and {"device": "0"}. Stop a render with POST /stable_cancel. More in API Reference and chapter 8 of the manual.

Clone this wiki locally