How to Use the Stable Audio 3 Medium API

Stable Audio 3 Medium is available on EmpirioLabs. It is Stability AI's text-to-audio model for music and sound effects: describe a track or a sound and it returns a 44.1 kHz stereo clip from 1 to 380 seconds long. It runs on POST /v1/audio/generations, in the API and in the Playground. Powered by Stability AI.

Listen

A 30-second beat from the prompt "Warm lo-fi hip hop beat with soft Rhodes piano, vinyl crackle, mellow bass and brushed drums, 85 BPM":

And a 10-second sound effect, "A steam train passing through a station: approaching chuffing, a long whistle, brakes squealing, then fading away":

Both prompts against Stable Audio 2.5 and ACE-Step 1.5 XL are in Stable Audio 3 Medium vs Stable Audio 2.5 vs ACE-Step 1.5 XL.

How it differs from Stable Audio 2.5 and ACE-Step 1.5 XL

Stable Audio 2.5 makes clips up to 190 seconds in MP3 or WAV. Stable Audio 3 Medium goes to 380 seconds and adds FLAC and Ogg. ACE-Step 1.5 XL writes songs with sung lyrics; Stable Audio 3 Medium makes instrumental music and sound effects and does not take lyrics.

Your first request

curl https://api.empiriolabs.ai/v1/audio/generations \
  -H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "stable-audio-3-medium",
    "prompt": "Upbeat acoustic folk with bright guitar and soft percussion, 110 BPM",
    "duration": 45
  }'

The response is a job_id. Poll it until it completes; the result carries a signed URL to the clip, MP3 by default:

curl https://api.empiriolabs.ai/v1/jobs/JOB_ID \
  -H "Authorization: Bearer $EMPIRIOLABS_API_KEY"

Controls

• Length. duration (or audio_duration) from 1 to 380 seconds; 30 by default.
• Steps. num_inference_steps from 4 to 8; 8 by default.
• Seed. seed sets the random seed for a clip.
• Formats. format takes mp3, wav, flac, or ogg, and response_format set to b64_json returns base64 instead of a URL.

Writing prompts

• Music. Name the genre, the lead instruments, the mood, and the tempo in BPM.
• Sound effects. Describe what happens in order, such as approaching, a whistle, then fading away.

Operational notes

• Every request is billed the same, at any length. A 380-second track costs what a 10-second one does, so ask for the length you need in one request.
• Refused fields. negative_prompt, lyrics, guidance settings, and audio inputs are refused with a 400, as is a duration above 380 seconds.
• Usage policy. Use of this model follows the Stability AI Acceptable Use Policy and the Gemma Prohibited Use Policy for its text encoder, both in section 18 of our Terms of Service.

Pricing

Billed per request. One request returns one clip at the length you ask for, and a refused request is not billed. Current rates are on the model page and the pricing page.

Start building

Try Stable Audio 3 Medium in the Playground and read the API documentation for every parameter.

Stable Audio 3 Medium is licensed under the Stability AI Community License, Copyright © Stability AI Ltd. All Rights Reserved. Powered by Stability AI.