Stable Audio 3 Medium (Text-to-Music)
Stable Audio 3 Medium is the mid-size open-weight text-to-music model in Stability AI's Stable Audio 3 family. It also makes solo instrument stems and sound effects, in clips up to 6 minutes 20 seconds.
Overview
Stable Audio 3 Medium is the mid-size, open-weight member of Stability AI's Stable Audio 3 family. It is a latent diffusion transformer built on SAME, Stability AI's semantic-acoustic autoencoder for stereo 44.1 kHz audio. Stability AI lists it at 1.4B parameters and up to 6 minutes 20 seconds per clip, aimed at longer pieces where structure and phrasing matter.
Adversarial post-training reduces the number of sampling steps the released checkpoint needs. This page defaults to 8.
As a text-to-music model it writes full tracks, and the same checkpoint also makes solo instrument stems and short sound effects. The lighter Stable Audio 3 Small Music and Small SFX models split music and sound effects between them and stop at 2 minutes.
What you can build
- Background music for videos, reels and trailers, set to the length of the edit
- Game soundtracks: menu themes, level music and ambient beds
- Solo instrument stems to arrange around in a DAW
- One-shot hits, risers and other short sound effects for apps and games
Inputs
- Prompt (required): an English description of the audio. Stability AI's guide suggests naming genre, instruments, mood and tempo in BPM, as the default does: "Cinematic neo-soul groove with electric piano, brushed drums, walking upright bass, smoky vibe 92 BPM". Start with "TrackType: Instrument" for a solo stem or "TrackType: SFX" for a sound effect.
- Duration: clip length in whole seconds, default 60.
- Inference Steps: 1 to 50, default 8.
- Guidance Scale: 1.0 to 2.0, default 1.0.
Outputs
Each run returns a stereo audio file. The two samples on this page are 44.1 kHz stereo MP3s, each about 66 seconds long.
Recommended settings
- Keep Inference Steps at 8 and Guidance Scale at 1.0. Stability AI's reference example uses the same pair, and higher guidance on this post-trained checkpoint adds high-frequency noise.
- Match Duration to the content: a few seconds for a one-shot effect, a full song length for an arranged track.
- Put the tempo in the prompt ("124 BPM") when the clip has to sync with picture.
Limitations
- Vocals are not intelligible. The model sometimes produces vocal-like textures without real words.
- Stability AI built the model for clips up to 380 seconds. Keep Duration at or below that.
- This page takes a text prompt only. Audio-to-audio editing, inpainting and continuation from Stability AI's own release are not offered here.
- Stability AI says the model won't output the same thing twice, and this page has no seed setting, so keep the takes you like.
Safety & compliance
- Stability AI reports training on 1,278,902 recordings: 806,284 licensed from AudioSparx and 472,618 from Freesound under CC0, CC BY and CC Sampling+ licenses, screened to remove copyrighted material.
- The weights ship under the Stability AI Community License. Stability AI states that users own the audio they generate and may commercialize it; organizations above the license's annual revenue threshold need an enterprise license.
- The T5Gemma text encoder is redistributed under Google's Gemma Terms of Use.
Example prompts
Great starting points for stable-audio-3-medium.
API quick start
Run stable-audio-3-medium with a single API call.
{
"prompt": "Cinematic neo-soul groove with electric p…",
"scale": 1.0,
"duration": 60,
"promptExpansion": "False"
}curl -X POST "https://api.wiro.ai/v1/Run/stabilityai/stable-audio-3-medium" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Cinematic neo-soul groove with electric p…",
"scale": 1.0,
"duration": 60,
"promptExpansion": "False"
}
JSONOther models you might like
Explore similar music generation models.
