Stable Audio Open
Open-weight audio for effects and loops, forty-seven seconds, with a commercial threshold.
Model profile
- Vendor
- Stability AI
- Form
- 本地模型
- Input
- 文本 · 音频
- Output
- 音频
- Licence
- Stability AI 社区许可：年收入低于门槛可商用，超过需企业许可；开放权重不等于无条件开源
- Verified
- 2026-10-31
Related
Open weights, not a free-for-all
Stable Audio Open can be downloaded, run locally and fine-tuned with LoRA, but the licence is not MIT-style. The Stability AI Community License allows commercial use below a revenue threshold; above it you need an enterprise licence, and redistribution, hosted services and using output to train other models carry further limits. Read the model card before it enters your delivery chain. This is the 2024 open-weight generation, not the same specification as the later Stable Audio 3.0 family.
Text to sound, forty-seven seconds
Input is a text description, optionally audio for audio-to-audio; output is 44.1kHz stereo, up to about forty-seven seconds per clip. It does not make songs or sing, and its strength is effects, ambience, drum loops and instrumental beds. "Warm analogue synth pad with light tape hiss" works better than "sad"; exact duration and sample rate belong to the model card.
Samples, loops and effects
Training audio comes from CC-licensed libraries, with a tilt toward electronic, ambient and field recording, which suits sample packs and game sound. Drum transients are clean and ambient beds have depth, but the ends are not seam-matched, so looping the file as-is clicks and needs trimming in a DAW. Low-end level varies between batches, so a set of assets needs per-file normalisation.
| Use | Fit | Watch |
|---|---|---|
| Ambient beds | High | Fades at both ends |
| Drum and percussion loops | High | Seams and low end |
| Instrumental motifs | Medium | Weak structure, repetition |
| Songs with vocals | Low | No words |
Local hardware and quality
A consumer GPU is enough, with VRAM depending on batch size and sampling steps; speed falls as length rises, and forty-seven seconds is a hard ceiling, so long beds need stitching. Prompt adherence is moderate and specific instruments and styles get averaged out; loudness and timbre drift between runs from one prompt. Treat it as a material generator, not a final source. For repeated DAW tests, fix a seed and a prompt template to reduce drift.
Licence threshold and alternatives
Confirm: the revenue threshold, enterprise licensing, whether output may be redistributed as a sample library, whether hosted services are allowed, and how fine-tuned weights must be labelled. The Stability AI Community License and model card are the authority. Alternatives: the Stable Audio 3.0 family for longer, newer open weights; YuE or MiniMax Music for songs and vocals; MusicGen as another local sound option.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- Stability AI 的 Stable Audio Open 模型卡（训练数据、时长、采样率与许可）。
- Stability AI 社区许可文本（商用门槛与企业授权）。
- 公开用户报告中的循环接缝、低频不稳与人声缺失问题，只取现象，不引用具体评测数据。