AudioGen vs Stable Audio Open
Both generate sound effects from text. One is a 2023 model with a single checkpoint and non-commercial weights; the other outputs 44.1 kHz stereo and can be used commercially. The choice is less balanced than it looks.
Side by side
| AudioGen | Stable Audio Open | |
|---|---|---|
| Output | 32 kHz mono | 44.1 kHz stereo |
| Max length | Short clips | Up to 47 s |
| Licence (weights) | CC-BY-NC 4.0 | Stability AI Community License |
| Commercial use | No | Conditional — revenue threshold |
| Checkpoints | One — audiogen-medium | Full and Small variants |
| Negative prompts | No | Yes — a real quality lever |
| Candidate ranking | No | num_waveforms_per_prompt, auto-ranked |
| Training data | Not fully disclosed | ~48k CC0/CC-BY recordings |
| VRAM | ~16 GB | ~8 GB at fp16 |
Where they actually differ
Sample rate decides it for production work
AudioGen outputs 32 kHz mono. Stable Audio Open outputs 44.1 kHz stereo. For sound design going into a mix, a game engine or a video timeline, that is not a subtle difference — it is the difference between an asset you can use and one you have to apologise for.
Negative prompts are a genuine advantage
Stable Audio Open responds strongly to negative prompts; the documented recommendation of "low quality, average quality" is a free quality improvement, and you can steer away from unwanted reverb or noise the same way. AudioGen has no equivalent control.
Training data cuts both ways
Stable Audio Open was trained largely on Freesound under CC0, CC BY and CC Sampling+ — which is exactly the material sound design needs, and also why it can be licensed for commercial use. AudioGen’s training data is not fully disclosed and its weights are non-commercial regardless.
AudioGen’s remaining argument is the API
If you are already running MusicGen or MAGNeT through AudioCraft, AudioGen is the same get_pretrained / set_generation_params / generate pattern with the same loudness-normalised writing. That consistency is worth something — but not 44.1 kHz stereo and a usable licence.
Which should you choose?
Choose AudioGen when…
- You are already using AudioCraft and want one API
- The work is non-commercial
- You want Meta’s specific sonic character
Choose Stable Audio Open when…
- You need 44.1 kHz stereo assets
- The work is commercial and you are below the revenue threshold
- You want negative prompts and automatic candidate ranking
- You are generating sample packs in batches
- You care about training-data provenance