Soundbox does not train models. It runs 46 open-weight models made by other people (sound effects, music, speech and stem separation) on your own machine through audio.cpp. This page records what a reading of each one's licence found, with a link to every source consulted.
audio.cpp is a ggml-based C++ inference engine — the audio equivalent of llama.cpp — with CUDA, HIP/ROCm, Metal, Vulkan and CPU backends. It is Apache-2.0 licensed, an independent project that Soundbox builds on.
The practical consequence is the whole reason this list looks the way it does: there is no Python in the inference path. No per-model torch pin, no virtualenv per model, no CUDA wheel to match against your driver. One native binary, one quantised GGUF file per model, and adding a model is a catalog entry rather than a new dependency tree. That is why a list this long is maintainable at all.
Licence details on these pages are summaries of the linked source documents, last checked 25 August 2026. They are not legal advice, and Soundbox is not affiliated with any model's authors — weights download from Hugging Face under the user's own account and terms. Company and model names identify whose work it is, and belong to their owners. The linked licence text is authoritative; commercial use should be checked against it, per model, before anything ships.
Text-to-audio for foley, impacts, ambience — the thing Soundbox was built for.
Text-to-music, some of it lyrics-conditioned, some able to extend or re-render a clip you already have.
Text-to-speech and voice conversion. The licences here are the most restrictive in the catalog, and the consent questions are the sharpest.
Not generators. These take audio you already have and pull it apart into stems.