Press kit
Soundbox, for anyone writing or filming about it
Everything on this page is free to use in coverage of Soundbox, and a credit is appreciated. The numbers come from the README, the model catalog and runs measured on one machine, and each file below downloads directly.
In a sentence
Quote either of these as written. Soundbox is one word, with a capital S.
Soundbox is a free, open-source app that generates sound effects, music and speech on your own computer from open-weight AI models, with no account and nothing uploaded.
Soundbox is a free, MIT-licensed app for Windows and macOS that generates sound effects, music, songs with lyrics and speech on your own computer. It runs open-weight models through audio.cpp, a C++ runtime, so there is no Python to install. Nothing is uploaded and there is no account. Four models are verified end to end and on by default; another 61, most of them text-to-speech, can be switched on, and each shows its licence before it downloads. On an RTX 5080 a five-second sound effect takes 1.3 seconds once the model is loaded. Clips can be arranged on a multi-track timeline, and a built-in MCP server lets coding agents such as Claude Code and Cursor make sounds and save them into a project.
Fact sheet
- Price
- Free, with no subscription, credits or account.
- Licence
- Soundbox itself is MIT. Each model's weights carry their own licence, shown before download, and that licence decides what you may do with the output.
- Runs on
- Your own computer. NVIDIA GPUs use CUDA and Apple silicon uses Metal, though none of the verified models has been confirmed on a Mac yet. Without a supported GPU it runs on the CPU, where a clip takes minutes rather than seconds. It works offline once the weights are downloaded.
- Runtime
- Every model runs through audio.cpp, a native C++ runtime, so neither Python nor the CUDA toolkit is needed.
- Models
- 4 verified end to end and on by default: Stable Audio 3.0 SFX, Stable Audio 3.0 Small Music, ACE-Step 1.5 Turbo and MiniMax Music 3. Another 61 can be switched on, 45 of them text-to-speech.
- Makes
- Sound effects and ambience, instrumentals, songs from a lyric sheet, speech, cloned voices, voice conversion and stem separation.
- Install
- Installers for Windows 10 and 11 (x64) and macOS (Apple silicon and Intel). Docker for a headless server, and Linux from source.
- For agents
- An MCP server with 24 tools for Claude Code, Cursor and any other MCP client, plus a local HTTP API.
- Data
- Audio is saved as WAV files under ~/.soundbox, and the library is one SQLite file beside them. Nothing is sent anywhere.
How long a generation takes
Measured on an RTX 5080 (16 GB) with audio.cpp 0.9.2. Warm means the model was already loaded. A smaller GPU is slower, and on the CPU alone the same clip takes minutes.
| Request | Model | Time |
|---|---|---|
| 5 s sound effect | Stable Audio 3.0 SFX | 1.3 s warm, 11.8 s cold |
| 10 s instrumental | Stable Audio 3.0 Small Music | 8.0 s |
| 15 s song with lyrics | ACE-Step 1.5 Turbo | 15 s warm, 151 s first load |
| 15 s orchestral piece | MiniMax Music 3 | 42 s |
Videos
Use them whole or cut down in coverage of Soundbox, and a credit is appreciated. Each is an H.264 MP4 under 10 MB, which X accepts as it is. In the "Bring your ___ to life" films, every sound was generated in Soundbox from the prompt shown on screen.
Vertical
The tour, under-the-hood and agents films end on the project's previous address, github.com/Gadurp1/soundbox, which redirects to github.com/HamSanders/soundbox. The under-the-hood film was cut at v0.2.0, so its version label is older than the current release.
Screenshots and GIFs
Taken from the running app in its dark theme. Click one for the full-size file.
The studio, with a script cast across three voices above the timeline.
1600 × 1000PNG
The sound effects library. Each take keeps its prompt, model and seed.
1600 × 1000PNG
The music page takes a style prompt or a lyric sheet.
1600 × 1000PNG
The models page shows each licence and VRAM figure before download.
1600 × 1000PNG
The agents page lists each MCP client that has connected and what it called.
1600 × 1000PNG
Sound effects, music and speech in one library.
1600 × 1000PNG
Saved and cloned voices in the Voice Lab.
1600 × 470PNG
GIFs
Mark, colours and type
The mark is a waveform on a rounded tile. Use whichever tile suits your background, and please don't recolour or redraw it.
Colours
Teal #2DD4BF
The mark, waveforms and links.
Ground #0A0C0D
The background of the app and every film.
Ink #E6ECEC
Text on the ground.
Muted #8E999D
Labels and secondary text.
Deep teal #0D6B64
Teal on a light background.
Type
Inter for headings and the interface (the films set titles in its 800 and 900 weights), and JetBrains Mono for prompts, labels and code. Both are free under the SIL Open Font License.
Links
Contact
For interviews, review questions or a file in another format, open an issue on GitHub.