audioGen
Music and speech — made on your own machine.
local AIYour audio will appear here.
About audioGen
audioGen is a local-first text-to-audio workspace for ClackOS. The interface is built from a shared model catalog and each model family is implemented by a worker adapter, so two music models and three speech models sit behind one set of controls. Prompts and generated audio stay on this device; model weights are downloaded from the Hugging Face Hub and cached by your browser, so only the first load of a model waits for the network.
Everything runs in WebAssembly on the processor, except MusicGen, which uses the graphics adapter where there is one able to take it. A short music clip can take a few minutes; speech takes seconds. Magenta RealTime 2 is the heaviest thing here by some way — about 1.6 GB to fetch once, and roughly twenty seconds of work per second of music.
Engine: Transformers.js 4.2.0 by Hugging Face, Apache-2.0 · ONNX Runtime Web by Microsoft, MIT, which Magenta RealTime 2 is driven on directly.