I Built a 100% Private, On-Device AI Audio Stem Splitter (No Servers!)

Most "AI" tools these days are just wrappers around an API. You upload your file, wait for a server to process it, and hope your data isn't being used to train the next big model. When I decided to add an Audio Stem Splitter to Kitmul , I had one non-negotiable rule: Zero server uploads . The result is a tool that can take any song and split it into Vocals, Drums, Bass, and Instruments entirely within your browser. The Problem with Traditional Audio Splitting If you've ever used tools like PhonicMind or LALAL.AI, you know the drill: Upload your MP3. Wait in a queue. Pay for "credits" or high-quality downloads. Your file sits on someone else's server. For musicians, producers, or just karaoke fans, this is slow and privacy-invasive. I wanted to see if we could bring the power of models like Demucs directly to the user's hardware using WebAssembly and Web Workers . How it Works: AI in the Browser The magic happens thanks to a few modern web technologies: WebAssembly (WASM): We run the he