Simon Edwardsson developed a 125 million-parameter transformer model designed to autocomplete piano performances in real time on devices like the iPhone 15, processing around 108 notes per second, according to simedw.com. The model powers the RollTab app, which is available for free on the Apple App Store for users with MIDI keyboards and iPhones or iPads.
Edwardsson began this project nearly a year ago with the goal of connecting a MIDI piano to a phone to enable AI-assisted song completion, similar to GitHub Copilot but for piano. After fourteen experiments, he achieved satisfactory results by optimizing the MIDI representation, aggressively cleaning the training data, and applying direct preference optimization (DPO) post-training. He shared audio samples demonstrating the model’s ability to continue well-known pieces such as "Für Elise" and "Terra's Theme."
This development highlights advances in on-device AI for music generation, allowing real-time interaction without cloud dependency. The model’s relatively modest size of 125 million parameters enables efficient performance on mobile hardware, contrasting with larger models that require extensive computational resources. RollTab’s approach could influence future music composition tools by providing accessible AI assistance directly to musicians’ devices.
RollTab is publicly available for download on the Apple App Store, enabling users with compatible MIDI keyboards and Apple devices to experiment with AI-driven piano autocompletion. Edwardsson’s blog post on simedw.com details the technical process and offers audio demonstrations of the model’s capabilities.