A new open-source project called AI Piano is bringing a genuinely useful edge case to Core ML: real-time musical autocompletion on iPhone without any network dependency. Published to DEV.to by developer leojulieta, the app generates β1-second music completions as you play, effectively finishing your improvised phrases before you've finished touching the last key.
How the Model Fits in Your Pocket
The implementation centers on a 125-million parameter model optimized for mobile inference. That's substantial enough to capture musical patterns and harmonic relationships, yet small enough to run locally on recent iPhones without thermal throttling every bar. The architecture appears designed around streaming audio input, processing your live playing and predicting the next sequence of notes in near-real-time.
Offline-First Architecture
What sets this apart from cloud-based AI music tools is the fully offline execution model. No API calls, no latency spikes, no privacy concerns about sending audio to external servers. The entire inference pipeline runs on-device using Apple's Neural Engine and GPU frameworks. For musicians who want creative assistance during flights, studio sessions without internet access, or simply don't want their musical ideas leaving their device, this approach solves a real problem.
Technical Considerations for Builders
The project raises interesting questions about model quantization and audio preprocessing on mobile. Getting sub-100ms latency from key press to predicted note output requires careful optimization of both the model architecture and the input pipeline. Developers interested in similar on-device generative audio applications will want to study how leojulieta handles buffer management and inference scheduling.
Key Takeaways
- 125M parameter model runs locallyβno cloud dependency for music generation
- Real-time autocompletion means the app predicts while you play, not after
- On-device processing addresses both latency and privacy concerns
- Open-source implementation available for developers to study and extend
The Bottom Line
This is exactly the kind of project that shows what on-device AI can do beyond chatbots. When inference happens locally with low enough latency to be musically useful, you get responsive creative tools that don't betray your data or require constant connectivity. That's a real win for working musicians and a solid reference architecture for anyone building interactive audio ML applications.