WaveNet Guitar Amp Model Pruning Enables Real-Time iOS Deployment

Ryota Sato and Eli Silverstein have published a paper on arXiv (arXiv:2607.10086) presenting a method to run WaveNet-style convolutional neural networks for guitar amplifier emulation in real time on iOS devices. Traditionally, these models produce high-fidelity tube amp and distortion pedal sounds but are computationally expensive, limiting them to desktops or dedicated DSP hardware. The researchers apply aggressive iterative magnitude pruning, removing 90% of the network weights with no perceptible loss in audio quality. They built a custom sparse C++ inference engine that directly converts this sparsity into compute savings, enabling low-latency, real-time operation on CPU-only iPhones where the dense model cannot run. The on-device output matches the full model within int16 quantization error.

The work is accepted as a demo at DAFx 2026, where attendees can play a guitar through the app on iPhone hardware and A/B test the pruned model against the physical pedal it emulates. Source code and audio examples are available online. This breakthrough makes high-quality neural amp modeling accessible on consumer mobile devices without requiring specialized hardware. It opens the door for guitarists to use realistic virtual amps and effects directly on their phones for practice, recording, or performance, potentially disrupting the traditional pedal and amp market.