Compile differentiable audio models to real-time DSP
ADAC is a tool that compiles differentiable audio models, such as neural audio effect models, into real-time DSP code. It enables running models trained in machine learning frameworks directly on embedded devices and plugins with minimal latency and overhead.
Background
- **ADAC** (Automatic Differentiable Audio Compiler) converts ML models (e.g., PyTorch) into optimized C++ code for real-time audio processing on hardware like guitar pedals, synths, and plugins.
- **Differentiable audio models** are neural nets trained via gradient descent for tasks like amp modeling or distortion. They normally require Python and GPUs, making them unsuitable for low-latency live performance on embedded devices.
- **Real-time DSP** demands every audio sample be processed within a strict time window (e.g., under 1 ms). ADAC automates the translation from ML frameworks to efficient C++, a task previously done by hand.
- The project targets musicians, plugin developers, and researchers who want to deploy AI audio effects on stage or in the studio without manual rewriting.