I put out an open source library to automatically compile trainable audio models into real-time targets (FPGA, embedded systems, audio plug-ins, the web).
It currently supports linear time-invariant (LTI) systems by traversing PyTorch models, extracting their parameters, and generating equivalent Faust code. Support for nonlinear models is something I’d like to add in the future.
The project originally came out of feedback delay network research, where it also enables hot-reloading into a DAW to audition optimisation whilst it trains and guarantees stability for every exported model.
I put out an open source library to automatically compile trainable audio models into real-time targets (FPGA, embedded systems, audio plug-ins, the web).
It currently supports linear time-invariant (LTI) systems by traversing PyTorch models, extracting their parameters, and generating equivalent Faust code. Support for nonlinear models is something I’d like to add in the future.
The project originally came out of feedback delay network research, where it also enables hot-reloading into a DAW to audition optimisation whilst it trains and guarantees stability for every exported model.