Methodology for a CNN-based AU audio feedback suppressor. Active research.

Overview

Reference materials for designing and iterating on an Audio Unit (AU v2) plugin whose feedback-suppression stage is a trained convolutional neural network on learned float32 weights, synchronised across two threads via Accelerate (BLAS) kernels, with a soft-limiter output stage.

Includes

• Architecture notes for a 7-layer temporal CNN with a learned 2-slope PReLU + first-order IIR activation between convs, compact enough for audio rates on Apple Silicon
• Runtime state-capture tooling for inspecting DSP layers on live audio during development
• Thread-synchronisation notes for an audio-thread + worker-thread split parallelising sgemm/sgemv inside a 512-sample render callback