// Noise reduction. Stub for Phase 2.
//// Two paths planned:
// 1. Basic DSP: spectral subtraction (open source, works but has artifacts)
// 2. ML model: custom trained ONNX model (proprietary, better quality, default)
//// Both compile to native and WASM from the same Rust code.