Active Noise Cancellation Engine.
A fixed-point Filtered-x LMS noise-cancellation engine in C — −34 dB at the 50 ms deadline, certified by a bit-reproducible test bench that drives the real compiled firmware.
An adaptive noise-cancellation engine of the kind that ships on earbuds and aircraft cabin systems, written in C.
The algorithm is Filtered-x LMS: a reference microphone hears the noise early, a 32-tap adaptive filter drives a cancellation speaker, and an error microphone’s residual continuously re-tunes the filter. Convergence was deadline-limited by the source’s weak harmonics, fixed with a decorrelating first-difference on the update path. A fixed step size diverged when the source got 8 dB louder, fixed with a division-free normalizer that halves the step per doubling of input power. The best bug: a truncating right-shift biased every weight update by half an LSB, which integrated into a −17.5 dB performance floor — rounding the shift moved the floor to −38 dB. One line, 21 dB.
Certified numbers, all from a deterministic, bit-reproducible test bench that drives the real compiled firmware sample-by-sample through its 16-bit interface: −34.0 dB noise reduction at the 50 ms deadline, a −41 dB steady-state floor, an 8-case robustness sweep plus fault-injection tests all passing, and a 60 s endurance run with zero watchdog trips. A two-tier watchdog distinguishes real divergence from external bangs by checking whether the engine’s own output dominates the error power.
There’s also a live bridge: laptop mic through the actual compiled engine to headphones at 44.1 kHz.