Mic-array direction estimates biased to one channel: per-channel ratio excess (v/b - 1) inflates a noisy low-baseline mic; use absolute gain-matched excess, spike-only frames and robust combining
Rule: for a 4-mic loudness array (front/right/back/left max-min levels per 25 ms) that keeps answering front, check how excess over baseline is computed. e = max(v/b - 1, 0) * gain with b the channel own baseline makes a mic with a small or noisy baseline produce a large excess from noise alone, so one channel dominates x=(r-l)/total, y=(f-b)/total; gain matching a ratio is not matching in common units. Better: robust per-channel noise floor (median + MAD of a 1-2 s quiet pre-roll, excluding motor-active windows), absolute excess (v - median)/gain, require excess > k*MAD per channel, use only frames near the spike peak (>=50% of the interval peak), combine with trimmed mean or Huber weights after dropping circular-outlier frames, and replace analytic 8-way rounding with nearest-centroid classification learned from labelled calibration trials (head and hat shadow asymmetrically), with a generic fallback when the margin is small. Evaluate with a confusion matrix, specifically front-bias (non-front called front) and cardinal-only accuracy for a 4-motor band. Also: baseline taken from frames outside the window can include a previous word, a motor buzz or a clap tail.