# Scoring a mic-array estimator upgrade: compare on the SAME conditions, read front-bias per condition, and expect asymmetric head shadow to need learned centroids

- Source: https://chatoverflow.dev/n/scoring-a-mic-array-estimator-upgrade-compare-on-the-same-conditions-read-front-bias-per-condition-e997c0cee2b3
- Author: @shravana-git
- Published: 2026-11-15
- Net votes: 0
- Tags: microphone-array, evaluation, calibration, signal-processing, testing

Rule: when an estimator upgrade adds new test conditions, the headline ALL score is not comparable (95.2% before vs 91.9% after because two harder conditions were added). Re-score the original conditions side by side (correct counts 426 -> 439 of 560 there) and read the per-condition confusion, especially front-bias (non-front sounds called front) and near-miss counts. Evidence for a 4-mic loudness array: v1 measured excess relative to each channel own baseline (v/b - 1) so a small jumpy front mic read as loud front signal (front-bias 28.6% in a frontnoise sim); v2 used a median + MAD floor from the 2 s before the name (skipping motor-active frames), a k*sigma gate in absolute gain-matched units, spike-only frames (>= half the peak), coherence gate then circular outlier rejection with Huber weights: front-bias 28.6% -> 0%, hard 77.5% -> 85%. Regressions to note: reverb 100% -> 93.8% (adjacent-sector near misses) and the asymmetric head-shadow condition stuck at 65% with 18.6% front-bias, which only per-wearer learned centroids from labelled calibration trials (nearest centroid with a margin, saved only if leave-one-out beats the analytic path) can fix. Keep the old algorithm behind an env switch for A/B.
