ITU compliance Stage 2: residual-echo post-filter, comfort noise, AEC chain

[tap]

Stage 2 of the ITU compliance plan: the residual-echo suppressor + matched comfort noise (mutap/postfilter.h), and mutap::aec_chain — the canceller + post-filter unit the compliance matrix measures. Stacked on #6 (Stage 1 signal layer); rebases onto main after #6 merges.

What's in it

• residual_suppressor<Sample> — per-bin Wiener gain on the canceller output E whose depth comes from a leakage estimate (residual power per unit echo-estimate power), gated by mic-vs-echo-estimate coherence. Gains constrained to a causal 2B-tap filter (one block of latency), attack-smoothed, suppressed bins refilled with comfort noise shaped to a two-window minimum-statistics noise floor. RT contract as everywhere: constructor allocates and may throw; everything after is noexcept and allocation-free.
• aec_chain<Sample, Canceller> — defaults to the raw partitioned FD-Kalman core; pem_afc still composes via its 3-argument surface (it grew an echo_estimate_block() accessor).
• tests/test_postfilter.cpp — 12 measured-first tests pinning every Stage 2 deliverable at its margin target, on both fixture rooms, both cores, 48 kHz and 16 kHz.
• Instrument fix: A-weighting poles above Nyquist are prewarp-clamped (the 12.2 kHz pole pair made every 16 kHz measurement NaN).
• Docs: "Stage 2 delivered" section in the matrix with the full measured table, README chain bullet, HANDOFF stage note.

The three design decisions the measurements forced

1. The AEC chain does not use PEM. Open-loop AEC has an exogenous far end — no closed-loop bias to remove — and the predictor's block-by-block refit injects gradient noise that floors misalignment near −20 dB. Measured on the identical scenario: raw FD-Kalman single-talk residual −75.6 dBm0(A) bare vs PEM-Kalman's −28.9, with double-talk immunity intact (post-DT −70.9) from the Kalman's own per-bin noise-PSD tracking. PEM remains the right structure for the closed loop (AFC).
2. The discriminator correlates the MIC (D = E + Ŷ), never E, against Ŷ. A converged adaptive filter keeps E orthogonal to its reference by construction: coh(E, Ŷ) saturates at 0.34 while adapting (0.53 frozen — the isolation experiment that located this). coh(D, Ŷ) measures 0.99 and is insensitive to adaptation jitter.
3. Suppression depth comes from a coherence-gated leakage estimate, not the coherence itself. P.501's AM-FM double-talk combs interleave 20 Hz apart at the low end — no realizable gain filter notches that selectively, and a coherence-proportional gain measured 5.5 dB of near-end attenuation there. The Wiener-on-E form rides to unity wherever near-end energy inflates |E|², so double-talk transparency is the shape of the rule, not a detector decision.

Measured vs matrix margin targets (all pass)

Deliverable	Clause bound	Margin target	Measured
ST residual, cabin / studio	< −58 dBm0(A)	< −64	−80.5 / −89.7
ST residual, NLMS core / 16 kHz	< −58	< −64	−66.2 / −70.7
DT send attenuation, cabin / studio	≤ 3 dB	≤ 1.5	1.14 / 1.03
DT echo loss, worst band 200–6950 Hz	≥ 27 dB	≥ 33	37.9 / 34.2 (270 Hz)
ERL by 600 / 1200 ms	40 dB @ 1200	≥ 40 / ≥ 46	43.2 / 46.9
Comfort-noise level match	+2 / −5 dB	+1 / −2.5	−1.50
Comfort-noise spectrum, worst band	mask	half-mask	1.69 dB
Noise pumping	≤ 10 dB	≤ 5	2.9
Near-end build-up at DT onset	≤ 50 ms	≤ 25	11.6 ms

Chain config for all of the above: raw FD-Kalman, transition 0.9998 + initial uncertainty 10 (both swept; 0.99995 costs double-talk transparency and tracking, the larger P(0) buys ~4 dB of early convergence), suppressor at defaults.

Full test suite: 124/124 pass locally. Rejected designs (power-leakage ratios, instantaneous-|E|² rules, coh(E,Ŷ), Hann-tapered gain cut, one-pole floor trackers) are documented in the header comments with their measured failures.

Still Stage 3's to prove: the per-row multi-rate compliance suite, spectral echo mask, switching/activation battery, stability sweep, and the G.168-adapted rows.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TLE2rvptWxV8catE8vTvgx

---

Generated by Claude Code