[AlxcBTW]

Follow-up to #132 for the question whether the 0.36 double-talk correlation was real.

• crates/voice/tests/echo_speech.rs: four 6 s LibriSpeech clips (CC BY 4.0, attribution in tests/fixtures/librispeech/README.md), two speakers, 28 s, far end and player at equal level.
• Result: with real speech about 0.75 of the player's voice correlation survives double talk (0.36 was the synthetic periodic signals), echo in the player's pauses is 20 dB down, a fixed extra device delay of 20 or 40 ms changes nothing. Under 20 to 40 ms of clock drift over 28 s (about ten times a typical laptop) far-end-only ERLE is 22 to 34 dB, ERLE in the player's pauses drops to 11 and 5 dB, and the voice kept falls to 0.73 and 0.64.
• Tuning: sonora only exposes AEC3's transparent mode, which changed nothing; the rest (suppressor, delay estimator, clock-drift mode) needs sonora-aec3 directly or a patched copy, which this evidence does not justify.
• ADR 0040 has an Update section with the table and the recommendation: default stays on.
• sonora* crates are built optimised in the dev profile so the tests take seconds, not minutes.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}