"Every attempt was logged, and every attempt failed for the same physical reason. That is when you stop building detectors."
The Feature Everyone Wants
In a spoken conversation you interrupt by talking. Voice assistants call it barge-in: the moment you speak over the reply, the reply stops. Firekeeper built it on day one of Stage 2, and the trouble was immediate. The microphone hears the speakers. If the app cannot tell Pippa's voice coming back through the room from Dad's voice, she interrupts herself.
A Night of Measurements
The first two days were a chain of honest attempts, each logged per reply so the next attempt started from data. Run the reply on the microphone's own audio engine so the system can cancel echo. Vote which mic channel to listen to per utterance. Learn the lag between her voice leaving the speakers and arriving at the mic, and gate on Dad's voice band. Each helped somewhere; none held on office. Office's microphone is a studio interface with 32 inputs, in a room where the monitors reach the mic as loud as Dad does. Apple's voice processing — the echo cancellation every phone call relies on — refuses a 32-channel device. A linear echo canceller topped out near 11 dB. The voice-band gate caught Dad, and still mistook her for him.
attempt result on office's room mic
------- ---------------------------
reply on the mic's engine (system AEC) AEC refuses a 32-input device
per-utterance channel vote helped, did not hold
linear echo canceller ~11 dB, not enough
Dad's voice band, aligned to echo lag caught Dad, mistook her
-------
ruling: voice cut-in OFF by default; the key cuts; no more detector work
The Ruling
Echo cancellation runs only on the system's default devices. A USB mic would change nothing unless it became the system default, and switching the default for every talk is not a way to live — Dad had avoided another vendor's voice mode for exactly that, and because it cannot even turn barge-in off. So the ruling closed the hunt instead of extending it: voice cut-in stays, off by default; a key is the sure cut; no more detector or echo work without Dad. The panel shows it as a "Voice cut-in" switch, off. Turn it on and talking over her stops her too, with the caveat its own tooltip states: a room mic that hears the speakers can mistake her for you.
The Key That Takes Nothing
The default cut-in key is Left ⌘ tapped alone. It is only watched, never held: a modifier tap counts when it is pressed and released with nothing in between — no other key, no click, no other modifier joining it, held no longer than a second — so ⌘C, ⌘Tab and ⌘-click never cut her off, and watching it needs no Input Monitoring. Any other key can be chosen; an ordinary key is registered as a hotkey only while she is speaking, so while she speaks it cuts her and never reaches the app in front, and at other times it types normally. Pause is Space, taken only while the panel is the key window, after a Space typed in another app resumed a paused talk. Clicking the face also cuts in while she speaks.
Where the System Does Cancel Echo
On a MacBook's own mic or AirPods, voice processing runs, and voice cut-in rides Apple's cancellation like the phone. Two guards were added there later. A loud chunk in her pitch range (205–360 Hz) does not count, because cancellation lets her through whole for half a second when she starts again after a pause. And since 2026-10-08 a voice trip does not cut at once: her reading drops to a quarter of its volume, and one second later the latest 2.3 seconds go to cwkPippa's speaker check, which cuts only for Dad's voice. An ear that is away or slow cuts anyway — the safe failure for a cut is to stop talking.