Microphone calibration for voice barge-in (tune sensitivity per mic)

What it is

Voice barge-in lets you interrupt the AI mid-sentence by talking over it, so a spoken read-back stops the moment you start speaking. To decide "the user is talking now," Omniscio listens to your microphone and asks a small on-device model how speech-like each moment of audio is. Whether a given level counts as "you talking" versus "background noise" depends heavily on your specific microphone and room. A laptop mic in a quiet office, a USB headset next to a loud fan, and a phone propped across the desk all produce very different baselines.

Microphone calibration is a short guided test that measures your mic and room once, then tunes the barge-in sensitivity to them. It exists to solve two opposite failure modes at the same time:

• Without tuning, a sensitive setup treats background noise (a fan, a keyboard, an air conditioner) as speech and falsely interrupts the AI.
• An over-cautious setting does the reverse: a quiet talker speaks but never registers, so barge-in never fires.

Calibration finds a threshold that sits comfortably above your room's normal background noise yet below your normal speaking voice, so real speech reliably interrupts and ambient noise does not.

Where to find it

Open Settings -> Voice Control. Voice Control is a top-level Settings section with accordion cards; the relevant one is Voice Report Back. Expand it and turn on Enable Voice Report Back to reveal the inner controls.

Inside that card, when the Lab flag is enabled you see, in order:

1. Enable barge-in (Lab) toggle - the voiceBargeInLabFlag switch described above, with the description "Interrupt the AI mid-sentence by speaking. Internal testing flag." This is intentionally not searchable in Settings search; you reach it only by scrolling this card.
2. A first-toggle nudge - the first time you turn the Lab flag on while your current microphone has never been calibrated, a small highlighted banner appears reading "Calibrate your microphone for best results" with two buttons, Calibrate now and Skip for now. It is a discoverability prompt, not a wall: you can ignore it.
3. A "Calibrate microphone" row - always available whenever the Lab flag is on (independent of the nudge), with a Calibrate button on the right. This is how you re-run calibration any time, including after the nudge is gone.

The nudge disappears on its own once the current mic has been calibrated, and immediately if you click "Skip for now." (If you skip, the nudge can reappear in a later session until you actually calibrate that mic; calibrating it makes the nudge stop for good for that mic.)

How it behaves

Current status (experimental / Lab flag, OFF by default)

This is an internal, experimental feature. It is hidden behind a Lab flag named voiceBargeInLabFlag that is off by default. Nothing about calibration is visible until you turn that Lab flag on, and barge-in itself does not run for most users yet.

This honesty matters: in this release (Slice 2.2) calibration is a dogfooding / internal-testing capability, not a finished setting that ships to everyone. The general-availability switch that turns barge-in on for all users (voiceBargeInEnabled) arrives in a later release (Slice 2.4). Until then, calibration is reachable only by enabling the Lab flag.

The calibration flow (what you actually do)

Clicking Calibrate now or Calibrate opens a small dialog. The whole test takes about 15 seconds and runs in two short steps:

1. Stay quiet. A 3-2-1 countdown ("Get ready to stay quiet...") leads into roughly 5 seconds where you simply say nothing. A microphone icon, a live level meter, and a countdown show that Omniscio is listening. This measures your room's background noise.

2. Read one sentence. A second 3-2-1 countdown ("Now get ready to read out loud...") leads into roughly 5 seconds where you read one fixed sentence at a normal speaking volume:

   The quick brown fox jumps over the lazy dog.

   (It is a pangram chosen because it covers a broad range of speech sounds and is short enough to read inside the window.) This measures your normal speaking voice.

After the two samples, Omniscio shows a brief "Working out your settings..." step, then a success screen: a green check, "Microphone calibrated," the note "Barge-in is tuned for this mic. You can recalibrate any time from this screen," and a plain-language Sensitivity: low / medium / high label (no raw numbers, which would only confuse). Click Done to save.

While the two samples and the countdowns are running, the dialog cannot be closed (no X, no Escape, no click-outside) so a half-finished measurement can never be saved. Before you start, and on the success or error screens, you can close it freely.

"Skip for now" and the uncalibrated default

You never have to calibrate. If you click Skip for now (on the nudge or on a recoverable error screen), or you simply never open the wizard, Omniscio uses a sensible built-in default sensitivity and saves nothing. That default is deliberately identical to the single fixed sensitivity barge-in used before calibration existed, so skipping changes nothing about how barge-in behaves compared to before this feature. Calibration is an opt-in improvement on top of a working default, not a prerequisite.

It is per-microphone

Calibration is stored per microphone, keyed to the specific input device. Each mic you use is calibrated once; the result is remembered and reused every time that mic is the active one. If you switch microphones (plug in a headset, move to a different machine, change your default input), the new mic starts uncalibrated and uses the default until you calibrate it too. There is no global "recalibrate everything" - you tune each mic when you first use it, and the saved values persist across restarts.

Edge cases you might hit

Calibration is honest about failure: if it cannot measure cleanly, it blocks with a clear message and saves nothing (your previous setting, or the default, stays in effect). It never silently writes a bad result. Each case below shows a short explanation plus either a "Try again" button, a "Skip for now" button (use the default), or a "Close" button, as appropriate:

• No microphone. "No microphone detected. Connect a mic and try again." Nothing is saved.
• Microphone access blocked. "Microphone access is blocked. Allow mic access in your system settings, then try again." Grant permission, then retry.
• Too much background noise. "There is too much background noise to calibrate reliably. Move somewhere quieter and try again, or skip for now to use the default settings." A reading taken in a noisy room would push the threshold so high that your real speech might never register, so a bad calibration is worse than the default.
• Speech too quiet. "We could not hear clear speech. Read the sentence at a normal volume and try again." Too little clear speech makes the measurement unreliable.
• Calibration could not start. A generic failure (for example the on-device speech model failed to load): "Calibration could not start. Please restart Omniscio and try again."

In every failure case you can retry or, where it makes sense, skip to keep the default. You are never left in a broken state, and you can always reopen the wizard later from the "Calibrate microphone" row.

For agents

How it works (code references)

Everything here runs in the renderer; there is no separate background service and no new IPC channel for calibration. The saved values ride on the existing settings save path.

• Saved data. Calibration results live in a single AppSettings field, vadThresholdsByDeviceId, a map keyed by the microphone's device id. Each entry holds two numbers, silenceFloor and speechThreshold, both voiced-probability values in the range 0 to 1 produced by the speech model. An absent entry means "uncalibrated" and resolves to the built-in default at read time; "Skip for now" writes nothing. The field defaults to an empty object and is validated by the settings Zod schema; saving merges the new device entry into the existing map via the normal SETTINGS_UPDATE settings write (no dedicated calibrate channel).
• The math. The pure, dependency-free measurement logic lives in src/renderer/src/features/voice/barge-in-calibration.ts (computeVadProfile, percentile, DEFAULT_VAD_PROFILE, SAFETY_MARGIN). The silence floor is the 95th percentile of the quiet sample (deliberately not the average, so one stray noise spike cannot drag the floor down). The speech threshold is the average voiced level during the spoken regions minus a safety margin, then clamped so it always sits a minimum distance above the floor and leaves headroom at the top. DEFAULT_VAD_PROFILE.speechThreshold is 0.75 - byte-identical to the fixed sensitivity barge-in used before calibration, so an uncalibrated mic behaves exactly as it did before.
• The wizard UI. src/renderer/src/features/voice/BargeInCalibrationWizard.tsx is the 2-step dialog (built on DialogShell), lazy-loaded because it pulls in the on-device speech model. Each 5-second sample is captured inside its own short-lived AudioContext (a deliberate discipline; a shared/singleton context was tried and reverted because it stopped delivering audio after a few cycles).
• The nudge. src/renderer/src/features/voice/BargeInFirstToggleNudge.tsx is the inline banner; all "should it show" logic lives in its parent so it can later be re-pointed at the general-availability toggle with no change.
• Where it attaches + the live read. The Lab toggle, nudge, and Calibrate row live in src/renderer/src/features/settings/sections/voice/VoiceReportBackSettings.tsx. The live barge-in detector in src/renderer/src/components/ui/VoiceInput.tsx reads vadThresholdsByDeviceId[currentDeviceId]?.speechThreshold (falling back to the default) so the calibrated value actually drives the interrupt decision; the stored silenceFloor is calibration-time only and does not feed the live comparison.

Spec and contract

• Design spec (rationale, algorithm detail, edge-case table, sign-off history): docs/superpowers/specs/2026-05-29-voice-barge-in-2-2-calibration-design.md.
• Feature contract (test-locked invariants, safe-change checklist): .claude/memory/contracts/voice-barge-in-calibration-contract.md.

Related

• voice-report-back.md - the read-back feature barge-in is meant to interrupt; the calibration controls live inside its Settings card.
• voice-and-tts.md - the underlying Voice Control, dictation, and TTS stack.

Public help-site note: this feature is still behind a Lab flag, so the public help site (docs.omniscio.com) is intentionally not updated yet. Public coverage lands when barge-in reaches general availability (Slice 2.4).