Activating voice mode

The conversational listener boots inactive — cicero start brings the daemon up listening for commands, but it does not open the microphone for a conversation until voice mode is armed. There are four ways to arm it:

Method	How	Notes
Type it	Type voice (or listen) at the cicero> prompt	Deterministic — matched by keyword before the LLM router
Global hotkey	Press Ctrl+Shift+Space	Fixed chord in the current macOS helper; grant Accessibility permission to the host app when prompted
Dashboard button	Click Start listening at http://127.0.0.1:8086	Works from any browser on this machine; no terminal focus needed
Double-clap	Clap twice, hands-free	On by default; see tuning below

Say "stop listening", "goodbye", press the hotkey again, or click Stop listening to turn it back off.

wake_word_enabled and wispr_hotkey are retired. They never enabled acoustic wake-word detection — their only effect was selecting a listener that drove the paid macOS app Wispr Flow. That listener has been replaced by native dictation (see dictation.md). Both keys are still accepted so an existing config keeps starting, and both are ignored with a warning.

The daemon accepts a hotkey label for compatibility, but the current native helper is hardcoded to Ctrl+Shift+Space. cicero status warns when a different value is configured. The helper checks System Settings → Privacy & Security → Accessibility; some macOS releases may also show an Input Monitoring prompt for the terminal host.

"It can't hear me" — check the microphone permission first

The single most common cause is that the process running the daemon lacks Microphone access. macOS grants mic permission per host app (Terminal, iTerm, VS Code, …), not per CLI. Without it, rec/sox opens no input and every recording is empty.

Cicero now surfaces this instead of failing silently — you'll see:

🎙️  Microphone capture failed: <reason>
Grant mic access to your terminal: System Settings → Privacy & Security → Microphone, then re-activate voice.

Fix: System Settings → Privacy & Security → Microphone, enable your terminal app, then re-activate voice mode. The dashboard button is the quickest way to confirm the fix — click it, watch the pill flip to LISTENING, and speak.

Double-clap activation

While voice mode is off, Cicero streams the mic and watches for two quick claps (a sharp energy transient — no model needed). It releases the mic the instant voice mode turns on, so it never contends with the conversational recorder, and resumes automatically when voice mode turns back off.

Because it holds the mic whenever the daemon is idle, the macOS mic indicator stays on. Tune or disable it in ~/.cicero/config.yaml:

clap:
  enabled: true      # set false to turn off clap-to-activate (frees the idle mic)
  threshold: 0.5     # peak amplitude 0..1 that counts as a clap; raise if false triggers
  min_gap_ms: 80     # ignore a 2nd clap faster than this (one clap's own ring)
  max_gap_ms: 600    # the 2nd clap must land within this of the 1st

If claps aren't registering, lower threshold (e.g. 0.35); if ordinary noise arms it, raise threshold and/or narrow the gap window.

End-of-turn: streaming VAD (default), not a volume threshold

A turn ends when Cicero decides you've stopped talking. By default this uses a streaming voice-activity detector, the way Gemini Live / pipecat / LiveKit do it — not the legacy sox absolute-volume gate (silence_threshold), which only stops when the raw signal drops below a fixed loudness and therefore never fires in a room whose noise floor sits above that level (it then records to the 30s cap every turn).

The VAD streams raw mic PCM, learns your room's noise floor during the first ~300ms and carries that estimate across later turns, opens when speech rises a factor above it (so it self-adjusts without swallowing the start of every turn), and ends the turn a short hangover after you stop — the "slight pause after you finish talking" that makes it feel like a conversation. Tune in ~/.cicero/config.yaml:

vad:
  enabled: true        # set false to fall back to the legacy sox silence gate
  hangover_ms: 500     # silence after speech that ends the turn — lower = snappier, higher = more patient
  open_factor: 3       # open threshold = noise floor × this; raise if noise trips it, lower if it misses soft speech
  min_speech_ms: 120   # ignore voiced blips shorter than this (clicks, taps)
  calibration_ms: 300  # initial noise-floor calibration window; the learned floor carries across turns
  preroll_ms: 240      # audio kept before the detected onset so the first phoneme isn't clipped

If it cuts you off at natural pauses, raise hangover_ms (e.g. 700). If it feels laggy, lower it (e.g. 350). The legacy silence_threshold / silence_duration keys only apply when vad.enabled: false.

Muting the beeps (earcons)

The activate/ready/thinking/success/error beeps are on by default. Turn them off for a quieter, less robotic loop (this also keeps earcon bleed out of the VAD's noise-floor calibration):

earcons: false

Dashboard control endpoint

The dashboard's button posts to POST /api/voice on the in-process server. That endpoint can arm the microphone, so it is hardened against CSRF / DNS-rebinding: the server binds to loopback only, and a control request must carry an X-Cicero-Dashboard header (which a cross-origin page cannot set without a CORS preflight the server never approves), a loopback Host, and a loopback Origin when one is present. The dashboard remains read-only over WebSocket for activity.

Deciding what was meant for Cicero

Activation decides when Cicero is listening. It does not decide whether a given utterance was aimed at Cicero — once active, every intelligible non-echo utterance becomes a command.

In a room with other people talking or a podcast playing, the optional intent judge adds a per-utterance veto over that. It is off by default, needs a classifier model, and can only ever decline an utterance — never cause one to be taken.