claudeaceae/.claude/skills/voice-call

/voice-call - FaceTime Voice Calling

Place and receive FaceTime Audio calls with live transcription and voice/text responses.

Quick Start

# Full voice conversation (responds via FaceTime audio)
~/.claude-mind/system/bin/voice-call --voice-response

# Call with text responses (via iMessage)
~/.claude-mind/system/bin/voice-call --text-response

# Call a specific number
~/.claude-mind/system/bin/voice-call +15551234567 --voice-response

Prerequisites

Run audio-setup --check to verify:

• sox — audio recording and playback
• SwitchAudioSource — audio device routing
• whisper-cli — speech-to-text transcription
• Loopback — virtual audio device routing (must be running)
• Claude Mic — virtual device for TTS → FaceTime input
• Call Capture — virtual device for FaceTime output → recording
• aggregate-device — Swift tool for creating CoreAudio aggregate devices

Scripts

| Script | Purpose | |--------|---------| | voice-call | High-level orchestrator (setup → call → listen → respond → teardown) | | facetime | FaceTime control (open/close/status/call/answer/hangup) | | facetime-incoming-watcher | Polls for incoming calls, auto-answers collaborator | | call-record | Aggregate device management + sox recording | | call-transcribe | Whisper transcription of WAV files | | call-listen | Recording + transcription + response loop | | call-speak | Play TTS audio to "Claude Mic" device | | audio-setup | Verify/configure audio prerequisites | | aggregate-device | Create/destroy CoreAudio aggregate devices |

Options

voice-call [target] [options]

  --text-response    Respond via iMessage (default)
  --voice-response   Respond with TTS through FaceTime
  --greeting TEXT    Custom greeting when call connects
  --no-greeting      Skip initial greeting
  --timeout SECS     Max call duration (default: 600)
  --answer           Answer mode (incoming call, skip dialing)

How It Works

1. Aggregate device: Creates a CoreAudio multi-output device ("Call Monitor") combining physical speakers + Call Capture. On macOS 26, FaceTime call audio routes through the system default output — a pure virtual device receives silence, but an aggregate containing physical speakers gets call audio on ALL sub-devices.
2. Timing: The aggregate device MUST be set as system output BEFORE the FaceTime call connects. Audio routing is locked at connection time.
3. Call initiation: Uses AppleScript UI automation (New → type number → switch dropdown to Audio → Return). The facetime-audio:// URL scheme only shows a notification on macOS 26+, not an actual call.
4. Hangup: Clicks NotificationCenter's "End" button. On macOS 26, calls persist as system-level calls even after quitting FaceTime.
5. Recording: sox rec with 35dB gain amplification + 0.1% silence threshold. Call audio from the aggregate device is very low amplitude; gain boost is required for silence detection.
6. Transcription: whisper-cpp batch mode on each WAV chunk (48kHz stereo → 16kHz mono conversion).
7. Turn-taking: Silence-based (1.5s threshold). Recording pauses during TTS playback to avoid self-echo.
8. Exit detection: Recognizes "goodbye", "hang up", "end call", "bye bye", "talk to you later".

Lifecycle

Outgoing Call

call-record setup     →  Create aggregate device, set system output
facetime call         →  UI automation dials FaceTime Audio
call-speak greeting   →  TTS greeting through Claude Mic
call-record start     →  Begin sox recording from Call Capture
  ↕ call-listen loop: record → transcribe → respond → pause/resume
call-record stop      →  Stop sox recording
call-record teardown  →  Destroy aggregate, restore audio devices
facetime hangup       →  Click NotificationCenter End button

Incoming Call (auto-answer)

facetime-incoming-watcher  →  Polls NotificationCenter every 5s
  detect Accept/Decline    →  Incoming call found
  identify caller          →  Match against collaborator name/phone
  call-record setup        →  Create aggregate device (before accepting!)
  facetime answer          →  Click Accept + set audio devices
  voice-call --answer      →  Greeting → call-listen loop → cleanup

The watcher runs as a launchd service (com.claude.facetime-incoming). It auto-answers calls from the collaborator (~2-3s pickup) and lets unknown callers ring.

Configuration

Stored at ~/.claude-mind/state/voice-call-config.json:

{
  "micDevice": "Claude Mic",
  "captureDevice": "Call Capture",
  "whisperModel": "~/.claude-mind/system/models/ggml-base.en.bin",
  "sampleRate": 16000,
  "silenceThreshold": 1.0,
  "silenceLevel": "0.1%",
  "callGain": 35
}

• silenceThreshold — seconds of silence before splitting chunks (default: 1.5)
• silenceLevel — amplitude threshold for silence detection (default: 0.1%)
• callGain — dB gain applied before silence detection (default: 35)

Service Toggle

service-toggle voiceCall on|off|status

Trigger Words

Use /voice-call or natural phrases like "call me", "FaceTime me", "let's talk on the phone".