Files
clawdbot/Swabble/README.md
2025-12-06 03:55:46 +01:00

4.0 KiB

🎙️ swabble — Speech.framework wake-word hook daemon (macOS 26)

swabble is a Swift 6.2, macOS 26-only rewrite of the brabble voice daemon. It listens on your mic, gates on a wake word, transcribes locally using Apple's new SpeechAnalyzer + SpeechTranscriber, then fires a shell hook with the transcript. No cloud calls, no Whisper binaries.

  • Local-only: Speech.framework on-device models; zero network usage.
  • Wake word: Default clawd (aliases claude), optional --no-wake bypass.
  • Hooks: Run any command with prefix/env, cooldown, min_chars, timeout.
  • Services: launchd helper stubs for start/stop/install.
  • File transcribe: TXT or SRT with time ranges (using AttributedString splits).

Quick start

# Install deps
brew install swiftformat swiftlint

# Build
swift build

# Write default config (~/.config/swabble/config.json)
swift run swabble setup

# Run foreground daemon
swift run swabble serve

# Test your hook
swift run swabble test-hook "hello world"

# Transcribe a file to SRT
swift run swabble transcribe /path/to/audio.m4a --format srt --output out.srt

Use as a library

Add swabble as a SwiftPM dependency and import the Swabble product to reuse the Speech pipeline, config loader, hook runner, and transcript store in your own app:

// Package.swift
dependencies: [
    .package(url: "https://github.com/steipete/swabble.git", branch: "main"),
],
targets: [
    .target(name: "MyApp", dependencies: [.product(name: "Swabble", package: "swabble")]),
]

CLI

  • serve — foreground loop (mic → wake → hook)
  • transcribe <file> — offline transcription (txt|srt)
  • test-hook "text" — invoke configured hook
  • mic list|set <index> — enumerate/select input device
  • setup — write default config JSON
  • doctor — check Speech auth & device availability
  • health — prints ok
  • tail-log — last 10 transcripts
  • status — show wake state + recent transcripts
  • service install|uninstall|status — user launchd plist (stub: prints launchctl commands)
  • start|stop|restart — placeholders until full launchd wiring

All commands accept Commander runtime flags (-v/--verbose, --json-output, --log-level), plus --config where applicable.

Config

~/.config/swabble/config.json (auto-created by setup):

{
  "audio": {"deviceName": "", "deviceIndex": -1, "sampleRate": 16000, "channels": 1},
  "wake": {"enabled": true, "word": "clawd", "aliases": ["claude"]},
  "hook": {
    "command": "",
    "args": [],
    "prefix": "Voice swabble from ${hostname}: ",
    "cooldownSeconds": 1,
    "minCharacters": 24,
    "timeoutSeconds": 5,
    "env": {}
  },
  "logging": {"level": "info", "format": "text"},
  "transcripts": {"enabled": true, "maxEntries": 50},
  "speech": {"localeIdentifier": "en_US", "etiquetteReplacements": false}
}
  • Config path override: --config /path/to/config.json on relevant commands.
  • Transcripts persist to ~/Library/Application Support/swabble/transcripts.log.

Hook protocol

When a wake-gated transcript passes min_chars & cooldown, swabble runs:

<command> <args...> "<prefix><text>"

Environment variables:

  • SWABBLE_TEXT — stripped transcript (wake word removed)
  • SWABBLE_PREFIX — rendered prefix (hostname substituted)
  • plus any hook.env key/values

Speech pipeline

  • AVAudioEngine tap → BufferConverterAnalyzerInputSpeechAnalyzer with a SpeechTranscriber module.
  • Requests volatile + final results; wake gating is string match on partial/final.
  • Authorization requested at first start; requires macOS 26 + new Speech.framework APIs.

Development

  • Format: ./scripts/format.sh (uses ../peekaboo/.swiftformat if present)
  • Lint: ./scripts/lint.sh (uses ../peekaboo/.swiftlint.yml if present)
  • Tests: swift test (uses swift-testing package)

Roadmap

  • launchd control (load/bootout, PID + status socket)
  • JSON logging + PII redaction toggle
  • Stronger wake-word detection and control socket status/health