Timbre

A free, fully-local dictation app for macOS. Hold right-Option, speak, release — cleaned-up text appears in whatever app you're using.

Everything runs on-device: transcription on Apple's SpeechAnalyzer, cleanup on the FoundationModels on-device LLM. No account, no network, no subscription. See GDR-0001 for why that's the whole point.

Requirements

• macOS 26.0+ on Apple Silicon, with Apple Intelligence enabled (System Settings › Apple Intelligence & Siri)
• Xcode 26+ to build

Building

# The pipeline library + tests (fast, no app launch):
cd packages/TimbreKit && swift test

# The app:
xcodebuild -project apps/Timbre/Timbre.xcodeproj -scheme Timbre build

Or open apps/Timbre/Timbre.xcodeproj in Xcode and ⌘R.

First run: grant Microphone when prompted. For Accessibility, approve the app in System Settings › Privacy & Security › Accessibility, then quit and relaunch — the trusted state is only read at launch.

Repository map

apps/Timbre/        Xcode app: scenes and glue views, nothing else
packages/TimbreKit/ The pipeline: audio → transcription → polish → insertion
docs/decisions/    ADRs (architecture) and GDRs (product) — start here
docs/learning/     Timbre University: macOS development from first principles
web/               Website (placeholder until the app ships)

The pipeline, in one breath: HotkeyMonitor emits press/release events → DictationController starts AudioCapture (mic → AnalyzerInput stream) → Transcriber (on-device ASR) streams live text into the OverlayController pill → on release, TextPolisher (on-device LLM) cleans the transcript → TextInserter pastes it where the user was typing.

Contributing to decisions

Significant choices are recorded in docs/decisions/ — architecture in adr/, product in gdr/. The records are immutable; to reverse one, write a new record that supersedes it. The index and process live in docs/decisions/README.md.