Kaze

Hold a global hotkey, speak, and the transcribed text is automatically pasted into whatever app you're using. Everything runs locally on your Mac -- no cloud, no API keys, no data leaves your machine.

Kaze.mp4

Download

Grab the latest .dmg from GitHub Releases.

How it works

1. Press your global hotkey (default: Option + Command) to start recording.
2. Speak while Kaze captures audio and shows a floating waveform overlay.
3. Stop recording (release in Hold mode, or press again in Toggle mode).
4. Kaze pastes the transcription into the focused app while preserving your clipboard.

The app lives entirely in the menu bar with no Dock icon. On first launch, a guided onboarding wizard walks you through hotkey setup and engine selection.

Features

Transcription engines

Personal recommendation -- use Parakeet v3 (NVIDIA) for the best overall results.

Kaze ships with 4 fully on-device transcription engines:

Engine	Framework	Notes
Direct Dictation	Apple SFSpeechRecognizer	Zero setup, real-time streaming, uses device locale
Whisper (OpenAI)	WhisperKit	Local model variants: Tiny, Base, Small, Large v3 Turbo
Parakeet v3 (NVIDIA)	FluidAudio	Fast, high-accuracy English ASR (~600 MB CoreML model)

Model management

• One-click download/remove in Settings -- view readiness status and model size on disk
• Cancel in-progress downloads at any time
• Idle model unloading -- models automatically free memory after 90 seconds of inactivity
• Graceful fallback -- if a selected model is unavailable, Kaze falls back to Direct Dictation

Recording overlay

• Dynamic Island / notch mode -- a recording indicator that extends from the MacBook notch at the top of the screen, with animated expand/collapse transitions
• Pill mode -- traditional floating pill at the bottom-center of the screen
• Real-time waveform bars driven by audio level
• Live scrolling transcription text with leading fade mask
• Processing state -- shimmer animation + spinner while model inference or text enhancement runs

Apple Intelligence enhancement

• Post-process transcriptions with on-device Foundation Models to fix grammar, punctuation, and formatting (macOS 26.0+)
• Customizable system prompt -- edit the enhancement instructions or reset to defaults
• Custom vocabulary words are injected into the enhancement prompt for better accuracy
• Only applies to Direct Dictation (AI model engines already produce clean output)

Global hotkey

• Configurable shortcut with support for key + modifier and modifier-only combos
• Two modes: Hold to Talk and Press to Toggle
• Default: Option + Command
• Visual shortcut recorder in both Settings and Onboarding

Microphone selection

• Pick a specific audio input device or use the system default
• Real-time device list updates when hardware is connected/disconnected
• Selection persists across sessions and is validated on launch

Onboarding

• 4-step guided setup on first launch: Welcome, Hotkey configuration, Engine selection, and Completion summary
• Preferences are saved automatically as you complete each step

Other features

• Custom vocabulary/keywords -- add names, abbreviations, and domain terms to improve recognition across all engines
• Transcription history -- persistent local history (latest 50 entries) with engine labels, "Enhanced" badge, relative timestamps, and one-click copy
• Clipboard-safe auto-paste -- saves and restores your clipboard contents around each paste
• Trailing space option -- optionally append a space after each transcription
• Launch at login -- start Kaze automatically when you log in
• About dialog -- version info, links to GitHub and Releases
• Menu bar status indicator -- icon dims when no model is loaded, animates during model loading

Tech stack

Layer	Technology
UI	SwiftUI + AppKit -- SwiftUI for Settings/Onboarding/Overlay views; AppKit for menu bar, floating panel, clipboard, and simulated key events
Speech	Apple Speech framework (SFSpeechRecognizer) for real-time streaming dictation
Whisper	WhisperKit for local OpenAI Whisper transcription
Parakeet	FluidAudio for Parakeet v3 CoreML runtime
Enhancement	Foundation Models (Apple Intelligence on-device LLM) for text cleanup
Hotkey	CGEvent tap for low-level global hotkey detection
Audio	AVCaptureSession + Accelerate/vDSP for microphone capture, format conversion, and resampling
State	Combine for reactive state bridging between transcription engines and the UI
Login item	SMAppService for launch-at-login registration

Requirements

• macOS 26.0+
• Xcode 26+ (for building from source)
• Accessibility permission (for global hotkey)
• Microphone permission
• Speech Recognition permission (used by Direct Dictation)

Building from source

git clone https://github.com/fayazara/Kaze.git
cd Kaze
open Kaze.xcodeproj

Build and run in Xcode. Dependencies (WhisperKit + FluidAudio) are resolved automatically via Swift Package Manager.

License

MIT