ZeroWhisper

Free, local-only, push-to-talk speech-to-text for macOS. Zero cloud. Zero cost. Zero telemetry.

Unlike SuperWhisper et al, this costs zero. Runs fully locally. No API keys needed.

Features

• Push-to-talk: Hold Right Option to record, release to transcribe
• Dictate mode: Raw transcript via Parakeet STT (Right Option)
• Polish mode: Transcript cleaned up by Qwen LLM (Right Option + Shift)
• 100% local: All processing on-device, nothing leaves your machine
• Menubar app: Lives in the menubar, no dock icon
• Clipboard safe: Transcript always stays on your clipboard

Requirements

• macOS 13+ (Ventura or later)
• Apple Silicon Mac (M1/M2/M3/M4)
• Xcode (for building)
• Python 3.10+ with pip

Setup

git clone <repo-url> && cd zero-whisper
./Scripts/setup.sh

That's it. Installs Python dependencies, downloads ML models (~4GB), builds the app, and outputs build/ZeroWhisper.app.

Then:

open build/ZeroWhisper.app

Or copy to Applications:

cp -r build/ZeroWhisper.app /Applications/

On first launch, grant Microphone (prompted automatically) and Accessibility (System Settings > Privacy & Security > Accessibility > toggle ZeroWhisper on).

Usage

Action	Hotkey
Dictate (raw transcript)	Hold Right Option (⌥)
Polish (cleaned up text)	Hold Right Option (⌥) + Shift (⇧)

1. Focus the app where you want text inserted
2. Hold the hotkey
3. Speak
4. Release — text is transcribed and pasted

Text is always on your clipboard. If the paste misses, just Cmd+V manually.

Architecture

Swift app (menubar)          HTTP            Python server (localhost:8426)
┌──────────────────┐    POST /transcribe    ┌─────────────────────┐
│ HotkeyManager    │ ─────────────────────> │ Parakeet STT        │
│ AudioRecorder    │ <───────────────────── │ Qwen rewrite (opt)  │
│ PasteManager     │       JSON response    │ Models stay warm    │
└──────────────────┘                        └─────────────────────┘

Troubleshooting

• Hotkey not working: Check Accessibility permission in System Settings
• No audio: Check Microphone permission
• Slow first use: Models are loading into memory. Subsequent uses are fast.
• Helper not starting: Run ./Scripts/setup.sh to rebuild