Whisper Smart

Hold a key. Speak. Release.
Voice-to-text for macOS and Linux — transcribed on-device and injected straight at your cursor, in any app.

---

01 — How it works

1. Hold your hotkey (Right Command on macOS, Right Control on Linux)
2. Speak — the app records from your mic
3. Release — transcription lands at your cursor

Works everywhere you type: browsers, editors, terminals, Slack, email.

Two apps, one product. macOS is Swift and SwiftUI with MLX for speech; Linux is Rust and GTK4 with whisper.cpp. They are not a cross-compile of each other — neither platform's UI toolkit or speech runtime exists on the other — but the dictation lifecycle, hotkey semantics, insertion strategies and settings model are the same, and every release ships both from one tag.

02 — Install

Both platforms are published on the same release.

macOS

1. Download Whisper-Smart-mac.dmg
2. Drag Whisper Smart into Applications
3. Launch — it lives as a mic icon in your menu bar
4. Grant Accessibility and Microphone permissions when prompted

Requires macOS 14 (Sonoma) or later. Updates arrive automatically through Sparkle.

Linux

# Prebuilt tarball (built on Arch, installs under ~/.local, no root)
tar xzf whisper-smart-X.Y.Z-linux-x86_64.tar.gz
./whisper-smart-X.Y.Z-linux-x86_64/install.sh

# Or from source, on any distro
git clone https://github.com/itisrmk/whisper-smart
cd whisper-smart/linux && bash packaging/install.sh

whisper-smart --check                     # reports anything still missing
systemctl --user enable --now whisper-smart

Needs a GTK 4 desktop and membership of the input group (sudo usermod -aG input "$USER"), which is how a global hotkey works at all under Wayland. whisper-smart --check prints the exact command for anything missing. Full setup: linux/README.md.

03 — Speech providers

Pick the engine that fits your workflow in Settings → Provider. Every local engine keeps audio on your machine; cloud is opt-in and never a silent fallback.

	macOS	Linux
Default local	Whisper Large-v3 Turbo (MLX)	Whisper Large-v3 Turbo (whisper.cpp)
Fast local	Whisper Tiny/Base (MLX)	faster-whisper
Alternative local	Parakeet TDT 0.6B (MLX)	Parakeet TDT (ONNX)
Cloud	OpenAI Whisper API	OpenAI Whisper API
Zero-setup fallback	Apple Speech (built in)	—
GPU	Apple Silicon via MLX	NVIDIA via ggml-cuda

On macOS an unready provider falls back to Apple Speech, so you are never stuck. Linux has no such system engine, so it says what is missing and how to fix it instead — and never silently sends your audio to the cloud.

04 — Features

• Press-and-hold or one-shot — hold the hotkey while speaking, or start dictation from the menu for toggle mode
• Left/right key aware — bind the right-hand modifier without the left one triggering it
• Configurable hotkey — pick a preset or record any modifier combo in Settings
• Terminal-friendly — paste handling tuned for Terminal.app, iTerm2, Kitty, Warp, Ghostty, Alacritty, foot
• Smart silence detection — no speech means instant return to idle, no waiting
• Writing styles — neutral, formal, concise, casual, or developer mode with per-app overrides
• History — recent transcripts, re-insertable with one click
• Privacy-first — local providers keep audio on your machine; cloud is opt-in

05 — Customizing the hotkey

Open Settings → Hotkey to pick a preset or record a custom combo. Each physical key is tracked independently, so Right Command and Left Command are different bindings.

	macOS	Linux
Default	Right Command hold	Right Control hold
Mechanism	CGEvent tap (Accessibility permission)	/dev/input via evdev (input group)
Insertion	Accessibility AXValue, then ⌘V paste	wtype, then clipboard paste

06 — Building from source

# macOS
swift build && .build/debug/Whisper\ Smart
bash scripts/run_qa_smoke.sh

# Linux
cd linux
cargo build --release && ./target/release/whisper-smart --check
bash scripts/run_qa_smoke.sh

Architecture, and the file-by-file mapping between the two ports, is in CLAUDE.md, docs/ARCHITECTURE.md and linux/README.md.

07 — Releases

One tag, both platforms. Every vX.Y.Z carries the macOS DMG, the Sparkle appcast, the Linux tarball and its checksum, built from the same commit — the tag is only created once both builds pass. See docs/RELEASE.md.

License

See LICENSE for details.