Push-to-Talk Dictation for Linux, Entirely Offline

What it does

Voxtype is a local voice-to-text daemon for Linux that types transcription directly at your cursor. Hold a hotkey, speak, and release to inject text. It targets Wayland compositors first—Hyprland, Sway, River, Niri, GNOME, and KDE—with an evdev fallback for X11.

The interesting bit

Instead of bolting a generic speech API onto the desktop, it wires itself into the compositor: it pauses MPRIS music while you speak, shows a floating waveform OSD aligned with swayosd, and can pipe raw output through a local LLM or shell script to clean up grammar before typing. It also juggles seven ASR engines (Whisper, Cohere, Parakeet, and others) with dynamic loading so you only pay RAM for the one in use.

Key highlights

• Runs fully offline by default with no telemetry; optional remote Whisper servers if you explicitly want them.
• Cohere Transcribe runs at 9–11× realtime on a Zen 4 CPU, while Parakeet targets AMD and NVIDIA GPUs via MIGraphX and CUDA.
• Meeting mode offers continuous transcription with speaker attribution and export to Markdown, JSON, SRT, or VTT.
• Built-in text processing handles spoken punctuation and per-user word replacements, plus an optional post-processing command for grammar or filler-word removal.

Caveats

• KDE’s KWin does not expose key-release events, so you must use toggle mode instead of true push-to-talk.
• Non-US keyboard layouts may need fallback to clipboard paste mode, as direct typing backends can struggle with variants.
• Releasing multi-modifier keybindings slowly on some compositors can cause the typed transcription to trigger window manager shortcuts instead of inserting text.

Verdict

Wayland-native developers and heavy Linux desktop users who want private, low-latency dictation should look here. If you expect a polished cross-platform app or prefer zero-configuration cloud APIs, this is not your tool.