Vocalinux

An open-source, system-wide voice-dictation app for Linux — the first source in this wiki about the application layer that sits on top of STT engines rather than a model itself. It runs in the background, activates on a configurable hotkey (toggle or push-to-talk), and injects recognized text into whatever app has focus: browsers, editors, terminals, chat clients. The pitch is native voice typing on Linux comparable to the built-in dictation on Windows/macOS, with no cloud and no subscription.

What it actually is

Vocalinux is a wrapper and desktop-integration layer, not a recognizer. The recognition is done by one of three local backends you choose:

• whisper.cpp (default) — the efficient C/C++ port of Whisper.
• OpenAI Whisper — the reference Python implementation.
• vosk — the lightweight Kaldi-based offline engine.

Models run on-device and work offline after the initial download, which is the whole point: “No internet required. No data leaves your machine.” It exposes model-size choices so weaker hardware can trade accuracy for speed, and gets Vulkan GPU acceleration across AMD, Intel, and NVIDIA — the repo notes this works “without requiring CUDA or PyTorch” with the whisper.cpp backend, which is what keeps the install light on a mixed-GPU Linux fleet. A neural Voice Activity Detection stage (Silero VAD via ONNX Runtime, when available) gates the recognizer so it isn’t transcribing silence. Beyond dictation it recognizes editing voice commands (“new line,” “delete that,” “capitalize”) and supports languages past English (French, German, Russian).

The Linux-desktop hard parts it handles

The interesting engineering is the OS integration, not the ASR. It supports both X11 and Wayland, where text injection works differently (Wayland’s security model makes synthetic input harder) — handling both is the app’s real value-add over “run Whisper in a terminal.” It adds a system-tray indicator with idle / listening / processing states, audible cues, and non-US keyboard layouts, and ships via shell installer, Flatpak, and the Arch AUR across Ubuntu/Debian/Fedora/Arch/openSUSE.

Stack is Python-dominant (~73%) with some TypeScript and shell, GTK for the tray/settings UI, tested across GNOME/KDE/Xfce/Cinnamon/MATE/LXQt. It’s the Linux arm of a “Voca” ecosystem — VocaMac (macOS, beta) and VocaWin (Windows, planned) — maintained by Jatin Malik (@jatinkrmalik); ~650 stars, 24 releases at refresh.

Why it’s in the wiki

Every prior STT page here is about the models and their WER/latency (speech-to-text, whisper, canary-qwen, open-source-stt-models). Vocalinux is the consumption side: it shows what the open-weight STT wedge is for on a personal machine — private, offline dictation assembled from off-the-shelf engines, where the differentiator is desktop plumbing and hardware reach, not the recognizer. It also anchors vosk and whisper-cpp as the “small/local” pole of the STT field that the accuracy-leaderboard pages don’t cover.

Caveats

Refreshed 2026-07-23 from the primary repo (jatinkrmalik/vocalinux), upgrading the earlier Linuxiac-only writeup — so T1 for the repo facts, though feature claims are still vendor self-description (no independent benchmark or security audit). The primary source resolves the earlier open license question: it is GPL-3.0. Still no WER/accuracy numbers, as expected — accuracy is inherited from whichever backend (whisper.cpp / Whisper / VOSK) and model size the user runs, so Vocalinux is untestable as a recognizer by design.

Related

speech-to-text · whisper · whisper-cpp · vosk · open-source-stt-models · speech-audio-ai