Skip to content

Speech to text on Linux: download a file, make it executable, speak.

WhisPaste ships prebuilt for Linux on x86_64. You download one file, mark it executable, and press your hotkey. The transcript is pasted straight at your cursor once automatic paste is set up: that needs a one-time permission rule for /dev/uinput: the .deb package sets it up during installation, and with the AppImage or tar.gz you set it up once with sudo. Without it, the text lands in your clipboard, ready to paste with Ctrl+V into the editor, terminal or browser field you are working in.

Download WhisPaste for Linux

Short answer: yes, there is a single executable file. WhisPaste publishes a Linux AppImage that runs without an install step, a .deb for Debian and Ubuntu, and a plain tar.gz of the release bundle. The speech-recognition engine is compiled into the download — there is no Python environment to create, no pip install, and no compile step.

The three Linux files, by name

Every WhisPaste release publishes the same three Linux x86_64 artefacts on GitHub. The version placeholder is the release number (for example 1.3.0); the AppImage and the .deb are additionally published under a stable, unversioned name that always points at the newest release.

  • WhisPaste-<version>-linux-x64.AppImage

    Single executable file, no installation. Distribution-independent — this is the one to take if you just want to run WhisPaste.

    Unversioned alias, always the newest release: WhisPaste-linux-x64.AppImage

  • WhisPaste-<version>-linux-x64.deb

    Debian package for Debian and Ubuntu. Installs into /opt/whispaste with a /usr/bin/whispaste symlink, a desktop entry and an icon.

    Unversioned alias, always the newest release: WhisPaste-linux-x64.deb

  • WhisPaste-<version>-linux-x64.tar.gz

    The plain release bundle: executable plus its lib/ and data/ directories. Unpack it anywhere, run the binary from the folder. Also the source the AppImage, the .deb and the packaging manifests are built from.

AppImage: one file, no install step

Download the AppImage, mark it executable, run it. In a file manager: right-click the file → Properties → Permissions → "Allow executing file as program", then double-click. In a terminal it is chmod +x on the file followed by running it. Running it writes nothing into system directories (only the optional automatic-paste setup adds a uinput rule once with sudo); the app keeps its models and settings under ~/.config/whispaste. To remove WhisPaste again, delete the file.

.deb: the system package for Debian and Ubuntu

If you would rather have WhisPaste in your package manager, install the .deb with apt install ./WhisPaste-<version>-linux-x64.deb. It declares its runtime dependencies (libgtk-3-0, libsecret-1-0, libnotify4, libayatana-appindicator3-1, libasound2, libcurl4, libkeybinder-3.0-0, liblzma5, libc6, zlib1g), so apt pulls in anything missing. The app itself lands in /opt/whispaste and stays read-only there; models and settings still live in your home directory.

No Python, no build, no separate engine download

WhisPaste runs whisper.cpp in-process through a bundled libwhisper library that sits next to the executable inside the download. There is no virtualenv to create, no pip install, no CUDA toolkit, and no compiler invocation — the GPU acceleration ships with it as well (Vulkan on Linux, covering NVIDIA, AMD and Intel with a single build; plain CPU when no usable driver is present). Recent NVIDIA cards can add an optional CUDA pack from the settings.

One thing is downloaded once: the speech model itself, when you start the model download during setup, into ~/.config/whispaste/models/stt. After that WhisPaste transcribes with no network connection at all.

Voice-input shortcut on Wayland (GNOME and KDE)

Wayland does not let apps grab global hotkeys themselves, so let your desktop do it: bind a custom shortcut to the command below. It forwards the call to the running WhisPaste and exits right away. --toggle starts a recording or stops the running one (and starts WhisPaste first if it is not running yet), --cancel discards the running recording. With the AppImage, use the full path to the file instead of whispaste. Push-to-talk, which records while you hold the key, is available on Linux as an experimental option on X11 and on Wayland (GNOME 48+, Plasma 6+).

whispaste --toggle
whispaste --cancel

GNOME: Settings → Keyboard → View and Customize Shortcuts → Custom Shortcuts → +. Name it WhisPaste, enter the command, then set the shortcut.

KDE Plasma 6: System Settings → Keyboard → Shortcuts → Add New → Command or Script…, enter the command and assign the shortcut. On Plasma 5: System Settings → Shortcuts → Custom Shortcuts → Edit → New → Global Shortcut → Command/URL.

For scripts: sending SIGUSR2 to the running instance toggles as well, without starting a second process.

pkill -USR2 -x whispaste

What your machine needs

A 64-bit x86_64 CPU and 8 GB of RAM on Ubuntu Linux or a comparable distribution: that is enough for the compact model on the CPU alone. A GPU is never required; with a dedicated card, Vulkan makes the same transcript come back noticeably faster. Prefer to compile it yourself? The full source is on GitHub. Besides Flutter it needs the speech-engine build step from BUILD.md.

All download options, including Windows and macOS

Before you get started

Is WhisPaste available as an AppImage?
Yes. Every release publishes WhisPaste-<version>-linux-x64.AppImage for x86_64, plus a stable unversioned WhisPaste-linux-x64.AppImage that always resolves to the newest release. It is a single executable file: mark it executable and run it, no installer and no package manager involved.
Can I run WhisPaste on Ubuntu without installing it?
Yes — that is exactly what the AppImage is for. Running it needs no root and touches no system directories, so it can live on a USB stick or in your downloads folder. Only the optional automatic paste needs a one-time permission rule for /dev/uinput, which you set up with sudo. Some newer distributions also need libfuse2 to start it (see the download page). If you would rather install it properly, WhisPaste-<version>-linux-x64.deb is the package for Debian and Ubuntu, and WhisPaste-<version>-linux-x64.tar.gz is the plain bundle you can unpack anywhere.
Do I need Python for WhisPaste on Linux?
No. The speech-recognition engine is whisper.cpp, compiled into a libwhisper library that ships inside the download and is called in-process. There is no Python interpreter, no virtualenv, no pip install, and no build step anywhere in the setup.
Does speech to text work offline on Ubuntu?
Yes. The one thing that needs the network is the initial model download on first launch. Once the model file is in ~/.config/whispaste/models/stt, transcription runs entirely on your machine: no account, no cloud round-trip, and no drop in quality on an air-gapped workstation.
Does it need an NVIDIA card or a CUDA install?
Neither. On Linux WhisPaste accelerates through Vulkan, which covers NVIDIA, AMD and Intel with one single build, and there is no CUDA toolkit to install. Owners of a recent NVIDIA card (RTX or GTX 16 series) can optionally add the CUDA pack from the settings. Without a usable GPU driver everything falls back to the CPU, which is a supported configuration rather than a failure mode.