Skip to content

Packaging

EasySpeak ships distro .deb and .rpm packages, split into three:

  • easyspeak (amd64) — the application: a self-contained Python runtime (standalone CPython + all wheels, including the piper engine and the hey_jarvis wake word) under /opt/easyspeak. No pip/uv step and no compiler at install time. Desktop integration (launcher, GNOME Shell extension, AT-SPI dictation helper) is wired to your system's PyGObject / AT-SPI packages, declared as dependencies. The graphical About dialog uses GTK4 / libadwaita where installed, a suggested dependency that any GNOME desktop already has.
  • easyspeak-stt-parakeet (noarch) — the speech recognition model: NVIDIA's multilingual Parakeet TDT v3, one int8 ONNX model for the 25 European languages it covers, under /opt/easyspeak/models/parakeet. EasySpeak uses it by default.
  • easyspeak-lang-en (noarch) — English speech data: the Piper en_US-amy-medium voice and the Whisper base.en recognition model, which stands in when the Parakeet pack is not installed, under /opt/easyspeak/models/en. More languages ship as easyspeak-lang-* packages, each under its own /opt/easyspeak/models/<language>.

Splitting keeps the app small and lets you pick (or add) languages independently without re-downloading the runtime or the shared recognition model. See Language.

EasySpeak integrates deeply with the host (raw input devices, the AT-SPI bus, session D-Bus, a GNOME Shell extension, and spawning host binaries). That is why the supported formats are unconfined distro packages — sandboxed formats (Flatpak/Snap) fight every one of those integration points.

Install

Download the latest packages from the Releases page — the app, the Parakeet pack and a language pack — and install them together. Just take the latest of each: a pack is a separate file (not auto-fetched from GitHub) and is versioned by its models, so the app and pack version numbers are independent and need not match:

sudo apt install ./easyspeak_*_amd64.deb ./easyspeak-stt-parakeet_*_all.deb ./easyspeak-lang-en_*_all.deb

sudo dnf install ./easyspeak-*.x86_64.rpm ./easyspeak-stt-parakeet-*.noarch.rpm ./easyspeak-lang-en-*.noarch.rpm

This is fully offline (assuming the system dependencies — PortAudio, PyGObject, … — are already present, as on any GNOME desktop). Then launch EasySpeak from the applications menu, or run easyspeak.

Dictation additionally needs wl-clipboard, since it places text by pasting rather than through the accessibility bridge, which Chromium-based applications ignore. It is a hard dependency, so it comes with the package.

Installing the app without the packs also works: the wake word is built in, the speech model then downloads on first run if you allow it (see EASYSPEAK_OFFLINE in the configuration), and voice feedback stays off until you add a Piper voice. Without the Parakeet pack and with downloads off, a language pack's Whisper model does the recognition, more slowly.

One non-package step remains: log out and back in once after the first launch so GNOME loads the bundled Shell extension (mouse grid + panel indicator).

To use hold-to-dictate, add yourself to the input group (raw /dev/input access) and log back in:

sudo usermod -aG input "$USER"

Speech recognition

easyspeak-stt-parakeet carries NVIDIA's Parakeet TDT v3 model, converted to an int8 ONNX model, which recognizes speech in 25 European languages and tells them apart by itself. It is one pack for every language, under /opt/easyspeak/models/parakeet, and the recognizer EasySpeak uses by default. The model is CC BY 4.0; the package ships the attribution as /usr/share/doc/easyspeak-stt-parakeet/copyright, next to the full license text.

Language

The default easyspeak-lang-en pack is US English — Piper en_US-amy-medium (feedback) and Whisper base.en (recognition without the Parakeet pack). The hey_jarvis wake word is built into the app (it ships inside the pyopen-wakeword wheel), so it stays English regardless of language pack.

easyspeak-lang-de is German — the multilingual Whisper small and Piper de_DE-thorsten-medium: German dictation and German replies in the German voice. easyspeak-lang-it is Italian the same way, with Piper it_IT-paola-medium, easyspeak-lang-fr French, with Piper fr_FR-siwis-medium, and easyspeak-lang-es Spanish, with Piper es_ES-davefx-medium. Most commands take that language's words too; the English ones work everywhere.

EASYSPEAK_LANGUAGE (default en) is the language you dictate in, and it picks the pack installed under /opt/easyspeak/models/<language>. To use a language we ship no package for, install/drop its models and point these environment variables at them (an explicit value always wins; set them in a systemd user override or your shell profile):

  • EASYSPEAK_WHISPER_MODEL — a multilingual faster-whisper model name (e.g. base, small) or a local model directory. The English-only *.en models can't transcribe other languages.
  • EASYSPEAK_PIPER_MODEL — path to a Piper voice .onnx (with its .onnx.json next to it).

So adding a language needs no root and no package from us: download any faster-whisper model and Piper voice, then set the two variables. A different wake word, however, needs a custom-trained model that pyopen-wakeword can load.

Build the packages yourself

The packages are built in a clean Ubuntu container, so the result is identical whether produced locally or in CI — handy on NixOS, where the bundle's standalone CPython and manylinux wheels can't run on the host:

just package-app              # -> ./dist/ : the easyspeak app .deb and .rpm
just package-app 1.2.3        # set an explicit app version
just package-lang             # -> ./dist/ : every easyspeak-lang-* .deb and .rpm
just package-lang en          # ... or just the named pack(s)
just package-stt              # -> ./dist/ : the easyspeak-stt-parakeet .deb and .rpm
just package-distro           # all of the above (every distro package)

The app and the packs build independently. The app build needs a compiler and the standalone CPython bundle; a pack is only downloaded speech models, so it builds in a lighter container with no compiler and carries its own version from pins.toml, independent of the app release. They release on their own cadence too, but all go through the same release.yml workflow on a published GitHub Release: an application tag builds and attaches the app packages, while a lang-<code>-<version> or stt-<code>-<version> tag skips those jobs and instead builds that pack and attaches it to that release.

Other distributions

These formats are maintained outside this repository. The notes below capture what each one needs.

Nix / NixOS

The repository's flake.nix is a development environment only — a uv run wrapper. It is not a deployable package; treat it purely as a reference. A real Nix package belongs upstream (e.g. nixpkgs) and must, in particular, prepare the GTK4/libadwaita "About" dialog separately: on NixOS its PyGObject and the GTK4/libadwaita typelibs can't ride inside the app venv and must be wired in explicitly (the same pattern the flake already uses for the AT-SPI dictation helper via atspiPython / giTypelibPath).

Arch (AUR)

A PKGBUILD can reproduce the bundled-venv layout under /opt/easyspeak, or build against Arch's own Python. Declare the same runtime dependencies as the .deb/.rpm (PortAudio, PyGObject, AT-SPI, and the optional GTK4/libadwaita, piper/qutebrowser/PipeWire/PulseAudio tooling).

Fedora COPR

A source build from an RPM .spec — effectively the .rpm recipe, built on Fedora's infrastructure for multiple Fedora/EPEL targets.

openSUSE Build Service (OBS)

OBS can build both .deb and .rpm for many distributions from a single recipe; a good option if broad distro coverage is wanted later.

Flatpak / Snap

Not provided. EasySpeak needs raw /dev/input, the AT-SPI accessibility bus, broad session D-Bus access (Mutter RemoteDesktop, GNOME Shell, MPRIS), the ability to install a GNOME Shell extension into your home directory, and to spawn host binaries (piper, qutebrowser, gnome-extensions, gsettings, …). Granting all of that to a sandbox removes the confinement that is the whole point of those formats.