Speak a messy ramble. A voice-native language model hears the audio and returns clean text, written in your style. Runs on Android 12+ and Linux.
Ramble Voice is text-only: it never executes commands, and spoken instructions are transcribed, not obeyed. There is no backend, no telemetry and no analytics. Nothing leaves your device except Gemini requests, and only if you turn on cloud mode.
License: GPL-3.0-or-later. Repository: https://github.com/luohoa97/ramble-voice
Android (Oppo Find X2 Neo, Android 12).
| Record | Cleaned result | History |
|---|---|---|
 |
 |
 |
| Floating mic next to Gboard | Ramble keyboard, listening | Ramble keyboard |
|---|---|---|
 |
 |
 |
- Dictate freely, with pauses, false starts and filler words, and get clean text.
- On-device by default with Gemma 4. Your audio never leaves the phone or PC.
- Optional Gemini cloud mode with your own free API key. If Gemini is rate limited, offline or out of quota, the app falls back to the on-device model when it is downloaded.
- Long recordings: Gemma 4 accepts 30 seconds of audio per request, so audio is split at silences (about 28 seconds) and the results are stitched together.
- Android: a keyboard with hold-to-talk, a processing state, text inserted directly into the focused field, and a quick switch back to your previous keyboard.
- Linux: global hold-to-talk hotkey, text insertion into the focused app, an overlay, and a tray icon, depending on your desktop (see the matrix below).
- GPU acceleration when available: OpenCL on Android (Snapdragon and Adreno work), Vulkan on Linux with a real vendor driver, otherwise CPU.
Android: Android 12 or newer (API 31), arm64-v8a only. Newer features (Android 16 Live Update download progress, the Android 14 threshold haptic) are used where the phone supports them.
Linux uses a custom GTK4 runner. Legend: "verified" means tested on real hardware by the maintainer. "Expected" means it should work from how the desktop is built, but nobody has tested it yet. Reports are welcome, please use the Platform report issue template.
| Desktop and session | Hotkey | Text insertion | Overlay | Tray |
|---|---|---|---|---|
| GNOME 50, Wayland, x86_64 | verified (GlobalShortcuts portal) | verified (RemoteDesktop portal) | notification only, no layer-shell on GNOME | verified (needs AppIndicator extension) |
| GNOME, X11 | expected | expected (xdotool) | notification only | expected |
| KDE Plasma, Wayland | expected (GlobalShortcuts portal) | expected (RemoteDesktop portal) | expected (layer-shell) | expected |
| KDE Plasma, X11 | expected | expected (xdotool) | expected | expected |
| Hyprland, Wayland | expected (GlobalShortcuts portal) | expected (wtype) | expected (layer-shell) | expected |
| Sway, Wayland | expected, via compositor binding to --start and --stop |
expected (wtype) | expected (layer-shell) | expected |
Details, fallbacks and config snippets are in docs/platform-support.md. When no insertion method works, the text is copied to the clipboard and a notification is shown.
- No backend, no telemetry, no analytics, no accounts.
- Audio and text stay on your device in on-device mode.
- The only network traffic is: downloading model weights from Hugging Face (once, verified by SHA-256), and Gemini requests when you enable cloud mode.
- In cloud mode your audio is sent to Google's Gemini API using your own key. On the free tier, Google may use content to improve its products and human reviewers may read it (users in the EEA, UK and Switzerland get paid-tier terms). The app shows this warning: "Free-tier requests may be used by Google to improve its products."
- Your API key is stored in the platform secret store (libsecret on Linux, Android Keystore-backed storage on Android).
More in docs/PRIVACY.md.
Choose one engine in Settings. You can set up both.
- Open Settings, then Engine.
- Pick a model and tap Download.
- Wait for the download and verification to finish. Then dictate.
| Model | Download size | Recommended free RAM |
|---|---|---|
| Gemma 4 E4B (recommended) | 3.66 GB | about 6 GB |
| Gemma 4 E2B | 2.59 GB | about 4 GB |
The RAM figures are our own estimates. Weights come from the ungated litert-community repositories on Hugging Face (Apache-2.0), are downloaded at runtime, are checked against a SHA-256 hash, and are never part of this repository.
- Go to https://aistudio.google.com and sign in.
- Choose Get API key, then Create API key, and copy the key.
- Open Settings in Ramble Voice, paste the key, and tap Test key.
The app lists the models available to your key and always uses the newest
gemini-*-flash-lite generation.
Warning: free-tier requests may be used by Google to improve its products, and human reviewers may read them. Do not dictate anything you would not want Google to see.
On Linux you can also provide the key through the GEMINI_API_KEY environment
variable. It takes precedence over a stored key.
# Fish
set -Ux GEMINI_API_KEY your_key
# Bash
echo 'export GEMINI_API_KEY=your_key' >> ~/.bashrc
# Zsh
echo 'export GEMINI_API_KEY=your_key' >> ~/.zshrcFor the Flatpak, set it with an override:
flatpak override --user --env=GEMINI_API_KEY=your_key io.github.luohoa97.RambleVoiceNote that this stores the key in plain text in the override file.
Download files from the
Releases page and compare
them with SHA256SUMS.
Linux (Flatpak, x86_64 and aarch64):
flatpak install --user ramble-voice-x86_64.flatpak
flatpak run io.github.luohoa97.RambleVoiceAndroid: install the APK (ramble-voice-arm64-v8a.apk), then enable
Ramble Voice under Settings, System, Keyboard, On-screen keyboard. If a release
APK is named -debug-signed, it was not signed with the project's release key.
Optional host helper for the ydotool text insertion route:
packaging/linux/host-setup.sh (see docs/platform-support.md).
Ramble Voice needs Flutter 3.47.7.
Flutter's Linux embedding is GTK3 only, so Ramble Voice has its own GTK4 runner on the Flutter Embedder API. The Flutter engine binaries come from ardera/flutter-ci and are pinned by SHA-256.
Fedora:
sudo dnf install clang cmake ninja-build pkgconf-pkg-config gtk4-devel \
gtk4-layer-shell-devel pipewire-devel pulseaudio-libs-devel libsecret-devel \
libepoxy-develUbuntu 24.10 or newer:
sudo apt install clang cmake ninja-build pkg-config libgtk-4-dev \
libgtk4-layer-shell-dev libpipewire-0.3-dev libpulse-dev libsecret-1-dev \
libepoxy-devUbuntu 24.04 does not package gtk4-layer-shell. Build it from source (https://github.com/wmww/gtk4-layer-shell) or build without the overlay.
Then:
flutter pub get
linux/engine/fetch_engine.sh x64 release # or aarch64 (CMake also fetches it automatically)
flutter build linux --releaseThe bundle is in build/linux/<arch>/release/bundle/.
Requirements: Flutter 3.47.7, JDK 17, Android SDK 36.
flutter pub get
flutter build apk --release --split-per-abiThe arm64-v8a APK is in build/app/outputs/flutter-apk/.
See CONTRIBUTING.md, docs/ARCHITECTURE.md and the Code of Conduct. Report vulnerabilities as described in SECURITY.md.
Ramble Voice is licensed under GPL-3.0-or-later, see LICENSE and NOTICE for third-party components.
The Gemma model weights are not part of this project. They are downloaded at runtime from Hugging Face and are licensed by Google under the Apache License 2.0, not under the GPL.