Live voice dictation for Linux on Wayland.
Press a key, talk, and your words are typed into the app you're using.
|
|
| First run: one page, three things | Settings, with a check of what your system provides |
Status: alpha, v0.1.0 released 2026-11-11. Works on KDE Plasma 6 and GNOME 50.
flatpak install --user https://gabus.github.io/murmur/murmur.flatpakrefOr click Install on the project page to open it in Discover or GNOME Software. Check what a distro must have first.
| Like Mac dictation | One key to start, the same key to stop. |
| Private | Runs on your computer. Nothing you say is sent anywhere. |
| Any app | Types wherever your cursor is. |
| Quiet | No popups, no sounds. Keep typing while it listens. |
| Light | Mic and model are released when you're not dictating. |
| Easy install | One Flatpak. No sudo, no setup scripts. |
flowchart LR
key(["Press<br/>Super+H"]) --> talk["Talk"]
talk --> rec["Recognized on<br/>your computer"]
rec -->|"about 1 s behind your voice"| app(["Typed into<br/>your app"])
Typed words are never changed afterwards, so nothing breaks if you move the cursor or type yourself.
Murmur doesn't depend on a particular desktop, only on these parts. Settings โ Your system checks each of them on your computer and says what's missing.
| Part | Murmur uses it for | Provided by, for example |
|---|---|---|
| Wayland compositor | Everything: X11 isn't supported | KWin, Mutter, Hyprland |
| xdg-desktop-portal 1.18+ | Talking to the desktop | the xdg-desktop-portal package |
| GlobalShortcuts portal | The start/stop key | xdg-desktop-portal-kde, -gnome 48+, -hyprland |
| RemoteDesktop portal | Permission to type | xdg-desktop-portal-kde, -gnome |
| Input emulation (libei) | Typing into apps | KWin 6.1+, Mutter 46+ |
| Clipboard portal | Characters not on your keyboard layout | xdg-desktop-portal-kde, -gnome |
| PipeWire or PulseAudio | The microphone | pipewire-pulse, pulseaudio |
| x86_64 computer | Running the app | no ARM build yet |
| System tray host (optional) | The tray icon | Plasma's panel, GNOME's AppIndicator extension, waybar |
| Notification portal (optional) | Error messages | xdg-desktop-portal-kde, -gnome |
| Background portal (optional) | Start at login | xdg-desktop-portal-kde, -gnome |
Tested on: KDE Plasma 6.7 and 6.8 (everything found) and GNOME 50 (everything but the tray until the AppIndicator extension is installed). On GNOME the suggested key is Super+Alt+H, it's changed in GNOME Settings, and the permission dialog needs Allow Remote Interaction switched on.
About this project: Murmur's code, tests and packaging were written with AI assistance (Anthropic's Claude), directed and reviewed by the maintainer. It's distributed from its own Flatpak repo, not Flathub.
| You | Murmur |
|---|---|
| Press the shortcut | The icon turns red and the mic opens. Nothing you say is lost, even on the first press. |
| Speak | Words appear at the cursor about a second behind your voice. |
| Pause | The sentence is finished, with its final punctuation. |
| Type, click, switch apps | Keeps listening. The next words go wherever the cursor is. |
| Press the shortcut again | Finishes the last words and closes the mic. |
| Click the tray icon | Same as the shortcut. Right-click for the menu. |
 |
  |
 |
 |
|---|---|---|---|
| Idle | Listening grows with your voice |
Loading or waiting waiting: let go of Ctrl/Alt/Super |
Problem hover for details |
| Model | Download | Languages | |
|---|---|---|---|
| Parakeet v3 | 490 MB | 25, detected automatically | Default. Most accurate |
| Moonshine base | 110 MB | English | Smaller and lighter |
Both add punctuation and capitals. The model downloads during first-run setup; switch any time in Settings or the tray menu.
Parakeet understands:
| ๐ง๐ฌ Bulgarian | ๐ญ๐ท Croatian | ๐จ๐ฟ Czech | ๐ฉ๐ฐ Danish | ๐ณ๐ฑ Dutch |
| ๐ฌ๐ง English | ๐ช๐ช Estonian | ๐ซ๐ฎ Finnish | ๐ซ๐ท French | ๐ฉ๐ช German |
| ๐ฌ๐ท Greek | ๐ญ๐บ Hungarian | ๐ฎ๐น Italian | ๐ฑ๐ป Latvian | ๐ฑ๐น Lithuanian |
| ๐ฒ๐น Maltese | ๐ต๐ฑ Polish | ๐ต๐น Portuguese | ๐ท๐ด Romanian | ๐ท๐บ Russian |
| ๐ธ๐ฐ Slovak | ๐ธ๐ฎ Slovenian | ๐ช๐ธ Spanish | ๐ธ๐ช Swedish | ๐บ๐ฆ Ukrainian |
For contributors: the rules the code follows (decision log, benchmarks). Each section expands.
Principles
- The user is in control. Listening starts and stops only on the shortcut or a tray click. Murmur never stops on its own (no silence timeout) and never starts on its own.
- Nothing is lost. Audio is recorded from the moment the key is pressed, even while the model is still loading, and is transcribed once the model is ready.
- No interruptions. No popups, preview windows, sounds or notifications in normal use. Notifications appear only for errors that need action (no mic, permission revoked, model missing).
- Typing and talking mix freely. Keyboard and mouse work normally while listening. Each word is inserted at the cursor, wherever focus is at that moment.
- Readable output by default. The default model must produce punctuation and capitals.
- Setup happens once. Shortcut, typing permission and model download are handled in first-run setup, never during the first dictation.
First run and Settings
One setup window, all on one page (D-009):
| Row | What happens |
|---|---|
| Shortcut | The portal's dialog, where the user picks a key. Murmur suggests Super+H, or Super+Alt+H on GNOME, where Super+H hides the window (D-016). Changeโฆ reopens the desktop's dialog where the portal can (v2). GNOME's portal is v1, so there it's changed in GNOME Settings. |
| Typing permission | Allowโฆ opens the RemoteDesktop portal dialog. The permission is remembered and never asked again. |
| Speech model | The default starts downloading as soon as the window opens, with a progress bar and an option to pick another. |
| Start at login | Checkbox, off by default. |
Then the window closes and Murmur lives in the tray. Settingsโฆ in the tray menu opens a window with the same look (D-010):
- shortcut, microphone, model, the pause that ends a sentence, idle unload time, start at login;
- Your system (D-019, D-022): each part from What a distro must have, with what Murmur uses it for and a green tick, or a red cross and what to install. Checked read-only each time the window opens (no dialogs). Input emulation is known only once Murmur connects to type, so until typing is allowed it shows as pending.
Real-time typing
Words are typed about 1 second behind your voice, and once typed they are never changed. The delay is deliberate: it gives the model time to settle, so most corrections ("I scream" โ "ice cream") happen before a word is typed.
| Step | Detail |
|---|---|
| Re-check | Every ~0.5 s the offline model re-transcribes the sentence so far. |
| Commit | A word is typed once two consecutive runs agree on it (the "local agreement" policy from whisper_streaming). |
| Punctuation | Comes from the model itself. |
| Sentence end | The audio buffer is trimmed at each finished sentence. On a pause, or when the user stops, the remaining words and final punctuation are typed. |
| Cost | CPU while speaking, nothing while idle. The delay is configurable (default ~1 s). |
Difference from macOS: macOS shows provisional words and can still correct them afterwards. Murmur catches most corrections during the 1 s wait; anything the model changes its mind about later stays as typed.
Known limitations of keystroke output
Murmur types by sending keystrokes through the RemoteDesktop portal. That works in every app but can't be fully seamless:
| Limitation | What happens |
|---|---|
| Collision | Each word is one burst (under 1 ms), so your own keystrokes can land between Murmur's words but not inside them. |
| Layout | Characters missing from the active layout (e.g. ฤ
on US) are pasted through the clipboard, which is restored afterwards. A clipboard manager may still record them. |
| Held modifiers | With Ctrl/Alt/Super held, letters would act as shortcuts. Murmur sees held, latched and locked modifiers through EIS and waits until they're released. |
| No context | Murmur can't read the text around the cursor, so spacing and capitals after your own edits are a guess. |
| Password fields | Murmur can't detect them. |
| No corrections | Words are typed once. Earlier words can't be fixed when later context arrives. |
An input-method engine (Fcitx5 on KDE, IBus on GNOME) would solve all of these: text is committed in one piece through the text-input protocol, the field type and surrounding text are known, and live text can be shown underlined in the field. See the open questions in TODO.md.
Functional requirements
| Area | Requirement |
|---|---|
| Hotkey | Global toggle shortcut via the org.freedesktop.portal.GlobalShortcuts portal. Rebindable through the desktop's settings. |
| Audio | Capture the default PipeWire/Pulse source at 16 kHz mono. The stream is open only while listening. Optional source picker. |
| VAD | Silero VAD for end-of-phrase detection only. Never stops listening. |
| ASR | sherpa-onnx. The offline model re-runs every ~0.5 s on the current sentence; words two runs agree on are typed. |
| Output | Keystrokes through the org.freedesktop.portal.RemoteDesktop portal's EIS channel (ConnectToEIS, libei), resolved through the compositor's XKB keymap and active layout, with Murmur pressing Shift/AltGr itself (D-005). Words wait while held, latched or locked modifiers could change what's typed (D-007); after ~2 s the tray shows Waiting. If typing is denied, Murmur asks again only on the next dictation. A word with characters missing from the active layout is pasted whole (Ctrl+V; paste_keys = "ctrl+shift+v" in the config for terminals), then the previous clipboard is restored and served by Murmur until something else is copied (D-011). Every typed word is followed by one space (D-002). |
| Indicator | StatusNotifierItem tray icon: idle, loading, listening (follows the mic level), waiting, error. Menu: start/stop, model picker, settings, quit. |
| Models | Download on demand into $XDG_DATA_HOME/murmur/models with checksum verification. Size shown before download. Delete from the menu. |
| Resources | Idle: no audio stream, no model in memory, ~0 % CPU. The model unloads after a configurable idle time (default 10 min). |
| Autostart | Optional start at login via the org.freedesktop.portal.Background portal. |
| Config | One TOML file, $XDG_CONFIG_HOME/murmur/config.toml: model, pause_secs, idle_unload_minutes, microphone, autostart, paste_keys, setup_done. Most are also in Settings. |
Flatpak permissions
| Permission | For |
|---|---|
--socket=wayland |
The setup and settings windows |
--socket=pulseaudio |
The microphone |
--share=network |
The model download only |
--talk-name=org.kde.StatusNotifierWatcher |
The tray icon |
Shortcut, typing, clipboard, notifications and autostart go through portals and need no
permission. Never requested: --device=all, --filesystem=host, X11. Never used:
/dev/uinput, evdev, dotool/ydotool, the input group or udev rules: they would break the
one-click install.
flowchart LR
subgraph sandbox["Flatpak sandbox"]
subgraph daemon["murmur daemon (tokio)"]
app["app.rs<br/>Idle โ Listening"]
audio["audio thread<br/>cpal, 16 kHz"]
asr["ASR worker thread<br/>sherpa-onnx"]
out["output<br/>EIS typing"]
tray["tray<br/>ksni"]
end
ui["window process<br/>murmur --ui (Slint)"]
end
portal[("xdg-desktop-portal")]
comp[("compositor")]
pw[("PipeWire")]
pw --> audio
audio --> asr
asr -->|words| app
app --> out
app <-->|JSON lines| ui
app <-->|shortcut, permission| portal
out -->|keystrokes over EIS| comp
tray <-->|D-Bus| panel[("panel")]
The daemon is a tokio app. Blocking work runs on dedicated threads: audio capture, and the ASR worker that owns the model. The windows run in a separate short-lived process, so the daemon never loads a GUI toolkit while idle.
| Concern | Choice |
|---|---|
| Language | Rust (stable, edition 2024), tokio |
| Portals | ashpd: GlobalShortcuts, RemoteDesktop, Clipboard, Notification, Background |
| Typing | libei protocol via reis, keymaps via xkbcommon |
| Tray | ksni (StatusNotifierItem, no Qt/GTK) |
| Windows | Slint, software renderer: no GPU permission (D-001) |
| Audio | cpal with its pure-Rust PulseAudio client (PipeWire serves it too) |
| ASR + VAD | sherpa-onnx via the official sherpa-onnx crate, ONNX Runtime on the CPU. Dev builds use sherpa-onnx's prebuilt static libs (sherpa-static, default); the Flatpak builds ONNX Runtime and the sherpa-onnx C API from source as shared libs (sherpa-shared, D-014) |
| Config / download / logs | serde + toml ยท reqwest + rustls + sha2 ยท tracing |
| Packaging | Flatpak on org.freedesktop.Platform 26.08, built offline from source (docs/packaging.md) |
| App ID | io.github.gabus.Murmur |
Source tree
src/
main.rs # startup, wiring; `--ui` runs the window process, `--check` the system check
app.rs # state machine Idle โ Listening; window state and requests
instance.rs # single instance (D-Bus name), Activate/Toggle methods
hotkey.rs # GlobalShortcuts portal (bind, change, activations)
output/ # typing: RemoteDesktop portal โ EIS (libei), keymap planning,
# modifier hold-back, clipboard paste fallback
audio.rs # mic capture (cpal), 16 kHz mono, open only while listening
agc.rs # gain control for quiet microphones
vad.rs # Silero VAD (end of phrase only)
asr/ # re-check engine (Parakeet, Moonshine), local-agreement
# committer, worker thread
models.rs # registry (URLs, sizes, sha256), engine loading
model_manager.rs # active model, downloads, config persistence
download.rs # streaming download + sha256 + atomic extract
tray.rs, icons.rs# StatusNotifierItem tray, drawn state icons
notify.rs # error notifications (Notification portal)
autostart.rs # Background portal
syscheck.rs # probes what the host provides (Settings โ Your system)
portal.rs # shared portal helpers
config.rs # $XDG_CONFIG_HOME/murmur/config.toml
ui/ # Slint windows (setup, settings), daemonโUI protocol,
# checks.rs: the Your system checklist
ui_host.rs # spawns and drives the window process
Nothing is installed on the host: build tools live in a toolbox
(murmur-dev) and in Flatpak. The toolbox shares the home directory, Wayland session, D-Bus and
PipeWire, so Murmur run from it reaches the portals and the mic.
T="toolbox run -c murmur-dev"
$T cargo run # the daemon (first run opens the setup window)
$T cargo run -- --check # what this system provides
bash .claude/skills/murmur-gate/gate.sh # fmt + clippy -D warnings + tests
flatpak/install-local.sh # Flatpak: offline build into ./repo, install from "murmur-local"One-time setup (Fedora)
toolbox create murmur-dev
# Rust + C++ toolchain
toolbox run -c murmur-dev sudo dnf install -y rust cargo rustfmt clippy gcc-c++ cmake clang pkg-config \
git alsa-lib-devel pulseaudio-libs-devel libxkbcommon-devel wayland-devel fontconfig-devel \
python3-aiohttp python3-tomlkit # last two: flatpak-cargo-generator
# headless compositor tests (nested KWin runs from the host; GNOME runs in the toolbox)
toolbox run -c murmur-dev sudo dnf install -y mutter gnome-shell dbus-daemon \
gsettings-desktop-schemas dconf systemd libei-utils python3-gobject-base wev
# GNOME portal tests (tests/portal_headless_gnome.rs; gjs + GTK 4 come with gnome-shell)
toolbox run -c murmur-dev sudo dnf install -y xdg-desktop-portal xdg-desktop-portal-gnome \
gnome-control-center wl-clipboard wayland-utils
# Flatpak builds
flatpak install flathub org.flatpak.Builder \
org.freedesktop.Sdk//26.08 org.freedesktop.Sdk.Extension.rust-stable//26.08Delete the toolbox any time with toolbox rm -f murmur-dev.
Tests and tools
$T cargo test # unit tests
# integration tests in invisible compositors (never touch the real desktop);
# the nested-KWin ones need the spike test pad: (cd spike && $T cargo build)
$T cargo test --test output_nested_kwin -- --ignored
$T cargo test --test portal_nested_kde -- --ignored # real KDE portal: dialog, token, paste
$T cargo test --test portal_headless_gnome -- --ignored # real GNOME portals, headless
$T cargo test --test ui_nested_kwin -- --ignored # Settings window size, 100 % and 150 %
$T cargo test --features dev-harness --test e2e_nested_kwin -- --ignored # needs models
$T cargo run --release --example simulate -- --model parakeet-v3 file.wav # engine on WAVs
$T cargo run --example ui_shots -- /tmp/shots # window screenshots
$T flatpak/update-cargo-sources.sh # after every Cargo.lock changeThe dev-harness cargo feature (never shipped) adds MURMUR_DEV_AUDIO=<wav> (a WAV instead of
the mic), MURMUR_DEV_EIS=kwin-direct, MURMUR_DEV_NO_HOTKEY and MURMUR_DEV_AUTORUN.
The first Flatpak build compiles ONNX Runtime and sherpa-onnx (about 10 minutes on 16 threads);
flatpak-builder caches them in .flatpak-builder/. CI builds the same bundle on every push
(.github/workflows/flatpak.yml). Details and Flathub notes: docs/packaging.md. Cutting a release:
docs/release.md.
GPL-3.0-or-later (LICENSE, D-013). The speech models download separately and carry their own licenses: Parakeet CC-BY-4.0, Moonshine English MIT.