Live captions for Linux. NowCaption shows what is being said in videos, calls, streams and games, or into your microphone, in a small window that floats over your desktop. It is made for people who are deaf or hard of hearing.
- Private. Everything runs on your computer, offline. No account, and audio never leaves it.
- Accurate. Punctuation and capital letters included. NowCaption's own speech model makes 2.7% word errors on clear speech, a third of upstream Live Captions' 8.1%, and in 1.2.0 about a fifth fewer than before on meetings, calls, accents and noisy speech (benchmark).
- Light. Runs on the CPU as speech arrives. No GPU needed.
flatpak install --user --from https://techlm77.github.io/NowCaption/nowcaption.flatpakrefUpdates then arrive with flatpak update or your software centre. Works on x86_64
and ARM64. You can also download a .flatpak file from the
latest release and run
flatpak install --user NowCaption-*.flatpak.
To uninstall: flatpak uninstall io.github.Techlm77.NowCaption.
AppImage: download NowCaption-*-x86_64.AppImage (or aarch64) from the
latest release, then run
chmod +x NowCaption-*.AppImage and ./NowCaption-*.AppImage. It runs on Ubuntu 22.04,
Debian 12 and anything newer: on systems older than 2024 it brings its own C library.
Gear Lever or AppImageUpdate can keep it up to date.
Captions start straight away from whatever your computer is playing. Drag them to move, drag an edge to resize. Right-click for the menu:
- Appearance: text size, background and text colour
- Position: floating, or docked along an edge of the screen; reset the position and size
- Transcript: open the transcript, keep history, or clear this session
- More options…: what to listen to, desktop audio or your microphone, and how new words appear
On a touchscreen, touch and hold the captions for the menu, and right-click NowCaption in your app launcher for Open transcript, More options and Quit.
Transcript. Open transcript shows everything captioned, in a window you can scroll, select and copy from, and search with Ctrl+F. Earlier sessions are listed under their date, and the current one updates as each line is finished. Copy all and Save give the whole transcript; Delete history removes it.
History. NowCaption keeps each session's transcript for 7 days, then deletes it,
so it never fills up your disk. It stays on your computer, readable only by you, in
~/.var/app/io.github.Techlm77.NowCaption/data/nowcaption/history (Flatpak) or
~/.local/share/nowcaption/history (AppImage). Untick Keep 7 days of history in
the menu to stop keeping it. Clear this session also removes the current session
from the history.
Text colour. Captions are white by default. If another colour is easier for you to read, pick one of the dots under Text colour: yellow, lime, cyan, orange, pink, lavender, brown or red. The captions change as soon as you click, so you can compare colours before closing the menu. NowCaption remembers your choice, like your text size, background and position, every time it starts and after updates.
New words. New words fade in by default. Under More options… they can appear at once instead, or arrive as pixels: Pixel assemble flies them together from left to right, Pixel dissolve switches them on bit by bit, Mosaic sharpen brings them into focus from big blocks, and Pixel rain drops them into place, each at a fast, normal or slow speed. A sample caption in that window shows each one as you pick it, and Apply uses your choices for the captions.
Motion. New lines glide up instead of jumping. The captions fade in when NowCaption starts and out when it quits, and "Listening…" gently pulses while NowCaption waits for speech. For no motion at all, turn animations off for your desktop (on Plasma, set Animation speed to Instant; on GNOME, turn on Reduce Animation) and NowCaption follows.
On KDE Plasma, Hyprland and Sway the captions float above all windows. On GNOME, press Alt+Space on the window and choose Always on Top.
Position. On KDE Plasma, Hyprland and Sway, the little screens under Position in the menu choose where the captions go. Floating, they stay wherever you drag them, over any window. Docked at the top, bottom, left or right, they take that strip of the screen like a panel, next to any panels already there, and maximised or snapped windows make room for them, so nothing is covered. Drag the edge between the captions and your windows to resize both together. Choosing another position slides the captions into their edge and out of the new one, and your windows move along with them. On KDE Plasma 5.27 and 6, while a fullscreen game or video is open on their screen, docked captions float above it and dock again when it closes; elsewhere, choose Floating for fullscreen apps. On KDE Plasma 6, a floating panel on the captions' edge stops floating while they're docked against it, as it does when a window touches it, and floats again when they move away or NowCaption quits. Plasma 5.27 has no way for apps to do that: there, turn off Floating for that panel in Edit Mode, or dock the captions on another edge. Floating captions keep their bottom edge in place when you change the text size.
KWin permission. On KDE Plasma, the Flatpak may talk to KWin, so docked captions can
float over fullscreen apps. That lets it load a KWin script: NowCaption loads a small one
while the captions are docked, to learn when a fullscreen app opens on their screen. If
you'd rather not allow it, run
flatpak override --user --no-talk-name=org.kde.KWin io.github.Techlm77.NowCaption;
docked captions then stay docked under fullscreen apps.
Plasma permission. The Flatpak may also talk to Plasma, to turn a floating panel's
floating off while docked captions touch it (and back on after). If you'd rather not allow
it, run flatpak override --user --no-talk-name=org.kde.plasmashell io.github.Techlm77.NowCaption;
the panel then keeps floating next to docked captions.
No captions? Make sure sound plays through your default output device. After switching to headphones, or after a problem, open More options… and press Restart.
First word missing? When speech is hard to make out, such as strong accents or unclear recordings, the first word after a pause is the one most often missed or misheard, because the speech model has nothing yet to go on. 1.2.0's speech model misses far fewer of them than before, but it still happens most there. Clear speech is rarely affected.
Needs GTK 4.12+ and libadwaita 1.5+ (Debian 13, Ubuntu 24.04, Fedora 40, Arch).
# Arch
sudo pacman -S python-gobject python-cairo gtk4 libadwaita gst-plugins-base gst-plugins-good gtk4-layer-shell
# Debian / Ubuntu
sudo apt install python3-gi python3-gi-cairo python3-venv gir1.2-gtk-4.0 gir1.2-adw-1 \
gir1.2-gstreamer-1.0 gir1.2-gst-plugins-base-1.0 gstreamer1.0-plugins-good gstreamer1.0-pulseaudio
# Fedora
sudo dnf install python3-gobject python3-cairo gtk4 libadwaita gstreamer1-plugins-base gstreamer1-plugins-good
./tools/run-dev.sh # downloads the models on first run
./tools/build-flatpak.sh # or build the FlatpakTests: PYTHONPATH=src .venv/bin/python tests/test_reliability.py (fast, no models);
the other files in tests/ say what they cover.
| Part | Model | License |
|---|---|---|
| Speech recognition | NowCaption's own, fine-tuned from NVIDIA's FastConformer streaming | CC-BY-4.0 |
| Punctuation and capital letters | Edge-Punct-Casing by frankyoujian | Apache-2.0 |
NowCaption's own code is MIT licensed. Everything else it ships with, and the speech its model was trained on, is listed in MODEL-LICENSES.md. How speech model v3 was trained, chosen and exported is in its model card.