Skip to content
urbanvcodePublic

About

Real-time subtitles and translation for any video in your browser, made locally on your own GPU. Private, offline, free. Whisper, Parakeet, SenseVoice + llama.cpp.

Topics

Resources

Stars

0 stars

Watchers

1 watching

Forks

Repository files navigation

Subtle icon

Subtle

Real-time subtitles and translation for any video in your browser, made on your own PC.
Free, private and offline: nothing you watch leaves your computer.

Latest release Windows, Linux, macOS Chrome, Edge, Brave, Opera, Vivaldi, Firefox NVIDIA GPU Install test MIT license

Subtle translating English speech into Spanish subtitles over a film, with the original line underneath
Real Subtle output: English speech turned into Spanish subtitles while the film plays, the original line underneath. Frame from Tears of Steel, (CC) Blender Foundation | mango.blender.org.

What it does

  • Subtitles for any video on any website: YouTube, streaming sites, anime, lectures, live streams.
  • Translates while you watch: Japanese, Chinese, Korean and ~95 other languages into English (or into any language with the Best engine).
  • Fast: a line appears about 0.2 s after the speaker stops.
  • Private: speech recognition and translation both run on your own graphics card. No account, no cloud, no API keys.
  • Light: uses no video memory until you start subtitles, and unloads by itself when you stop.
  • Easy: one-click install, a normal Windows app to manage models, and a browser extension with a start button.
  • Subtitles for any app: Discord, Zoom, games, VLC, a film on your PC. Subtle listens to what your speakers play and shows subtitles at the bottom of your screen.
  • Learning mode: the original line with furigana or romaji (Japanese), pinyin (Chinese) or romanization (Korean). Click a word to see what it means; the video waits for you.
  • Save subtitles as .srt or .vtt files that work in VLC and other players.
  • Catch me up: missed something? A short summary of the last few minutes appears on the video.
  • Yours to arrange: drag the subtitles anywhere on the video (remembered per site), hide them or show the last line again with a shortcut, and pick colours, box, outline and size.

Examples

Real lines from Subtle (English film audio, Best engine, Spanish subtitles), with how long each took after the speaker stopped:

Heard (English) Subtitle (Spanish) Delay
Look, Celia, we have to follow our passion. Mira, Celia, tenemos que seguir nuestra pasión. 0.2 s
You have your robotics and I just want to be awesome in space. Tienes tus robóticos y yo solo quiero ser genial en el espacio. 0.2 s
Why don't you just admit that you're freaked out by my robot hand? ¿Por qué no admites que te asusta mi mano robótica? 0.2 s
It's not my fault, you know? No es mi culpa, ¿sabes? 0.2 s
Of course you're ready. You're a rock star. Claro que estás listo. Eres una rock star. 0.2 s

Screenshots

Subtle app, Home page: status, subtitle delay and memory use Browser extension popup: start button, languages, Fast or Best engine
Subtle app. Status, subtitle delay in milliseconds, and exactly how much memory each part uses. Browser extension. Five tabs: start and tools, languages, display, style and more.
Models page with recommended downloads for your graphics card Subtitle style options with live preview
Models. Every model with accuracy, speed and memory at a glance; recommended ones for your graphics card download with one click. Subtitle style. Colours, box, outline and size, with a live preview.

Shortcuts

Keys What it does
Alt+Shift+S Start or stop subtitles on the current tab
Alt+Shift+H Hide or show the subtitles
Alt+Shift+L Show the last line again
Alt+Shift+U Catch me up: a short summary of the last few minutes
Drag a subtitle Move the subtitles (remembered for that website); double-click puts them back

You can change the keys at chrome://extensions/shortcuts.

Subtitles for any app

For Discord, Zoom, games, VLC or a film on your PC (not in the browser): open the Subtle window, pick the languages under Subtitles for any app on the Home page and press Start. Subtle listens to what your speakers play and shows the subtitles at the bottom of your screen, on top of every window (clicks go straight through them). Size, position, box and screen can be changed right there.

  • Only one program: under Listen to, pick Only Discord (or any program that is playing sound) and Subtle hears just that program; music or videos elsewhere are left out. (Windows 10 version 2004 or newer. Refresh this list shows programs that started playing later.)
  • Games in exclusive fullscreen draw over everything: use borderless or windowed mode.
  • If a browser tab is also subtitled, both get subtitles: stop one of them.

How it works

video in a browser tab ─► Subtle extension ─► Subtle app on your PC ─► subtitles over the video
                                              ├─ Fast: Whisper translates straight to English
                                              └─ Best: a speech model writes the speech down, a built-in
                                                       translator model turns it into any language

What you need

Minimum Recommended
System Windows 10 or 11 (64-bit), Linux (64-bit), or macOS on an Apple M-series Mac (Experimental) Windows 11
Browser Any Chromium browser: Chrome, Edge, Brave, Opera / Opera GX, Vivaldi, Arc and others. Firefox works too (Experimental, see below). Not Safari
Graphics card NVIDIA GTX 16-series / RTX 20-series or newer, 6 GB video memory NVIDIA RTX 3060 / 4060 or better, 8 GB+
Memory (RAM) 8 GB 16 GB
Disk 8 GB + models (2-10 GB)
Internet only for the install and model downloads
  • Older NVIDIA cards (GTX 10-series) work, but more slowly.
  • AMD and Intel graphics (Experimental):
    • Linux + AMD: speech and translation both run on the graphics card (ROCm).
    • Windows + AMD or Intel, or Linux + Intel: the translator runs on the graphics card (Vulkan), but speech recognition runs on the processor. Use the small or medium speech model and the Fast engine.
  • Apple M-series Macs (Experimental): speech runs on the Apple GPU where it can, and the translator on Metal.
  • No graphics card: Subtle still works on the processor, slowly. Use small models and Fast.
  • Keep your NVIDIA driver reasonably up to date. The installer tells you if it's too old.

Experimental means it is installed and tested automatically on clean Linux and Mac machines with every change (on the Mac, speech runs on its Apple GPU), but not yet used day to day on that hardware. If you have one, please try it and tell us how it went.

Install (about 10 minutes)

Windows

  1. Download Subtle from Assets:

    • Subtle-Setup-<version>.exe: double-click it. It asks where to put Subtle and installs it.
    • or Subtle-<version>.zip: unzip it somewhere with free space (e.g. D:\Subtle) and double-click Install Subtle.bat.
    • Want only the basics? Take Subtle-Lite-<version>.zip: just subtitles and translation for the browser, no extra features, about 260 MB smaller. (The full version can hide its extras too: Settings → Simple mode.)

    Either way, everything goes inside that one folder (a private Python, the translator engine), plus a Subtle shortcut on the desktop and in the Start menu. Nothing is installed system-wide.

  2. The Subtle window opens on its Help page. Follow the three steps there:

    • Download the recommended models. It picks them for your graphics card, one click.
    • Add the browser extension: open chrome://extensions in your browser (this address works in Chrome, Edge, Brave, Opera and Vivaldi), turn on Developer mode (top right; in Edge it's on the left), click Load unpacked, and choose the extension folder inside Subtle. (The Help page has a button that copies the exact path.)
  3. Open a video, click the Subtle icon in the browser toolbar, then Start subtitles on this tab (or press Alt+Shift+S).

From then on, just start Subtle from the desktop shortcut. Closing its window keeps it running in the tray (bottom right, next to the clock). It uses no video memory until you turn subtitles on; click the tray icon to open the window again. Want it always ready? Settings → Start Subtle when Windows starts.

Windows SmartScreen or your antivirus may ask about Subtle-Setup.exe or Install Subtle.bat the first time, because it's a new, unsigned download. The installer only downloads from GitHub (the uv Python manager and llama.cpp, both checksum-checked), PyTorch and PyPI (pinned package versions). Models come from Hugging Face at pinned versions and are checksum-checked too.

Linux and macOS (Apple M-series)

  1. Download Subtle-<version>.zip (or the Lite zip) and unzip it, e.g. into ~/Subtle.
  2. In a terminal, in that folder: ./install.sh (it detects your graphics card; to choose yourself, use ./install.sh --gpu nvidia, or nvidia-old, amd, intel or cpu).
  3. Start Subtle: Linux: from your applications menu, or ./subtle.sh. macOS: double-click Subtle.command in the Subtle folder.
  4. Add the browser extension as on Windows (step 2 above).

Linux needs curl and, for the tray icon, an AppIndicator-capable desktop. Subtitles for any app uses PulseAudio or PipeWire. On macOS the first start may need System Settings → Privacy & Security → Open anyway, because the download is not signed.

Firefox (Experimental)

Firefox can't record a whole tab like Chromium browsers can, so it subtitles the video on the page itself. That works on YouTube and most players, but not on protected (DRM) sites such as Netflix.

Tip: no extension needed. The Any app page of the Subtle window subtitles whatever your speakers play, so it works in Firefox on every site, usually protected ones too, because it listens to the sound and not to the page. Choose Only Firefox under What to listen to to leave other sounds out, and drag the subtitles over the video. The extension adds learning mode, saving subtitles and Catch me up.

  1. Open about:debugging#/runtime/this-firefox and click Load Temporary Add-on….
  2. Choose manifest.json in the extension-firefox folder inside Subtle (or the Subtle-Firefox-extension-<version>.zip from the release).
  3. Firefox removes temporary add-ons when it closes; a signed version for addons.mozilla.org is coming.

Which models should I use?

The Models page and the Help page recommend a set for your graphics card automatically.

Your graphics card Speech (Whisper) Translator (for Best) Video memory in use
4 GB medium - (use Fast) ~2 GB
6 GB large-v3 + turbo Qwen3.5 2B Fast ~4.1 GB · Best ~4.3 GB
8-11 GB (recommended) large-v3 + turbo Qwen3 4B Instruct 2507 Fast ~4.1 GB · Best ~5.7 GB
12-15 GB large-v3 + turbo Qwen3.5 9B Best ~8.7 GB
16 GB+ large-v3 + turbo Gemma 4 12B Best ~10 GB

And by spoken language (Subtle picks these by itself when they are installed):

Spoken language Fast Best
Japanese Whisper large-v3 Parakeet 0.6B + translator
Chinese Whisper large-v3 SenseVoice Small + translator
Korean and every other language Whisper large-v3 Whisper large-v3 turbo + translator

These come from our own benchmark (benchmarks/, 120 test sentences per language, cut into lines the way Subtle cuts them; wrong characters, lower is better):

Japanese Chinese Korean Speech step per line
Whisper large-v3 6.1 % 10.5 % 5.4 % ~95 ms
Whisper large-v3 turbo 7.0 % 11.7 % 6.3 % ~40 ms
Parakeet 0.6B (Japanese) 6.2 % - - ~8 ms
SenseVoice Small 11.7 % 9.3 % 9.4 % ~6 ms

The test sentences are read aloud (FLEURS), not film or anime dialogue, so treat the numbers as a guide. With the Best engine the translator adds about 60 ms per line.

  • Fast (Whisper only): straight to English, the least memory, about 70 ms per line.
  • Best (Whisper + translator): more natural lines, uses the previous lines as context, and works for any subtitle language, at about 95 ms per line.
  • Japanese with Best: download Parakeet 0.6B (Japanese) on the Models page and Subtle uses it by itself. It reads a whole line in one pass (about 8 ms instead of 43 ms for the speech step) and was a little more accurate than Whisper turbo on test sentences. Any other speech model can still be picked in the popup at any time.
  • Chinese with Best: download SenseVoice Small (Chinese) and Subtle uses it by itself. It was clearly more accurate than Whisper turbo on Chinese test sentences (9.3 % against 11.7 % wrong characters) and takes about 6 ms instead of 40 ms for the speech step.
  • Parakeet and SenseVoice only write speech down; they can't translate. So they are used where the translator follows (Best). With Fast, Subtle keeps Whisper large-v3, which translates by itself.
  • Languages → Speech model for each language changes what "Auto" means for Chinese, Japanese, Korean and all other languages. A model picked in the popup still wins.
  • You can add your own models. Put a Whisper model folder into models\whisper\, or any instruct .gguf file into models\llm\, then press Rescan. Models you already downloaded with LM Studio are found automatically, so there's no need to download them twice.

How Subtle saves memory

Subtle only uses memory while you're actually watching something with subtitles:

State Video memory (VRAM) Memory (RAM)
Waiting in the tray (window open or closed) 0 ~40 MB
Fast (Whisper large-v3) ~4.1 GB ~0.1 GB
Best (Whisper turbo + Qwen3 4B) ~5.7 GB ~0.15 GB
Best, Japanese (Parakeet + Qwen3 4B) ~5.0 GB ~0.15 GB
Best, Chinese (SenseVoice + Qwen3 4B) ~4.8 GB ~0.15 GB
  • Nothing is loaded until a tab starts subtitles. Starting takes about 3-5 seconds; the sentence being spoken meanwhile still gets its subtitle.
  • The Subtle window is a plain Windows window (no browser inside), about 15 MB.
  • Starting the speech engine takes about 1.5 GB of RAM for a moment. A few seconds after a model has loaded, the engine hands that start-up memory back to Windows and keeps about 50 MB in use.
  • The speech model is loaded into exactly-sized blocks of video memory, without the gaps a normal load leaves behind (0.3-0.5 GB less).
  • Five minutes after the last subtitle everything is unloaded (Settings → Memory and speed: 1-30 minutes, or keep models loaded for instant starts). Free memory now on the Home page or in the tray menu does it at once.
  • Only one speech model is in memory at a time, and the translator only runs for the Best engine.
  • The translator is set up for subtitles: a small context and a compact 8-bit cache. That's about 25% less video memory than a general chat setup.
  • Previews (the line updating while someone speaks) are the main GPU work. Settings → Memory and speed can make them less frequent, or turn them off.
  • The Home page shows exactly how much video memory and RAM each part of Subtle is using.

Privacy

  • Nothing you watch leaves your PC. The extension only talks to the Subtle app on your PC (127.0.0.1). The Subtle window and its connection only work on this PC.
  • Subtitles live in memory. No subtitle text or audio is written to disk, unless you turn on Save transcripts in the extension popup (one text file per session in transcripts\).
  • Speaker voiceprints (Identify speakers, off by default) are kept in memory only.
  • Error log: logs\errors.log keeps errors only, never subtitle text (about 1 MB at most).
  • Catch me up summaries are made by the translator on your PC, from subtitles kept in memory.
  • Subtitles for any app record what your speakers play only while it is switched on, and send it only to the speech engine on your PC. Nothing is saved.
  • Learning mode word lookups are answered by the translator on your PC.
  • Saved subtitle files (.srt / .vtt) are only made when you press Save subtitles.
  • The internet is used only by Install Subtle.bat, the Download buttons on the Models page, and a daily check for new versions: Subtle asks GitHub for the newest version number and sends nothing else (turn it off in Settings → Tell me about new versions).

Troubleshooting

Problem What to do
The popup says "Subtle isn't running" Start Subtle from the desktop or Start-menu shortcut. Look for its icon in the tray.
The first subtitle takes a few seconds Normal: models load when you start (3-5 s). After that lines take ~0.1 s. To skip it, turn on Settings → Keep models loaded.
"No translator model" Download one on the Models page (Best needs a translator; Fast doesn't).
Netflix, Prime Video, Disney+... show nothing with Auto-start DRM video can't be read from the page. Click the Subtle icon (or press Alt+Shift+S) to use tab capture instead.
Out of video memory / games stutter Use Fast, pick a smaller translator, set Previews to "every second", or press Free memory now.
Installer: "NVIDIA driver too old" Update it from nvidia.com/drivers or the NVIDIA app, then run Install Subtle.bat again.
Anything else Subtle window → Settings → Show activity log, and the install log in logs\install.log.

Update: when a new version is out, the Subtle window and the extension popup show Update now. One click downloads it (checked against GitHub's checksum), keeps your models and settings, and restarts Subtle; if anything goes wrong, the old version is put back. The browser extension reloads itself afterwards. Coming from 0.5-0.8, update by hand once: replace the Subtle files with the new version (keep models\, runtime\ and settings.json), run Install Subtle.bat again, and reload the extension in chrome://extensions.

Uninstall: quit Subtle (tray icon → Quit), untick "Start with Windows" if you turned it on, delete the Subtle folder and the two shortcuts, and remove the extension from the browser.


Popup settings (browser extension)

Setting What it does
Spoken → Subtitles in Source language (or auto-detect) and subtitle language
Fast / Best See above
Speech model Auto picks large-v3 for Fast and turbo for Best (with Best: Parakeet for Japanese and SenseVoice for Chinese, if installed; changeable in the Subtle window's Settings)
Translator Any installed translator; the default is set on the Languages page of the Subtle window
Show original-language line With Best, shows the transcribed original under the translation
Identify speakers Gives each voice its own colour (26 MB voiceprint model, Models page)
Auto-start when a video plays Subtitles start by themselves on any tab where a video plays with sound (not for DRM video)
Catch me up Shows a short summary of the last few minutes on the video (uses the translator model)
Save subtitles Saves this video's subtitles as .srt or .vtt, with the original line too if Original-language line is on
Learning mode Shows the original line with readings (furigana or romaji, pinyin, romanization). Click a word for its meaning. Needs the Best engine

For developers

Part File Job
Supervisor server/supervisor.py Always-on app (no torch, ~40 MB with its window): tray, relays tabs to the engine, starts and stops the engine and translator, idle unload
Window server/ui_qt.py, server/tray.py The Subtle window (Qt; ui.py is a Tkinter fallback) and tray icon; they talk to the supervisor in-process
Engine server/server.py, engines.py, whisper_core.py, parakeet_core.py, sensevoice_core.py VAD segmentation, previews and finished lines. Whisper, Parakeet and SenseVoice are loaded and run by Subtle's own code (no transformers, NeMo or FunASR): CUDA graphs on the GPU, eager on the CPU. Started on demand by the supervisor
Translator server/llm_runner.py, llm_client.py Runs bin/llama/llama-server.exe; any OpenAI-compatible server works as "external"
Models server/modelstore.py, downloads.py, models/catalog.json Finds models (models/, HF cache, LM Studio); pinned, checksummed downloads
Extension extension/ Captures tab audio and draws the overlay
Installer Install Subtle.bat, installer/ uv + private Python, pinned packages (requirements.txt), pinned llama.cpp build (versions.json)

Testing:

python server\smoke_test.py
python server\supervisor.py --no-window --no-tray --port 8799
python server\test_client.py clip.f32 --url ws://127.0.0.1:8799 --language ja
python server\test_client.py clip.f32 --url ws://127.0.0.1:8799 --mode llm --target fr

Through the supervisor, the engine and translator start by themselves. server.py --verbose runs the engine alone (it then accepts extension origins on port 8765 itself). clip.f32 is 16 kHz mono float32 PCM (ffmpeg -i clip.mp4 -ac 1 -ar 16000 -f f32le clip.f32).

Licenses

Subtle's code is MIT-licensed (see LICENSE). Models are downloaded from their authors on Hugging Face under their own licenses, shown on the Models page:

  • Whisper large-v3, medium, small (OpenAI): Apache-2.0; Whisper large-v3 turbo (OpenAI): MIT
  • Parakeet-TDT-CTC 0.6B ja (NVIDIA): CC-BY-4.0
  • SenseVoice Small (FunAudioLLM, Alibaba): FunASR Model License (see its model page)
  • Kotoba-Whisper v2.0 and Bilingual v1.0 (Kotoba Technologies): Apache-2.0
  • Anime-Whisper (litagin): MIT
  • Qwen3 and Qwen3.5 (Alibaba Cloud): Apache-2.0
  • Gemma 4 (Google): Apache-2.0
  • WeSpeaker ResNet34 voiceprints (WeSpeaker): CC-BY-4.0

Subtle also uses llama.cpp (MIT) for the translator, Silero VAD (MIT) to hear when someone speaks, and PyTorch (BSD-3-Clause). The benchmark numbers above were measured on the FLEURS test set (Google, CC-BY-4.0); scripts and full results are in benchmarks/.

The demo image and example lines use Tears of Steel by the Blender Foundation (mango.blender.org), CC BY 3.0.

Changelog

See CHANGELOG.md.

About

Real-time subtitles and translation for any video in your browser, made locally on your own GPU. Private, offline, free. Whisper, Parakeet, SenseVoice + llama.cpp.

Topics

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages