EchoSub

Live, translated captions for anything your PC plays.
Videos, streams, calls, games — heard, transcribed and translated on your own GPU.

---

What is EchoSub?

EchoSub listens to whatever comes out of your speakers, recognizes the speech in almost 100 languages, translates it into the language you choose, and shows it in a transparent, always-on-top caption box. It works with any app — a YouTube video, a Twitch stream, a Discord or Teams call, a game, a movie — because it captures the audio after it leaves the app, not inside it.

Everything runs locally: speech recognition and translation happen on your own machine (with an optional Google Translate engine if you prefer), and nothing you hear is uploaded.

Features

Recognition & translation

• Captures system audio through WASAPI loopback and follows you when you switch speakers/headphones; recovers by itself if a device is unplugged.
• Speech recognition with faster-whisper (Whisper large-v3-turbo by default), automatic language detection, almost 100 languages.
• Offline translation with NLLB-200 (600M or 1.3B), or Google Translate.
• Captions appear the moment a sentence is recognized; the translation replaces it a fraction of a second later.
• Speech already in your caption language can still be "translated", turning dialect or casual speech into the standard language.
• Moroccan Arabic (Darija) as a spoken and caption language.
• Optional Arabic diacritics (تشكيل) on the translation and/or the original text.
• Model downloads show a progress window with speed and time left, can be canceled, and resume where they stopped.

Speakers

• Recognizes different voices: every person gets a new line and their own color (up to 8 voices).

Caption box

• Language labels (code, name, flag) on the original text and on the translation — before, after, above or below.
• Fonts, sizes, colors, bold, line height and spacing for original and translated text separately.
• Text alignment, 9 screen positions (or drag it anywhere), multi-monitor, autosize, corner radius and padding, and a size (zoom) control — Shift + mouse wheel over the box.
• Smooth slide / fade animations for new lines, changed text, and the box resizing.
• Hides itself when nobody is talking; stays while your mouse is over it.
• Click-through lock, global hotkeys (Ctrl+Alt+H show/hide, Ctrl+Alt+P pause, Ctrl+Alt+L lock).

EchoSub Lens — translate the screen

• A tool of its own, with its own tray icon: choose an area of the screen or a whole screen and every piece of text in it is shown translated right on top of the original.
• Clicks pass through to the page or program underneath; hold Ctrl to select and copy a translation. Ctrl+Alt+T starts and stops it, Ctrl+Alt+R chooses a new area.
• Text is read offline with Windows' own OCR; translation uses NLLB (offline) or Google.
• Your own look: font, size, bold, text and background colours, opacity and rounded corners.
• Shows in screenshots and screen sharing (Discord, OBS), or only on your own screen if you prefer.
• A lens around the mouse (Ctrl+Alt+M): only the text under the pointer is translated, and the lens follows the mouse.
• Translates files too: pictures, PDF (scanned ones as well), Word (.docx, formatting kept) and text or subtitle files.
• Any key for any option: Settings → Keyboard shortcuts.

Everything else

• Caption history window (copy / save) and optional automatic transcripts.
• Live preview while you change settings; searchable dropdowns; restore defaults.
• Tray icon with status (loading, listening, paused, error), log file, single instance.

Screenshots

Any language in, your language out	Arabic diacritics (تشكيل)

Language & engine	Text & colors

Position, alignment & animation	Speakers

Your microphone, captioned beside the PC's sound

Caption history	Model download	About

System requirements

	Minimum	Recommended
Operating system	Windows 10, 64-bit	Windows 11, 64-bit
Processor	64-bit, 4 cores (e.g. Intel Core i3-8100 / AMD Ryzen 3 2200G)	6+ cores (e.g. Intel Core i5-9400F / AMD Ryzen 5 3600)
Memory (RAM)	8 GB	16 GB
Graphics	None required — runs on the CPU with the Small or Base speech model (captions lag several seconds)	NVIDIA GeForce GTX 1060 6 GB or better (GTX 10-series and newer); RTX cards are faster. AMD Radeon and Intel graphics work through Run on: Graphics card, any brand (Vulkan)
Graphics memory	4 GB for the default models on an NVIDIA GPU	6 GB or more (needed for NLLB 1.3B or Large v3)
NVIDIA driver	527.41 or newer (CUDA 12)	Latest Game Ready or Studio driver
Free disk space	6 GB (app 2.3 GB + default models 2.3 GB)	10 GB on an SSD (room for all models)
Internet	Once, to download the models (~2.3 GB)	Also needed if you use the Google Translate engine
Audio	Any Windows playback device (speakers, headphones, virtual devices)	—
Display	1280 × 720	1920 × 1080 or higher

EchoSub picks the fastest precision for your GPU automatically (for example int8 on GTX 10-series cards, which have no fast float16).

NVIDIA's libraries (cuBLAS, cuDNN, NVRTC — about 1.3 GB) are not in the installer: a PC with an NVIDIA card downloads them the first time the card is used, from NVIDIA's packages on PyPI, checked against pinned hashes. Other PCs never need them, which keeps the installer small.

AMD and Intel graphics: choose Settings → Language & Engine → Run on → Graphics card, any brand: AMD, Intel or NVIDIA (Vulkan). It runs a second speech engine, whisper.cpp, on any card with a current driver, and downloads its own compressed copy of the chosen model the first time (574 MB for Large v3 Turbo, 264 MB for Small). On a GTX 1060 it ran Small about 7× faster than real time against under 1× on the processor; on NVIDIA cards the NVIDIA-only option is still the fastest. It has been tested on NVIDIA through Vulkan — reports from AMD and Intel owners are welcome.

Models

Choose the models in Settings → Language & Engine. Download sizes are exact; graphics memory is an estimate for the loaded model; speeds were measured on an Intel Core i5-9400F with a GTX 1060 6 GB while a game was using the GPU at 99% — an idle GPU is faster.

Model	Download	Graphics memory (est.)	Time to caption a 9-second sentence
Speech — Base	141 MB	~0.3 GB	1.3 s (GPU)
Speech — Small	464 MB	~0.5 GB	2.5 s (GPU) · 10.7 s (CPU)
Speech — Medium	1.5 GB	~1.0 GB	2.9 s (GPU)
Speech — Large v3 Turbo (default)	1.5 GB	~1.2 GB	2.2 s (GPU) · 64 s (CPU — not usable live)
Speech — Large v3	3.1 GB	~2.0 GB	slower than Turbo

Translation model	Download	Graphics memory (est.)	Time to translate a sentence
NLLB 600M (default)	617 MB	~0.8 GB	0.22 s (GPU) · 1.4 s (CPU)
NLLB 1.3B	1.3 GB	~1.6 GB	0.47 s (GPU)
Google Translate	—	—	depends on your connection

Also downloaded: the speaker detection model (27 MB, runs on the CPU) and, only if you turn on Arabic diacritics, the diacritics model (69 MB, CPU).

Default setup (Large v3 Turbo + NLLB 600M + speaker detection): about 2.3 GB to download and about 2 GB of graphics memory.

Install

1. Download EchoSub-Setup-<version>.exe from the Releases page.
2. Run it and follow the steps (you can install for all users or just for you).
3. Start EchoSub. The first start downloads the AI models (~2.3 GB) in a progress window that you can cancel at any time — the download resumes where it stopped next time. After that EchoSub works offline (unless you choose Google Translate).

Settings, models, logs and transcripts are kept in %LOCALAPPDATA%\EchoSub. The uninstaller asks whether to remove them.

Using EchoSub

• Menu: right-click the caption box, or click the EchoSub icon in the system tray (next to the clock).
• Translation language: menu → Translation language, or Settings… → Language & Engine for all languages.
• Move / resize: menu → Adjust window position and size, then drag the box or its corner; or pick a spot under Settings… → Position & Alignment.
• One program or one window only: Settings… → Language & Engine → Only this program or window captions just Discord, say, or the program behind one window — pick it from the list or with Pick a window…, then switch to it. Windows keeps sound apart per program, so two windows of the same program (two Chrome windows) cannot be told apart, but two copies of a program can.
• Light mode: tray menu → Light mode, for when a game or another program is using the GPU: captions arrive sooner with a smaller speech model.
• Delayed captions: Settings… → Advanced → Catch-up audio buffer temporarily keeps delayed audio on disk so captions can recover after a short slowdown. It does not run another AI model or create a permanent recording.
• Updates: the installed app checks once a day for an official release. Use Check for updates… from the tray menu to check now; EchoSub opens the installer only after you choose it.
• Your microphone: Settings… → Language & Engine → Microphone captions what you say as well as what the PC plays, in its own colour and with your own label.
• On-screen answers: for a shared screen — someone says a phrase you chose, in any letter case and with no wake word needed, and EchoSub shows your own prepared line in the box. Off by default, and nothing is sent anywhere.
• Voice commands: tray menu → Voice commands (off by default). Say the wake word ("echo sub") and then, for example, "open chrome", "close paint" or "pause captions" — in English, Arabic, French, Spanish, German, Turkish, Russian, Chinese and more. Only a fixed list of everyday apps and EchoSub's own controls can be reached this way.
• A moment to change your mind: whatever EchoSub understood appears at the top of the screen with a countdown, and one click on it calls the command off. Three seconds by default, adjustable or switched off in Settings… → Commands.
• Your own microphone needs no wake word: say "open the calculator" and it happens, as long as the command begins the sentence. What the PC plays always needs the wake word.
• Go to a place, or play a file: "go to www.example.com" opens in your browser and "go to downloads" opens that folder. In Settings… → Commands you can write a word of your own beside any folder or file — say the word and a song plays in your usual music player. Only plain http and https addresses are opened, and a spoken word never becomes a path: it only picks one you wrote down.
• Ask an AI about what was said: off by default — Settings… → AI. Say your wake word and “ask …”, and Claude, ChatGPT, Gemini, DeepSeek or LM Studio answers in the caption box, using the recent captions you choose to send. Nothing is sent until you ask, and with LM Studio nothing leaves your PC at all. See PRIVACY.md before switching it on with a cloud service.
• Box position: right-click the caption box or the tray icon → Box position and pick a spot.
• Typing what you say: off by default — say “type” and the rest is typed into the window you are using, with “press enter” as a separate command. Anything the PC plays must say the wake word first, the text is shown in the countdown before any key is pressed, and a typed line can never contain a new line.
• Your own commands: give a phrase one of the approved actions, or a program and its arguments that you type yourself. That line is started directly rather than through a command shell, and nothing that was said is ever added to it.
• Copy a caption: hover the box and click the copy button next to a line.
• Bigger / smaller: hold Shift and turn the mouse wheel over the box (also Settings… → Position & Alignment → Size (zoom)).
• Lock: menu → Lock window (click-through) makes clicks pass through the box; unlock from the tray icon or Ctrl+Alt+L.
• History: menu → Caption history….
• Translate what is on the screen: start EchoSub Lens from the Start menu (or EchoSub Lens: translate the screen… in the tray menu). Its own icon appears by the clock: click it to start or stop, right-click it to choose an area or a screen, the language and the translator. Hold Ctrl to select and copy a translation; otherwise clicks go through to what is underneath.
• Tray icon dot: blue = listening, amber = loading or reconnecting, grey = paused, red = error.

If captions lag, choose a smaller speech model (Settings → Language & Engine → Small or Medium). The full User Guide explains every setting and common fixes.

Build from source

You need Windows, Python 3.10 (64-bit) and, for the installer, Inno Setup 6.

# run EchoSub from source (creates .venv and installs dependencies on first run)
powershell -ExecutionPolicy Bypass -File scripts\dev.ps1

# build the app into dist\EchoSub\EchoSub.exe
powershell -ExecutionPolicy Bypass -File scripts\build.ps1

# build the app and the installer dist\installer\EchoSub-Setup-<version>.exe
powershell -ExecutionPolicy Bypass -File scripts\build-installer.ps1

See docs/DEVELOPMENT.md for the project layout, how the pipeline works, and all script options.

Privacy

EchoSub processes audio on your computer. Audio and captions are never uploaded — except the caption text sent to Google if you select the Google Translate engine. See PRIVACY.md.

Support EchoSub

EchoSub is free and always will be. If it helps you follow a video, a call or a game, you can support its development with a donation through PayPal — it is also in the app, under Support EchoSub in the tray menu and the About window. It is entirely optional: nothing in EchoSub is held back, and the app never asks.

Contributing

Bug reports, ideas and pull requests are welcome — please read CONTRIBUTING.md and the Code of Conduct first. Security issues: see SECURITY.md.

License

Copyright © 2026 Mohammad Al-Safadi.

EchoSub is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License version 3 as published by the Free Software Foundation. It is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See LICENSE for the full text.

EchoSub uses third-party libraries and downloads AI models that have their own licenses — see THIRD_PARTY_NOTICES.md. Note: the default offline translation model (NLLB-200) is licensed for non-commercial use only; for commercial use, select the Google Translate engine or another model.

Author

Mohammad Al-Safadi

GitHub · LinkedIn · Discord · mPhpMaster@gmail.com