easysub

Project Overview

EasySub is a fully offline browser extension that provides real-time subtitles for any audio playing in a web page, such as English videos, online courses, and livestreams. Its key feature is running the speech recognition model locally on your own computer: it uses WASM to load the sherpa-onnx offline inference engine, transcribes the sound from the current browser tab in real time, and overlays bilingual (Chinese-English) subtitles on the page. No account registration, no connection to any server, and no audio upload are required, which solves the privacy leaks, mandatory internet connection, and subscription costs of online subtitle services. It is already published on the Microsoft Edge Add-ons store.

Key Features

• Real-time subtitles: captures the audio of the current tab and performs low-latency streaming recognition, showing captions as soon as someone speaks, usually with latency between 200 ms and 1 second.
• Fully offline: after the model is downloaded it works without the internet; all computation is local and audio is never uploaded.
• Bilingual recognition: uses a Zipformer Chinese-English model, suitable for English videos and mixed content.
• Subtitle overlay: subtitles appear as an overlay on the page and can be dragged freely; locking keeps only the text, and the locked state persists across restarts.
• Recent-line replay: click the top-left arrow to review the last 10 subtitle lines, useful when you miss something, and it works even when locked.
• Latency indicator: a green/yellow/red dot in the bottom-right shows recognition latency, or the exact millisecond value when unlocked.
• History search: filter, highlight, and count historical subtitles in real time from the panel.
• Punctuation restoration: punctuation is restored when a sentence finishes, making subtitles easier to read without blocking real-time performance.
• Bilingual help: advanced settings have a question-mark icon with detailed Chinese-English tooltips.
• Toggleable features: recent-line replay and the latency indicator can be turned off for zero performance overhead.

Target Audience

• Learners who often watch English videos, courses, talks, or livestreams and need real-time captions.
• Privacy-conscious users who do not want their audio uploaded to the cloud.
• Users with unstable networks or those who want offline subtitle capability.
• General users looking for a free subtitle solution.

Quick Start

1. Search for and install the EasySub extension from the Chrome or Edge extension store (it is available on Edge, and you can also build it yourself).
2. After installation, click the EasySub icon in the browser toolbar to open the control panel.
3. Click the Start button in the panel.
4. Play audio in the current tab (a video, course, etc.) and a live subtitle overlay will appear on the page.
5. Drag the subtitle layer to reposition it; click lock to keep only the text; click the top-left arrow to replay the last 10 lines.
6. To search history, type keywords in the panel search box to filter in real time.
7. To build it yourself: run npm install, npm run download-wasm (downloads the model; full version about 357MB, lite version about 150MB), and npm run build, then load the unpacked dist directory from the Chrome extensions page.

Core Advantages

• Privacy and safety: fully local inference, audio never leaves the device, no registration or network required.
• Completely free: no payment, subscription, or ads.
• Low latency: streaming architecture with AudioWorklet buffering on a separate thread, so main-thread blocking does not drop frames.
• Offline capable: works without the internet once the model is loaded.
• Open source: released under the MIT license with reviewable code.

System Requirements

• OS: desktop systems that can run Chrome or Edge, such as Windows, macOS, and Linux.
• Architecture: a Chrome/Edge extension (Manifest V3) running inside the browser.
• Additional dependencies: the recognition model must be downloaded for the first build or install; about 357MB for the full version and about 150MB for the lite version.

Notes

• The model must finish loading on first use; the full model is fairly large.
• Very old Chrome versions may not support AudioWorklet and will fall back to ScriptProcessorNode.
• Recognition accuracy depends on audio quality, speaking speed, and accent.