YouMoji — On-Device AI Emojis

Real-time floating emotion emojis for YouTube videos, powered by the offline Emo AI model.

YouMoji is a Chrome extension (Manifest V3) that parses YouTube video captions locally using an on-device AI model to display real-time floating emotion emojis over the video player. It operates completely offline within your browser without sending any subtitle data to external servers.

Why YouMoji?

100% On-Device AI Execution

YouMoji runs an 11MB LiteRT (TensorFlow Lite) model and tokenizer directly inside a Chrome Offscreen WebAssembly worker. It infers emotional intent and maps subtitle text to emojis in under 2ms per line, supporting 22 languages natively.

Zero Network Dependencies

No network requests are initiated during AI inference. All model weights, WebAssembly binaries, scripts, and visual assets are 100% bundled within the extension package. There are no API keys, accounts, or cloud backend servers required.

Bilibili-Style Danmaku Overlay

Captions are rendered as floating emojis across 8 physics-based trajectories with 3D perspective tilt. The extension integrates with YouTube's player without interfering with native CC buttons or video controls.

Key Features

On-Device Emo AI Model — Powered by Emo (by Desert Ant Labs) with a curated 812-emoji vocabulary tuned for short intent-carrying text
Sub-2ms Inference — Instant local execution per line with zero network latency
Multilingual Support — 22 languages supported natively, including English, Chinese, Japanese, and Korean
900+ Offline Visual Assets — Includes Microsoft Animated 3D, Twemoji, Fluent 2D, and JoyPixels
No Outbound Network Requests — Subtitle text never leaves your browser sandbox
Sidepanel Controls — Customize confidence thresholds, toggle master switches, choose visual styles, and test sentences in an offline playground
Native Subtitle Rollback — Intercepts player caption streams seamlessly while keeping native YouTube CC functional

How It Works

Open Video: Open any YouTube video that includes captions.
Local Parsing: YouMoji captures the subtitle stream in real time and passes text lines to the local WASM worker.
On-Device Inference: The Emo model predicts emotion scores and matches corresponding emoji assets in under 2ms.
Floating Rendering: Emojis float dynamically over the video player according to your configured position and threshold settings.

Technical Overview

Framework: WXT (Web Extension Toolbox, Manifest V3)
Build Tool: Vite
UI Stack: React 19, TypeScript, Tailwind CSS
AI Engine: @desert-ant-labs/emo + @litertjs/core (WebAssembly)
Execution Context: Chrome Offscreen Document
Storage: chrome.storage.local
Permissions: sidePanel, storage, offscreen, *://www.youtube.com/*

Version: 2.3.0