TLD2 v1.0.0 🎉

TLD2 uses Mozilla's industry-leading Readability.js library to extract clean article content from cluttered web pages:

• Intelligent parsing - Identifies main article content, ignoring ads, navigation, and sidebars
• Wide compatibility - Works on news sites, blogs, documentation, and most article pages
• Metadata extraction - Preserves title, author byline, and excerpt information
• Text-only output - Strips HTML formatting for clean AI processing

Supported content types: News articles, blog posts, medium stories, technical documentation, Wikipedia pages, and more.

Local AI Mode (Default)

Powered by Chrome AI Summarizer API—Google's built-in, on-device summarization model:

• ✅ 100% local processing (zero network requests)
• ✅ Complete privacy (no data sent to servers)
• ✅ Free and unlimited usage
• ✅ Works offline after initial setup
• ⚠️ Requires Chrome 120+ (availability varies by region)

Cloud AI Mode (Optional)

Powered by Google Gemini 2.0 Flash Lite for enhanced accuracy:

• ✅ Higher quality summaries for complex content
• ✅ Better context understanding
• ✅ Free tier: 15 requests/minute, 1M tokens/day
• ⚠️ Requires user-provided API key
• ⚠️ Article text sent to Google servers

Summary Length Options

Length	Target	Best For
Short	50-100 words	Quick overviews, executive summaries
Medium	150-300 words	Balanced detail (default)
Long	300-500 words	Comprehensive summaries with context

TLD2's TTS is powered by StreamingKokoroJS—a browser-based neural TTS engine using the Kokoro 82M parameter model:

Voice Selection (4 Natural Voices)

Voice ID	Gender	Accent	Description
af_sky	Female	American	Clear, professional tone (default)
af_nicole	Female	American	Warm, friendly conversational style
bm_fable	Male	British	Authoritative, narrator style
bm_lewis	Male	British	Casual, conversational tone

Audio Quality

• Sample Rate: 24 kHz (high-quality, natural sound)
• Model: Kokoro-82M-v1.0-ONNX (quantized to 86MB)
• Format: Float32 PCM audio via Web Audio API
• Latency: 200-500ms first chunk (WebGPU), 1-3s (WASM)

Playback Controls

• Speed: 0.5x to 2.0x (default: 1.0x)
• Pitch: 0.5x to 2.0x (default: 1.0x)
• Pitch Correction: Auto-adjust pitch to match speed (toggle)
• Auto-play: Start playback immediately after summarization
• Progress Tracking: Visual progress bar with time remaining

Streaming Architecture

TLD2 implements real-time streaming for immediate playback:

• Audio chunks generate in parallel while playing
• Starts playing after 2-3 sentence chunks (no wait for full summary)
• 120 tokens/second display rate mimics LLM streaming UX
• Buffered playback prevents audio dropouts

WebGPU (Recommended)

GPU-accelerated neural inference via WebGPU delivers 2-10x speedup:

• Requirements: Chrome 113+, modern GPU (NVIDIA, AMD, Intel, Apple Silicon)
• Performance: 200-500ms first chunk latency on typical hardware
• Detection: Automatic—TLD2 checks navigator.gpu on load

WASM Fallback

CPU-only fallback for universal compatibility:

• Requirements: Any modern browser (Chrome 90+)
• Performance: 1-3s first chunk latency (3-8x slower than WebGPU)
• Optimizations: Multi-threading (4 threads), SIMD instructions

Performance Comparison

Hardware	Backend	Latency	Speedup
NVIDIA RTX 3080	WebGPU	150-250ms	8-10x
Apple M1 Pro	WebGPU	200-350ms	5-7x
Intel Iris Xe	WebGPU	400-600ms	3-5x
CPU-only (8-core)	WASM	1500-2500ms	1x (baseline)

TLD2 provides comprehensive settings for personalized experience:

Voice & Playback

• Voice selection (4 voices: af_sky, af_nicole, bm_fable, bm_lewis)
• Speed control (0.5x - 2.0x, step: 0.1)
• Pitch adjustment (0.5x - 2.0x, step: 0.1)
• Pitch correction toggle (auto-adjust pitch with speed)
• Auto-play toggle (start TTS immediately after summary)

Summarization

• Summary length (short, medium, long)
• AI provider (Chrome AI, Gemini API)
• Gemini API key configuration (optional)

Appearance

• Theme selection (light, dark, auto)
• Dark mode optimized for VS Code aesthetics

Storage

Settings persist across sessions via chrome.storage.sync:

• Synced across devices (if Chrome Sync enabled)
• Encrypted by Chrome's storage API
• Locally stored (not accessible to TLD2 servers—we have none!)

TLD2 uses Chrome's sidePanel API for a non-intrusive, always-available interface:

• Persistent UI: Sidebar stays open across page navigations
• Resizable: Drag to adjust width (200px - 800px)
• Non-blocking: Read original article while TLD2 processes
• Collapsible: Close sidebar when not needed

Local AI Mode (Default)

Complete data privacy with zero external network requests:

Web Page (your browser)
  ↓
TLD2 Content Script (local)
  ↓
Chrome AI Summarizer (local)
  ↓
StreamingKokoroJS TTS (local)
  ↓
Audio Playback (local)

❌ No data leaves your device
❌ No tracking or telemetry
✅ 100% private processing

Cloud AI Mode (Optional)

Transparent cloud usage with user control:

• ⚠️ Article text sent to Google Gemini API
• ✅ TTS still processes locally
• ✅ Opt-in only (must provide your own API key)
• ✅ Clear UI indication when cloud mode active

Manifest V3 Structure

• Background Service Worker: Handles initialization, message routing, and TTS model management
• Content Scripts: Dynamically injected to extract article content with Readability.js
• Sidebar UI: Chrome sidePanel for summary display and TTS controls
• Message Passing: chrome.runtime.sendMessage for component communication

Data Flow

1. User clicks extension icon
2. Background worker injects content script
3. Content script extracts article with Readability.js
4. Background worker sends text to AI (Chrome AI or Gemini)
5. Summary returned and displayed in sidebar
6. User clicks Play → Background worker generates TTS
7. Audio chunks stream to sidebar for playback