LM Subs

Synchronized closed captions, SRT/WebVTT/TXT/JSON export, and opt-in encrypted cloud sync for Google NotebookLM Video Overviews.

Download lm-subs-v0.2.8.zip | Beginner Setup Guide | Neon vs Supabase & Firebase Guide (DEPLOYMENT_GUIDE.md) | About (@sarang-cmd)

---

Install guide (no coding or credit card needed)

You do not need terminal commands, developer tools, or a paid subscription to use LM Subs. Follow these four steps in Google Chrome, Brave, Microsoft Edge, or Arc:

Step 1: Download and unzip the extension folder

1. Download lm-subs-v0.2.8.zip to your computer.
2. Unzip it into a normal folder:
   • macOS: Open Downloads in Finder and double-click lm-subs-v0.2.8.zip. Move the unzipped folder into Documents so it stays in a permanent spot.
   • Windows: Open Downloads in File Explorer, right-click lm-subs-v0.2.8.zip, click Extract All..., and click Extract. Keep the unzipped folder somewhere permanent such as Documents.

Step 2: Load the folder in Chrome

1. Open a new browser tab, type chrome://extensions into the address bar, and press Enter (brave://extensions in Brave, edge://extensions in Edge).
2. Turn ON the Developer mode switch in the top-right corner of the page.
3. Click the Load unpacked button in the top-left corner and select the unzipped lm-subs folder from Step 1.
4. Click the puzzle-piece icon in the top-right Chrome toolbar and click the pin icon beside LM Subs.

Step 3: Paste one free Groq transcription key (60 seconds)

LM Subs connects directly from your browser to a speech-to-text provider so nobody charges you a subscription fee. Groq offers a free developer account with no credit card that transcribes up to 8 hours of audio per day:

1. Open console.groq.com/keys and sign in with Google or GitHub.
2. Click Create API Key, enter LM Subs as the name, and copy the key starting with gsk_.
3. Click the LM Subs icon in your Chrome toolbar, paste that gsk_... code into the API key box, and click Save.

Step 4: Watch any NotebookLM Video Overview

1. Open notebooklm.google.com, open any notebook, and play a Video Overview.
2. Within about three seconds, synchronized captions with active-word highlighting appear inside the lower frame of the video player (accompanied by an optional completion chime).
3. Use the compact CC, Transcript, and ••• controls docked inside the bottom play bar to toggle captions (C), open the click-to-seek Transcript drawer (T), adjust word-for-word highlight delay ({ / } or the Style slider), switch caption backdrop (B), resize text (A- / A+), shift cue timing ([ / ]), view keyboard shortcuts (?), or export .srt, .vtt, .txt, .md, .csv, or .json files.

---

Product tour and screenshots

Compact bottom play-bar controls and highlighted spoken-word captions inside NotebookLM

Synchronized closed captions with 3-state progressive word highlighting, adjustable word-highlight delay (-500 ms to +1500 ms), compact CC and Transcript controls integrated directly into the bottom play bar, and a secondary tools popover for styling, study loops, and .srt/.vtt/.txt/.md/.csv/.json export.

Statistics that match the extension

Every number on the website is the same number the extension reports, computed on your device and merged with your account when you sign in: watch time, listening time, exports, and per provider usage.

Chrome toolbar popup

Configure a free Groq or Mistral key, inspect account sync state, trigger captions on the active NotebookLM tab, or export .srt, .vtt, .txt, and .json files in two clicks. Includes an inline first-time setup guide when no key is saved yet.

Settings, opt-in privacy switches, statistics, and priority queue

Sign in with Google, GitHub, or a 6-digit email code, toggle granular cloud sync switches (Sync API keys, Sync transcriptions, Sync preferences), inspect real transcription metrics, and order all 12 speech-to-text providers.

Page that opens after install

Setup finishes itself: pin the extension, paste one free API key, then press CC on any overview.

---

Why LM Subs?

Google NotebookLM generates narrated Studio Video Overviews, yet its built-in player ships without closed captions, downloadable subtitle files, or cross-browser transcript sync.

Naive transcription tools upload entire 50 to 120 MB HD .mp4 video files or inflate compressed audio into 40+ MB of uncompressed .wav data, hitting API upload limits and turning a 3-minute Video Overview into a multi-minute wait.

Metric	Full Video / WAV Upload	LM Subs (v0.2.8)
Media payload	50 to 120 MB .mp4 + 42 MB .wav	~3.4 MB =m140 AAC (or HTTP 206 audio byte ranges)
Audio prep overhead	45 to 90s full download + decode	0 ms direct =m140 passthrough (<1.2s Range remux)
Whisper timestamp mode	Slow word-level DTW alignment	Segment timestamps (~200x real-time) + clause balancer
Long-video chunking	Sequential 1-by-1 upload	3x concurrent workers + live progressive cues
Second-browser open	Re-downloads and re-transcribes	~50 ms Neon Postgres cache hit (0 media bytes fetched)

---

How LM Subs delivers captions in seconds

1. Zero-decode =m140 audio stream passthrough (0 ms IPC): Google's FIFE =m140 stream serves a ~3 to 6 MB audio/mp4 payload. service-worker.js inspects inspectMp4Tracks() and Content-Type and forwards pure-audio payloads (<=24 MB) directly to your speech-to-text provider without inflating them into PCM .wav.
2. Parallel HTTP 206 Partial Content audio byte-range extraction: When a CDN edge serves a muxed 50 to 120 MB =m22 video/mp4, LM Subs aborts the full download after reading headers, probes the moov atom via Range: bytes=..., and fetches only the 3 to 6 MB soun audio slices over 6 parallel HTTP 206 requests.
3. Segment-level Whisper timestamps (~200x real-time): Requesting segment timestamps instead of word-level DTW alignment and balancing two-line cues client-side via cuesFromSegments() cuts transcription latency by 4 to 8x.
4. Concurrent chunk workers + progressive cues (SUBTITLES_PARTIAL): Long recordings (>24 MB) transcribe 3 audio chunks concurrently and render live cues on the <video> element as soon as chunk 1 completes.
5. Pre-download Neon Postgres cache hit (~50 ms): Before downloading media from Google's CDN, service-worker.js checks local storage and GET /api/sync/subtitle?cacheKey=... so previously captioned videos load instantaneously across your signed-in browsers.

---

System architecture

• Page Main World (src/content/page-hook.js at document_start): Observes NotebookLM SSR AF_initDataCallback payloads and batchexecute RPCs (R7cb6c, gArtLc) without touching storage or credentials.
• Isolated Content Script (src/content/content.js + player-overlay.js): Attaches both a native <track kind="subtitles"> (for OS Picture-in-Picture) and an isolated closed ShadowRoot caption bar inside NotebookLM's <video> container.
• MV3 Service Worker (src/background/service-worker.js): Sole reader of secretApiKeys and cloudSyncSecret in chrome.storage.local. Executes audio stream extraction, 12-provider failover, and opt-in cloud sync.
• Hosted Web App & Neon Postgres API (src/web/, server/, api/index.js): React + Radix UI web application (/, /guide, /library, /stats, /about) and Hono REST API backed by Neon Serverless PostgreSQL (db/schema.sql). Includes Vercel Web Analytics and Vercel Speed Insights.

---

Hosted account and opt-in cloud sync

End users never self-host a database or configure a server URL. Sign in from the extension or web app using Google, GitHub, or a passwordless 6-digit email code:

• Granular opt-in switches: Control three independent switches in both the web library and extension settings:
  1. Sync API keys (AES-256-GCM per-user HKDF-derived encryption bound to ${userId}:${providerId} as AAD).
  2. Sync transcriptions (searchable caption archive with .srt, .vtt, .txt, and .json downloads).
  3. Sync preferences (provider priority order, caption font scale, and target language).
• One-click cloud purge: Turning off any switch can immediately delete your stored cloud rows for that category in Neon Postgres while keeping your local Chrome data untouched.
• Live SQL statistics (/stats): Displays real COUNT, SUM, AVG, 14-day daily transcription volume, upload bandwidth saved, and per-provider latency queried directly from lmsubs_subtitles and lmsubs_transcription_events.

Deploying your own fork (Neon correlated with Supabase & Firebase)

Read DEPLOYMENT_GUIDE.md for:

• A first-principles guide correlating Neon with Supabase and Firebase (SQL Editor migrations, why no storage buckets are needed, why Firebase auto-configures Google OAuth while Supabase and Vercel + Neon use Google Cloud Console OAuth credentials, and server-vault encryption vs client RLS).
• Exact step-by-step instructions to connect Neon (falling-cherry-16443194), Google OAuth, GitHub OAuth, and Resend on Vercel.

---

Supported speech-to-text providers

Configure one key or stack multiple providers in priority order for automatic HTTP 429 rate-limit failover:

Rank	Provider	Default model	Free tier and billing notes
1	Groq (Recommended)	whisper-large-v3-turbo	Free developer tier (~8 hrs/day, no credit card)
2	Mistral AI (Recommended)	voxtral-mini-latest	Free La Plateforme tier, 48 MB file cap
3	Google Gemini	gemini-2.5-flash	Free Google AI Studio tier, native translation
4	Deepgram	nova-3	$200 free starter credit
5	AssemblyAI	universal	$50 free starter credit
6	Gladia	solaria-1	10 free hours per month
7	Cloudflare Workers AI	@cf/openai/whisper-large-v3-turbo	10,000 free neurons per day
8	ElevenLabs	scribe_v1	Free starter quota, 99 languages
9	Hugging Face	openai/whisper-large-v3-turbo	Free monthly inference credits
10	OpenRouter	google/gemini-2.5-flash	Unified multi-model router
11	OpenAI	whisper-1	Pay-as-you-go ($0.006/min)
12	Local / Custom	Systran/faster-whisper-small	100% free and offline (faster-whisper-server / whisper.cpp)

---

Security and privacy architecture

• Minimal Manifest V3 permissions: Requests only "storage" in extension/manifest.json, plus narrow host permissions for NotebookLM and Google media CDNs.
• Service-worker secret isolation: Automated static-analysis tests (tests/e2e.test.js) verify that content.js, page.js, popup.js, and options.js never import secret-storage getters.
• Diagnostic redaction: GET_DEBUG_REPORT scrubs Bearer headers, lms_... session tokens, gsk_.../sk-.../AIza... keys, and signed FIFE query parameters.

---

Development and testing

npm install          # Install dependencies (0 vulnerabilities)
npm run assets       # Render SVG/PNG brand marks, favicons, badges, diagram, and 4 screenshots
npm run build        # Bundle all 5 MV3 extension entrypoints + React web app (docs/app.js)
npm run zip          # Rebuild bundles and package docs/releases/lm-subs-v0.2.8.zip
npm test             # Run unit, security, speed, export, and PGlite Postgres E2E tests
npm start            # Start unified Web App + /api/* Hono server on http://0.0.0.0:4173

---

About the creator

Created by Sarang (@sarang-cmd), a computer science student and agentic engineer based in Germany.

---

License

Released under the MIT License. Independent open-source project, unaffiliated with Google LLC.