Speaker Notes

A floating speaker-notes overlay for Windows that is completely invisible to screen capture — Teams, Meet, Zoom, OBS, all of them. No admin rights. No drivers. No DLL injection. Just a Python window that hides itself using a single Win32 API call.

---

Origin

This started with Netflix.

Like most streaming services, Netflix blocks screen recording — try to capture it and you get a black rectangle where the video should be. I got curious about how that worked. After digging into it, I found that Windows exposes a per-window flag called SetWindowDisplayAffinity. When you set it to WDA_EXCLUDEFROMCAPTURE, the window is rendered normally on your screen but the capture pipeline sees only a transparent void.

Netflix's player sets this flag on its video surface. That's the whole trick.

The obvious next thought: what if I set that flag on a window I own? A window full of presenter notes, floating over everything, visible only to me — invisible to whoever is watching my screen share.

That is exactly what this project is.

---

What it does

• Displays your speaker notes as slide cards in a borderless, always-on-top window
• The window is excluded from all screen capture at the OS level — it simply does not exist as far as Teams, Meet, Zoom, or any recording software is concerned
• Notes sync live from your phone via Firebase or a local WebSocket — type on your phone, see it on your laptop instantly
• Three-card layout shows previous slide (faded), current slide (active), and next slide (preview) simultaneously
• Supports rich formatting: headings, subheadings, bullets, bold, and speaker-prompt callout boxes

---

How the cloaking works

Windows has a display stack with multiple layers. Applications like Netflix render into a protected layer that the capture APIs cannot read. The mechanism is a single function:

SetWindowDisplayAffinity(hwnd, WDA_EXCLUDEFROMCAPTURE)

The critical constraint: the calling process must own the window. You cannot cloak someone else's window — Windows enforces this. This is why Netflix's protection cannot be trivially bypassed by an external tool, and it is also why self-cloaking works perfectly: our Python process creates its own tkinter window and immediately cloaks it.

There is one subtle gotcha: tkinter.winfo_id() returns an inner child HWND, not the top-level window handle. SetWindowDisplayAffinity rejects child windows with error 87 (invalid parameter). The fix is GetParent(winfo_id()) to reach the real top-level HWND.

inner = window.winfo_id()
hwnd  = ctypes.windll.user32.GetParent(inner) or inner
ctypes.windll.user32.SetWindowDisplayAffinity(hwnd, 0x00000011)

No admin. No elevated privileges. No side effects on the rest of the system.

---

Features

• Screen-capture invisible — works against Teams, Meet, Zoom, OBS, Windows Game Bar, PrintScreen
• Three-card viewport — prev / active / next slide always visible; click to navigate
• Live sync via Firebase — type notes on your phone, updates appear on your laptop in under a second
• Live sync via local WebSocket — same-network mode with no internet dependency
• Rich note formatting — # headings, ## subheadings, **bold**, - bullets, > speaker prompts
• Speaker prompt boxes — > lines render as a distinct highlighted callout box, separate from the main notes
• Teleprompter mode — auto-scroll with live word highlight, comfortable reading pace
• Slide switcher — sidebar logs panel lists all slide titles; click to jump
• Resize + drag — 8-direction resize handles, drag from title bar
• Opacity control — adjustable from the settings popup
• Meeting timer — counts up from 00:00:00 on launch
• Font size control — A− / A+ buttons in the status bar
• No admin, no DLL injection, no external drivers

---

AI Copilot mode (100% on-device)

A built-in live interview copilot — like Parakeet, but nothing leaves your machine. Click the ✦ COPILOT sidebar tab to arm it. It listens to the other side's audio (the call output, not your mic), transcribes it locally, and when you press F9 it streams an AI-generated answer straight into the same cloaked card — invisible to the screen share.

• Fully local — system audio is captured via WASAPI loopback, transcribed on-device with faster-whisper, and answered by a local LLM running in Ollama. The transcript and the answer never touch the network.
• Manual trigger — you decide when to answer (F9); nothing is sent automatically.
• Your voice — drop a short résumé summary in copilot_profile.md (git-ignored) and answers come back in first person, grounded in your background.

Setup: install Ollama and ollama pull qwen3:8b, then pip install -r requirements.txt (first run downloads the Whisper model). See CLAUDE.md for env overrides and GPU notes.

Privacy & security notes

• On-device guarantee: the copilot only ever talks to http://localhost:11434 (Ollama). No audio, transcript, or answer is uploaded anywhere. Model weights are downloaded once from Hugging Face and then run offline.
• While copilot mode is active, two system-level things happen, and stop the moment you leave the tab or close the app: (1) a global F9 hotkey is registered so the trigger works even when the meeting window is focused, and (2) system output audio is captured (loopback). Both are torn down on exit.
• The existing Firebase / Local-WebSocket sync (for notes) does send your notes off-machine by design — that is separate from the copilot, which is never routed into those sync paths. The Local WebSocket server is unauthenticated and LAN-exposed; use it on trusted networks.

---

Web UI (phone_ui/index.html)

A single HTML file — no build step, no framework. Open it directly in your phone browser or deploy it to GitHub Pages / Vercel.

• Per-slide editor with syntax highlighting overlay
• Smart paste: copies from Word or Google Docs auto-convert to Markdown
• H1 / H2 / Bold / Bullet / Prompt toolbar buttons
• Dark and light theme
• Collapsible connection panel (collapses automatically after connecting)
• Firebase and Local WebSocket modes

---

Requirements

• Windows 10 version 2004+ or Windows 11 (earlier builds lack WDA_EXCLUDEFROMCAPTURE)
• Python 3.9+
• pywin32 (usually pre-installed or available via pip)

pip install -r requirements.txt

---

Setup

1. Clone and create a virtual environment

git clone https://github.com/Sid009192/speaker-notes-prompter.git
cd speaker-notes-prompter
python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt

2. Configure your sync method

Copy .env.example to .env and fill in your Firebase URL:

FIREBASE_URL=https://your-project-default-rtdb.firebaseio.com

Or leave it blank and use the --firebase / --local flags at runtime (see below).

---

Running

# Activate venv first
.\.venv\Scripts\activate

# Auto-connect using .env
python overlay.py

# Local text file
python overlay.py sample_notes.txt

# Firebase (URL overrides .env)
python overlay.py --firebase https://your-project-default-rtdb.firebaseio.com

# Local WebSocket (same network as phone)
python overlay.py --local
# Prints the LAN IP — enter it in the web UI

# Custom port
python overlay.py --local --port 9000

---

Notes file format

# Slide Heading

## Subheading

Regular paragraph text. **Bold inline.**

- Bullet point one
- Bullet point two

> This is a speaker prompt — renders as a highlighted callout box.
> Visible only in the overlay, great for reminders and cues.

---

# Next Slide

--- on its own line separates slides.

---

Firebase setup (for live phone sync)

1. Go to console.firebase.google.com and create a project
2. Build → Realtime Database → Create Database → choose a region → Start in test mode
3. Copy the database URL (looks like https://your-app-default-rtdb.firebaseio.com)
4. Paste it into .env as FIREBASE_URL=... or enter it in the web UI

Test mode rules expire after 30 days. Renew in Firebase Console → Rules if sync stops working.

---

Project structure

overlay.py              # Main application — OverlayApp class
copilot/                # Local, on-device AI copilot (capture, transcribe, LLM)
  capture.py            # WASAPI loopback audio capture
  transcriber.py        # faster-whisper STT (RealtimeSTT + Silero VAD)
  llm.py                # Ollama streaming client (local LLM)
  engine.py             # Orchestrator: audio -> STT -> buffer -> F9 -> answer
phone_ui/
  index.html            # Web UI for phone (no build step)
sample_notes.txt        # Example notes file
requirements.txt        # requests, websockets, + copilot deps
copilot_profile.example.md  # Template for your private résumé summary
.env.example            # Config template — copy to .env

---

Keyboard shortcuts

Key	Action
← / →	Previous / next slide
Ctrl + Shift + ` or Ctrl + `	Toggle show / hide overlay (system-wide global hotkey + in-app)
Ctrl + Shift + H or Ctrl + H	Toggle show / hide overlay (alternative global hotkey)
F2 or Ctrl + E	Toggle Read-Only / Editable mode (locks click-to-edit; enables content space drag)
Alt + D / Alt + I (or Ctrl + [ / ])	Decrease / increase window opacity by 10%
F9	Copilot: answer the latest transcribed question (global while copilot mode is on)
Ctrl + Shift + Q	Screen capture: take screenshot of presenter's screen & sync to client
Escape	Close settings popup / slide switcher / stop copilot / exit

---

Limitations

• Windows only — SetWindowDisplayAffinity is a Win32 API with no equivalent on macOS or Linux
• Requires Python — not a standalone executable (though pyinstaller packaging is straightforward)
• Firebase test-mode rules are open read/write — suitable for personal use; add Firebase Auth for shared environments

---

License

MIT