Skip to content

RELEASE_NOTES.md

Latest commit

 

History

History
145 lines (114 loc) · 6.33 KB

RELEASE_NOTES.md

File metadata and controls

145 lines (114 loc) · 6.33 KB
Raw

FloatNote 1.0.0 🎙️

Be present in the meeting — FloatNote remembers it for you.

FloatNote is a Windows desktop app that captures a meeting as it happens — every word, who said it, and what's on screen — and turns it into a clean transcript, an on-demand summary, and a chatbot you can actually ask about it. It's built local-first: the recording, transcription, speaker separation, and screen reading all run on your own machine, and only the final summary/chat step reaches a language model.

⭐ At a glance

Capability Powered by Runs
Live transcription (mic + system audio) OpenAI Whisper 🖥️ On your PC
Speech / silence gating Silero VAD 🖥️ On your PC
Speaker separation & labels Resemblyzer voice embeddings 🖥️ On your PC
Screen Reader (OCR) OpenCV + Tesseract (bundled) 🖥️ On your PC
Keywords & action items spaCy 🖥️ On your PC
Summaries · chat · keyword cleanup Qwen2.5-7B-Instruct ☁️ HuggingFace
Meeting history SQLite 🖥️ On your PC

🧩 A closer look at the features

🎤 Two-way live transcription

Captures your microphone and the meeting's system audio at the same time, so your side and the remote participants both land in a single transcript. You can mute either stream independently mid-meeting without stopping the recording.

🗣️ Automatic speaker separation

Tells voices apart as the meeting runs — no need to process the whole recording afterward — and tags each line (Speaker 1, Speaker 2, …). You can rename a speaker to a real name and it sticks for that meeting.

🖥️ Screen Reader (OCR)

Watches your screen and, only when the content actually changes, reads the text off it — slides, shared docs, whiteboards — and adds it to the meeting as its own stream, separate from the spoken transcript. Tesseract is bundled, so there's nothing extra to install.

✅ Keywords & action items

A lightweight language pipeline picks out the key topics and detects tasks and who they're for in real time, so follow-ups don't slip through.

🧠 Summaries & chat — one model, zero fuss

A single language model (Qwen2.5-7B via HuggingFace) powers the on-demand summary, the chatbot, and keyword cleanup — so there's just one setup and one place your data goes. The chatbot answers only from the saved meeting, so it stays grounded instead of guessing.

💾 Everything saved locally

Meetings, transcripts (tagged by mic / speaker / screen), speaker names, and action items are stored in a local database you own — ready to summarize or revisit any time.

🚀 Install & first run

  1. Download FloatNote Setup 1.0.0.exe from the Assets below and run it.
  2. SmartScreen? If Windows warns "unknown publisher," it's just because the build isn't code-signed yet — click More info → Run anyway.
  3. First launch downloads the speech models once (~250 MB) — give it a minute.
  4. Add a token when prompted: a free HuggingFace token powers the summary and chat. Create one at huggingface.co/settings/tokens with the "Make calls to Inference Providers" permission. It's stored only on your device.

🎬 Using it, step by step

  1. Optionally name the session, then hit ▶ Start Meeting.
  2. The transcript streams in with speaker labels; keywords and action items appear live. Have a slide up? The Screen Reader grabs its text too.
  3. Mute the mic or system audio whenever you need to go off the record.
  4. Hit ⏹ Stop to end the session.
  5. Press Generate Summary for the recap, then chat with the meeting to pull out specifics.
  6. Your next meeting starts on a clean slate.

🏗️ How it's built

FloatNote ships as one installer that carries both the interface and the whole AI engine — no Python, no manual setup on the user's side.

  • Desktop shell — Electron. On launch it starts the local engine, waits for it, then loads the interface; on exit it shuts the engine down.
  • Interface — React + Vite + Tailwind CSS.
  • Engine — a FastAPI (Python) service frozen into a standalone .exe with PyInstaller, streaming to the UI over a WebSocket.
  • On-device AI — Whisper (transcription), Silero (voice detection), Resemblyzer (speaker matching), OpenCV + Tesseract (screen OCR), and spaCy (action items) all run offline.
  • Cloud AI — only summaries and chat call out, to the Qwen2.5-7B model on HuggingFace.
  • Storage — a local SQLite database via async SQLAlchemy.
 speak / share screen
        │
        ▼
 ┌──────────────── local engine (on your PC) ────────────────┐
 │  Whisper + Silero + Resemblyzer   →  transcript + speakers │
 │  OpenCV + Tesseract               →  on-screen text        │
 │  spaCy                            →  keywords + action items│
 └───────────────┬───────────────────────────┬───────────────┘
                 ▼                           ▼
          live to the app             saved to SQLite
                 │
                 ▼
   "summarize" / "ask a question"  →  Qwen2.5-7B (HuggingFace)  →  answer

✅ Requirements

  • Windows 10 or 11, 64-bit
  • ~2.5 GB free disk space (the app bundles its full AI runtime)
  • Internet — for the one-time model download and for summaries/chat
  • A microphone for your own voice (remote audio is captured automatically)

🔒 Privacy

  • Nothing is captured until you press Start.
  • Audio, speaker analysis, screen text, and meeting history all stay on your device. Only the text needed for a summary or chat reply is sent to the model, and your token never leaves your machine otherwise.
  • The FloatNote window is excluded from screen capture, so the Screen Reader never reads its own interface and the app won't appear in recordings.

⚠️ Known limitations

  • Windows only for now (system-audio capture is Windows-specific).
  • Not code-signed yet — expect the one-time SmartScreen prompt.
  • Summaries and chat need an internet connection and a HuggingFace token.