Product & engineering
Capture roadmap debates and design decisions with the slides that drove them. Query 'why did we pick Postgres?' six weeks later.
Desktop-first meeting assistant
Your meetings, turned into searchable memory.
FloatNote runs quietly in the background — capturing your mic, reading your screen, and turning every meeting into live transcripts, AI summaries, and a chatbot that actually knows what was said.
Free & open source · Windows 10/11 · Your transcripts stay on your machine
Recording · Live
- MIC
Let's align on the Q3 roadmap before we ship.
- OCR
Slide: Roadmap Overview — 3 priorities
- MIC
Priya will share the draft by Friday.
Action item detected
Share roadmap draft → assigned to Priya, due Fri
- Live transcription & VAD
- Real-timeLive transcription & VAD
- Transcripts stored locally
- 100%Transcripts stored locally
- Meeting bots to invite
- 0Meeting bots to invite
- Source, self-hosted
- OpenSource, self-hosted
Features
Everything a meeting leaves behind — captured
Seven systems working together, from the moment audio hits your mic to the question you ask three weeks later.
Mic & System Audio Capture
Streams audio from your mic and remote participants (via Windows WASAPI loopback) through OpenAI Whisper base model, gated by Silero VAD for real-time speech detection.
Screen OCR
Reads slides and shared screens as they change via Tesseract OCR, extracting on-screen text and keywords so your meeting notes capture what was shown.
AI Summarization
Turns meeting recordings into structured summaries using Qwen2.5-7B via Hugging Face Inference Providers, with a local fallback when offline.
Meeting Chatbot
Ask questions about any past meeting. Answers are grounded in a local FAISS vector store via retrieval-augmented generation (RAG) — no hallucinated recaps.
Speaker Diarization
Runs fully offline — streaming Resemblyzer d-vector embeddings cluster utterances in real time, assigning consistent speaker labels (SPEAKER_00 / SPEAKER_01) across the live call.
Action Item Extraction
An NLP pipeline (spaCy) detects tasks and who they're assigned to straight from spoken language — so nothing agreed on gets lost.
Persistent Local Memory
Transcripts, speaker aliases, OCR captures, and action items are saved locally to SQLite via async SQLAlchemy — searchable meeting memory on your machine.
How it works
No bot to invite. No workflow to change.
FloatNote sits on your desktop and does the remembering for you.
- 01
It runs in the background
Launch FloatNote and it quietly captures your mic, system audio, and shared slides. No bot joins the call, no awkward 'recording started' banner.
- 02
Everything becomes memory
Speech becomes transcript, slides become text, and both are indexed into a local FAISS vector store as the meeting happens.
- 03
Ask it anything, after
Get an instant AI summary, pull the action items, or chat with the meeting — 'What action items were assigned to me?' — and get grounded answers.
Use cases
Built for anyone whose work happens in meetings
Sales & customer calls
Never scribble notes mid-call again. Get the summary, the commitments, and the follow-ups extracted automatically from mic and loopback audio.
Research & interviews
Transcribe user interviews verbatim with speaker labels, then chat across every session to find patterns without re-watching recordings.
Students & lectures
Record lectures with the on-screen slides captured via Tesseract OCR, then ask the chatbot to explain any concept from class.
Under the hood
Local-first AI, wired together
Transcription runs on your machine with Whisper. Slides are read with Tesseract OCR. A LangChain + FAISS pipeline turns it all into a queryable vector store — so the chatbot answers from what actually happened, not a guess.
Fully open source and configurable — bring your own models and API keys.
FastAPI
Async server + WebSockets
Whisper + Silero VAD
Local speech-to-text & VAD
WASAPI Loopback
Mic & system audio capture
Resemblyzer
Offline speaker diarization
Tesseract OCR
Screen & slide reading
Qwen2.5-7B (HuggingFace)
Summaries, chat & keywords
LangChain + FAISS
RAG retrieval pipeline
spaCy NLP
Action item extraction
SQLite + SQLAlchemy
Local persistent storage
React 19 + Vite
Dashboard UI
Electron
Desktop app wrapper
FAQ
Questions, answered
Does my audio get sent to the cloud?
Transcription (Whisper), VAD (Silero), speaker diarization (Resemblyzer), and RAG vector storage (FAISS) all run locally on your machine. Transcripts and meeting data stay in a local SQLite database. Summaries and the chatbot use Hugging Face Inference Providers (Qwen2.5-7B), so only text you choose to query or summarize is sent.
Do I have to invite a bot to my meeting?
No. FloatNote captures your microphone and system audio (via Windows WASAPI loopback) and reads your screen directly from your desktop. There's no meeting bot, no join link, and no participant to explain — it runs quietly in the background.
What can the meeting chatbot actually answer?
Anything grounded in the meeting. It uses retrieval-augmented generation (RAG) over a local FAISS vector store of your transcript and OCR captures, surfacing decisions, action items, and context tied to what was actually said or shown.
Which platforms does it run on?
FloatNote is desktop-first, built with a FastAPI backend and a React dashboard, with an optional Electron wrapper for a native window. Windows is required for system audio loopback capture; mic-only capture works on all platforms.
Is it free and open source?
Yes. FloatNote is open source — you run it yourself with a Hugging Face API token for AI summaries and chat (with a local fallback if no token is provided). The full source code, architecture, and setup instructions are on GitHub.
Stop taking notes. Start asking your meetings.
Download FloatNote for Windows, or clone the repo and run it in minutes — free and fully yours.