GitHub - DatanoiseTV/meeting-assistant: Meeting Assistant transforms your spoken conversations into structured, actionable knowledge. It handles real-time transcription, deep AI analysis, and generates professional reports tailored for your specific role.

Meeting Assistant

Meeting Assistant is a high-performance terminal application that transforms spoken conversations into structured knowledge. It combines real-time local transcription with deep AI analysis to generate professional reports, visual mind maps, and insights tailored to your specific professional role.

Why Meeting Assistant?

Manual note-taking is a cognitive burden that distracts from active participation. Meeting Assistant solves this by:

Eliminating Cognitive Load: Focus entirely on the conversation while the AI handles the documentation.
Role-Specific Filtering: Specialized personas (Dev, PM, Exec) ensure you only get the insights relevant to your role.
Truly Offline & Private: Speech-to-text is powered by whisper.cpp and happens entirely offline on your machine. Your raw audio never leaves your local environment.
Flexible AI Intelligence: Use high-performance cloud models (Gemini, OpenAI) or maintain a 100% offline workflow by connecting to a local Ollama instance.

Real-World Examples

1. The Daily Standup (Persona: PM)

Focus on identifying blockers and ensuring the timeline is on track.

# Start a session focused on deliverables and blockers
meeting_assistant -l --ui -p gemini --persona pm

2. Technical Architecture Review (Persona: Dev + Research)

Focus on capturing complex logic and fact-checking external libraries.

# Capture technical details and research mentioned libraries/APIs
meeting_assistant -l --ui -p gemini --persona dev --research

3. Fully Offline Confidential Meeting (Persona: General + Ollama)

When privacy is paramount, run everything on your own hardware.

# Local transcription + Local LLM analysis
meeting_assistant -l --ui -p ollama -L llama3

Core Capabilities

Active Intelligence

Live AI Copilot: Press [Space] during a meeting to query the AI about the current context.
Contextual Continuity: Whisper retains a rolling memory of the last 200 characters to maintain accuracy across ongoing sentences.
Visual Mapping: Every meeting generates a Mermaid.js diagram to visualize topics and decisions.

Seamless Integration

Obsidian v3: Notes use modern Properties and semantic callouts to integrate directly into your second brain.
Standalone HTML: Generates tidy, CSS-styled reports perfect for sharing via email or Slack.

Installation

Prerequisites

CMake: 3.14 or higher.
PortAudio: Required for live microphone input (brew install portaudio on macOS).

1. Download Whisper Model

The application requires a Whisper model in ggml format. Choose a model based on your hardware and accuracy needs:

Model	Size	Speed	Accuracy	Recommended For
tiny.en	75 MB	Fastest	Lowest	Real-time testing / Low-power
base.en	142 MB	Very Fast	Good	Standard laptops / Most meetings
small.en	466 MB	Fast	Great	High-accuracy requirements
medium.en	1.5 GB	Slow	Excellent	Post-meeting batch processing
large-v3	2.9 GB	Slowest	State-of-the-art	Maximum precision (Requires GPU)

Using curl to download:

mkdir -p models
# Example: Downloading the 'small.en' model
curl -L https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.bin -o models/ggml-small.en.bin

2. Build

mkdir build && cd build
cmake -DCMAKE_PREFIX_PATH=/opt/homebrew ..
make
sudo make install

Dashboard Hotkeys

[Space]: Open AI Copilot to ask a question during the meeting.
[N]: Finalize current session and start a New Meeting immediately.
[Q / ESC]: Save all reports and Quit.

Configuration

Settings are persisted in ~/.meeting_assistant/config.json.

Template: Copy the provided config.json.example to ~/.meeting_assistant/config.json.
CLI: Alternatively, update settings via the command line using the --save-config flag.

Refer to config.json.example for a full list of supported fields including GitHub/GitLab integration.

License

Apache License 2.0 - See LICENSE for details.