MeetingNotes - Intelligent Meeting Analysis with Voxtral

Web application using Voxtral AI from Mistral AI to automatically analyze your audio/video meetings with:

• Direct analysis: Transcription and structured summary in one step
• 3 processing modes: Local (Transformers), MLX (Apple Silicon), API (Cloud)
• Quantized models: 4bit/8bit support for memory efficiency
• Smart diarization: Speaker identification and renaming
• Customizable summaries: Modular sections according to your needs
• Language-adaptive: Automatically detects and responds in meeting language
• Centralized UI: Clean English interface with multilingual analysis
• 🚀 Hugging Face Spaces: Try the simplified version online at VincentGOURBIN/MeetingNotes-Voxtral-Analysis

🚀 Quick Start

🌐 Online Version (Hugging Face Spaces)

Try MeetingNotes directly in your browser: VincentGOURBIN/MeetingNotes-Voxtral-Analysis

This simplified version uses standard Mistral Voxtral models optimized for Zero GPU with automatic chunk processing.

💻 Local Installation (Full Version)

1. Clone the repository and install dependencies:

git clone <repository-url>
cd meetingnotes
pip install -r requirements.txt

2. Configure your Hugging Face token:

cp .env.example .env
# Edit .env and add your Hugging Face token

3. Launch the application:

python main.py

The web interface will be accessible at http://localhost:7860

⚙️ Configuration

Hugging Face Token (Required)

Get an access token from Hugging Face and add it to .env:

HUGGINGFACE_TOKEN=your_token_here

Mistral API Key (Optional)

To use cloud API mode, get a key from Mistral AI:

MISTRAL_API_KEY=your_mistral_api_key

🎯 Features

Processing Modes

Choose the mode that best fits your hardware and needs:

Local Mode (Transformers)

• Local processing: Everything runs on your machine with PyTorch
• Privacy: No data sent to external servers
• GPU acceleration: Automatic CUDA/MPS detection

MLX Mode (Apple Silicon)

• Optimized for Mac: M1/M2/M3 processors with MLX Framework
• Best performance: Native Apple Silicon acceleration
• Memory efficient: Optimized quantized models

API Mode (Cloud)

• Cloud processing: Uses Mistral Cloud API
• No local resources: Minimal memory usage
• Always up-to-date: Latest models and improvements

Models and Quantization

Model	Precision	Repository	Memory Usage
Voxtral Mini	Default	mistralai/Voxtral-Mini-3B-2507	~6GB
Voxtral Mini	8bit	mzbac/voxtral-mini-3b-8bit	~3.5GB
Voxtral Mini	4bit	mzbac/voxtral-mini-3b-4bit-mixed	~2GB
Voxtral Small	Default	mistralai/Voxtral-Small-24B-2507	~48GB
Voxtral Small	8bit	VincentGOURBIN/voxtral-small-8bit	~24GB
Voxtral Small	4bit	VincentGOURBIN/voxtral-small-4bit-mixed	~12GB

Speaker Diarization

• Automatic identification: Detection of different speakers with pyannote.audio
• Reference segments: Listen to audio samples for each speaker
• Custom renaming: Assign human names to speakers
• Context integration: Use speaker information in summaries

Customizable Summaries

Modular sections: Choose the sections to include according to your needs

• 📄 Executive Summary: Global overview of the meeting
• 💬 Main Discussions: Main topics addressed
• ✅ Action Plan: Actions, responsibilities, deadlines
• ⚖️ Decisions Made: Validated decisions
• ⏭️ Next Steps: Follow-up actions
• 📌 Main Topics: Information presented
• ⭐ Key Points: Insights and key data
• ❓ Questions & Discussions: Questions asked and answers
• 📝 Follow-up Elements: Clarifications needed

Predefined Profiles:

• 🎯 Action Profile: Focus on tasks and decisions
• 📊 Information Profile: Focus on data and insights
• 📋 Complete Profile: All sections activated

Supported Formats

• Audio: WAV, MP3, M4A, OGG, FLAC
• Video: MP4, AVI, MOV, MKV (automatic audio extraction)

🔧 Usage

1. Processing Mode Configuration

1. Choose the mode: Local, MLX or API
2. Select the model: Mini or Small according to your needs
3. Choose precision: Default, 8bit or 4bit to optimize memory

2. Upload and Options

• File: Direct audio or video (automatic extraction)
• Optional trimming: Start/end trimming (leave empty for 0)

• Chunk size: Processing duration (5-25 minutes)

3. Diarization (Optional)

1. Analyze speakers with pyannote.audio
2. Listen to reference segments of each speaker
3. Rename speakers with custom names
4. Apply renamings for enriched context

4. Summary Customization

• Modular sections: Enable only necessary sections
• Quick profiles: Action, Information or Complete
• Flexible configuration: Adapt summary to your usage

5. Analysis and Results

Click "Analyze Meeting" to get a customized structured summary.

🏗️ Architecture

The project follows a modular architecture with two versions:

Main Project (src/meetingnotes/)

src/meetingnotes/
├── ai/                    # Artificial Intelligence
│   ├── voxtral_analyzer.py      # Local Voxtral analyzer (Transformers)
│   ├── voxtral_api_analyzer.py  # Voxtral API analyzer
│   ├── voxtral_mlx_analyzer.py  # Voxtral MLX analyzer (Apple Silicon)
│   ├── diarization.py           # Speaker diarization (pyannote)
│   ├── memory_manager.py        # Optimized memory management
│   └── prompts_config.py        # Centralized prompt configuration
├── audio/                 # Audio Processing
│   ├── wav_converter.py         # Format conversion
│   └── normalizer.py            # Volume normalization
├── core/                  # Business Logic
│   ├── voxtral_direct.py        # Direct processing (Transformers)
│   ├── voxtral_api.py           # Mistral API interface
│   └── voxtral_mlx.py           # MLX Apple Silicon interface
├── ui/                    # User Interface
│   ├── main.py                  # Main Gradio interface
│   ├── handlers.py              # Event handlers
│   └── labels.py                # UI labels and text constants
└── utils/                 # Utilities
    ├── __init__.py              # Utils module
    ├── time_formatter.py        # Duration formatting
    └── token_tracker.py          # Token usage tracking

Hugging Face Spaces Version (huggingface-space/)

Simplified version deployed at VincentGOURBIN/MeetingNotes-Voxtral-Analysis:

huggingface-space/
├── src/
│   ├── ai/
│   │   ├── voxtral_spaces_analyzer.py   # HF Spaces optimized analyzer
│   │   └── prompts_config.py            # Shared prompts configuration
│   ├── ui/
│   │   ├── spaces_interface.py          # Simplified Gradio interface
│   │   └── labels.py                    # UI labels
│   └── utils/
│       ├── zero_gpu_manager.py          # Zero GPU management
│       └── token_tracker.py             # Token tracking
├── app.py                               # HF Spaces entry point
├── requirements.txt                     # HF Spaces dependencies
└── deploy.py                           # Deployment script

Key differences in HF Spaces version:

• Only Transformers backend (no MLX/API modes)
• Standard Mistral Voxtral models (optimized for Zero GPU)
• No speaker diarization (simplified interface)
• Progress bar with chunk-based tracking
• Automatic chunk duration optimization (15min Mini, 10min Small)

For more details, see ARCHITECTURE.md.

🔧 Advanced Configuration

Environment Variables

# Required for all modes
HUGGINGFACE_TOKEN=your_hf_token

# Optional for API mode
MISTRAL_API_KEY=your_mistral_key

Hardware Optimizations

• Mac M1/M2/M3: Use MLX mode for better performance
• NVIDIA GPU: Local mode with automatic CUDA acceleration
• CPU only: Prefer 4bit models to save memory
• Limited memory: Mini 4bit (~2GB) or Small 4bit (~12GB)

🔍 Technical Features

Memory Optimizations

• Pre-quantized models: 4bit and 8bit for memory reduction
• Memory manager: Automatic cleanup between chunks
• Multi-platform support: MPS (Apple), CUDA (NVIDIA), optimized CPU

Intelligent Processing

• 3 inference modes: Direct audio-chat without intermediate transcription
• Language-adaptive: Automatically detects meeting language and responds accordingly
• Adaptive chunks: Smart division of long files with synthesis
• Modular prompts: Customizable summary sections with centralized configuration
• Enriched context: Integration of diarization in analyses
• Token tracking: Comprehensive usage statistics across all processing modes

Modern Interface

• Centralized UI labels: Clean English interface with maintainable text management
• Improved API layout: API key positioned next to model selection for better UX
• Interactive diarization: Speaker listening and renaming
• Modular sections: Advanced summary customization with preset profiles
• Real-time feedback: Detailed progress indicators and token consumption tracking

📦 Main Dependencies

• gradio: Modern web user interface
• torch/torchaudio: Deep learning framework (Local mode)
• transformers: Hugging Face and Voxtral models
• mlx/mlx-voxtral: MLX framework optimized for Apple Silicon (macOS only)
• pyannote.audio: Speaker diarization
• pydub: Audio processing and conversion
• requests: Communication with Mistral API
• python-dotenv: Environment variables management

🔒 Security and Privacy

• Local processing: Option for entirely on-machine processing
• Environment variables: Secure tokens via .env
• No cloud storage: Your files remain local
• Automatic cleanup: Temporary files removal

🚦 Project Status

✅ Version v2.2 - Hugging Face Spaces Integration

• 🚀 HF Spaces version: Simplified online version at VincentGOURBIN/MeetingNotes-Voxtral-Analysis
• 3 processing modes: Local, MLX, API with improved layout (main version)
• Standard Mistral models: Original Voxtral models optimized for Zero GPU (HF Spaces)
• 6 model configurations: Mini/Small + Default/8bit/4bit (main version)
• Complete diarization: Speaker identification and renaming (main version only)
• Modular summaries: 9 customizable sections with preset profiles
• Language-adaptive AI: Automatically responds in detected meeting language
• Progress tracking: Real-time progress bar with chunk-based updates
• Centralized UI management: Clean English interface with maintainable labels
• Token tracking: Comprehensive usage statistics for all modes
• Improved UX: Better API mode layout and visual organization
• Multi-platform support: Windows, macOS, Linux (main), Zero GPU (HF Spaces)

🤝 Contributing

To contribute to the project:

1. Fork the repository
2. Create a feature branch
3. Implement your changes
4. Add tests if necessary
5. Open a Pull Request

📄 License

This project is under MIT license. See the LICENSE file for more details.

---

MeetingNotes - Powered by Voxtral from Mistral AI | 🚀 Intelligent meeting analysis | 💾 Secure local and cloud processing