Offline Meeting Transcriber

Privacy-first, AI-powered meeting transcription with speaker diarization, completely offline!

Transform your meetings into searchable, speaker-labeled transcripts without sending any data to the cloud. This tool combines advanced speech recognition with speaker diarization to create professional meeting records.

Features

• 100% Offline - No internet required, complete privacy
• Dual Audio Capture - Records both microphone and system audio
• Speaker Diarization - Automatically identifies different speakers
• Multi-language Support - Configurable language detection
• Cross-platform - Works on Windows and Linux
• Fast Processing - Optimized for CPU inference
• Multiple Output Formats - Plain text and speaker-labeled transcripts
• Easy Setup - Simple CLI interface

Quick Start

Prerequisites

• Python 3.8 or higher
• Git LFS (for downloading models)
• Audio input device (microphone)
• ~2GB free disk space for models

Installation

Quick Install (Recommended)

Linux/Mac:

git clone https://github.com/whitehole07/offline-meeting-transcriber.git
cd offline-meeting-transcriber
chmod +x install.sh
./install.sh

Windows:

git clone https://github.com/whitehole07/offline-meeting-transcriber.git
cd offline-meeting-transcriber
install.bat

The installation script will automatically:

• Create a virtual environment
• Install all Python dependencies
• Download Whisper model (~1.5GB) via git LFS
• Download SpeechBrain speaker model (~100MB) via git LFS
• Set up the complete environment
• Test the installation

Manual Installation

If you prefer to install manually:

1. Clone the repository

   git clone https://github.com/whitehole07/offline-meeting-transcriber.git
   cd offline-meeting-transcriber

2. Create virtual environment

   python -m venv venv
   
   # Windows
   venv\Scripts\activate
   
   # Linux/Mac
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

4. Download models

   • Whisper Model: Systran/faster-whisper-medium (~1.5GB)
     • Speech-to-text transcription model
     • Optimized for CPU inference
     • Supports multiple languages
   • SpeechBrain Model: speechbrain/spkrec-ecapa-voxceleb (~100MB)
     • Speaker diarization model
     • ECAPA-TDNN architecture
     • Trained on VoxCeleb dataset

Model Details and Manual Download

If you need to download models manually:

Whisper Model (Systran/faster-whisper-medium):

git clone https://huggingface.co/Systran/faster-whisper-medium models/faster-whisper-medium

• Purpose: Converts speech to text
• Architecture: Transformer-based encoder-decoder
• Size: ~1.5GB
• Languages: 99 languages supported
• Optimization: CPU-optimized version of OpenAI Whisper

SpeechBrain Model (speechbrain/spkrec-ecapa-voxceleb):

git clone https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb models/spkrec-ecapa-voxceleb

• Purpose: Identifies different speakers in audio
• Architecture: ECAPA-TDNN (Extended Context Aggregation)
• Size: ~100MB
• Training: VoxCeleb dataset (1M+ utterances from 7K+ speakers)
• Features: Speaker embedding extraction and clustering

Usage

Start Recording

# Record with microphone + system audio
python cli.py start

# Record system audio only (no microphone)
python cli.py start --no-mic

Stop Recording and Process

Press Ctrl+C to stop recording and automatically process the audio.

Output Files

After processing, you'll find these files in output/YYYYMMDD/:

• recording_HHMMSS.wav - Original audio recording
• transcription_HHMMSS.txt - Plain text transcription
• diarized_HHMMSS.txt - Speaker-labeled transcription

Example Output

Plain Transcription:

When you become fascist or communist or anarchist in those years, you can also simply be someone who never reasons...

Diarized Transcription:

SPEAKER_01 (00:00-01:10): When you become fascist or communist or anarchist in those years, you can also simply be someone who never reasons and therefore it's true yes I go with my friends and beat up those others...

SPEAKER_00 (01:10-02:14): We're not doing that well, no we're not doing that well guys, if it goes well you make me laugh...

Configuration

Edit config.py to customize:

WHISPER_MODEL = "medium"  # base, small, medium, large
WHISPER_LANGUAGE = "en"   # Language code

Troubleshooting

Model download fails:

python -c "from faster_whisper import WhisperModel; WhisperModel('medium')"

Poor transcription quality:

• Ensure clear audio input
• Check microphone levels
• Try different Whisper model sizes
• Verify language setting matches audio

Recording doesn't stop:

• Press Ctrl+C to stop recording and process audio
• The system will automatically transcribe and diarize after stopping

Installation issues:

• Make sure you have Python 3.8+ installed
• Ensure you have Git LFS installed (required for model downloads)
  • Linux/Mac: git lfs install (after installing git-lfs package)
  • Windows: Install Git LFS from https://git-lfs.github.io
• Ensure you have internet connection for model downloads
• If models fail to download, try running the installation script again
• Check that you have sufficient disk space (~2GB for models)
• On Windows, make sure you're running the batch file as administrator if needed
• If git clone fails, try: git lfs pull in the model directories

Platform-Specific Notes

Windows:

• Requires pyaudiowpatch for system audio capture
• WASAPI loopback support needed
• May need audio driver updates

Linux:

• Requires PulseAudio or PipeWire
• May need additional audio packages
• Check device permissions

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• OpenAI Whisper - Speech recognition
• SpeechBrain - Speaker diarization
• faster-whisper - Optimized Whisper implementation
• PyAudio - Audio I/O

Project Stats

---

Found a bug? Open an issue

Have a feature request? Start a discussion