Audio Transcription with Speaker Diarization

A Python script that transcribes audio files using OpenAI's Whisper and identifies different speakers using machine learning clustering techniques.

Features

• 🎙️ Accurate transcription using OpenAI Whisper (local processing, no API required)
• 👥 Speaker diarization - automatically identifies different speakers
• ⏰ Timestamp support - precise timing for each speech segment
• 🗣️ Multi-language support - currently optimized for Russian
• 🔧 Configurable - adjustable chunk size, model quality, and number of speakers
• 💾 No external dependencies - runs completely offline after initial setup

Requirements

• Python 3.8+
• FFmpeg (for audio processing)
• macOS, Linux, or Windows

Installation

1. Clone the repository:

   git clone <your-repo-url>
   cd whisper-speaker-diarization

2. Create and activate virtual environment:

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

4. Install FFmpeg:

   macOS (using Homebrew):

   brew install ffmpeg

   Ubuntu/Debian:

   sudo apt update
   sudo apt install ffmpeg

   Windows: Download from https://ffmpeg.org/download.html

Usage

Basic Usage

python transcribe_audio.py path/to/your/audio.wav

Interactive Mode

python transcribe_audio.py
# Script will prompt you to enter the audio file path

Configuration

Edit the script to adjust settings:

CHUNK_LENGTH_MIN = 10      # Audio chunk size in minutes
MODEL_NAME = "base"        # Whisper model: tiny, base, small, medium, large
NUM_SPEAKERS = 2           # Expected number of speakers
OUTPUT_FILE = "transcription.txt"  # Output filename

Whisper Models

Model	Size	Speed	Accuracy
tiny	39MB	Fastest	Basic
base	142MB	Fast	Good
small	466MB	Medium	Better
medium	1.5GB	Slow	Very Good
large	2.9GB	Slowest	Best

Output Format

The script generates a timestamped transcription with speaker identification:

[00:00 - 00:05] Спикер 1: У нас будет группы 5 человек и по 4.
[00:05 - 00:11] Спикер 2: Мы группы образуем спортивным образом.
[00:11 - 00:14] Спикер 1: Почитаемся на первую, вторую, первую и первую.

How It Works

1. Audio Loading: Uses Whisper's built-in audio loading with FFmpeg
2. Chunking: Splits long audio into manageable chunks (default 10 minutes)
3. Transcription: Each chunk is transcribed using Whisper
4. Feature Extraction: Extracts audio features for each speech segment:
   • Signal energy
   • Zero crossing rate
   • Spectral centroid and bandwidth
   • Mel-frequency coefficients
5. Speaker Clustering: Uses K-means clustering to group segments by speaker
6. Output Generation: Combines transcription with speaker labels and timestamps

Troubleshooting

FFmpeg Not Found

[ERROR] FFmpeg не найден!

Solution: Install FFmpeg following the installation instructions above.

SSL Certificate Errors

The script automatically handles SSL certificate issues during model download.

Memory Issues

For very long audio files, consider:

• Using a smaller Whisper model (tiny or base)
• Reducing CHUNK_LENGTH_MIN
• Processing shorter segments

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

License

MIT License - see LICENSE file for details.

Acknowledgments

• OpenAI Whisper for speech recognition
• scikit-learn for machine learning clustering
• FFmpeg for audio processing

Roadmap

• GUI interface
• Real-time processing
• More sophisticated speaker diarization
• Support for more languages
• Docker containerization
• Web interface