🇺🇸 English | 🇧🇷 Português
Offline Speaker Diarization & Transcription Powerhouse
Built with Tauri, WhisperX, and Pyannote.audio
Koalizer is a high-performance desktop application designed for researchers, journalists, and developers who need to process audio with maximum privacy. Unlike cloud-based solutions, Koalizer runs entirely offline, performing state-of-the-art transcription and speaker identification on your local hardware.
- Transcription (WhisperX): Sub-segmental precision using OpenAI's Whisper models.
- Speaker Diarization (Pyannote): Distinguish "Who spoke when" with high accuracy.
- Word-level Alignment: Forced alignment for perfectly synced timestamps.
- YouTube Integration: Seamlessly fetch and process content via
yt-dlp. - Hardware Acceleration: Full support for NVIDIA GPUs (CUDA) to speed up inference.
The project utilizes a Sidecar Architecture, separating the lightweight UI from the heavy-duty AI processing.
| Component | Technology |
|---|---|
| Frontend | React, TypeScript, Tailwind CSS |
| App Shell | Tauri v2 (Rust) |
| AI Engine | Python 3.10, PyTorch, WhisperX, Pyannote.audio |
| Inter-Process | Flask-based Local API |
Koalizer/
├── frontend/ # Tauri + React Application
│ └── src-tauri/ # Rust backend & Sidecar config
├── backend/ # Python Intelligence
│ ├── app.py # Flask Entry point
│ ├── models/ # Pyannote/Whisper local weights
│ └── bin/ # External binaries (FFmpeg, yt-dlp)
└── build.py # Unified build automation script
- C++ Build Tools: Required for
pyannote.audioandwhisperXdependencies. - NVIDIA Drivers: (Recommended) For CUDA acceleration.
cd backend
python -m venv venv
source venv/bin/activate # or venv\Scripts\activate on Win
pip install -r requirements.txt
python app.py
cd frontend
npm install
npm run tauri dev
Koalizer features a "One-Click Build" system that bundles the entire Python environment into a standalone binary using PyInstaller before wrapping it into the Tauri installer.
cd backend
python build.py
This will generate a production-ready installer in frontend/src-tauri/target/release/bundle/.
Koalizer is standing on the shoulders of giants. If you use this tool in academic research, please consider citing the underlying models:
- Pyannote.audio: Bredin et al. (2020) "pyannote.audio: neural building blocks for speaker diarization".
- WhisperX: Bain et al. (2022) "WhisperX: Time-Accurate Speech Transcription of Long Audio".
This project is licensed under the MIT License. Note that the AI models (Pyannote/Whisper) are subject to their own respective licenses (MIT and Apache 2.0).
This project was developed as part of my undergraduate thesis (TCC). The full thesis document is available in Portuguese only.
The version presented in this work corresponds to Koalizer v0.1.1.