Privacy-first, voice-controlled AI assistant for Windows – fully offline, no cloud required.
Nexus AI listens to your voice, transcribes speech locally using Vosk (small German model), and generates smart responses or executes system commands via a local Qwen2.5-7B LLM (llama.cpp backend). Everything runs offline – your data never leaves your machine.
| Feature | Details |
|---|---|
| 🎤 Offline Speech Recognition | German, Vosk small model (vosk-model-small-de-0.15) |
| 🧠 Local LLM | Qwen2.5-7B-Instruct, quantised GGUF via llama.cpp |
| 💬 Natural Conversation | Bilingual – German & English |
| ⚙️ Command Execution | Open apps by voice: „Öffne Browser", „Öffne Rechner", „Öffne Spotify" |
| 🔊 Audio Feedback | Beep/tone on wake-word detection |
| 🖥️ Native UI | Built with Qt 6, C++17 – lightweight and responsive |
| 💬 Try it by saying 'Hallo Nexus, öffne Firefox/Explorer/Brave Browser' |
| Component | Minimum |
|---|---|
| OS | Windows 10 / 11 (64-bit) |
| RAM | 8 GB (12 GB recommended) |
| Disk | ~6 GB for models + ~500 MB for app |
| Compiler | MinGW 64-bit or MSVC (tested with Qt 6.11.0 MinGW 64-bit) |
| Build tools | CMake 3.20+, Qt Creator |
| Input | Microphone |
https://github.com/KernelPhantom-010/Nexus-AI-Desktop-Assistant.git
cd Nexus-AIWell, actually that's already it!
- Start the executable (Either
Nexus_listener.exeto clap 2 times after your pc reboots, to start Nexus-AI, orOwnAI-Release.exeif you want to use Nexus directly) - Allow microphone access if prompted by Windows
- Speak – Nexus AI starts listening immediately after launch (slower PC's may take a little longer to load the models. Check the buttom right corner if the microphone symbol is up!)
- Use natural German or English sentences:
- „Öffne den Browser" → opens your default browser
- „Öffne den Rechner" → opens Windows Calculator
- „Öffne Spotify" → launches Spotify
- „Was ist das Wetter heute?" → answered by the local LLM NOTE - If you want to use Nexus in English language, you need to use the prompt field on the right. Nexus is, for now, only concepted for german VTT recognition. The assistant beeps once when it detects speech and processes your input. Responses are displayed in the UI.
| Library | Version | Purpose |
|---|---|---|
| Qt (Optionaly if you are an developer or researcher) | 6.x | UI framework |
All dependencies are either included as submodules or linked as pre-built libraries for Windows.
Microphone not detected
- Check Windows sound settings → make sure your mic is set as the default recording device
- Run the app as Administrator once to grant permissions
LLM responses are very slow
- Increase
n_threadsin the source code - Use a smaller quantisation level (e.g.
q3_k_m) for lower RAM usage at the cost of quality
Speech not recognised / wrong language
- Make sure the Vosk model is correct. If you really want to use the english model, simply download the english Vosk model and paste the content into the
modelsfolder. - Speak clearly and at a normal pace; the small model has limited accuracy with strong accents
Build fails with missing Qt modules
- Ensure Qt 6 Multimedia and Qt 6 Widgets are installed via the Qt Maintenance Tool (if you are a developer/researcher)
- [✔️] Wake-mechanic -> Listener: Clap 2 times with your microphone activated to start Nexus
- [✔️] Text-to-speech output (offline, e.g. Piper TTS)
- Plugin system for custom commands
- English Vosk model support
- System tray mode
Pull requests are welcome! Please open an issue first to discuss larger changes.
- Fork the repository
- Create a feature branch:
git checkout -b feature/my-feature - Commit your changes:
git commit -m 'Add my feature' - Push and open a pull request
This project is licensed under the MIT License – see LICENSE for details.
- Vosk – offline speech recognition toolkit
- llama.cpp – efficient LLM inference in C++
- Qwen2.5 – open-weight language model by Alibaba DAMO Academy
- Qt Framework – cross-platform application framework