2026
Voice Assistant
100% local voice AI
Full-duplex voice assistant with screen vision for Windows, with interruptible conversation and screen-aware answers, fully offline, no cloud and no API keys.
A voice assistant for Windows with real full-duplex audio: you can interrupt it mid-sentence, and the interruption becomes the next question. The moment you speak, it captures the screen and uses the image as context for its answer.
The whole pipeline runs locally on a single GPU stack (GGML + Vulkan) shared by speech recognition and the LLM, which keeps it portable across AMD, NVIDIA and Intel GPUs. The target is under 2 seconds from the end of your speech to the first syllable of the reply.
Natural interruption
Barge-in under 100 ms, detected by Silero VAD while the assistant is still speaking.
Screen vision
A screenshot goes to the model with your question; it can be turned off in settings for privacy.
Streaming voice
The transcript arrives in real time and the answer is spoken sentence by sentence as it is generated, with Piper.
No cloud
whisper.cpp and llama.cpp on Vulkan, with models downloaded from Hugging Face on first run. No API key and no data leaving the machine.