VoiceAssistant
A fully local, full-duplex voice assistant for Windows that lets a user interrupt it mid-sentence and answers using visual context captured from the desktop the instant the user speaks.
VoiceAssistant supports real full-duplex audio: a user can interrupt the assistant at any moment with natural barge-in, interruption latency under 100 milliseconds, and the interrupting speech becomes the next question. A screenshot is captured the instant the user speaks and sent to the LLM for visual context, toggleable for privacy.
The entire pipeline runs locally on a single GPU shared between speech-to-text and the LLM through GGML and Vulkan, keeping the design portable across AMD, NVIDIA and Intel GPUs, with a target end-to-end latency under two seconds from the end of speech to the first spoken syllable of the reply. Transcripts stream in real time and the reply is spoken sentence by sentence, in a Brazilian Portuguese voice.
Binaries and models are downloaded on first run, with a visible progress bar; the window opens instantly, and servers and models only load once the user presses the central button. Settings such as voice-activity thresholds, speech speed, the vision toggle and volume apply live, without a restart.