Love it! Solving the latency problem is essential to making voice ai usable and comfortable. Your point on VAD is interesting - hadn't thought about that.
Amazing, Im happy that you like it! The voice agent indeed uses Silero VAD(v5) - There is an onnx wasm file available for the underlying VAD model, so we can perfectly run that with the onnxruntime-web, just like we run the TTS and STT models!
Maker here. A few technical notes that might be interesting to this crowd:
The Voice Agent chains three models in the browser: Whisper for STT → a local LLM → Kokoro/SpeechT5 for TTS. All inference runs on-device via WebGPU. The latency isn't amazing yet, but the fact that it works at all with zero backend is kind of wild.
The landing page has an auto-playing demo that generates speech locally as soon as you visit — you'll hear it typewrite and speak three sentences. That was important to me because "runs in your browser" sounds like marketing until you actually hear it happen.
Happy to go deep on the WebGPU inference pipeline, model conversion process, or anything else.