Instant voice interruption: Cut off mid-response and it listens immediately (no lag)
Fast: Optimized for 50-token, 3-sentence replies (2–4s latency)
High accuracy: ~85% transcription accuracy with adaptive VAD
Fully customizable: Add your own personalities, tweak response styles, or integrate other AI models
No install: Runs in the browser with Web Speech API + Gemini API
Demo: miles-voice-ai.vercel.app Code: github.com/mandarwagh9/miles