I kept wanting to talk to my local models instead of typing, but every voice setup wanted a GPU, shipped my audio to the cloud, or was macOS-only. So I built one that's none of those — and I benchmarked it, so these are real measured numbers, not vibes. One command installs the whole stack and wires it into your agent. Then you just talk. Everything runs on CPU and stays off your GPU (your GPU is busy running the actual LLM): Silero VAD — knows when you start/stop talking, no push-...