pipecat-ai/smart-turn ? reverse-engineered prompt

Build me a Python project that can listen to microphone audio and tell me when a person is really done speaking, not just when the sound gets quiet. I want it to work like a smart turn detector for voice assistants, using raw PCM audio and handling short pauses, filler words, and natural speech patterns better than simple silence detection.

Please include a simple command line demo I can run locally to record from my mic and print turn predictions, plus an inference script I can call from my own code. It should work on CPU and also support a faster GPU version if available. If you can, add clear setup steps, sample usage, and a way to train or fine tune the model on my own audio data too. Look up any current docs online if you need to, but keep the result easy to run and understand.