This was my halloween costume for our party this year. Basically there's a microphone listening to everything I say, which it's sending to a small LLM running on my laptop, which is told to try to output any insecurities or character flaws in the person speaking.
Some more specifics:
- The screen is just my phone, inside of a transparent pice of plastic sewed into a shirt. I wrote a simple app which takes the audio data, sends it via a stream to the laptop, then recieves text and displays it.
- The server is a small Python script, using LMStudio to run a local LLM (qwen3-4b-2507 in this case) and whisper (small) to transcribe the audio.
- The microphone is a "YRF Wireless Lapel Microphone".
It worked well! It transcribed things well even when stood next to the speakers, especially since all it needed was a few keywords to be able to guess the topic (e.g. it hearing "forgot to..." would make it output "POOR PLANNING"). It was also quite fun the day after the party to read the transcript of everything I had said.