chaplin

Why we included this project

Chaplin reads your lips through a webcam and types the words you mouth silently, then passes the raw recognition through a local LLM to fix errors before inserting the text at your cursor. It is a different approach to speech-to-text, useful for hands-free typing in quiet spaces, accessibility experiments, or just a fun proof of concept for visual speech recognition. Setup is more involved than a typical app: you download the Auto-AVSR model weights, run Ollama with a small Qwen model, and launch everything from the command line. That makes it a better fit for developers comfortable wiring local components together than for end users expecting a polished installer. Since everything runs locally, it also works as a reference for combining a VSR model with an LLM correction stage in one pipeline.