xclgordon/plaud-pipeline

This Python project automates the process of turning meeting recordings into a searchable knowledge base. The pipeline takes audio files, processes them through several stages, splitting the audio into segments by speaker (called "diarization"), identifying who each speaker is, generating AI-written summaries, and finally building a wiki-style website you can browse and search. The tool is designed to handle recordings from a device called Plaud, though it can work with audio files from other sources. It uses speaker-recognition AI to figure out which participant is talking at any given moment, with a reported 94% accuracy rate for identifying named speakers. It merges notes from the recording device with AI-generated summaries to produce structured meeting notes that include decision tables, action items assigned to specific people, and key discussion points. The resulting wiki is a web interface you can run locally, it supports full-text search and a question-answering mode (called RAG, short for Retrieval-Augmented Generation) where you can ask questions and get answers drawn from your meeting history with cited sources. Optionally, you can connect a Telegram bot to receive daily digest notifications. The pipeline is idempotent, meaning if it's interrupted partway through, you can restart it without losing progress. It is written in Python and requires a GPU with at least 8GB of memory for the speaker-detection step. It is currently optimized for Chinese-language recordings.