AI Meeting Summarizer: Free Local Pipeline with Whisper and Ollama

A free local meeting summarizer: faster-whisper for transcription, Ollama for summaries, n8n for the pipeline.

A local meeting summarizer that stays offline

Pipeline: recording, faster-whisper transcription with timestamps, chunking by a 15 minute window, a local LLM extracting decisions and action items per chunk, then a reduce pass merging everything into one Markdown summary, with n8n watching a folder and pushing the result to Slack or Notion.

Why chunk: a two hour meeting does not fit a small model's useful context. Splitting at 15 minutes keeps each request at 2000 to 3000 characters, where 7B and 8B models stay coherent.

Two prompt rules carry most of the quality: instruct the model to extract only what is in the transcript, and require "未定" (undecided) for missing owners and deadlines. Invented deadlines are the main failure mode of AI meeting notes.

Hardware: 16GB RAM plus an 8GB GPU is comfortable for large-v3 transcription and 8B summarization. CPU-only works but is much slower. Speaker labels need whisperX diarization or FunASR, since Whisper alone does not separate speakers.

Keep the raw transcript. Summaries are for reading, transcripts are for disputes.