AI会议纪要2026：免费自托管转写+总结（Whisper方案）
会议纪要总是迟到又总漏掉待办事项。AI转写方案能在会议结束后几分钟出准确纪要——免费。
AI Meeting Notes: The Self-Hosted Stack That Works
Meeting transcription is one of the most solved AI problems in 2026. The paid tools (Otter.ai, Fireflies, Teams/Meet built-ins) work, but they send your conversations to third parties and charge per seat. The self-hosted alternative - Whisper for transcription plus an LLM for summarization - is free, private, and surprisingly easy to run.
The Architecture
- Capture. Record the meeting audio (or use a bot that joins the call and records).
- Transcribe. Whisper (106,367 GitHub stars) converts audio to text with timestamps. Use faster-whisper for speed; a 1-hour meeting transcribes in ~5 minutes on a consumer GPU, or overnight on CPU.
- Summarize. Send the transcript to an LLM (local or API) with a prompt asking for: decisions, action items with owners, and open questions.
- Store. Save the notes to a knowledge base - the accumulated notes become a searchable company memory.
The Prompt That Gets Good Notes
You are a meeting summarizer. From this transcript extract:
1. Decisions made (with context)
2. Action items (who, what, by when)
3. Open questions
4. A 3-sentence summary
Be specific. Use names as spoken.
That single prompt pattern beats generic summarize-this for most teams - the structured output is what makes notes useful.
Tools You Can Run Today
- Meeting bot - many open-source projects (and n8n workflows) join Zoom/Teams/Meet, record, and drop the audio in a folder.
- faster-whisper server - expose transcription as an API for any tool to use.
- Open WebUI or similar - a chat frontend where teammates can ask questions about meeting history.
The Privacy Angle
Self-hosting means meeting audio never leaves your infrastructure. For legal, medical, and HR meetings that is not a nice-to-have - it is the difference between recording and not recording. Whisper supports 99 languages, so multilingual teams get notes in their own language, and translation is a prompt away.
FAQ
Is Whisper good enough for meetings? Yes - with speaker diarization it handles real conversations well; heavy accents and crosstalk are the weak spots.
Do I need a GPU? Recommended for speed; CPU works for overnight batch transcription.
Can I get speaker names? Whisper diarizes speakers (Speaker 1, Speaker 2); mapping to names needs a small post-processing step.
Is this really free? The software is free; you pay for compute and storage only.
