Offline meeting minutes with whisper.cpp + Ollama

What they built

Meeting recordings contain client names, figures and sometimes HR data. This guide reproduces what Otter or Fireflies sell, but on your own machine: whisper.cpp for speech-to-text, then an Ollama model for a structured summary.

Setup

whisper.cpp built from source (CUDA, Metal or ROCm optional)
Whisper medium by default, large-v3 if the GPU allows
Ollama with qwen3.5:9b (or mistral-small)
ffmpeg; 8 GB RAM minimum, 16 GB comfortable

How it works

Convert audio to 16 kHz mono WAV with ffmpeg
Transcribe with whisper-cli (-otxt or -ojson for timestamps)
POST the transcript to Ollama /api/chat with a minutes template
Write the Markdown report

Results & numbers

RTX 3060 12 GB + medium: ~3 min transcription + ~30 s summary for 1 h
MacBook M2 16 GB: ~6 min total
RTX 3060 transcribes 1 h in ~4 min with large-v3

Gotchas

Ollama's default 2,048-token context silently truncates transcripts; set num_ctx
Whisper hallucinates during silences, so enable VAD
Over 90 minutes, use map-reduce chunking
whisper.cpp does not do speaker diarization

Do it yourself

Build whisper.cpp and download the medium model
ffmpeg -i meeting.m4a -ar 16000 -ac 1 -c:a pcm_s16le meeting.wav
./build/bin/whisper-cli -m models/ggml-medium.bin -f meeting.wav -otxt -of meeting
ollama pull qwen3.5:9b
Send the transcript with num_ctx 16384 and a 'do not invent names or decisions' rule

curl -s http://localhost:11434/api/chat -d '{
  "model": "qwen3.5:9b", "stream": false,
  "options": {"temperature": 0.2, "num_ctx": 16384},
  "messages": [{"role":"user","content":"Summarize decisions, action items and open questions:\n<transcript>"}]
}' | jq -r '.message.content'