Posted on
• Artificial Intelligence

Artificial Intelligence Meeting Summaries

Author
  • User
    linuxbash
    Posts by this author
    Posts by this author

AI-Powered Meeting Summaries from the Linux Terminal (Bash-first Guide)

Ever left a meeting with a full recording but no time to write minutes? Or wished someone else would pull out the decisions and action items for you? With today’s speech-to-text and language models, you can automate accurate, private meeting summaries right from your Linux terminal—no SaaS required.

In this guide, you’ll build a simple, reliable Bash pipeline that:

  • Converts your meeting audio to clean text

  • Summarizes it into clear, structured minutes (Agenda, Key Points, Decisions, Action Items)

  • Automates the process for any new recording that lands in a folder

All with open tools you control.

Why this matters (and works)

  • Time savings: Manual minutes consume hours per week. Automation turns recordings into minutes in minutes.

  • Better recall: Transcripts let you search, link, and verify what was said.

  • Privacy-first: Run transcription and summarization locally—keep sensitive data off third-party servers.

  • Reliability: Whisper-based transcription is state-of-the-art; local LLMs can generate structured outcomes if you give them a good prompt.

  • Bash-native: The workflow glues together standard Linux tools you already use.

1) Install prerequisites

We’ll need:

  • ffmpeg for audio processing

  • inotify-tools to watch a folder for new files (optional but useful)

  • Python + pip for Whisper CLI

  • Ollama for local LLM summarization

Pick your package manager below and run the commands.

Debian/Ubuntu (apt)

sudo apt update
sudo apt install -y ffmpeg inotify-tools python3 python3-pip curl git

Fedora/RHEL/CentOS (dnf)

sudo dnf install -y ffmpeg inotify-tools python3 python3-pip curl git

If ffmpeg isn’t found on your Fedora/RHEL variant, enable RPM Fusion (then rerun the ffmpeg install):

sudo dnf install -y https://download1.rpmfusion.org/free/fedora/rpmfusion-free-release-$(rpm -E %fedora).noarch.rpm
sudo dnf install -y ffmpeg

openSUSE (zypper)

sudo zypper refresh
sudo zypper install -y ffmpeg inotify-tools python3 python3-pip curl git

Install Whisper CLI (transcription)

Whisper requires PyTorch; pip will install a CPU wheel automatically on most Linux distros.

python3 -m pip install --upgrade pip
python3 -m pip install --upgrade openai-whisper torch

Validate:

whisper --help

Install Ollama (local LLM for summarization)

curl -fsSL https://ollama.com/install.sh | sh

Start the service and pull a model. Llama 3 8B is a good balance; Mistral 7B also works.

ollama serve &
ollama pull llama3:8b
# or
ollama pull mistral

Note: CPU works but is slower; a GPU will accelerate both Whisper (if you later install CUDA-enabled torch) and Ollama.

Optional fallback summarizer (if you don’t want a local LLM yet):

python3 -m pip install --upgrade sumy

2) Transcribe any meeting audio to text

Clean and normalize the audio with ffmpeg, then transcribe with Whisper.

# Convert to mono, 16kHz WAV for best results
ffmpeg -y -i meeting.mp4 -ac 1 -ar 16000 -af loudnorm meeting.wav

# Transcribe to plain text (choose a model: tiny, base, small, medium, large)
whisper meeting.wav --model small --language en --output_format txt --output_dir out
# This produces: out/meeting.txt

Tips:

  • If your meeting language isn’t English, omit --language en and let Whisper auto-detect.

  • Larger models increase accuracy but need more compute.

3) Summarize the transcript with a local LLM

Pipe the transcript into Ollama and prompt for structured minutes.

cat out/meeting.txt | ollama run llama3:8b -p 'You are a meticulous meeting-notes assistant.
Summarize the meeting transcript in {{ .input }} into a concise Markdown document with:

- Title (+ date if found)

- Attendees (guess if necessary)

- Agenda (bulleted)

- Key Points (bulleted)

- Decisions (D1, D2, ...)

- Action Items (A1, A2, ... with owner and due date if mentioned)

- Risks/Blockers

- Next Steps
Be specific, avoid fluff, and quote exact numbers/dates where present.' > out/meeting.summary.md

Open out/meeting.summary.md and you should see clean, structured minutes.

No LLM? Use a classic extractive fallback (less rich but instant):

sumy lex-rank --length=15% --file=out/meeting.txt > out/meeting.summary.md

4) One-command workflow script

Create a single script that:

  • Normalizes audio

  • Transcribes with Whisper

  • Summarizes with Ollama (or falls back to Sumy)

  • Writes a Markdown file with a timestamped name

Save as ai-minutes.sh and make it executable.

#!/usr/bin/env bash
set -euo pipefail

# Config
WHISPER_MODEL="${WHISPER_MODEL:-small}"
LANG="${LANG:-en}"  # set to "" to auto-detect
SUM_MODEL="${SUM_MODEL:-llama3:8b}"

infile="${1:-}"
if [[ -z "${infile}" || ! -f "${infile}" ]]; then
  echo "Usage: $0 <audio_or_video_file>"
  exit 1
fi

outdir="${2:-out}"
mkdir -p "$outdir"

base="$(basename "$infile")"
stem="${base%.*}"
ts="$(date +%Y%m%d-%H%M%S)"
work="${outdir}/${stem}-${ts}"
mkdir -p "$work"

echo "[1/4] Normalizing audio..."
ffmpeg -y -i "$infile" -ac 1 -ar 16000 -af loudnorm "$work/audio.wav" >/dev/null 2>&1

echo "[2/4] Transcribing with Whisper ($WHISPER_MODEL)..."
lang_flag=()
[[ -n "$LANG" ]] && lang_flag+=(--language "$LANG")
whisper "$work/audio.wav" --model "$WHISPER_MODEL" --output_format txt --output_dir "$work" "${lang_flag[@]}" >/dev/null 2>&1

transcript="$work/audio.txt"
if [[ ! -s "$transcript" ]]; then
  echo "Transcription failed or produced empty output."
  exit 2
fi

echo "[3/4] Summarizing..."
summary="$work/summary.md"
if command -v ollama >/dev/null 2>&1; then
  cat "$transcript" | ollama run "$SUM_MODEL" -p 'You are a meticulous meeting-notes assistant.
Summarize the meeting transcript in {{ .input }} into a concise Markdown document with:

- Title (+ date if found)

- Attendees (guess if necessary)

- Agenda (bulleted)

- Key Points (bulleted)

- Decisions (D1, D2, ...)

- Action Items (A1, A2, ... with owner and due date if mentioned)

- Risks/Blockers

- Next Steps
Be specific, avoid fluff, and quote exact numbers/dates where present.' > "$summary"
else
  echo "Ollama not found; using extractive fallback (sumy)."
  python3 -m pip show sumy >/dev/null 2>&1 || python3 -m pip install --user sumy >/dev/null 2>&1
  sumy lex-rank --length=15% --file="$transcript" > "$summary"
fi

echo "[4/4] Done."
echo "Transcript: $transcript"
echo "Summary:    $summary"

Usage:

chmod +x ai-minutes.sh
./ai-minutes.sh meeting.mp4

5) Auto-summarize every new recording dropped into a folder

Use inotify-tools to watch a directory and run the pipeline for each completed file.

Save as watch-minutes.sh:

#!/usr/bin/env bash
set -euo pipefail

WATCH_DIR="${1:-incoming}"
OUT_DIR="${2:-out}"

mkdir -p "$WATCH_DIR" "$OUT_DIR"

echo "Watching $WATCH_DIR for new files..."
inotifywait -m -e close_write --format '%w%f' "$WATCH_DIR" | while read -r file; do
  case "$file" in
    *.mp3|*.wav|*.m4a|*.mp4|*.mkv|*.webm)
      echo "Detected new media: $file"
      ./ai-minutes.sh "$file" "$OUT_DIR" || echo "Failed to process: $file"
      ;;
    *)
      ;;
  esac
done

Run it:

chmod +x watch-minutes.sh
./watch-minutes.sh ./incoming ./out
# Now drop files into ./incoming and watch summaries appear in ./out

Real-world tips

  • Improve accuracy: Use a better Whisper model (e.g., medium) and ensure clean audio (headsets > room mics).

  • Multilingual: Omit --language to auto-detect or set LANG="" in the script.

  • Customize style: Edit the prompt in ai-minutes.sh to match your team’s format (e.g., include Jira ticket patterns).

  • GPU acceleration: Install CUDA-enabled torch for Whisper and ensure Ollama uses your GPU.

  • Privacy and policy: Even local pipelines handle sensitive data; set disk permissions and clean up raw audio if required.

Conclusion and next steps

With a few commands, you now have a private, Linux-native system that turns meetings into actionable minutes—accurately and on autopilot.

Your next steps:

  • Run ai-minutes.sh on your last meeting and review the result.

  • Tune the prompt to your team’s template.

  • Put watch-minutes.sh on a server or CI box and point your recorder uploads to the watch folder.

If this saved you time, share the script with your team—or extend it with calendar integration and automatic attendee detection. Happy automating!

Related posts