Skip to content

Meeting transcription (meet)

Shipped: 2026-05-31 (PR #699, closes #698) Module: modules/services/meeting-transcribe.nix Status: deployed on razer (client) and p620 (client + processor)

Overview

One-button meeting recording, transcription, and AI summarization. Press SUPER+SHIFT+M to start recording mic + system audio; press again to stop. ~2–5 minutes later a desktop notification announces a markdown brief at ~/meetings/YYYY-MM-DD-HHMM.md containing:

  • TL;DR — 2–3 sentence summary
  • 🎯 Your action items — - [ ] checkboxes assigned to you (Ollama separates "you" from other speakers using context)
  • 📋 Action items (others) — assigned to other speakers
  • 🔑 Key decisions
  • ❓ Open questions
  • 🚩 Flagged moments — timestamped mentions of configurable keywords (default: blocker, deadline, urgent, incident, risk, escalate)
  • 🗺️ Topic timeline — meeting carved into 3–7 segments
  • 👥 Participants — diarized speakers with talk-time % estimates
  • 📜 Full transcript — diarized when a HuggingFace token is configured

All processing runs on the user's own hardware — no SaaS, no audio leaves the local network.

Pipeline

record  ──▶  transcribe              ──▶  summarize          ──▶  render
ffmpeg       whisperX large-v3            ollama mistral          bash + jq
+ pulse      + pyannote diarize           format=json             checkboxes
~1 MB/min    ~10x realtime (16C CPU)      ~30s for 1h transcript  instant
  1. Record — ffmpeg with two PulseAudio inputs (mic + monitor of default sink), mixed via amix, encoded to 16 kHz mono Opus at 24 kbps. Detached via setsid + nohup so it survives the keybind shell exiting. SIGINT on stop, 5-second poll, SIGKILL fallback.
  2. Transcribe — whisperX (pkgs.whisperx 3.8.5) on CPU, int8 quantized. Diarization via pyannote/speaker-diarization-3.1 when HF token is available.
  3. Summarize — Ollama (mistral-small3.1) via /api/chat with format: "json" and a strict schema embedded in the user prompt.
  4. Render — bash + jq queries pull the JSON apart and emit the markdown brief.

Topology

Two roles, both configured by the same module option:

Host Role processHost installProcessor
razer Client only "p620" false
p620 Client + processor "local" true

razer records locally and offloads heavy work to p620 over Tailscale SSH (rsync up → ssh meet-process → rsync brief back). p620 records AND processes locally. The same meet CLI runs on both; behaviour is determined at runtime by cfg.processHost.

Configuration

Client-only host (razer)

features.meetingTranscribe = {
  enable = true;
  processHost = "p620";        # SSH-reachable host where meet-process lives
  installProcessor = false;
  userName = "Olaf";
  userEmail = "olaf@freundcloud.com";
};

Client + processor host (p620)

age.secrets = lib.mkIf (builtins.pathExists ../../secrets/api-huggingface.age) {
  api-huggingface.file = ../../secrets/api-huggingface.age;
};
features.meetingTranscribe = {
  enable = true;
  processHost = "local";
  installProcessor = true;
  huggingfaceTokenFile =
    if builtins.pathExists ../../secrets/api-huggingface.age
    then config.age.secrets."api-huggingface".path
    else null;
  ollamaUrl = "http://localhost:11434";
  userName = "Olaf";
  userEmail = "olaf@freundcloud.com";
};

Available options

Option Type Default Notes
enable bool false Installs the meet CLI on this host.
processHost string "local" "local" runs whisperX + Ollama here. Anything else is an SSH host name.
installProcessor bool false Installs whisperX + meet-process. Must be true if processHost = "local".
huggingfaceTokenFile path or null null Path to HF token file. Required on processor for diarization; gracefully degrades when missing.
ollamaUrl string "http://p620:11434" Ollama API base URL. Override to http://localhost:11434 on p620.
ollamaModel string "mistral-small3.1" Must be pulled on the Ollama host.
whisperModel string "large-v3" One of tiny, base, small, medium, large-v3.
language string "en" e.g. en, no, da.
outputDir string "~/meetings" Per-user; tilde expanded at runtime.
userName string required Helps Ollama identify "you" in the transcript.
userEmail string required Same.
flagKeywords list of string [ "blocker" "deadline" "urgent" "incident" "risk" "escalate" ] Timestamped into the Flagged section.

Setup

One-time, after first deploy

For diarization, you need a HuggingFace account + accepted EULAs on two pyannote models. The pipeline works without this — it just falls back to plain transcription with no speaker labels.

  1. Sign up at huggingface.co/join.
  2. Accept the terms on:
  3. pyannote/speaker-diarization-3.1
  4. pyannote/segmentation-3.0
  5. Generate a read token at huggingface.co/settings/tokens.
  6. Add it to agenix on a machine that has the user key:
./scripts/manage-secrets.sh edit api-huggingface
# paste the real token, save, exit
  1. Deploy:
just quick-deploy p620

Recipients

The HF token is encrypted to allUsers ++ [ p620 ] — only p620 needs it at runtime; razer never sees it.

Usage

Commands

meet start      # Start recording mic + system audio
meet stop       # Stop recording, dispatch transcription, return immediately
meet toggle     # Start if idle, stop if recording (used by the keybind)
meet status     # Show current recording state (PID, elapsed time)
meet process F  # Process an existing audio file F
meet help       # Show subcommands

Keybind

SUPER+SHIFT+M is wired in home/desktop/gnome/keybindings.nix (slot custom5):

"org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/custom5" = {
  binding = "<Super><Shift>m";
  command = "meet toggle";
  name = "Meeting record/transcribe/summarize";
};

Verify after deploy:

gsettings get org.gnome.settings-daemon.plugins.media-keys custom-keybindings

Should list custom0/ through custom5/.

Per-user state

Recording state (PID, start timestamp, audio path) lives in $XDG_RUNTIME_DIR/meet/ — auto-cleaned on logout, so no stale state across sessions.

Troubleshooting

"pactl unavailable" on start

pulseaudio package must be installed (the module includes it). Confirm with pactl info. On NixOS with services.pipewire.pulse.enable = true the binary comes from pipewire-pulse.

Recording is empty / silent

Default sink might not have a monitor. Check:

pactl list sources short | grep monitor

You should see ${default_sink}.monitor. If not, your default sink is something exotic (e.g. a hardware loopback) — switch the default and retry.

whisperX hangs at "pyannote/speaker-diarization-3.1"

The HF token file is missing or the EULAs aren't accepted. Either:

  • Accept both pyannote models' EULAs and add the token, OR
  • Remove huggingfaceTokenFile from the module config — the pipeline will drop diarization and produce a plain transcript.

"Remote processing failed" on razer

p620 isn't reachable, or meet-process isn't installed there. Check:

ssh p620 'which meet-process'
ssh p620 'systemctl is-active ollama'

If meet-process is missing, p620 needs installProcessor = true and a redeploy.

Brief is empty or LLM returned invalid JSON

meet-process falls back to a transcript-only brief when Ollama returns unparsable JSON. Logs are in the SSH-side /tmp/meet-remote.log (when running remotely) or stderr (when running locally). The audio file is kept on disk so you can retry with meet process <file>.

Caveats

  • HuggingFace EULA dance — three clicks across two models plus a token generation. One-time, but annoying.
  • CPU whisperX — ~10× realtime on a 16-core CPU. A 1-hour meeting takes ~6 minutes to process. ROCm CTranslate2 wheels aren't yet in nixpkgs; when they land, switch with --device rocm.
  • Speaker-identity heuristic — the LLM identifies which SPEAKER_NN is "you" using context (who others address by name, who hosts, etc.). Reliable for 5+ people; occasionally misfires in 1-on-1s.