meeting-transcription

name	meeting-transcription
description	Create diarized meeting transcripts from a user-provided local audio recording using configured pyannoteAI diarization and Mistral speech-to-text models, then save an Obsidian-ready Transcript.md. Use when the user asks to transcribe, diarize, process, or generate a transcript for meeting audio and provides or points to the audio file and the desired existing or new Obsidian Meetings folder.

Meeting Transcription

Use this skill to turn a local meeting audio file into an Obsidian Transcript.md with named speakers where possible.

Required Inputs

Collect or infer these before running the CLI. Do not assume a recordings folder, vault path, speaker names, or meeting title.

• Audio source file path. It must be accepted by both Mistral STT and pyannoteAI media upload; current verified formats include .mp3 and .m4a.
• Speaker count. Prefer exact --num-speakers; use --min-speakers / --max-speakers only when the exact count is unknown.
• Speaker names if known. If names are not known, run once without --speaker-map, inspect the first clear turns, then rerun with --skip-existing and a map.
• Spoken language(s), usually nl or en. Pass --language for documentation and future modes; the current Mistral timestamped call auto-detects because language hints are not compatible with segment timestamps.
• Model configuration. Prefer repo-local config/local.env values MEETING_TRANSCRIPTION_PYANNOTE_MODEL and MEETING_TRANSCRIPTION_MISTRAL_MODEL; override per run with --pyannote-model and --mistral-model when needed.
• Output destination:
  • Existing Obsidian meeting package: pass --meeting-dir.
  • New ad hoc meeting package: pass --create-meeting --meeting-date "YYYY-MM-DD HH:MM" --meeting-title "Title" and either pass --meetings-root or rely on local MEETING_TRANSCRIPTION_MEETINGS_ROOT, MEETING_WORKFLOWS_OBSIDIAN_VAULT, or discovered vault config.
  • Non-meeting scratch output: pass --output-dir.

Workflow

1. Confirm the repo exists:

cd /path/to/meeting-workflow-skills/plugins/meeting-workflows/skills/meeting-transcription
./bin/diarized-transcript secrets-check

If the parent meeting-workflows plugin is installed and its bin/ directory is on PATH, diarized-transcript secrets-check is equivalent.

2. Use the audio file path the user gives you. If they point to a folder, search only that folder unless there is a clear reason to expand.

3. Run the transcript job.

Existing meeting folder:

./bin/diarized-transcript run "/path/to/audio.m4a" \
  --language nl \
  --num-speakers 2 \
  --speaker-map "SPEAKER_01=Speaker Name,SPEAKER_00=Other Speaker" \
  --meeting-dir "/path/to/Obsidian/Meetings/YYYY/YYYY-MM/YYYY-MM-DD HH.mm - Meeting Title"

New ad hoc meeting folder:

./bin/diarized-transcript run "/path/to/audio.m4a" \
  --language nl \
  --num-speakers 2 \
  --speaker-map "SPEAKER_01=Speaker Name,SPEAKER_00=Other Speaker" \
  --create-meeting \
  --meetings-root "/path/to/Obsidian/Meetings" \
  --meeting-date "YYYY-MM-DD HH:MM" \
  --meeting-title "Meeting Title"

If MEETING_TRANSCRIPTION_MEETINGS_ROOT is set in the environment or in repo-local config/local.env, --meetings-root may be omitted. The CLI can also derive the meetings root from MEETING_WORKFLOWS_OBSIDIAN_VAULT plus MEETING_WORKFLOWS_MEETINGS_FOLDER (default Meetings).

Use config/local.env for non-secret install defaults. Keep API keys in environment variables, Bitwarden Secrets Manager, or an ignored repo-local .env; do not put API keys in config/local.env.

First pass when speaker mapping is unknown:

./bin/diarized-transcript run "/path/to/audio.m4a" \
  --language nl \
  --num-speakers 2 \
  --output-dir "/tmp/meeting-transcription-check"

Then inspect the first minutes of Transcript.md, infer the map, and rerun with --skip-existing.

4. Review:

• Transcript.md is the user-facing transcript.
• uncertainty_report.md lists timestamp-overlap conflicts and short ambiguous turns.
• aligned_segments.json is the structured alignment output.
• Raw paid API outputs are cached under raw/ or Attachments/diarization/raw/.

Use --skip-existing whenever rerunning only to change speaker names, formatting, or Obsidian placement.

Obsidian Rules

The folder layout below is this plugin's opinionated Obsidian convention, not an Obsidian standard. Read ../../docs/obsidian-meeting-package.md when creating or explaining a folder-based meeting package, including how to configure or discover the vault path.

Store meeting notes under:

Meetings/
  YYYY/
    YYYY-MM/
      YYYY-MM-DD HH.mm - Meeting title/
        Meeting.md
        Preparation.md
        Transcript.md
        Attachments/
        Materials/

When the user says the meeting already exists in Obsidian, write into that folder as Transcript.md.

When there is no meeting folder, create the canonical folder using the meeting date/time and a concise title inferred from the transcript or provided by the user. Create Meeting.md, Attachments/, and Materials/; keep decisions, action items, and follow-up in Meeting.md, not in Transcript.md.

Use real tab or at least four-space indentation for nested Markdown lists in Obsidian files.

Guardrails

• Do not print API keys or Bitwarden values.
• Do not call the paid APIs again when raw outputs are already cached and --skip-existing can solve the task.
• Do not claim the transcript is perfect. Say it is a best-effort Precision-2 + Mistral merge with explicit uncertainty flags.
• If diarization labels look wrong, share examples and rerun from cache with corrected --speaker-map.