Obsidian Transcribe Meeting

1. Run obsidian vault info=path to get the vault root.
2. Determine the target date: use the argument if a date is provided, otherwise use today.
3. Determine context: if the user provides additional info (project name, participants, meeting topic), note it. Otherwise, these will be inferred from any surrounding daily note context or left generic.

Phase 1: Discover & Match Recordings

Try sources in order:

Option A: Auto-Detect (Primary)

If no explicit URL or file path was given:

1. Discover screen recordings:

   bash skills/obsidian-transcribe-meeting/scripts/find-screenrecordings.sh "{date}"
   

   Save the JSON output to a temp file (e.g., /tmp/screenrecs-{date}.json).

2. Discover Rodecaster recordings:

   OBSIDIAN_RODECASTER_MOUNT=/run/media/olivier/RodeCaster/RODECaster bash skills/obsidian-transcribe-meeting/scripts/find-recordings.sh "{date}"
   

   If OBSIDIAN_RODECASTER_MOUNT fails (directory not found), fall back without it to try auto-detect. Save the JSON output to a temp file (e.g., /tmp/rodecaster-{date}.json).

3. Match recordings by time overlap:

   bash skills/obsidian-transcribe-meeting/scripts/match-recordings.sh /tmp/screenrecs-{date}.json /tmp/rodecaster-{date}.json
   

   This produces groups with mode (omarchy+rodecaster, omarchy-only, rodecaster-only), video, audio, and transcribe_from fields.

4. Check idempotency — for each group, search for existing meeting notes:

   • By recording: field (for groups with Rodecaster audio):

     obsidian search query='recording: "{folder}"' path=Meetings
     

   • By video_file: field (for groups with screen recordings):

     obsidian search query='video_file: "{filename}"' path=Meetings
     

   Skip any group that already has a meeting note.

5. Present findings to the user with mode info for each group:

   Found 2 recording groups:

Option B: Google Drive URL

If the input contains drive.google.com:

1. Run the download script:

   bash skills/obsidian-transcribe-meeting/scripts/download-gdrive.sh "<url>"
   

2. Capture the output — it's the local file path.
3. Check idempotency via audio_url field (see Idempotency section).

Option C: Local File Path

If the input is a local file path, use it directly.

Phase 1.5: Gather Context

Before transcription, gather participant names and project vocabulary to improve whisper accuracy and action item attribution. Run the three tracks below in parallel to minimize added time.

Track 1: Screenshots (authoritative for participants)

Find screenshots taken during the meeting window:

bash skills/obsidian-transcribe-meeting/scripts/find-screenshots.sh "{date}" "{start_time}" "{duration_secs}"

Save the JSON result for reuse in Phase 3.5. For each screenshot found, read the image and look for:

• Participant names visible in the Teams/Zoom/Meet UI (participant panel, name labels on video tiles)
• Any on-screen text that reveals the meeting title or topic

Collect all names found across all screenshots. These are the authoritative participants. If screenshots show who is in the meeting, that list takes priority over all other sources. Do not add names from project pages or other indirect sources that contradict or dilute what the screenshots show.

Track 2: Vault context (vocabulary only, not participants)

1. Read the daily note for the target date (Daily Notes/{date}.md)
2. From the time entry lines, identify which project this recording likely belongs to (match by time proximity to the recording start time)
3. If a project is identified, read the project page (Projects/{project}.md) and extract:
   • Project-specific terminology (product names, acronyms, technical terms)
4. Also check if participants are mentioned in the time entry line itself (e.g., "sync with Bhrugen"). Names from time entry lines can be added to the participant list since they are specific to this meeting.

Important: Do NOT pull team member names from project pages to use as participants. A project page lists everyone on the team, not who attended this specific meeting. Using project-level names causes misattribution (e.g., listing a client in an internal standup).

Track 3: Quick first-pass

1. Extract the first 90 seconds of audio:

   ffmpeg -i "{audio-file}" -t 90 -c:a pcm_s16le -ar 16000 -ac 1 /tmp/firstpass_$$.wav -y -loglevel warning
   

2. Transcribe just that clip (no prompt, fast since it's tiny):

   bash skills/obsidian-transcribe-meeting/scripts/transcribe.sh /tmp/firstpass_$$.wav "{engine}"
   

3. Scan the resulting text for proper nouns and name-like words
4. Clean up: rm /tmp/firstpass_$$.wav

Combine results

Participant list priority (highest to lowest):

1. Screenshots showing meeting UI with participant names (authoritative)
2. Time entry line names (e.g., "sync with Bhrugen" means Bhrugen is a participant)
3. First-pass transcript names spoken in the audio (supplementary)

Do NOT add names from project pages to the participant list. Project pages provide vocabulary for the whisper prompt only.

Build the whisper prompt:

Meeting participants: Olivier, Kanish, Tara, Dinesh, Adam. Project: KHov. Topics: deployment pipelines, QA issues, container apps.

Store this as WHISPER_PROMPT for Phase 2 and keep the participant list for Phase 3.

If no context was gathered from any track (no screenshots, no daily note match, no names in first-pass), proceed without a prompt. This is not an error.

Phase 2: Transcribe

1. Determine the engine: check echo $OBSIDIAN_WHISPER_ENGINE — defaults to openai if unset.
2. Run the transcription script:

   WHISPER_PROMPT="{prompt from Phase 1.5}" bash skills/obsidian-transcribe-meeting/scripts/transcribe.sh "<audio-file>" "<engine>"
   

   If no prompt was gathered in Phase 1.5, omit the WHISPER_PROMPT variable.
3. Capture the JSON output — an array of {start, end, text} segments.
   • With pyannote VAD: segments also include speaker field (e.g. "SPEAKER_00")

VAD (Voice Activity Detection) is controlled by OBSIDIAN_VAD_MODEL:

• silero (default) — Silero VAD strips non-speech segments before transcription. Fast, CPU-only. Reduces Whisper hallucination and improves chunking.
• none — no VAD, uses silence trimming + fixed chunking
• pyannote — pyannote.audio speaker diarization. Provides VAD + speaker labels so the transcript is attributed per-speaker. Requires GPU (ROCm/CUDA) and HF_TOKEN for HuggingFace model access.
• pyannote,silero — comma-separated fallback chain. Tries pyannote first; if it fails, tries silero; if both fail, falls back to default chunking.

On systems with externally-managed Python (e.g. Arch), install torch and pyannote in a venv and set OBSIDIAN_VAD_VENV to its path. The script will use $OBSIDIAN_VAD_VENV/bin/python3 instead of system python3.

When VAD provides speaker labels (pyannote mode), Phase 3 should use them to attribute speech in the transcript and improve summarization (decisions, action items attributed to specific speakers).

Phase 3: Summarize & Create Meeting Note

Using the transcript segments and the participant list from Phase 1.5, generate:

1. Title: A concise meeting title (inferred from transcript content and any provided context)
2. Summary: 2-3 paragraph overview of what was discussed
3. Decisions: Key decisions made (bullet list, omit section if none)
4. Action Items: Tasks assigned with @Name attribution where possible (checklist, omit if none)
5. Open Questions: Unresolved items (bullet list, omit if none)

Action item attribution rules:

• Use the participant list from Phase 1.5 as the source of truth for who is in the meeting
• Attribute based on verbal cues in the transcript: "I'll handle that", "Can you...", "{Name}, will you..."
• When a name is mentioned right before or after a commitment, attribute to that person
• If attribution is genuinely ambiguous (no verbal cue, no name mention), use @(Team) rather than guessing
• Olivier is always a participant; attribute to him when he says "I'll..." or someone asks him directly

Format the transcript with timestamps:

[H:MM:SS] Text of the segment...

When speaker labels are present (pyannote mode), format as:

[H:MM:SS] **Speaker A**: Text of the segment...
[H:MM:SS] **Speaker B**: Text of the segment...

Speaker labels from pyannote are generic (SPEAKER_00, SPEAKER_01). Replace with real names if participants are known from context, daily notes, or people-context.md. Otherwise use Speaker A, Speaker B, etc.

Create the meeting note with obsidian create path="Meetings/{date} {Title}.md" content="{formatted content}":

---