Utility Transcribe

Voice Memo Transcriber

Transcribes audio using OpenAI Whisper, structures it into meeting notes, and writes one markdown file per memo to the Obsidian wiki.

Inbox folder (scan mode)

/Users/<user>/Library/CloudStorage/GoogleDrive-<your-email>@gmail.com/My Drive/99 Voice memos

Voice files arrive here automatically via Samsung → Autosync → Google Drive → Drive Desktop on Mac. After a successful wiki write, files move to the processed/ subfolder so the inbox always shows pending work.

Output target

Voice notes are written as markdown files to:

<workspace>/1 - Knowledge/99-Voice notes/

One file per memo. Filename: YYYY-MM-DD - <slug>.md (5–8 word slug derived from the BLUF). YAML frontmatter so Dataview / search-by-area works.

Step 0 — Determine mode

If scan mode finds no audio files in the inbox, tell the user "Inbox is empty" and stop.

Step 1 — Get the transcript

Check the file extension first:

For audio files, run the transcription script:

source <workspace>/.venv/bin/activate && \
python3 <workspace>/.claude/skills/utility-transcribe/scripts/transcribe.py "<audio_file_path>"

If the venv doesn't have openai or requests, install first:

pip install openai requests

Step 2 — Determine output mode

Check the user's request for a mode flag. Default to Summary if not specified.

Summary mode

Produce this structure. Be concise and executive-ready throughout.

Summary (BLUF)

2–3 sentences. Lead with the single most important point. What was decided or what matters most?

Actions

Bullet list. Each item: what, who (if named), by when (if stated). Use "TBC" if unclear.

Discussion & Notes

Full context in cleaned prose. Remove filler, preserve substance. Organise by topic if multiple threads.

Risks & Blockers

Bullet list. Only include if genuinely present in the recording — don't fabricate. If none, write "None identified."

Stakeholder Pulse

1–2 lines per person mentioned. Tone, position, concerns, alignment. If no named stakeholders, omit this section.

Detailed mode

Use when the recording contains technical depth, methodology, systems context, or analysis recommendations that would be lost in a summary. Produce an expanded structure:

Summary (BLUF)

2–3 sentences. Lead with the single most important point.

Actions

Bullet list. Each item: what, who, by when. Use "TBC" if unclear.

Discussion & Notes

Expand each topic thread fully. Include:

• Background and context
• System/data landscape (tools, sources, access status)
• What's measurable vs not — and why
• Recommended analysis approach (phased if appropriate)
• Data requirements and access blockers

Risks & Blockers

Bullet list. Flag structural, data, access, and timeline risks explicitly.

Stakeholder Pulse

1–2 lines per person. Tone, position, concerns, alignment.

Step 3 — Write to Obsidian wiki

Create one markdown file at: <workspace>/1 - Knowledge/99-Voice notes/<YYYY-MM-DD> - <slug>.md

• Date: Use the audio file's mtime (when the recording happened), not today's date.
• Slug: 5–8 words derived from the BLUF. Plain ASCII, spaces between words, no punctuation. e.g. Plan maintenance fire pilot Teeth.
• Collisions: If the target filename exists, append (2), (3), etc.

Frontmatter (YAML):

---
title: "<slug>"
type: voice-note
date: <YYYY-MM-DD>
source: "<original audio filename>"
area: <Telecommunications|D&SI|Transport|Infrastructure|PMO|Team|General>
people: [Name1, Name2]   # named participants — omit field entirely if none
tags: [voice-note, <2-4 topical tags in kebab-case>]
---

Body: Use the structured format matching the output mode (Summary or Detailed) — same section headings (## Summary (BLUF), ## Actions, ## Discussion & Notes, ## Risks & Blockers, ## Stakeholder Pulse). Detailed mode just has fuller content in each section.

Do NOT include a top-level # heading in the body — the frontmatter title and filename carry the title; Obsidian renders the filename as the H1.

Step 4 — Move processed file

After a successful wiki write, move the source audio to the processed/ subfolder of the inbox:

mv "<source_path>" "/Users/<user>/Library/CloudStorage/GoogleDrive-<your-email>@gmail.com/My Drive/99 Voice memos/processed/"

Only move on success. If the wiki write failed, leave the file in the inbox so the next scan retries it. Tell the user which files succeeded and which failed.

For single-file mode where the source isn't in the Drive inbox (e.g. an ad-hoc path the user gave you), skip the move — just tell the user the path of the new wiki note.

API Key

The script loads OPENAI_API_KEY from .env in this skill directory:

<workspace>/.claude/skills/utility-transcribe/.env

If the key is missing, prompt: "Add your OpenAI API key to .claude/skills/utility-transcribe/.env — see .env.example for format."

Supported formats

m4a, mp3, mp4, mpeg, mpga, m4a, wav, webm, ogg (OpenAI Whisper limits: 25MB per file)

Large files (auto-handled)

The script handles oversize files automatically:

1. >25 MB → compress to 32 kbps mono AAC via afconvert
2. Still >25 MB after compression (calls >~1.5 hrs) → split into 60-min chunks via ffmpeg, transcribe each, concatenate with \n\n separators No manual chunking required. Requires ffmpeg on PATH (installed via brew install ffmpeg).