Meeting Transcribe

Turns a raw audio recording into a complete text transcript with timestamps. Priority is on transcript completeness and accuracy — no speaker identification.

The output is ready to be fed into the meeting-sync skill.

When to Use

• The user attaches or references an audio file (.m4a, .mp3, .wav, .aac, .ogg, .flac)
• The user mentions an iPhone voice memo or a meeting recording
• Preparing material before running meeting-sync

Engine

faster-whisper with large-v3 (CTranslate2, CPU int8) — fast on Apple Silicon, no token required, excellent French + English accuracy. Built-in Silero VAD trims silences and reduces hallucinations.

iPhone: Record with Voice Memos → AirDrop the .m4a to the Mac → process with transcribe.py. See setup.md.

Outputs Produced

The .md file is the artefact intended for meeting-sync.

Procedure

Step 0 — Collect Context

1. Audio path — confirm the absolute path. If multiple files are attached, ask which one.
2. Language — default to auto-detect (omit --language). If the filename or conversation clearly indicates a language, pass --language fr or --language en to speed things up and improve accuracy.
3. Output folder — default to transcripts/ at the workspace root. Created if missing.

Do not start processing until the audio path is known.

Step 1 — Verify / Bootstrap the Environment

Run ./scripts/check_env.sh. It will:

• Check for ffmpeg (install via brew install ffmpeg if missing)
• Create/reuse a dedicated venv at ~/.cache/meeting-transcribe/venv
• Install faster-whisper if needed

Step 2 — Run Transcription

Invoke ./scripts/transcribe.py:

transcribe.py <audio_path> \
  --language fr \
  --model large-v3 \
  --output-dir transcripts/

Key flags:

• --language — fr or en. Omit for auto-detection.
• --model — large-v3 (default, best accuracy). Use medium if the user reports the run is too slow.
• --beam-size — default 5. Raise to 8 for slightly higher accuracy on hard recordings.
• --compute-type — default int8 (fastest on Apple Silicon).

The first run downloads large-v3 (~1.5 GB) to ~/.cache/. Subsequent runs are instant to start.

Step 3 — Quality Check & Summary

Before reporting back:

• Open the first ~30 lines of the .md and confirm:
  • Timestamps are monotonic and cover the full duration
  • No long stretches of repeated text (Whisper hallucination — rerun with --beam-size 8 if present)
  • Language matches expectation

Report to the user:

• Path to the .md and .json
• Total duration & detected language
• Approximate word count
• A 5–10 line preview
• Suggest the next step: "Run meeting-sync on this transcript to update the project docs."

Markdown Output Format

The .md file produced in Step 2 has this shape:

# Meeting Transcript — <basename>

- **Source audio:** <path>
- **Duration:** HH:MM:SS
- **Language:** fr
- **Model:** faster-whisper / large-v3
- **Segments:** N

---

**[00:00:03]** Bonjour à tous, on va commencer par …

**[00:00:11]** Oui, juste avant, je voulais ajouter …

…

Anti-patterns (Do Not Do)

• Do not call OpenAI's plain Whisper API — the local engine is free, private, and equally accurate.
• Do not add speaker labels or guess who said what — this skill intentionally does not do diarization.
• Do not truncate or summarise the transcript — completeness is the priority.
• Do not commit the audio file to the repo.
• Do not silently swallow errors — if transcription fails, surface the message.

References

• setup.md — one-time setup (ffmpeg, optional iPhone workflow)
• transcribe.py — transcription script
• check_env.sh — environment bootstrap