Transcribe Diarize

Transcribe Diarize

Metadata

• Trigger when: the user explicitly wants speaker labels, diarization, or structured speaker-aware transcript output.
• Do not use when: a plain fast transcript is enough.

Skill Purpose

Produce a diarized transcript with the narrow settings required for speaker-aware output instead of treating every transcription job as a plain text task.

Instructions

1. Gather the audio path, desired output location, and any known-speaker reference files. Check that OPENAI_API_KEY is set before doing anything else.
2. Run ~/.codex/skills/transcribe/scripts/transcribe_diarize.py with gpt-4o-transcribe-diarize, diarized_json, and any provided --known-speaker hints under a stable output path such as output/transcribe/.
3. Validate the transcript, speaker labels, and segment boundaries. State clearly when speaker confidence is weak because references were missing or the audio is noisy.

Non-Negotiable Acceptance Criteria

• The task explicitly needs diarization or speaker labels.
• Output format is diarization-capable and the chosen model matches that need.
• Known-speaker hints are used only when they actually exist; they are never invented.
• Confidence or segmentation caveats are called out when labels are weak.

Output

• The diarized transcript path.
• The model, response format, and any known-speaker hints used.
• A short note on label confidence, segmentation quality, or caveats.
• Next skill options (only if needed): $transcribe-basic — use when the user decides speaker labeling is unnecessary after all.