Transcribe

Step 1: Transcribe

Run the transcription script bundled with this skill:

source .venv/bin/activate && python ${CLAUDE_SKILL_DIR}/scripts/transcribe.py $ARGUMENTS

The script will:

• Send the file to Deepgram API (model nova-3, language: ru, diarization enabled)
• Cache the raw API response as <file>.deepgram.json (reused on subsequent runs — no duplicate API charges)
• Generate a draft transcript as <file>.transcript.md with generic speaker labels (Speaker 0, Speaker 1, ...)

If a .deepgram.json cache already exists, the script skips the API call and reuses it.

Step 2: Auto-identify Speakers

Read the generated .transcript.md file and analyze the conversation to identify speakers:

1. Look for self-introductions: "Привет, я Никита", "Меня зовут Мария", etc.
2. Look for how speakers address each other: "Никита, расскажи...", "Как думаешь, Арсений?"
3. Look for role mentions: "я ведущий разработчик", "мы в Positive Technologies"
4. Cross-reference with known project participants if available (check .assistant/submissions/ for speaker names)

Build a proposed mapping like:

Speaker 0 → Арсений Савин (track curator, leads the conversation)
Speaker 1 → Никита Гурняк (speaker, discusses their talk topic)

Present the proposed mapping to the user and ask for confirmation or corrections using AskUserQuestion. Show a few representative quotes from each speaker to help the user verify.

Step 3: Relabel

Once the user confirms the speaker mapping, run the relabel command:

source .venv/bin/activate && python ${CLAUDE_SKILL_DIR}/scripts/transcribe.py $ARGUMENTS --relabel '{"Speaker 0": "Confirmed Name", "Speaker 1": "Another Name"}'

This rewrites .transcript.md with real names. The cached .deepgram.json is reused — no API call is made.

Step 4: Save

Ask the user where to save the final transcript. Suggest a sensible default based on the project structure (e.g., a transcripts/ directory).

Move the final .transcript.md to the chosen location with a descriptive filename.

Report the final file path to the user.

Notes

• Supported formats: mp4, mp3, wav, m4a, webm, ogg, flac, and other common audio/video formats
• Transcription of a 1-hour recording takes ~1-2 minutes
• The Deepgram nova-3 model works well with Russian language
• Cached .deepgram.json files can be large (10-50 MB) — consider adding them to .gitignore