generate-audio

name	generate-audio
description	Generate TTS audio from speaker_notes and upload to Vercel Blob
user-invocable	false
allowed-tools	Read, Write, Edit, Glob, Grep, Bash

Generate Audio

Generate TTS narration for all page cards with speaker_notes in a course.

Usage

/generate-audio <course-slug>

Example: /generate-audio claude-architect

Workflow

1. Read .env to load OPENAI_API_KEY and BLOB_READ_WRITE_TOKEN.
2. Scan all cards in courses/<slug>/cards/ for page cards with speaker_notes.
3. Load courses/<slug>/.enrichment-hashes.json if it exists to check for already-generated audio.
4. For each page card that needs audio (no audio field, or hash has changed): a. Call OpenAI TTS API (tts-1-hd model, voice alloy) with the speaker_notes text. b. Upload the MP3 to Vercel Blob at path courses/<slug>/audio/<card-id>.mp3. c. Update the card's frontmatter to add audio: "<blob-url>". d. Update .enrichment-hashes.json with the MD5 hash of the speaker_notes.
5. Use a consistent voice per course (default: alloy).
6. Run the media validator after completion:

   npx tsx scripts/validate-media.ts courses/<slug>
   

7. Fix any failures before completing.

OpenAI TTS API

curl https://api.openai.com/v1/audio/speech \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tts-1-hd",
    "input": "<speaker_notes text>",
    "voice": "alloy",
    "response_format": "mp3"
  }' \
  --output audio.mp3

Vercel Blob Upload

curl -X PUT "https://blob.vercel-storage.com/courses/<slug>/audio/<card-id>.mp3" \
  -H "Authorization: Bearer $BLOB_READ_WRITE_TOKEN" \
  -H "Content-Type: audio/mpeg" \
  -H "x-api-version: 7" \
  --data-binary @audio.mp3

Hash Tracking

.enrichment-hashes.json format:

{
  "01-welcome.md": { "audio_hash": "a1b2c3d4" },
  "02-agentic-loop.md": { "audio_hash": "e5f6g7h8" }
}

Hash is first 8 chars of MD5 of the speaker_notes content. Skip regeneration when hash matches.