ai-podcast-pipeline（AI 播客管道）

AI Podcast Pipeline

⚠️ Security Notice

This skill may trigger antivirus false positives due to legitimate use of:

• base64 decoding: Used ONLY to decode audio data from Gemini TTS API responses (standard practice for binary data in JSON)
• subprocess calls: Used ONLY to invoke ffmpeg for audio/video processing
• Environment variables: Reads API keys from user-configured environment (GEMINI_API_KEY)
• Network requests: Calls Google Gemini API for text-to-speech generation

All code is open source and auditable in this repository. No malicious behavior.

Build end-to-end podcast assets from Trend/QuickView-* content.

Core Workflow

1. Select source QuickView file.
2. Generate script (full or compressed mode).
3. Build dual-voice MP3 (Gemini multi-speaker, chunked for reliability).
4. Generate full-text Korean subtitles (no ellipsis truncation).
5. Render subtitle MP4 with tuned font/size/timing shift.
6. Build thumbnail + YouTube metadata.
7. Deliver final package.

Step 1) Select Source

Prefer weekly QuickView file from your configured Quartz root.

If user gives wk.aiee.app URL, map to local Quartz markdown first.

Step 2) Generate Script

Read and apply:

• references/podcast_prompt_template_ko.md

Modes:

• Full mode: 15~20 minutes
• Compressed mode: 5~7 minutes (core tips only)

Rules:

• no system/meta text in spoken lines
• host intro once at opening only
• conversational Korean, short sentences, actionable
• save script in archive/

Step 3) Build Audio (Gemini Multi-Speaker, Reliable)

Preferred: chunked builder (timeout-safe)

# Set API key via environment (required)
export GEMINI_API_KEY="<YOUR_KEY>"

# Run from skills/ai-podcast-pipeline/
python3 scripts/build_dualvoice_audio.py \
  --input <script.txt> \
  --outdir <outdir> \
  --basename podcast_full_dualvoice \
  --chunk-lines 6

Single-pass (short scripts)

python3 scripts/gemini_multispeaker_tts.py \
  --input-file <dialogue.txt> \
  --outdir <outdir> \
  --basename podcast_dualvoice \
  --retries 3 \
  --timeout-seconds 120

Default voice mapping (2026-02-10 fixed):

• Callie (female) → Kore
• Nick (male) → Puck

Output: MP3 (default delivery format)

Step 4) Build Korean Subtitles (Full Text)

Use full-text subtitle builder (no ... truncation):

python3 scripts/build_korean_srt.py \
  --script <script.txt> \
  --audio <final.mp3> \
  --output <outdir>/podcast.srt \
  --max-chars 22

Step 5) Render Subtitled MP4 (Font + Timing)

Use renderer with adjustable font and timing shift:

python3 scripts/render_subtitled_video.py \
  --image <thumbnail.png> \
  --audio <final.mp3> \
  --srt <podcast.srt> \
  --output <outdir>/final.mp4 \
  --font-name "Do Hyeon" \
  --font-size 27 \
  --shift-ms -250

Notes:

• shift-ms negative = subtitle earlier (for lag fixes)
• If text clipping occurs, lower font-size (e.g., 25~27)
• keep text inside safe area; avoid overlap with character/object

Step 6) Build Thumbnail + YouTube Metadata

# Set API key via environment (required)
export GEMINI_API_KEY="<YOUR_KEY>"

python3 scripts/build_podcast_assets.py \
  --source "<QuickView path or URL>"

Reference (layout/copy guardrails):

• references/thumbnail_guidelines_ko.md

Step 7) Final Delivery Checklist

Always include:

1. source used
2. final MP3 path
3. subtitle MP4 path + size
4. thumbnail path
5. YouTube title options (3)
6. YouTube description

Reliability Rules

• Gemini timeout on long input: use chunked builder (build_dualvoice_audio.py)
• Subtitle clipping: reduce font size and increase bottom margin
• Subtitle lag: adjust --shift-ms (usually -150 to -300)
• Keep generated assets under Telegram practical limits

Security Notes

• API keys must be passed via environment variables (GEMINI_API_KEY), not hardcoded.
• Never paste raw keys into prompts, logs, screenshots, or public posts.
• Recent hardening: thumbnail generation now passes keys via env (not CLI args).

References

• references/podcast_prompt_template_ko.md
• references/workflow_runbook.md
• references/thumbnail_guidelines_ko.md