content-to-podcast-video

name	content-to-podcast-video
description	Converts long-form text or Markdown articles into mobile-optimized 1080x1920 presentation slides, single-speaker podcast audio with auto voice selection, synchronized subtitles, and a 1.25x speedup MP4 video. Use this skill when asked to create a video, slide deck + podcast, or short video from an article or document.

Content to Podcast Video Pipeline

This skill automates the complete end-to-end pipeline for converting an article or text document into a vertical (1080x1920) mobile-optimized video featuring:

1. Mobile Presentation Slides: Large, readable vertical cards (1.png, 2.png, ...).
2. Reserved Subtitle Area: Bottom 280px of every slide image left blank/clean for larger subtitles.
3. Host Portrait Overlay (Optional): Embedded host card on the bottom-right (circular avatar, name, title/intro, and soundwave indicator).
4. Single-Speaker Podcast: Conversational script with domain-tailored TTS voice (default 1.25x speed rate="+25%").
5. Text-Authoritative Video Synthesis: Content-driven slide image display (each image stays on screen exactly while its section is being spoken; no image looping) and high-precision TTS boundary-aligned ASS subtitles.

---

Pipeline Architecture & Steps

Step 1: Mobile Slide Generation (1080x1920)

• Cover Slide (1.png):
  • Title horizontally & vertically centered on the cover.
  • Adaptive font sizing based on title line count:
    • 1 line: font-size: 140px (extra-large, maximum impact).
    • 2 lines: font-size: 110px (still large and bold).
    • 3 lines (only when title cannot be shortened): font-size: 80px (slightly smaller but still prominent).
  • Title should be kept to 1–2 lines whenever possible. Only use 3 lines when the title absolutely cannot be further condensed.
  • Use text-align: center and flexbox centering (justify-content: center; align-items: center).
  • High-impact mobile hero layout with bold gradients.
• Content Slides (2.png, 3.png, ...):
  • Section title (50–60px), body bullet points (28–34px).
  • High contrast glassmorphic cards.
• Bottom Reserved Zone:
  • Keep padding-bottom: 300px or margin-bottom: 280px on all slide containers.
  • Do NOT place footers, text, or visual elements in the bottom 280px area.
  • This extra-large reserved zone ensures the enlarged subtitles (42px) never overlap slide content.
• Rendering: Uses Headless Chrome (--window-size=1080,1920 --screenshot=X.png).

Step 2: Podcast Script & Voice Selection

• Script Adaptation: Rewrites technical/long-form text into a natural 3–8 minute conversational single-speaker podcast script.
• Slide Mapping Markers: Insert explicit [SLIDE: 1], [SLIDE: 2], ... markers in script.txt corresponding to each generated slide image.

  [SLIDE: 1]
  大家好，欢迎收听本期播客。本期我们要讨论的是分布式数据库体系架构。
  
  [SLIDE: 2]
  第一部分，存储与计算分离...
  
  [SLIDE: 3]
  第二部分，多版本并发控制...
  

• Auto Voice Selection Matrix:

  Topic / Domain	Voice Code	Characteristics
  Tech / Database / Architecture / Science	zh-CN-YunjianNeural	Male, authoritative, steady
  Business / Product / News	zh-CN-YunxiNeural	Male, energetic, natural
  Storytelling / Lifestyle / General	zh-CN-XiaoxiaoNeural	Female, warm, expressive
  Regional / Casual Discussion	zh-CN-liaoning-XiaobeiNeural	Female, lively

• Audio Generation: Run edge_tts at 1.25x speed (rate="+25%" by default).

Step 3: High-Precision Subtitle Alignment & ASS Formatting

• TTS Sentence Boundary Alignment (Zero Timing Drift):
  • Subtitle timestamps are captured directly from Microsoft Edge TTS engine boundary events (SentenceBoundary), eliminating premature or delayed subtitle cues.
  • Long sentences are sub-divided into balanced multi-line cues strictly within the exact sentence start and end boundaries.
• Line Length & Formatting:
  • Limit lines to max 14–16 Chinese characters per line (to ensure text stays within 1080px with side margins).
  • Automatically break long clauses across max 3 lines (\N) when needed.
  • Two-line wrapping preferred; three-line only when content cannot fit in two.
  • Prevent splitting English words or punctuation pairs (——).
  • Both sides must have margin padding (MarginL/MarginR: 60px) to avoid text touching screen edges.
• ASS Styling:
  • PlayResX: 1080, PlayResY: 1920
  • Fontsize: 42 (enlarged for better readability on mobile).
  • Alignment: 2 (Bottom Center).
  • MarginL: 60, MarginR: 60 (side padding to keep text within safe area).
  • MarginV: 100 (positioned in the bottom 280px reserved area).
  • BorderStyle: 3, BackColour: &HA0080B15 (dark semi-transparent capsule box).

Step 3.5: Host Portrait Card Overlay (Optional)

If --avatar <image> is passed to build_video.py, a host portrait card is composited in the bottom-right corner of the video frame, sitting inside the subtitle reserved zone to the right of the subtitle text.

Card anatomy (420 × 220 px, RGBA transparent background):

Element	Details
Circular Avatar	Center-cropped from the provided image, masked to 140px circle with a cyan glowing ring
Host Name	Bold white text, 36px, PingFang SC
Host Title / Intro	Cyan text (#67E8F9), 26px, auto-wraps at ~10 chars
Soundwave Bars	5 bars of varying height in cyan, static decorative waveform
Background	Dark semi-transparent card (#080B15, 78% opacity) with rounded corners and cyan border

Position in video: x=640, y=1410 (20px from right edge, inside the 280px bottom zone).

Generation script:

• python3 scripts/make_avatar_overlay.py --avatar <path> --name <name> --title <title> --output <out.png>
• Requires Pillow (pip install pillow).
• Called automatically by build_video.py when --avatar is provided.

Fallback: If Pillow is missing or avatar file not found, the overlay is silently skipped and the video renders normally without it.

Step 4: Content-Driven Video Encoding (No Image Looping)

• Content-Synchronized Slide Display:
  • Images are NOT looped.
  • Each slide image (1.png, 2.png, ...) is shown only while speaking that specific page/section's content.
  • Slide i stays on screen from the speech start of [SLIDE: i] until the speech start of [SLIDE: i+1] (the last slide remains visible until the podcast ends).
• Audio Normalization: -af "loudnorm".
• FFmpeg Hardware Accelerated Encoding (Mac Videotoolbox):
  • Without avatar overlay:

    ffmpeg -y -hwaccel videotoolbox \
      -f concat -safe 0 -i concat.txt \
      -i podcast.mp3 \
      -vf "fps=24,scale=1080:1920,format=yuv420p,subtitles=podcast.ass" \
      -c:v h264_videotoolbox -b:v 2M -r 24 -tag:v avc1 \
      -c:a aac -ar 44100 -ac 2 -b:a 128k \
      -af "loudnorm" -shortest -movflags +faststart output.mp4
    

  • With avatar overlay (auto-generated _avatar_card.png, composited via movie + overlay filter):

    ffmpeg -y -hwaccel videotoolbox \
      -f concat -safe 0 -i concat.txt \
      -i podcast.mp3 \
      -vf "fps=24,scale=1080:1920,format=yuv420p,subtitles='podcast.ass' [base]; \
           movie='_avatar_card.png',format=rgba [ovrl]; \
           [base][ovrl] overlay=640:1410" \
      -c:v h264_videotoolbox -b:v 2M -r 24 -tag:v avc1 \
      -c:a aac -ar 44100 -ac 2 -b:a 128k \
      -af "loudnorm" -shortest -movflags +faststart output.mp4
    

---

Helper Automation Scripts

The skill provides two reusable Python scripts in scripts/:

1. generate_slides.py: Renders 1080x1920 slide PNGs (1.png..N.png) from a JSON spec.
   • python3 scripts/generate_slides.py --dir <out_dir> --spec <spec.json>
   • Auto-detects Chrome/Chromium/Edge; every slide keeps the bottom 300px subtitle zone clean.
2. build_video.py: TTS audio + high-precision ASS subtitles + default/custom host portrait overlay + content-driven non-looping subtitled MP4.
   • python3 scripts/build_video.py --script-file <script.txt> [--dir <d>] [--slides N] [--topic tech] [--avatar <path>] [--host-name <name>] [--host-title <title>] [--output output.mp4]
   • --dir defaults to the script file's directory (outputs land next to the script).
   • --avatar: Path to host avatar image (JPG/PNG). Defaults to pre-rendered references/_avatar_card.png (skips image synthesis). Pass none to disable.
   • --host-name: Display name of the host/speaker (default: digoal德哥).
   • --host-title: Title/intro line of the host (default: 数据库 & AI 专家).
   • Automatically parses [SLIDE: N] markers in script.txt to calculate exact slide display durations.
   • Prechecks that ffmpeg is on PATH; falls back to libx264 if h264_videotoolbox fails.

Slide Spec JSON (for generate_slides.py)

{
  "slides": [
    {"type": "cover", "kicker": "小标签", "title": "主标题<br>第二行",
     "subtitle": "副标题", "badges": ["标签1", "标签2", "标签3"]},
    {"type": "content", "tag": "第一笔账", "title": "标题 <span class='hl'>高亮</span>",
     "note": "一句说明",
     "stats": [{"big": "9.3万", "lbl": "标注<br>第二行"
         
                 
  

• Cover title font auto-sizes by <br> line count (1→140px, 2→120px, 3→88px). Keep it to 1–2 lines.
• Inline HTML allowed in text fields: <br>, <b>, <span class='hl'>, <span class='warn'>.
• Card style: accent (cyan) | purple | amber.

---

Usage Instructions

When executing this skill:

1. Read the input document/markdown file.
2. Formulate 5–8 key slide topics and write a slide spec JSON + the single-speaker podcast script formatted with [SLIDE: 1], [SLIDE: 2], ... section tags.
3. Choose the appropriate TTS voice based on topic (auto-selected by build_video.py --topic).
4. Generate slides: python3 scripts/generate_slides.py --dir <d> --spec <spec.json>.
5. Build the video (default uses pre-rendered references/_avatar_card.png host portrait overlay unless --avatar none is specified):
   • With default pre-rendered host card: python3 scripts/build_video.py --script-file <script.txt> --dir <d> --slides <N> --topic <t>
   • With custom host card: python3 scripts/build_video.py --script-file <script.txt> --dir <d> --slides <N> --topic <t> --avatar <photo.jpg> --host-name "姓名" --host-title "职位介绍"
   • Without host card: python3 scripts/build_video.py --script-file <script.txt> --dir <d> --slides <N> --topic <t> --avatar none
6. Verify outputs (1.png~N.png, podcast.mp3, podcast.ass, output.mp4). Inspect the video with ffmpeg -i output.mp4.

[!TIP] The host portrait card requires Pillow: pip install pillow. Pass --avatar none to disable the host portrait overlay.