---

name: paper-to-dialogue-video category: 内容创作 subject: 跨学科 kind: 内容创作 description: Turn a research-paper PDF into a source-grounded Chinese two-host podcast with a checked dialogue script, distinct voices, synchronized subtitles, paper figures, and a warm vertical HTML animation. Use when the user asks for 论文播客、双人讲论文、论文音频、论文视频、小红书论文动画、带字幕配音的论文解读, or wants an academic PDF converted into a conversational audiovisual episode.

Paper to Dialogue Video

Create a complete, inspectable episode rather than generating audio directly from an unchecked summary.

Workflow

1. Read the paper. Extract metadata, research question, method, essential variables, main results, author-stated limitations, and figure/table locations. Render and inspect the PDF pages when layout or figures matter.

2. Build an evidence sheet. Record every publishable number and claim with a page, section, figure, or table anchor. Separate model results, empirical observations, author interpretations, and later criticism. Read references/evidence-rules.md.

3. Plan the episode. Default to a 3–6 minute Chinese dialogue. Give the host concrete beginner questions and the explainer direct technical answers. Read references/dialogue-style.md.

4. Create episode.json. Follow references/episode-schema.md. Attach a source anchor, visual, and key point to every spoken segment.

5. Audit before speech. Check names, numbers, equations, causal wording, pronunciations, and whether each visual supports the current line. Do not synthesize an unverified draft.

6. Synthesize and assemble. Run:

   python3 scripts/build_episode.py \
     --episode /absolute/path/to/episode.json \
     --output /absolute/path/to/output-folder
   

   Prefer the free Edge neural backend for finished previews. Install it once with python3 -m pip install edge-tts imageio-ffmpeg, then run with --tts edge. It needs internet but no API key. Use --max-scenes 4 for a short voice check before synthesizing the full episode. The auto mode chooses Edge when both commands are available, otherwise it falls back to macOS say and afconvert. Use --no-audio only when neither backend is available.

7. Inspect the HTML. Open index.html; check voice identity, subtitle timing, image crops, text wrapping, mobile layout, keyboard playback, and reduced-motion behavior. Read references/visual-style.md.

8. Export the approved video. After approving the HTML and voices, run:

   python3 scripts/export_video.py \
     --episode-folder /absolute/path/to/output-folder \
     --output /absolute/path/to/output-folder/podcast.mp4
   

   Export a 1080×1440 H.264 MP4 with the complete 3:4 card background, translucent subtitle glass, active-speaker badges, evidence anchors, and synchronized audio.

9. Deliver the package. Include index.html, podcast.wav, podcast.mp4, subtitles.srt, timing.json, episode.json, 双人讲稿.md, and source-map.md.

Non-negotiable quality rules

• Keep the explanation plain, reliable, and technically sufficient. Shorten repetition, not the causal chain.
• Avoid opening with “不是……而是……”, decorative metaphors, manufactured surprise, and excessive banter.
• Let the host surface likely beginner confusion; let the explainer answer before adding nuance.
• State when a result is simulated, correlational, qualitative, or causally identified.
• Never invent a quote, number, experiment, figure, or limitation.
• Use the finished Xiaohongshu card as the full 3:4 scene background by default. Preserve it completely and place only translucent character, subtitle, section, and source layers above it.
• Use original paper crops only when explicitly requested; keep source and license notes.
• Keep one speaker per segment. Split long sentences before TTS.
• Keep stable character voices: 小鲸 uses zh-CN-YunxiNeural at about +8% rate and +2Hz pitch; 小柚 uses zh-CN-XiaoxiaoNeural at about -3% rate and -1Hz pitch. Adjust dialogue phrasing before making large pitch changes.
• Generate a 3–6 scene voice check before a full episode. Listen for role identity, unnatural numbers, English terms, and repeated mechanical pauses.
• Audit English terms, author names, symbols, and polyphonic Chinese characters separately.

Bundled resources

• scripts/build_episode.py: validate the episode, synthesize free Edge neural voices or local fallback voices, assemble WAV/SRT/timing, and render the HTML player.
• scripts/export_video.py: render the approved episode folder to a portable 3:4 H.264 MP4.
• assets/player-template.html: responsive vertical animated player.
• assets/ocean-fruit-background.png: default warm ocean-and-fruit background.
• references/: dialogue, evidence, schema, and visual rules loaded only when needed.