| name | edit-tedxseu-podcasts |
| description | Edit long-form TEDxSEU interview podcasts from raw recordings into versioned, reviewable releases. Use when a TEDxSEU episode needs ASR-driven paper editing, host tightening, coherent guest storytelling, noise analysis and denoise auditions, a multi-quote cold open with generated instrumental music, declarative episode plans, FFmpeg rendering, transcripts, mastering, or final QA. |
TEDxSEU Podcast Editing
Overview
Use this workflow to turn long Chinese interview recordings into a coherent, versioned podcast release without modifying the originals. Keep episode_plan.json as the source of truth so every cut, music cue, transcript, and render remains reproducible.
Required Workflow
-
Inspect the real project before editing.
- Inventory recordings, planning documents, API documentation, existing ASR, prior plans, denoise caches, music assets, and deliverables.
- Read API documentation locally, but never print or copy credentials into logs, plans, reports, or Skill files.
- Record source durations, channel layouts, sample rates, and obvious recording problems.
-
Protect source material.
- Never overwrite original recordings.
- Put derived audio, ASR, plans, renders, and reports in separate working or version directories.
- Preserve source file paths and source timecodes in every kept clip.
-
Build the editorial foundation.
- Normalize ASR into speaker-labeled segments with source-relative timestamps.
- Create a content map before making a final cut.
- Define the episode thesis, target duration, required story arcs, sensitive exclusions, and candidate cold-open quotes.
- Read editorial-rules.md before selecting or tightening dialogue.
-
Test restoration before full processing.
- Analyze representative speech, pauses, and noise-only regions.
- Render short A/B candidates before denoising the full episode.
- Reject processing that pumps, changes room tone abruptly, damages consonants, or makes pauses feel unnaturally gated.
- Keep the chosen denoise recipe and audition notes with the version.
-
Design the sound package.
- Build an approximately one-minute cold open from multiple quotes, not one uninterrupted answer.
- Generate several instrumental music candidates from a narrative brief. Do not imitate named artists.
- Select music for space around Chinese dialogue, usable openings/endings, restrained dynamics, and thematic fit.
- Place BGM under the montage and duck it under speech. Use music mainly for the intro, limited chapter transitions, and outro.
-
Create a declarative plan.
- Follow episode-plan-schema.md.
- Store dialogue clips, music clips, source ranges, timeline positions, chapters, exclusions, mastering targets, and provenance.
- Validate before rendering:
python3 scripts/validate_episode_plan.py episode_plan.json \
--selection work/editorial/selection.json
- Generate the readable script.
python3 scripts/build_plan_transcript.py episode_plan.json \
--asr-dir work/asr \
--output 剪辑文字版.md
- Render and master.
- Use the bundled renderer when the plan follows the schema:
python3 scripts/render_episode.py episode_plan.json --output-dir deliverables/VERSION
- Default master target is about
-16 LUFS integrated and no higher than -1.5 dBTP.
- Produce a 48 kHz, 24-bit WAV master and a 192 kbps MP3 review copy unless the project requires another format.
-
Package every version separately.
- Each version directory must contain its audio, plan, readable transcript, version notes, and render or QA report.
- Name folders by editorial purpose, not only by a number.
- Explain what changed relative to the original recording and the previous version.
-
Perform final QA.
- Run structural validation, loudness checks, silence checks, and source-range checks.
- Listen to the full-length final render. Automated checks do not replace full listening.
- Follow the checklist in workflow.md.
Decision Rules
- Prefer long coherent passages when the guest is explaining a process, telling a personal-growth story, or creating something live.
- Tighten hosts more aggressively than guests, but remove only filler, abandoned starts, redundant confirmations, and excessive thinking gaps that do not carry meaning.
- Preserve emotion, laughter, useful overlap, and natural cadence.
- Dialogue edits normally use
20-80 ms fades and retain roughly 80-250 ms of natural space. Do not use long crossfades between spoken sentences.
- Use sensitive exclusions as machine-checked ranges, not informal notes.
- When a user dislikes a processing artifact, keep the accepted earlier version and branch from it instead of stacking more processing.
References
