What problem does it solve? Creating accurate subtitles for course videos is tedious and error-prone: speech recognition introduces typos into narration whose script is already known, and burned-in subtitles are irreversible if wrong. This Skill builds a complete subtitle timeline for a ChatCut project by splitting known narration text directly (zero typos), running Whisper recognition only on segments with real human voice, and handing results to the user for proofreading before burning. ## Core Features & Use Cases - Dual-track subtitle generation: TTS narration segments use the known original script split by punctuation and display width; only screen recordings with real human audio go through Whisper ASR and are flagged as source: "asr". - Proofreading workflow: ASR-sourced cues are visually distinguished in the UI (dashed border, striped background, badge) so users only review machine-recognized lines; known text always overrides recognition on overlaps. - Burn-in and SRT export: After user confirmation, subtitles are burned into the video via libass with local CJK font files, and an identical standalone SRT file is exported. - Dependency handling: Checks for libass, CJK fonts, and Whisper; guides the user through downloading the whisper-base model (141 MiB from huggingface.co) with explicit consent. - Use Case: A course creator finishes narration synthesis in ChatCut, clicks "Generate Subtitles" in the workbench, reviews only the ASR-flagged cues from a demo recording segment, then burns the verified subtitles into the final video and exports an SRT for the course platform. ## Quick Start Ask the assistant to generate subtitles for the current ChatCut project, proofread the recognition-sourced cues it flags, then burn them in and export the SRT.
