Video Transcript Method

Problem

The core information in a video is usually carried by speech, but engineers still need to decide whether CC subtitles exist, download audio, run Whisper, and manually clean the transcript. Without a consistent workflow, the output can include filler, repeat segments, mechanical time-based cuts, or key points that drift from the original meaning.

How It Works

The skill treats video transcript extraction as a three-stage pipeline:
- Subtitle first: parse the URL, detect the platform and video ID, then check CC subtitles with yt-dlp --list-subs; if Chinese subtitles are available, use them directly instead of invoking speech recognition.
- Whisper fallback: when subtitles are missing, download the mp3 audio track and run Whisper to produce timestamped text; medium is the default, with FP32 fallback when FP16 is unavailable.
- Semantic cleanup: segment by content logic, add chapter titles and time ranges, extract a key-point table and exact quotes, and output .txt or .md.

Boundaries

It is useful for online videos supported by yt-dlp where speech carries the main information; danmaku are not CC subtitles, and login state or platform limits can affect subtitle access. Whisper accuracy depends on audio quality, and users should still verify completeness, timestamp alignment, and copyright.