文字起こしから議事録（タイムスタンプ引用つき）

What it does

Turns a meeting transcript into minutes in which every item can be traced to who said it and when.

• Auto-detects WebVTT, SRT and plain text and normalises them into numbered segments (start/end time, speaker, text)
• Speakers come from <v Name> tags, Name: prefixes, or Otter/Teams style header lines such as Jane Doe 0:15
• Collapses rolling duplicate captions and, with --merge, joins fragmented captions of the same speaker
• Rounds times down so a citation never points after the words start
• --cues marks candidate decisions, action items, deadlines and questions with English and Japanese keywords
• Outputs Markdown, JSON or TSV; untimed transcripts are cited by line number [L12]
• Ships a minutes template: decisions, action items, open questions

How to use

python scripts/transcript_segments.py meeting.vtt --merge 3 --cues --out segments.md

The agent reads everything and fills the template with decisions, action items with owner and due date, and open questions, each with its timestamp.

Requirements

• Python 3 (standard library only)
• If you only have audio, transcribe it first with a tool that outputs VTT or SRT

Caveats

• Keyword marks are hints: they miss implicit agreement and can fire on unrelated remarks, so items are decided from context.
• Owners and due dates that were never said are written as unassigned / not stated, never guessed.
• Speech-to-text errors in names and numbers are flagged as unclear instead of silently corrected.