meeting-copilot

name

meeting-copilot

description

A two-machine live meeting copilot. A Windows laptop streams two audio channels (your mic, and a loopback of the other side's voice) to a parent machine, which transcribes them, drives a teleprompter you read from, watches every incoming utterance against a ledger of pre-agreed facts, and answers off-script questions. Ships the parent daemons, the Windows capture agent, and worked config examples. Consumes raw transcripts only — never a meeting-AI summary. 会議に同席してリアルタイムで進行ナビ・前提監視・台本外質問への回答を行う2台構成のモニタ。子機(Windowsノート)が2系統の音声を親機へ送り、親機が文字起こし・カード生成・画面配信を行う。親機の常駐一式・子機の取り込み一式・設定の実例つき。入力は逐語のみで、議事録AIの要約は使わない。

meeting-copilot

会議に同席して、進行と前提を見張るモニタ一式

English abstract — Two machines. The child (a Windows laptop, in the meeting) captures two physical audio channels — T = your microphone, G = a WASAPI loopback of the default speaker — and streams both to the parent over TCP with a shared token. The parent transcribes each channel separately (speaker identity comes from the channel, never from diarization), then runs three layers over the transcript: a rule-only warden (copilot.py) that advances the agenda and fires boundary alarms with no LLM at all; a premise watcher (premise_watch.py) that classifies every guest utterance against a ledger of pre-agreed facts as contradiction / already-known / new; and an answerer (answerer.py) for off-script questions. A teleprompter (viewer2.py) serves two different pages — a prompt screen only you see on your phone, and a stage window you actually screen-share. Everything case-specific lives in config files; the code carries no customer data. Written mainly in Japanese; the structure is language-independent.

---

0. 全体像 — これは2台構成である

一番よくある取り違えが「1台で完結する道具だと思う」こと。必ず2台要る。