JA
← Back to portfolio

In production

Meeting minutes

A system that turns a meeting recording into minutes with a summary, agenda items, decisions and action items. It is in use for internal meetings at Ogaki Municipal Hospital. The demo on this page recreates the approach with fictional meetings.

For demonstration only, with fictional meetings. It is implemented separately from the system in actual use, and every meeting, attendee and remark is fictional.

The problem
Writing minutes means listening to the recording again, transcribing it and summarizing it, which takes several times as long as the meeting. An LLM is faster, but it can write decisions or numbers that nobody said, so its output cannot be used as is.
In production
It is used to write minutes of internal meetings at Ogaki Municipal Hospital.
My role
Planning, design, implementation and evaluation, on my own.
Technology in the real system
Python, MLX (inference on Apple Silicon), Whisper large-v3 (transcription), Llama-3.3-Swallow-70B and gpt-oss-120b (agenda extraction), Qwen3.8-27B (final write-up). It runs entirely on one Mac Studio; no audio or text leaves the machine.
How this demo differs
This demo is a small implementation written from scratch for display. Dictionary correction, checking against the transcript and merging agenda items are actually computed when this page is built. The LLM outputs (agenda extraction and the final write-up) are prepared examples. The real system also asks an LLM to confirm merges and checks the write-up against the transcript again.

Demo: from transcript to minutes

The system handles Japanese meetings, so the demo meetings are in Japanese. Every meeting, attendee and remark is fictional.

Both models extract the same agenda items from the same remarks. Every check passes, and items quoting the same remark are merged one by one.

Input

The title, date and attendees are taken from the input as is. The LLM does not write them.

Title
内科医局会
Date
2026-11-05
Attendees
医師A（司会）、医師B、医師C、医師D、医局秘書E

1. Transcript and dictionary correction

4 fixes by dictionary

Transcription errors are not fixed by the LLM; the dictionary only applies fixed replacements.

  1. 00:00:05それでは11月の医局会を始めます。最初に当直表Before: 当直票の件です。
  2. 00:00:2112月の当直表Before: 当直票の希望は、11月20日までに医局秘書に提出してください。締め切りを過ぎた分は調整できません。
  3. 00:01:02次に学会出張です。12月の学会に行く人は、出張届を2週間前までに出してください。
  4. 00:01:30術前Before: 術善の説明に使う同意書の書式が新しくなりました。来週から新しい書式を使います。
  5. 00:02:05麻酔科Before: ますい科とのカンファレンスは、これまでどおり毎週火曜です。
  6. 00:02:30ほかに無ければ終わります。お疲れさまでした。

2. Extraction by two models, and the check

Each model extracts agenda items with a supporting quote. Items whose quote is not in the transcript, and points with numbers not in the transcript, are dropped.

LLM outputs are prepared examples. Correction, checking and merging are actually computed.

Model A

  • 12月の当直表の希望提出

    Kept
    Quote: 「12月の当直表の希望は、11月20日までに医局秘書に提出してください」
    • 希望は11月20日までに医局秘書へ提出する
    • 締め切り後の希望は調整しない

  • 学会出張の届出

    Kept
    Quote: 「出張届を2週間前までに出してください」
    • 出張届は2週間前までに提出する

  • 術前説明の同意書の書式変更

    Kept
    Quote: 「来週から新しい書式を使います」
    • 来週から新しい書式を使う

Model B

  • 当直表の希望の締め切り

    Kept
    Quote: 「締め切りを過ぎた分は調整できません」
    • 締め切りは11月20日

  • 学会出張

    Kept
    Quote: 「12月の学会に行く人は」
    • 出張届は学会の2週間前までに出す

  • 同意書の新しい書式

    Kept
    Quote: 「術前の説明に使う同意書の書式が新しくなりました」
    • 来週から切り替える

Dropped by the check

Nothing was dropped.

3. Merging agenda items

Items that quote the same remark or have similar headings are merged into one.

  • 12月の当直表の希望提出

    Source: A + BSame remark quoted
    • Model A: 12月の当直表の希望提出
    • Model B: 当直表の希望の締め切り

  • 学会出張の届出

    Source: A + BSame remark quoted
    • Model A: 学会出張の届出
    • Model B: 学会出張

  • 術前説明の同意書の書式変更

    Source: A + BSame remark quoted
    • Model A: 術前説明の同意書の書式変更
    • Model B: 同意書の新しい書式

4. Minutes

The minutes are built from the merged items. Each item links back to the remark it is based on.

For demonstration only, with fictional meetings. It is implemented separately from the system in actual use, and every meeting, attendee and remark is fictional.

内科医局会

2026-11-05 ・ 医師A（司会）、医師B、医師C、医師D、医局秘書E

Summary

12月の当直表の希望は11月20日までに医局秘書へ提出する。学会出張の届出は2週間前まで。術前説明の同意書は来週から新しい書式に切り替える。

LLM outputs are prepared examples. Correction, checking and merging are actually computed.

Agenda items

  1. 1. 12月の当直表の希望提出

    Source 00:00:21
    • 希望は11月20日までに医局秘書へ提出する
    • 締め切り後の希望は調整しない
    • 締め切りは11月20日

  2. 2. 学会出張の届出

    Source 00:01:02
    • 出張届は2週間前までに提出する
    • 出張届は学会の2週間前までに出す

  3. 3. 術前説明の同意書の書式変更

    Source 00:01:30
    • 来週から新しい書式を使う
    • 来週から切り替える

Decisions

  • 来週から新しい同意書の書式を使う

Action items

TaskOwnerDue
12月の当直表の希望を提出する医師全員11月20日

Evaluation of the real system

Scored against minutes finished by people, the current ensemble approach does clearly better than the earlier single-model approach, in both content and structure.

My own evaluation, not computed by the demo on this page.