media-transcriber

Video/audio transcription + summarization

A skill that transcribes video or audio files with OpenAI Whisper (local) and outputs two files: full-text md and summary md.

Supported extensions: mp3 / m4a / wav / ogg / flac / mp4 / mov / mkv / webm / avi, etc. (whisper calls ffmpeg internally, so videos can be passed as-is).

Execution steps

1. Confirm the file path specified by the user (when drag-and-dropped, enclose in quotes)
2. Run transcription with the following command:

Basic (recommended default: turbo = large-v3-turbo + Whisper natural segmentation):

python .claude/skills/media-transcriber/scripts/transcribe.py "入力ファイル" --language ja

Highest accuracy (minutes with many proper nouns, important cases):

python .claude/skills/media-transcriber/scripts/transcribe.py "入力ファイル" --model large-v3 --language ja

Lightweight/fast (for drafts):

python .claude/skills/media-transcriber/scripts/transcribe.py "入力ファイル" --model base --language ja

Force a break every N seconds (when you want to capture in roughly 5-second increments):

python .claude/skills/media-transcriber/scripts/transcribe.py "入力ファイル" --language ja --segment-seconds 5

Split by sentence (punctuation detection + max 60-second cap):

python .claude/skills/media-transcriber/scripts/transcribe.py "入力ファイル" --language ja --sentence

Word by word (for speech timing analysis; note that the number of lines increases massively):

python .claude/skills/media-transcriber/scripts/transcribe.py "入力ファイル" --language ja --word-level

1. The output full-text md (<プロジェクトルート>/YYYYMMDD_HHMMSS_{basename}.md) を Read read with
2. Based on the content of the full text, Claude itself summary md Write create with. The file name is YYYYMMDD_HHMMSS_{basename}_summary.md(same timestamp, same directory as the full-text md). The summary structure defaults to the "summary md template" below; if the user gives instructions, follow them
3. Report both the full-text md and summary md paths to the user

Arguments

Argument	Description	Default
file	Input audio/video file path (required)	—
--model	whisper model name (tiny / base / small / medium / large-v3 / turbo)	turbo
--language	Language code (e.g.: ja, en). Auto-detected when unspecified	auto-detect
--output-format	Output format (md / json)	md
--segment-seconds	Split lines every N seconds	unspecified
--sentence	Split by sentence. Re-chunk with punctuation detection + upper-limit seconds cap	false
--word-level	Output one line per word. Highest priority	false
--sentence-max-seconds	--sentence Fallback upper-limit seconds for the time. Force flush so that one sentence does not become enormous in audio without punctuation	60.0

Priority order of split granularity (higher is stronger): --word-level > --segment-seconds > --sentence > Whisper natural segmentation (default)

Settings

Project root .env Set the following environment variables in (optional. If unset, output directly under the project root):

# 例: transcripts/ 配下に出したい場合
TRANSCRIPT_OUTPUT_PATH=transcripts

Variable name	Description
TRANSCRIPT_OUTPUT_PATH	Output destination directory (default: . = directly under the project root)

Output

Example of full-text md

# 文字起こし: meeting_20260421.mp4

- 元ファイル: /path/to/meeting_20260421.mp4
- モデル: medium
- 言語: ja (auto-detected)
- 実行日時: 2026-04-21 14:30:00
- 総尺: 00:45:12

## 全文

[00:00:00] 今日は四半期の振り返りから始めましょう。
[00:00:12] まず売上の件ですが、前年同月比で...

• File name: YYYYMMDD_HHMMSS_{basename}.md（basename is the name of the input file without the extension)
• Save location: directly under the project root (default)..env の TRANSCRIPT_OUTPUT_PATH can be overridden with

Summary md template

Claude outputs this template as the default skeleton:

# 要約: meeting_20260421.mp4

- 元ファイル: /path/to/meeting_20260421.mp4
- 総尺: 00:45:12
- 文字起こし: 20260421_143000_meeting_20260421.md

## 概要

3〜5行で全体像

## キーポイント

- 箇条書き5〜8個

## トピック別まとめ

### トピック1

...

## 次のアクション

- 担当 / 期限 / 内容

• File name: YYYYMMDD_HHMMSS_{basename}_summary.md(same timestamp as the full-text md)
• Save location: same directory as the full-text md

Notes on summaries

Assuming it will also be used as client meeting minutes/records, write in the following tone:

• Negative assertive words such as "wiped out," "immediate halt," "subordinate," "left to rot," and "failure"do not use
• Presenting facts with numbers is OK. Do not deny the other party with evaluative expressions about facts
• If the user gives a different tone instruction, prioritize that

Model selection guidelines

Model	Size	Speed	Japanese accuracy	Use case
tiny	~75MB	Ultra-fast	Low	Drafting, short English
base	~150MB	Fast	Below average	Short English, speed priority
small	~500MB	Medium speed	Medium	Simple Japanese
medium	~1.5GB	Somewhat slow	High	Former standard (can be replaced with turbo)
turbo	~1.5GB	Fast	High (equivalent to v2)	Recommended default(Japanese, general purpose)
large-v3	~3GB	Slow	Highest	Many proper nouns, important meeting minutes

2025 measured comparisonIn Japanese use cases, turbo (large-v3-turbo) has the best speed/accuracy balance.medium has equivalent memory and speed and equal or better accuracy, so normally use turbo as the first choice. Only when the highest accuracy is needed, large-v3。

On first run, the model is downloaded to ~/.cache/whisper/ (only once).turbo is about 1.5GB.

Check dependent packages

If an error occurs on first run, install with the following:

pip install -U openai-whisper python-dotenv

ffmpeg is required separately (whisper calls it internally):

brew install ffmpeg