AgentSkillsCN

transcribe

使用 Whisper 将音频和视频文件转录为韩语字幕。当需要进行转录、从媒体文件中提取语音或文本、将音频/视频转换为文字，或生成字幕时，可使用此技能。支持 ./sources/ 目录下的文件，并将输出保存至 ./output/。可处理 .mp4、.avi、.mov、.mkv、.webm 视频文件，以及 .mp3、.wav、.m4a、.aac、.ogg、.flac 音频文件。

SKILL.md
--- frontmatter
name: transcribe
description: Convert audio and video files to Korean transcripts using Whisper. Use when asked to transcribe, extract speech/text from media files, convert audio/video to text, or generate transcripts. Supports files in ./sources/ directory with output to ./output/. Works with .mp4, .avi, .mov, .mkv, .webm videos and .mp3, .wav, .m4a, .aac, .ogg, .flac audio files.

Transcribe

Convert audio and video files to Korean text transcripts using OpenAI Whisper.

Quick Start

Run the transcription script with one or more media files:

bash
.venv/bin/python .claude/skills/transcribe/scripts/transcribe.py sources/audio.m4a

The script:

  1. •Extracts audio from video files (skips for audio files)
  2. •Transcribes using Whisper with Korean language setting
  3. •Generates filename from transcript content + timestamp
  4. •Saves to ./output/ directory

Workflow

Single File Transcription

bash
.venv/bin/python .claude/skills/transcribe/scripts/transcribe.py sources/[FILE]

Example:

  • •User: "sources/audio.m4a를 전사해줘"
  • •Action: Run script with default base model

Multiple Files

bash
.venv/bin/python .claude/skills/transcribe/scripts/transcribe.py sources/file1.m4a sources/file2.mp4

Example:

  • •User: "sources/ 폴더의 모든 m4a 파일을 전사해줘"
  • •Action: Find all .m4a files, pass to script

Custom Model

Use --model flag to specify Whisper model (tiny, base, small, medium, large):

bash
.venv/bin/python .claude/skills/transcribe/scripts/transcribe.py sources/video.mp4 --model tiny

Model selection:

  • •tiny: Fastest, less accurate (good for testing, short files)
  • •base: Default, balanced speed/accuracy
  • •small: Better accuracy, slower
  • •medium/large: Best accuracy, very slow

Example:

  • •User: "tiny 모델로 sources/video.mp4를 전사해줘"
  • •Action: Run with --model tiny

File Duration Considerations

Check file duration before transcription to set user expectations:

bash
ffmpeg -i sources/[FILE] 2>&1 | grep Duration

Rough processing time estimates (base model):

  • •2-5 min audio: ~30 seconds
  • •10-20 min audio: ~1-2 minutes
  • •1+ hour audio: ~5-10 minutes

For long files (>30 min), inform user of expected wait time and suggest using tiny model for faster results.

Output Format

Generated files are saved to ./output/ with naming pattern:

code
[first 30 chars of transcript]_YYYYMMDD_HHMMSS.txt

Example: 네 엉덕션을 미야 사는 신의 순놈을_20260108_183045.txt

Supported File Types

Video: .mp4, .avi, .mov, .mkv, .flv, .wmv, .webm Audio: .mp3, .wav, .m4a, .aac, .ogg, .flac, .wma

Scripts

transcribe.py

Main script for transcription workflow. Handles:

  • •Video to audio extraction (ffmpeg)
  • •Whisper transcription
  • •Output file naming and saving

Usage:

bash
.venv/bin/python .claude/skills/transcribe/scripts/transcribe.py <files...> [--model MODEL] [--output-dir DIR]

extract_media_to_text.py

Lower-level utility for single file transcription. Use transcribe.py instead for standard workflows.