Video Transcribe

Extract text from video links (Douyin, Bilibili, YouTube, Kuaishou, etc.) with timestamps.

脚本位置

脚本文件相对于本 SKILL.md 所在目录：scripts/transcribe.py

执行时使用完整路径，例如：

python ~/.openclaw/skills/jeff-video-to-text-1.0.0/scripts/transcribe.py ""

Dependencies

pip install yt-dlp openai-whisper
winget install ffmpeg  # Windows
# brew install ffmpeg  # macOS
# sudo apt install ffmpeg  # Linux

First-time Setup

Model cache directory (recommended: non-system drive)

Whisper large-v3 model is ~2.9GB. Set cache to a non-system drive:

# Windows PowerShell (permanent, User-level):
[Environment]::SetEnvironmentVariable("WHISPER_DOWNLOAD_ROOT", "X:whisper_cache", "User")

# macOS/Linux (~/.zshrc or ~/.bashrc):
export WHISPER_DOWNLOAD_ROOT="~/whisper_cache"

Video download directory (optional)

# Windows PowerShell:
[Environment]::SetEnvironmentVariable("VIDEO_DOWNLOAD_DIR", "X:video_downloads", "User")

# macOS/Linux:
export VIDEO_DOWNLOAD_DIR="~/video_downloads"

yt-dlp global config (optional)

# Windows: Create %APPDATA%yt-dlpconfig (one line):
-o X:video_downloads%(title)s.%(ext)s

# macOS/Linux: Create ~/.config/yt-dlp/config (one line):
-o ~/video_downloads/%(title)s.%(ext)s

Usage

Agent call

python /scripts/transcribe.py "" [model] [prompt]

• video_url: 视频链接（必填）
• model: large-v3（默认，质量最好）或 turbo（更快，精度略低）
• prompt: 可选提示词，帮助 Whisper 识别特定领域术语

Manual use

python transcribe.py ""

Output

• Segmented text: with timestamps for editing

  [0.0s-8.0s] Text here...
  [8.0s-20.0s] More text...
  

• Full text: plain text for copying
• Video file: saved to download directory

Supported Platforms

yt-dlp supports 1000+ sites: Douyin, TikTok, Bilibili, YouTube, Kuaishou, Weibo, Xiaohongshu, Xigua Video, etc.

清理临时文件

脚本下载的视频文件会保存在 VIDEO_DOWNLOAD_DIR（默认为脚本所在目录），文件名为 temp_video.*。转录完成后如不需要保留视频，可主动删除：

# Windows
del <脚本目录>temp_video.*

# macOS/Linux
rm <脚本目录>/temp_video.*

常见问题

• 下载失败: 检查链接是否有效，部分平台可能需要 cookies（yt-dlp 支持 --cookies 参数）
• 转录很慢: CPU 模式下 large-v3 约 77 秒/分钟视频，可用 turbo 模型加速
• 中文识别差: 确保安装的是最新版 openai-whisper，且使用 large-v3 模型
• 内存不足: large-v3 需要约 10GB 内存，可用 turbo 模型减少占用

Notes

• CPU-only, no GPU required
• large-v3: ~77s per minute of video on CPU
• Chinese text optimized with punctuation
• Models stored in WHISPER_DOWNLOAD_ROOT directory
• If env vars not set, defaults to script directory