← Back to Skills Marketplace
whisper-transcribe-summarize
by
zhangjinjin-gitgit
· GitHub ↗
· v1.0.3
· MIT-0
83
Downloads
0
Stars
0
Active Installs
4
Versions
Install in OpenClaw
/install whisper-transcribe-summarize
Description
本地 Whisper 语音转文字，自动生成清理文本、书面整理稿和结构化总结稿。支持音频和视频输入，全程离线。
README (SKILL.md)
whisper-transcribe-summarize
本地语音转文字 + 文本清理 + 整理稿 + 总结稿。全程离线，不依赖外部接口。
适用场景
- 本地离线语音转文字
- 把 mp3 / wav / m4a / mp4 / mov 等音视频转成文字
- 转录后自动生成清理文本、书面整理稿、结构化总结稿
- 下载 Whisper 模型到本地，重复使用
包含脚本
scripts/download_whisper_model.py— 模型下载scripts/local_whisper_transcribe.py— 转录
依赖
python3ffmpegopenai-whisper（未安装时执行python3 -m pip install -U openai-whisper）
模型下载
python3 scripts/download_whisper_model.py medium
可选模型：tiny / base / small / medium / large
推荐：中文用 medium，追求速度用 base 或 small。模型下载到 ~/.cache/whisper。
转录命令
python3 scripts/local_whisper_transcribe.py "/路径/文件.mp4"
python3 scripts/local_whisper_transcribe.py "/路径/文件.mp3" --model medium
python3 scripts/local_whisper_transcribe.py "/路径/文件.mp3" --output "/路径/结果.txt"
支持音频和视频输入，视频通过 ffmpeg 自动提取音频。
输出
默认输出 \x3C源文件名>_whisper.txt，指定 --output 可自定义路径。
执行流程
- 确认文件存在。
- 未安装 whisper 则先安装。
- 模型未缓存则先下载。
- 执行转录。
- 对转录文本进行基础清理。
- 默认生成整理稿和总结稿（用户说"只转录"则跳过）。
转录文本基础清理
原始转录往往是繁体、无标点、有重复的粗糙文本，需要进行以下清理：
- 繁转简：繁体中文转简体。
- 加标点分段：补全标点，按语义分段。
- 去口语赘余：去除"嗯"、"啊"、"就是说"等无意义填充词。
- 去重复/幻觉：删除语音识别产生的重复片段。
- 修错别字：修正常见识别错字（如"维铭"→"文明"）。
- 保留原意：只做清理，不改写、不重组。
清理后覆盖保存到 \x3C源文件名>_whisper.txt。
整理稿
默认文件名：\x3C源文件名>_整理稿.txt
整理稿不是逐句清理，而是将口语转录完全重写为流畅的书面文章。
重写要求
- 去口语化：彻底去除口语痕迹（语气词、重复、卡顿等）。
- 第三人称：原文的"我"改为"作者/导演/分析者"或省略。
- 逻辑重组：按逻辑而非时间线组织，相关内容合并到同一段落。
- 散文体：连贯段落（每段 3-8 句），不用列表，不用清单式写法。
- 概括性标题：文章开头给一个概括主题的标题。
- 保留全部内容：专业术语和分析洞见完整保留，只是换更精炼的表达。
- 繁转简 + 修错字。
- 篇幅：约为原始转录的 60%-80%。
- 忠于原文：不编造、不添加原文未涉及的内容。
语言风格
- 精简克制：不堆砌修饰词，少用"极为"、"本质上"等加重词。
- 结构标记：用"第一幕"、"随着"、"接下来"、"至此"等词标记内容推进。
- 主题升华：关键段落要有总结性提炼，不只描述发生了什么，还要归纳意味着什么。避免过于直白的口语表达。
- 结尾总结：最后一段跳出具体分析，对全篇做整体评价，回扣主题。
- 段落饱满：完整发展一个想法再换段，小观点揉进同一段落。
- 最终效果：读起来像独立撰写的专业文章，而非清洗过的转录稿。
总结稿
默认文件名：\x3C源文件名>_总结稿.md + \x3C源文件名>_总结稿.html
在整理稿基础上生成结构化摘要，同时输出浏览器可直接查看的 html 版本。
格式要求
- 顶部引用块写一句话概括
- 关键术语和核心结论加粗高亮
- 用分割线划分大区块
- 章节标题用
### 第X部分 | 标题格式 - 原文摘录中的关键词也加粗
# 标题
> **一句话概括**：全文主题
---
## 核心摘要
- **要点一**：结论
- **要点二**：结论
---
## 结构拆解
### 第一部分 | 标题
- **关键点**加粗，其余正常
---
## 关键观点
### 1. 观点标题
简洁解释，**核心结论加粗**。
---
## 原文摘录
> 原文片段，**关键词加粗**
html 版本
用 Python markdown 库将 .md 转为带样式的 .html：
- 标题有层级大小
- 加粗文字红色高亮
- 引用块蓝色左边框 + 浅蓝背景
- 正文居中，最大宽度 860px
- 自动在浏览器中打开
注意事项
- 默认自动生成整理稿和总结稿，无需用户额外要求
- 用户说"只转录"、"不要整理"、"只要原文"时跳过整理和总结
- 不使用外部语音识别接口
- 不编造原文未涉及的信息
- 不在脚本中硬编码用户路径
回复格式
执行完成后只报告：
- 是否成功
- 使用的模型
- 转录的文件
- 转录文本路径
- 整理稿路径
- 总结稿路径
示例
用户：帮我把 whisper medium 模型下到本地
→ 执行 scripts/download_whisper_model.py medium
用户：用本地 whisper 把这个 mp4 转成文字
→ 执行 scripts/local_whisper_transcribe.py，返回转录路径
Usage Guidance
This skill appears safe for its stated purpose. Before installing, be comfortable with installing openai-whisper and downloading Whisper models, and choose transcript output paths carefully because generated files remain on your machine and may contain sensitive information.
Capability Analysis
Type: OpenClaw Skill
Name: whisper-transcribe-summarize
Version: 1.0.3
The skill bundle provides a legitimate toolset for local audio/video transcription and summarization using the OpenAI Whisper library. The Python scripts (`local_whisper_transcribe.py` and `download_whisper_model.py`) perform standard file operations and model loading without any evidence of data exfiltration, obfuscation, or malicious command execution. The instructions in `SKILL.md` are well-defined for text processing tasks and do not contain harmful prompt injections or requests for unauthorized access.
Capability Assessment
Purpose & Capability
The stated purpose matches the artifacts: user-selected audio/video is transcribed with local Whisper and written to output files. Cleanup, rewriting, summary, and HTML generation are mostly described as agent instructions rather than implemented in the provided Python scripts.
Instruction Scope
Instructions are scoped to downloading a Whisper model, transcribing a specified local media file, and producing local transcript/summary outputs. No prompt override, hidden goal change, or persistence instruction was found.
Install Mechanism
The registry lists no required binaries or install spec, while the skill documentation and _meta require python3, ffmpeg, and openai-whisper. The package install command is unpinned and model downloads are delegated to openai-whisper, which is expected for this purpose but should be reviewed by the user.
Credentials
The scripts read one user-provided media file and write transcript output either beside the source file or to a user-specified path. This is proportionate, but generated text files may contain sensitive audio content and will persist locally.
Persistence & Privilege
The skill caches Whisper models under ~/.cache/whisper and creates output files, both of which are disclosed and purpose-aligned. No credential use, autostart, background worker, or privileged persistence was found.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install whisper-transcribe-summarize - After installation, invoke the skill by name or use
/whisper-transcribe-summarize - Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.3
No code or logic changes detected in this version. No SKILL.md functional changes.
- No file or documentation changes in this release.
- Version number unchanged (1.1.0 in both SKILL.md's).
- Behavior and user-facing interface remain the same.
v1.0.2
- Added script: download_whisper_model.py for downloading Whisper models locally.
- Added script: local_whisper_transcribe.py for fully local audio/video transcription without external APIs.
v1.0.1
No user-visible changes in this version.
- Version updated with no code or documentation changes detected.
v1.0.0
**Major update with enhanced workflows and automatic structured output.**
- Adds automatic generation of a cleaned transcript, article-style rewrite (整理稿), and HTML-structured summary (总结稿) after local Whisper transcription.
- Details strict requirements for transcript cleaning (繁转简、加标点、去口语赘余、去ASR幻觉、错别字修正等).
- Introduces advanced guidelines for rewriting transcripts into professional, logically reorganized articles with specific language and style standards.
- Provides structured summary output in both Markdown and HTML, with visual styling and highlighted key points.
- Describes clear user workflows for model downloading, local transcription, and output handling—including support for multiple input formats.
- Automatically generates all outputs unless user requests only the raw transcript.
Metadata
Frequently Asked Questions
What is whisper-transcribe-summarize?
本地 Whisper 语音转文字，自动生成清理文本、书面整理稿和结构化总结稿。支持音频和视频输入，全程离线。 It is an AI Agent Skill for Claude Code / OpenClaw, with 83 downloads so far.
How do I install whisper-transcribe-summarize?
Run "/install whisper-transcribe-summarize" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is whisper-transcribe-summarize free?
Yes, whisper-transcribe-summarize is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does whisper-transcribe-summarize support?
whisper-transcribe-summarize is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created whisper-transcribe-summarize?
It is built and maintained by zhangjinjin-gitgit (@zhangjinjin-gitgit); the current version is v1.0.3.
More Skills