yaps-meeting-transcription

该 Skill 通过 Yaps 和 Node 辅助脚本，把会议音视频生成带说话人标签的转录项目；支持按片段校正文本、分配或合并说话人、导出 Markdown，以及生成回顾、章节和带来源引用的问答。它不代用户入会或实时录音，也不根据声音推断身份。

yaps

热度 18.59

最近更新 2026-09-25 09:16:41

分类：协作依赖：代码执行 / 读写文件 / 内部调用大模型应用场景：会议纪要 / 摘要总结 / 笔记整理 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add yaps-meeting-transcription

Metadata

name: yaps-meeting-transcription
description: Make a meeting transcript with speaker labels in Yaps. Correct it, export it, or ask questions. New users: install Yaps and sign in.

Yaps Meeting Transcription

Read the runtime guide before the first operation. It defines <yaps>, account readiness, local permissions, and file handling.

Required Yaps version: 2.3.124 or newer. Check installed command help for later capabilities.

Feature readiness: Meeting/Sherpa; optional MOSS on supported Apple Silicon systems. Inspect features and group help before selecting an engine. Recaps also need the local chat model.

Workflow

Write a private temporary JSON request with an absolute input, optional title, engine (default auto), and optional integer speakers from 1 to 20. Run node <skill-root>/runtime/run.mjs meeting-file <request.json>. Remove only the request file. The helper extracts a temporary WAV for recognized video types through Yaps, then creates the durable meeting project; Yaps retains its own project audio.

Use auto by default. Use sherpa for cross-platform execution or a supplied speaker count. Use moss only when installed and supported on Apple Silicon; it detects speakers itself and cannot take a speaker-count hint. Treat the returned engine and reason as authoritative. A project with no transcript segments is not a successful meeting.

Inspect <yaps> meeting show <id> and meeting speakers <id> before editing. Use stable segment IDs with meeting correct <id> --segment <id> --text-file <file> or meeting assign <id> --segment <id> --speaker <number>. Roster numbers are 1-based. Use rename-speaker for a name correction; use merge-speakers <id> --speakers 3,4 --into 2 only after the user identifies the same participant. Never infer identity from voice alone. Export using meeting export <id> --output <new.md>.

For requested recaps, use meeting summarize <id> and then meeting chapters <id>. Reuse a recap unless regeneration is requested or its transcript/template changed. For a grounded question, use meeting ask <question> --meeting <id> --scope meeting; --scope all requires an explicit cross-meeting request. Preserve returned source citations. Never edit project JSON directly.

Boundaries

No joining calls or starting live recording on the user's behalf.
Speaker labels are diarization labels until the user supplies identities.
Recaps are generated interpretations, not verbatim transcript. Preserve project audio and existing exports.