DSH HUB
HomePlugin StorePlugin PacksCommunityRankingsResourcesPublish Guide
Plugin source
Back to catalog

luczhang88-lang /

luczhang88-lang/dsh-meeting-recorder

Verified

This plugin has no description yet.

★ 0 Stars0 Forks0 IssuesN/A Community rating0 Confirmed installs
View on GitHub
READMESource: main@b39c9814

dsh-meeting-recorder

DeepSeek Harness 会议录音插件,带本机 FunASR 引擎源码。现场录音、导入音频、转写、说话人校对,并用 Harness 里已经配置好的模型写纪要、导出 Word。

Meeting recorder plugin for DeepSeek Harness: live or imported audio, FunASR transcription, and Word minutes using the model already configured in Harness.

版本 0.3.0。作者 Luc Zhang(GitHub luczhang88-lang)。

装上插件还不能转写。Python 虚拟环境、依赖和模型权重都要在第一次使用前单独准备,它们不在 git 里,也不在 npm 包里。

最短步骤见 INSTALL.md。模型许可见 THIRD_PARTY.md。


中文

功能

  • 现场录音:浏览器采集本机麦克风(需权限;HTTPS 或 localhost)
  • 后期导入:wav / mp3 / m4a / flac / ogg 等;非 wav 先转 16 kHz 单声道再转写(需 ffmpeg)
  • 本机转写:仓库自带 engine/speech_engine + FunASR,音频默认不出本机
  • 说话人分离:自动聚类;校对页可改姓名、改分段归属
  • 重点标记:录音中打标记,写入文字稿章节
  • 交付物:讲话文字稿 Word、发言要点提纲 Word、正式《会议纪要》Word(纪要正文走 Harness 已配置的大模型)

仓库里有什么 / 没有什么

进 Git / npm 包 不进仓库(首次 setup 或首次转写时下载)
Harness 插件(Host + Client) engine/.venv(约 1.3–1.5 GB)
engine/speech_engine Python 源码 FunASR 模型权重(约 2 GB+,默认落到 engine/models/)
engine/requirements.txt、scripts/setup.* __pycache__、本地会议录音、.env、密钥

系统要求

  1. DeepSeek Harness 0.2.x(engines.dsh: >=0.2.0-rc.1 <0.3.0)
  2. Node.js 20+(只用于 npx 调用 dsh 安装插件)
  3. Python 3.10+(setup 用它创建 engine/.venv)
  4. 磁盘约 4–6 GB 空闲(venv + 模型)
  5. 访问 PyPI 和 ModelScope 的网络(装依赖、下模型)
  6. ffmpeg(导入 mp3/m4a 等时;现场录音写的是 WAV,可不装)
  7. 麦克风权限(现场录音)

哪些步骤是自动的

步骤 谁来做
把插件登记进 Harness 你执行 dsh plugin add,或在插件页安装
创建 venv、安装 torch 与 FunASR 依赖 scripts/setup.*,或插件里的「只安装依赖」
下载约 2GB+ 模型 只有你不加 --skip-models / 不点「只安装依赖」时才会下载。插件按钮默认不下载。第一次转写也可能再试一次
现场录音、调用 Harness 大模型写纪要 依赖装好并且 Harness 已配置模型之后,在界面里操作

安装

引擎目录必须是 Harness 正在加载的那一份插件 旁边的 engine/。在另一份克隆里跑 setup,不会作用到已经装进 Harness 的副本。

1. 克隆

git clone https://github.com/luczhang88-lang/dsh-meeting-recorder.git
cd dsh-meeting-recorder

2. 在这份目录里准备引擎

Windows(会下载模型;失败则脚本失败):

scripts\setup.cmd

Windows,只要依赖、先不下载模型:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\setup.ps1 -SkipModels

macOS / Linux(会下载模型):

bash scripts/setup.sh

macOS / Linux,只要依赖:

bash scripts/setup.sh --skip-models

macOS 上的 torch 从 PyPI 安装。Linux 和 Windows 使用 https://download.pytorch.org/whl/cpu。发布环境没有在 macOS 上实际跑过安装。

3. 把同一目录装进 Harness

npx -y @deepseek-ai/dsh plugin --profile web add .

桌面版把 web 换成 desktop。然后完全退出并重启 Harness。

如果 dsh plugin add 复制了一份插件、没有带上 .venv,打开「会议录音」,看页面上的插件目录,用那里的按钮安装。按钮只会运行该目录里的 scripts/setup.sh 或 setup.ps1,不会执行你输入的命令。

dsh plugin add github:luczhang88-lang/dsh-meeting-recorder 同样只装插件代码。装完后要在加载出来的那份目录里 setup,或者用插件里的按钮。不要再另克隆一份来装环境。

路径

默认(可在 cordis.patch.yml 里改,也可用环境变量覆盖):

  • engineRoot → 本插件目录下的 engine/
  • modelCacheDir → engine/models/
  • Python → engine/.venv/Scripts/python.exe(Windows)或 engine/.venv/bin/python(macOS / Linux)

环境变量:MEETING_RECORDER_ENGINE_ROOT、MEETING_RECORDER_MODEL_CACHE、MEETING_RECORDER_PYTHON、MEETING_RECORDER_MEETINGS_ROOT。

一旦改了 engineRoot,界面里的安装按钮会拒绝执行,避免把依赖装到插件没在用的目录。

使用

  1. 侧边栏打开「会议录音」
  2. 若页面提示引擎未就绪,先安装依赖
  3. 「现场录音」或「导入录音」
  4. 停录后自动转写 →「校对」→「交付」

别人怎么发现这个插件

GitHub topic dsh-plugin 只是公开仓库上的社区标签,方便在 https://github.com/topics/dsh-plugin 里被找到。它不保证某个插件市场会自动收录。

可以这样安装(仓库公开之后):

npx -y @deepseek-ai/dsh plugin --profile web add github:luczhang88-lang/dsh-meeting-recorder

然后仍要完成本机 setup。market-entry.json 是给自己的 dsh-market 目录合并用的示例,本仓库不会写入你电脑上的市场文件。npm 包名是 dsh-meeting-recorder,这次发布没有执行 npm publish。

许可

插件与引擎封装代码是 MIT。FunASR 库是 MIT,ModelScope 库是 Apache-2.0。模型权重不在本仓库里,也不能把上面的 MIT 当成权重许可。细节见 THIRD_PARTY.md。


English

Features

  • Live recording via the browser microphone (HTTPS or localhost)
  • Import later: wav / mp3 / m4a / flac / ogg; non-wav needs ffmpeg
  • Local ASR via bundled engine/speech_engine + FunASR (audio stays on-device by default)
  • Speaker diarization with a proofreading UI
  • Markers during recording
  • Word deliverables: transcripts, talking points, and formal minutes (the minutes body uses the LLM already configured in Harness)

What ships / what downloads

In git and the npm package Not shipped (first setup or first transcription)
Harness plugin (Host + Client) engine/.venv (~1.3–1.5 GB)
engine/speech_engine sources FunASR weights (~2 GB+, default engine/models/)
requirements.txt, scripts/setup.* caches, recordings, secrets

Requirements

  1. DeepSeek Harness 0.2.x
  2. Node.js 20+ for the npx install command
  3. Python 3.10+
  4. ~4–6 GB free disk
  5. Network access to PyPI and ModelScope
  6. ffmpeg for compressed imports
  7. Microphone permission for live capture

What is automated

Adding the plugin does not install Python dependencies or download models. scripts/setup.* creates the venv and installs dependencies. Model download runs only when you omit --skip-models / -SkipModels. The in-plugin button does not download models unless you choose the separate action. A failed model download fails the script.

Install

The engine Harness uses is the engine/ directory next to the loaded plugin files. Setup in a different clone does not affect that copy.

git clone https://github.com/luczhang88-lang/dsh-meeting-recorder.git
cd dsh-meeting-recorder

Windows, including the model download:

scripts\setup.cmd

Windows, dependencies only:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\setup.ps1 -SkipModels

macOS / Linux, including the model download:

bash scripts/setup.sh

macOS / Linux, dependencies only:

bash scripts/setup.sh --skip-models

On macOS, torch comes from PyPI. On Linux and Windows it comes from https://download.pytorch.org/whl/cpu. The publishing environment did not run the macOS installer.

Then install this same directory and fully quit Harness:

npx -y @deepseek-ai/dsh plugin --profile web add .

Use desktop instead of web for the desktop app. If Harness copied the plugin without .venv, open Meeting Recorder and press 只安装依赖. The model download button is 安装依赖并下载模型. Both run only the bundled setup script in the loaded plugin directory.

dsh plugin add github:luczhang88-lang/dsh-meeting-recorder installs the plugin code only. Run setup in the loaded copy, or use the button. Do not point a second clone at it.

Paths

Defaults: engineRoot is <plugin>/engine, modelCacheDir is <plugin>/engine/models, Python is .venv inside that engine. Override with MEETING_RECORDER_ENGINE_ROOT, MEETING_RECORDER_MODEL_CACHE, MEETING_RECORDER_PYTHON, or MEETING_RECORDER_MEETINGS_ROOT. A custom engineRoot disables the in-plugin setup button.

Discovery

The GitHub topic dsh-plugin is a community label (https://github.com/topics/dsh-plugin). It does not by itself insert the plugin into every market index. market-entry.json is a sample record for a local catalog merge. This release does not publish to npm.

License

Project code is MIT. FunASR is MIT. ModelScope is Apache-2.0. Model weights are not in this repo and are not covered by this MIT license. See THIRD_PARTY.md.

—/ 5

No ratings yet

Verified DSH bundle

Commit b39c98144b1e

Community comments

No comments yet. Be the first to write one.

DSH HUB

A community index for DSH plugins. Not an official GitHub or DeepSeek AI product.

CommunityResourcesAPIAbout