Skip to content

h2fly/meetingscribe

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

19 Commits
tests
tests
 
 
.gitignore
.gitignore
 
 
CLAUDE.md
CLAUDE.md
 
 
LICENSE
LICENSE
 
 
README.md
README.md
 
 
config.jsonc
config.jsonc
 
 
install.sh
install.sh
 
 
meetingscribe.py
meetingscribe.py
 
 
requirements.txt
requirements.txt
 
 

Repository files navigation

MeetingScribe

Meeting scribe for macOS and Windows — captures both speaker and microphone audio, auto-transcribes, and produces AI-generated meeting notes or interview summaries.

Requirements

macOS

Item Detail
macOS 12+ Uses CoreAudio API
Python 3.9+ Verify with python3 --version
Node.js 18+ Required for Claude Code CLI
Homebrew Required to install BlackHole

Windows

Item Detail
Windows 10/11 64-bit
Python 3.9+ Download from python.org — check "Add to PATH" and "tcl/tk" during install
Node.js 18+ Required for Claude Code CLI

Installation — macOS

Step 1: Install Homebrew (skip if already installed)

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Step 2: Install system dependencies

brew install --cask blackhole-2ch   # virtual audio driver (captures system audio)
brew install ffmpeg                  # required by FunASR for audio decoding

Step 3: Configure macOS audio routing

This step routes system audio to both your speakers and BlackHole simultaneously.

  1. Open Audio MIDI Setup — search in Spotlight, or run:

    open "/System/Applications/Utilities/Audio MIDI Setup.app"

  2. Click "+" at the bottom left → "Create Multi-Output Device"

  3. Check both:

    • ✅ Your speakers (e.g. MacBook Air Speakers / external headphones)
    • ✅ BlackHole 2ch

  4. Check the "Drift Correction" checkbox on the BlackHole 2ch row

  5. Right-click the new Multi-Output Device → "Use This Device For Sound Output"

  6. Note the device name (default is Multi-Output Device; you can rename it)

After this setup, system audio will flow into BlackHole and MeetingScribe can capture it.

Step 4: Install Claude Code CLI

Claude Code CLI is used for AI polishing and generating meeting notes.

npm install -g @anthropic-ai/claude-code
claude login        # Follow the prompts to authorize
claude --version    # Confirm installation

Step 5: Install Python dependencies

cd meetingscribe

# Install PyTorch first (CPU version; visit https://pytorch.org for GPU options)
pip3 install torch torchaudio

# Install remaining dependencies
pip3 install -r requirements.txt

Step 6: Update config.jsonc

Open config.jsonc and update the following fields to match your device names:

"output_record":  "Multi-Output Device",   // Name of the Multi-Output Device from Step 3
"output_restore": "MacBook Air Speakers",  // Your speaker name (restored after recording)
"device_mic":     "MacBook Air Microphone" // Your microphone name

How to find device names:

python3 meetingscribe.py devices

Speaker names can be found in System Settings → Sound → Output.

Installation — Windows

Step 1: Install VB-Audio Virtual Cable

Download and install VB-Audio Virtual Cable (free): https://vb-audio.com/Cable/

This is the Windows equivalent of BlackHole — it creates a virtual audio device to capture system audio.

Step 2: Install VoiceMeeter Banana

Download and install VoiceMeeter Banana (free): https://vb-audio.com/Voicemeeter/banana.htm

VoiceMeeter routes system audio to both your speakers and Virtual Cable simultaneously (equivalent to macOS Multi-Output Device).

Configure VoiceMeeter:

  1. Open VoiceMeeter Banana
  2. In HARDWARE INPUT 1, select your microphone
  3. In HARDWARE OUT A1, select your real speakers / headphones
  4. In VIRTUAL INPUTS, enable routing to B1 (Virtual Cable)
  5. Right-click the Windows speaker icon in the taskbar → Sound settings → set output to VoiceMeeter Input

Once configured, VoiceMeeter runs in the background. No need to change audio devices before each recording.

Step 3: Install Claude Code CLI

npm install -g @anthropic-ai/claude-code
claude login

Step 4: Install Python dependencies

cd meetingscribe
pip install torch torchaudio
pip install -r requirements.txt

Step 5: Update config.jsonc

On Windows, audio device auto-switching is not supported — VoiceMeeter handles routing permanently. Set output_record and output_restore to empty strings, and update the device names:

"output_record":  "",                                    // leave empty on Windows
"output_restore": "",                                    // leave empty on Windows
"device_system_audio": "CABLE Output (VB-Audio Virtual Cable)",
"device_mic":     "Microphone (your mic name here)"

How to find device names:

python meetingscribe.py devices

Quick Start

GUI (recommended)

python3 meetingscribe.py ui
  1. Click ▶ Start Recording to begin
  2. Click ◼ Stop Recording when done
  3. Click Generate Meeting Notes or Generate Interview Summary
  4. When complete, click Open Result to view the Markdown output

Command Line

# Record and auto-generate meeting notes (Ctrl+C to stop)
python3 meetingscribe.py record

# Interview mode
python3 meetingscribe.py record --mode interview

# Process an existing recording
python3 meetingscribe.py transcribe /path/to/audio.wav
python3 meetingscribe.py transcribe /path/to/audio.wav --mode interview

Output Files

Recordings are saved to ~/Documents/meetingscribe/recordings/. All output files are placed alongside the recording:

Extension Content
.wav Recording (system audio + microphone, dual stream)
.raw.txt Raw transcript (with timestamps)
.proofread.txt AI-polished transcript
.md Meeting notes / interview summary (Markdown)

Re-running automatically detects completed steps and skips them — no need to re-transcribe.

Configuration

Edit config.jsonc in the project directory (supports // comments). Common options:

Key Default Description
mode meeting meeting or interview
transcribe_provider funasr funasr (local, default) / whisper / openai / gemini
polish_provider claude claude / openai / gemini
meeting_notes_provider claude claude / openai / gemini
stt.funasr.hotword "" Space-separated hotwords to boost recognition accuracy

FunASR models (configured in stt.funasr):

Key Default Description
model paraformer-zh ASR model — downloaded automatically on first run (~400 MB)
vad_model fsmn-vad Voice activity detection — enables long-audio support
punc_model ct-punc Punctuation restoration
hotword "" Domain-specific terms (space-separated) to improve accuracy
chunk_secs 300 Split audio into chunks of this length (seconds) for parallel transcription; 0 = always serial
workers 0 Parallel FunASR instances; 0 = auto (max(2, cpu_count / 2)); memory scales with this value

Optional: switch back to Whisper

Install faster-whisper and set transcribe_provider to whisper in config.jsonc:

pip3 install faster-whisper
Whisper Model Size Speed Accuracy
tiny ~75 MB Fastest Fair
base ~150 MB Fast Good
small ~480 MB Medium Better
medium ~1.5 GB Slow Very good
large-v3 ~3 GB Slowest Best

Troubleshooting

Q: No system audio captured (can't hear meeting app audio)

  • Check Audio MIDI Setup — confirm BlackHole 2ch is checked in the Multi-Output Device
  • Check System Settings → Sound → Output — confirm Multi-Output Device is selected
  • Check that output_record in config.jsonc exactly matches the device name (spaces matter)

Q: claude command not found

npm install -g @anthropic-ai/claude-code
claude login

Confirm Node.js is installed: node --version (requires 18+)

Q: First run is slow / downloading models FunASR downloads paraformer-zh, fsmn-vad, and ct-punc on the first run (~400 MB total). They are cached locally and not re-downloaded. Ensure a stable network connection for the first run.

Q: How to confirm device names

python3 meetingscribe.py devices

The listed names are exactly what you should put in config.jsonc. Names are case-sensitive.

Q: macOS requests microphone permission Click Allow when prompted. If you previously denied it, re-enable it in System Settings → Privacy & Security → Microphone.

MeetingScribe

支持 macOS 和 Windows 的会议记录助手 — 同时录制扬声器与麦克风声音，自动转写，最终由 AI 生成会议纪要或面试总结。

系统要求

macOS

条件 说明
macOS 12+ 使用 CoreAudio API
Python 3.9+ python3 --version 确认
Node.js 18+ 用于安装 Claude Code CLI
Homebrew 用于安装 BlackHole

Windows

条件 说明
Windows 10/11 64位
Python 3.9+ 从 python.org 下载，安装时勾选 "Add to PATH" 和 "tcl/tk"
Node.js 18+ 用于安装 Claude Code CLI

安装步骤 — macOS

第一步：安装 Homebrew（已安装可跳过）

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

第二步：安装系统依赖

brew install --cask blackhole-2ch   # 虚拟音频驱动，用于捕获系统播放的音频
brew install ffmpeg                  # FunASR 转写所需的音频解码工具

第三步：配置 macOS 音频路由

这一步让系统在播放声音给扬声器的同时，也把声音送入 BlackHole，从而实现录制。

  1. 打开「音频 MIDI 设置」：Spotlight 搜索，或运行

    open "/System/Applications/Utilities/Audio MIDI Setup.app"

  2. 左下角点击「+」→「创建多输出设备」

  3. 在右侧列表中勾选：

    • ✅ 你的扬声器（如 MacBook Air 扬声器 / 外接耳机）
    • ✅ BlackHole 2ch

  4. 勾选 BlackHole 2ch 那行的「漂移校正」复选框

  5. 右键新建的「多输出设备」→「将此设备用作系统声音输出」

  6. 记下该设备的名称（默认是「多输出设备」，可双击重命名）

完成后，系统播放声音时会同时进入 BlackHole，MeetingScribe 即可捕获。

第四步：安装 Claude Code CLI

Claude Code CLI 用于 AI 校对和生成会议纪要。

npm install -g @anthropic-ai/claude-code
claude login        # 按提示完成授权登录
claude --version    # 确认安装成功

第五步：安装 Python 依赖

cd meetingscribe

# 先安装 PyTorch（CPU 版；如需 GPU 版请访问 https://pytorch.org 选择对应命令）
pip3 install torch torchaudio

# 再安装其余依赖
pip3 install -r requirements.txt

第六步：修改配置文件

打开项目目录下的 config.jsonc，根据你自己的设备名称修改以下三项：

"output_record":  "多输出设备",             // 第三步创建的多输出设备名称
"output_restore": "MacBook Air Speakers",  // 录音结束后恢复的扬声器名称
"device_mic":     "MacBook Air Microphone" // 麦克风名称

如何查看设备名称：

python3 meetingscribe.py devices

扬声器名称可在系统「声音」设置 → 输出 中查看。

安装步骤 — Windows

第一步：安装 VB-Audio Virtual Cable

下载并安装 VB-Audio Virtual Cable（免费）： https://vb-audio.com/Cable/

这是 BlackHole 的 Windows 等价工具，用于捕获系统播放的音频。

第二步：安装 VoiceMeeter Banana

下载并安装 VoiceMeeter Banana（免费）： https://vb-audio.com/Voicemeeter/banana.htm

VoiceMeeter 可以将系统音频同时送到真实扬声器和 Virtual Cable，等价于 macOS 的「多输出设备」。

VoiceMeeter 配置方法：

  1. 打开 VoiceMeeter Banana
  2. 在 HARDWARE INPUT 1 中选择你的麦克风
  3. 在 HARDWARE OUT A1 中选择你的扬声器 / 耳机
  4. 在 VIRTUAL INPUTS 中，将路由启用到 B1（即 Virtual Cable）
  5. 右键任务栏扬声器图标 → 「声音设置」→ 将输出设备改为 VoiceMeeter Input

配置完成后 VoiceMeeter 后台运行，每次录音无需手动切换音频设备。

第三步：安装 Claude Code CLI

npm install -g @anthropic-ai/claude-code
claude login

第四步：安装 Python 依赖

cd meetingscribe
pip install torch torchaudio
pip install -r requirements.txt

第五步：修改配置文件

Windows 下不需要自动切换输出设备（VoiceMeeter 已常驻处理路由），将 output_record 和 output_restore 置空，并更新设备名称：

"output_record":  "",                                        // Windows 下留空
"output_restore": "",                                        // Windows 下留空
"device_system_audio": "CABLE Output (VB-Audio Virtual Cable)",
"device_mic":     "麦克风 (你的麦克风名称)"

查看设备名称：

python meetingscribe.py devices

快速开始

图形界面（推荐）

python3 meetingscribe.py ui

使用流程：

  1. 点击「▶ 开始录音」→ 开始会议
  2. 会议结束后点「◼ 停止录音」
  3. 点击「开始整理会议纪要」或「开始整理面试记录」
  4. 等待处理完成，点击「打开结果文件」查看 Markdown 纪要

命令行

# 录音 + 自动生成会议纪要（Ctrl+C 停止录音）
python3 meetingscribe.py record

# 面试模式
python3 meetingscribe.py record --mode interview

# 处理已有录音文件
python3 meetingscribe.py transcribe /path/to/audio.wav
python3 meetingscribe.py transcribe /path/to/audio.wav --mode interview

输出文件说明

录音文件默认保存在 ~/Documents/meetingscribe/recordings/，处理结果与录音同目录：

文件后缀 内容
.wav 录音文件（系统音频 + 麦克风双路）
.raw.txt 原始转写文本（含时间戳）
.proofread.txt AI 校对后文本
.md 会议纪要 / 面试总结（Markdown 格式）

重复运行时会自动检测已完成的步骤并跳过，无需重新转写。

配置说明

配置文件为项目目录下的 config.jsonc，支持 // 注释。常用配置项：

配置项 默认值 说明
mode meeting 默认模式：meeting（会议）/ interview（面试）
transcribe_provider funasr 转写引擎：funasr（本地，默认）/ whisper / openai / gemini
polish_provider claude 校对模型：claude / openai / gemini
meeting_notes_provider claude 纪要模型：claude / openai / gemini
stt.funasr.hotword "" 热词（空格分隔），提升专有名词识别率

FunASR 配置项（stt.funasr 下）：

配置项 默认值 说明
model paraformer-zh ASR 模型，首次运行自动下载（约 400 MB）
vad_model fsmn-vad 语音活动检测，支持长音频分句
punc_model ct-punc 标点恢复模型
hotword "" 热词（空格分隔），提升领域专有词识别率
chunk_secs 300 超过此时长自动分块并发转写（秒），0 = 始终串行
workers 0 并发实例数，0 = 自动（max(2, CPU核数/2)），内存随之线性增长

可选：切换回 Whisper

安装 faster-whisper 并在 config.jsonc 中将 transcribe_provider 改为 whisper：

pip3 install faster-whisper
Whisper 模型 大小 速度 准确度
tiny ~75 MB 最快 一般
base ~150 MB 快 较好
small ~480 MB 中等 好
medium ~1.5 GB 慢 很好
large-v3 ~3 GB 最慢 最佳

常见问题

Q：录不到会议软件的声音

  • 检查「音频 MIDI 设置」→ 多输出设备是否勾选了 BlackHole 2ch
  • 检查系统「声音」→「输出」是否选择了多输出设备
  • 检查 config.jsonc 中 output_record 的名称是否与多输出设备完全一致（包括空格）

Q：claude 命令找不到

npm install -g @anthropic-ai/claude-code
claude login

确认 Node.js 已安装：node --version（需要 18+）

Q：首次运行转写很慢 / 正在下载模型 FunASR 在首次运行时会自动下载 paraformer-zh、fsmn-vad、ct-punc 三个模型（合计约 400 MB），下载完成后本地缓存，后续无需重新下载。请确保首次运行时网络畅通。

Q：如何确认设备名称是否正确

python3 meetingscribe.py devices

列出的名称即为可填入 config.jsonc 的值。注意名称区分大小写。

Q：macOS 要求麦克风权限 首次运行时系统会弹出权限请求，点击「允许」即可。 如已拒绝，可在「系统设置」→「隐私与安全性」→「麦克风」中重新开启 Terminal / Python 的权限。

About

会议录音，转写为会议纪要，面试总结等

Resources

Readme

License

GPL-3.0 license
Activity

Stars

6 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages