Meeting-to-Text: Local Speaker-Separated Transcription Skill

The Problem

There is a frequent need to transcribe meeting audio or video recordings into text, especially when using AI assistants like OpenClaw that require structured, machine-readable formats. However:

• Commercial services (like Feishu/Lark Minutes or iFlytek) can be expensive.
• Open-source models are notoriously difficult to deploy.
• Video-to-audio extraction, speech-to-text (ASR), and speaker diarization (Speaker Separation) usually belong to entirely different modules and repositories, making integration frustrating.
• It is difficult for AI agents to call these fragmented tools seamlessly.

The Solution

Meeting-to-Text solves this by packaging these complex capabilities into a single, cohesive Agent Skill. By following the setup instructions to deploy the local dependencies and models, you can easily perform local speaker-separated transcription. Furthermore, the entire system is highly lightweight: after full deployment, the entire system (including all models) takes up less than 3GB of disk space.

Read this document in other languages: 中文(简体)

Features

• True Offline Processing: 100% local. No data leaves your machine.
• Auto Audio Extraction: Uses FFmpeg to extract audio streams from MP4, MKV, MOV, and more.
• Accurate ASR: Leverages Alibaba's SenseVoiceSmall for fast, accurate multilingual transcription.
• Speaker Diarization: Uses 3D-Speaker to cluster voices and generate separated transcripts with speaker labels (e.g., Speaker 1, Speaker 2).
• Universal Agent & CLI Ready: Structured as a .md skill definition and a single Python entrypoint. While we use Codex/OpenClaw as examples, it can be attached to any AI agent, copilot, or CLI automation flow.

Supported Formats

• Video: .mp4, .mkv, .mov, .avi, .webm
• Audio: .wav, .mp3, .m4a, .aac, .flac, .ogg

---

Quick Start (Agent Automated Install)

Instead of manually installing dependencies and models, simply copy and paste the following prompt to your AI assistant (e.g. Codex, OpenClaw, Cursor, etc.):

Prompt: Please follow the instructions in this repository to install this skill for me, including its required models and dependencies: https://github.com/henrCh1/meeting-to-text

---

Manual Installation Guide

1. Prerequisites

• Python 3.10+
• FFmpeg: Download a pre-compiled Windows build (e.g., from gyan.dev) and extract it. You will need the path to ffmpeg.exe.

2. Install Python Dependencies

It is highly recommended to use a virtual environment (venv or conda).

# Create and activate your environment
python -m venv envs\asr
.\envs\asr\Scripts\activate

# Install requirements
pip install -r requirements.txt

3. Clone Required Repositories

You need the 3D-Speaker repository available locally for speaker diarization.

mkdir repos
cd repos
git clone https://github.com/alibaba-damo-academy/3D-Speaker.git

4. Download Models

The system requires two local models from ModelScope/HuggingFace. Download them into your models/ directory:

• SenseVoiceSmall
• fsmn-vad

(Note: 3D-Speaker checkpoints will automatically cache to the hub directory on first run).

---

Configuration & Usage

The skill relies on absolute paths to function deterministically. You must configure the paths before using the skill.

Standalone Python Setup

You can run the script manually by injecting the paths as Environment Variables, or by modifying the fallback paths in skills/meeting-to-text/scripts/meeting_to_text.py:

$env:MEETING_TO_TEXT_FFMPEG="C:\path\to\ffmpeg\bin\ffmpeg.exe"
$env:MEETING_TO_TEXT_SENSEVOICE="C:\path\to\models\SenseVoiceSmall"
$env:MEETING_TO_TEXT_VAD="C:\path\to\models\fsmn-vad"
$env:MEETING_TO_TEXT_3D_SPEAKER="C:\path\to\repos\3D-Speaker"

python skills/meeting-to-text/scripts/meeting_to_text.py --input "C:\path\to\video.mp4" --output "C:\path\to\output.txt"

Using as an Agent Skill (Codex, OpenClaw, AutoGPT, etc.)

1. Open skills/meeting-to-text/SKILL.md.
2. Find the <YOUR_CONDA_ENV_PYTHON_PATH> placeholder and replace it with the absolute path to your virtual environment's python.exe.
3. Find the C:\path\to\your\meeting-to-text\scripts\meeting_to_text.py placeholder and replace it with the absolute path to the script.
4. Replace <YOUR_WORKSPACE_TEMP_PATH> with a temporary directory path (e.g., C:\temp\meeting_workspace).
5. Register the SKILL.md with your agent platform.

Output Format

The resulting .txt file will look like this:

[00:00:00 - 00:00:05] Speaker 1: Hello everyone, let's start the meeting.

[00:00:05 - 00:00:10] Speaker 2: Yes, I can hear you clearly. Let's begin.

License

Provided under the MIT License. See LICENSE for details.