MeetingAI Local v2.1.1 It is a "local-first" set of intelligent tools for real-time meetings, designed for scenarios such as real-time meeting transcription, meeting analysis, AI-powered Q&A, long-form audio processing, and local knowledge management.

It is not merely a “voice-to-text” tool, but rather a meeting dashboard that runs in a browser: the browser handles microphone input, while the Python server handles ASR, meeting intelligence, AI scheduling, history storage, and the knowledge base. The system can be used in real time during meetings and also supports offline transcription and analysis of existing long recordings in formats such as MP3, WAV, and M4A after they are imported.

Windows Public Download:

MeetingAI-Local_v2.1.1_Publicdownloading

Why We Created MeetingAI

MeetingAI is not designed merely to transcribe speech into long blocks of text, but rather to go a step further and enable participants in meetings toKey points, issues, actions, and risks naturally come to light.

Currently, MeetingAI's main workspace consists of three columns:

• Left: Real-time recording;
• Center: Smart Conferencing;
• On the right: AI Assistant.

The original verbatim transcript is retained, while the AI continuously generates updates on the current topic, real-time briefings, key points, and stage summaries alongside it, and can directly answer questions about the meeting.

v2.1.1 Key Changes: AI Response Gateway

The focus of v2.1.1 is not on adding a new surface button, but rather on making adjustments to the AI infrastructure layer.

More and more models now support reasoning and thinking. A typical response might be:

reasoning:
Internal analysis process of the model

final:
The answer that should actually be shown to the user

If the system fails to distinguish between these two parts correctly, a very strange problem may occur:

Even though the model was clearly just making an internal determination that “there is no new information, so it should return an empty string,” the system displayed this internal reasoning directly as a real-time alert.

Therefore, MeetingAI has added a unified layer between the AI provider and the meeting service, AI Response Gateway::

AI Provider
 ↓
AI Response Gateway
 ├─ Isolation of reasoning/thinking
 ├─ Extraction of final answer
 ├─  tag cleanup
 ├─ Text/JSON output contract
 └─ "final-only" retry if necessary
 ↓
MeetingAI Business Layer

This means:The model is allowed to think, but its thought process will not automatically become part of the meeting content.

In v2.1.1, this Gateway also added background observability:

• Number of AI calls processed;
• How much reasoning has been isolated;
• Valid empty output;
• Automatic retry;
• Abnormal state;
• Recently called tasks, models, elapsed time, and the number of characters in the "final" and "reasoning" sections.

Also available protect / observe / bypass Three modes.

It is important to note that even if full Gateway protection is disabled and the system enters safe bypass mode, the system will not revert to the dangerous behavior of providing a "fallback answer with an empty reasoning and 'final' status."

The original verbatim transcript and the AI's interpretation are separate.

MeetingAI applies a similar approach to ASR processing.

Original:

transcript.jsonl

Always preserve the original ASR/human-verified evidence so that AI cannot secretly alter it.

AI's semantic verification of recognition results is handled separately:

transcript_semantics.jsonl

In the middle.

For example, ASR takes Ollama If a term is identified as a homophone, the AI can suggest a correction with high confidence in the “AI-Assisted” view; however, higher thresholds are applied to facts such as amounts, dates, personnel, responsible parties, contract terms, and key decisions. If the AI cannot confirm a term, it is marked as “Pending Confirmation” rather than making an unfounded assumption.

So the entire process is:

Audio
→ ASR
→ Raw Transcript
→ AI Semantic Annotation Sidecar
→ Meeting Intelligence / AI Assistant / AI Notes

AI 出错或 Provider 暂时不可用时，系统会自动回退原始逐字稿，不阻塞会议主流程。

支持 MP3 等超长音频导入

除了实时麦克风，MeetingAI 也支持导入已有录音。

支持的常见格式包括：

MP3 / WAV / M4A / AAC / FLAC / OGG / OPUS / WEBM
MP4 / MOV / MKV 音轨

导入后不会把整个几小时音频一次性塞进内存，而是：

原文件落盘
→ FFmpeg 解码
→ 固定时长分段
→ 后台 ASR
→ 按时间轴写回 transcript
→ AI语义校验
→ 会议智能

目前采用的是固定时长切分，默认 60 秒，可以在 10–600 秒之间调整；单次导入还可以直接选择 30 / 60 / 90 / 120 秒或自定义。

当前版本暂未加入“智能静音边界切分”，优先采用更可预测、易于恢复且适合超长任务的固定时长切分方案。

本地 AI、知识库和回答范围

AI 助手支持多对话，并把“回答深度”和“知识范围”分开。

回答深度：

快速 / 标准 / 深入

回答范围：

仅会议 / AI知识 / 知识库 / 联网 / 全开

例如可按需要选择：

• 只根据本次会议回答；
• 使用 AI 自身知识补充；
• 使用指定本地知识库；
• 联网查询；
• 或全部开启。

知识库目前支持本地 TXT / Markdown 文档递归扫描和增量索引，不要求额外部署一个大型向量数据库才能开始使用。

运行环境、GPU 与国内网络

MeetingAI 不负责安装 Python 本体，但后台可以检测：

• Python；
• CPU；
• 内存；
• NVIDIA GPU；
• 显存；
• CUDA；
• PyTorch CUDA；
• FFmpeg；
• 核心依赖；
• ASR 引擎和模型。

同时支持中国大陆常用 Pip 镜像、ModelScope / HuggingFace 模型来源以及 HTTP(S) / SOCKS5 代理。

代理作用域是分开的，例如只让模型下载走代理，而不让本地 Ollama 请求绕一圈代理。

左侧栏现在可以折叠

v2.1.1 也做了一个比较简单但实用的 UI 调整：桌面端左侧导航可以折叠。

展开时可以看到：

MeetingAI
Meeting Intelligence

首页
历史会议
后台设置

折叠后只保留图标和状态灯，为三栏会议工作区腾出更多横向空间。

这个 Sidebar 状态和会议内部的三栏折叠/放大是两套完全独立的布局逻辑，不会因为收起左侧菜单把三栏工作区变成上下排列。

Windows 新手安装教程

下面按“第一次接触 Python 也能照着做”的方式写。

第一步：安装 Python

MeetingAI 要求：

Python 3.10+

Windows 64 位环境推荐使用 Python 3.11。

安装 Python 时一定勾选：

Add python.exe to PATH

安装结束后打开 CMD：

python --version

如果看到：

Python 3.11.x

说明正常。

第二步：下载并解压 MeetingAI

把 WordPress 下载包解压到一个简单目录，例如：

D:\MeetingAI

不要直接在 ZIP 压缩包内部双击程序。

第三步：安装核心依赖

直接双击：

install_core.bat

或者在 CMD 中执行：

cd /d D:\MeetingAI
python -m pip install -r requirements.txt

第四步：启动

双击：

start.bat

Or:

python meetingai.py

第一次启动会自动生成 Web UI、Prompt、Schema 等运行文件，并生成：

FIRST_RUN_PASSWORD.txt

使用这个文件里的密码登录。

默认本机地址通常是：

http://127.0.0.1:7777

最终以终端启动时显示的地址为准。

第五步：先跑一次自检

python meetingai.py --self-test

如果自检正常，再进入后台配置 ASR 和 AI Provider。

第六步：配置 ASR

进入：

后台设置 → 运行环境

先看系统检测结果。

Faster-Whisper

可以直接：

双击 install_asr_whisper.bat

Or:

python -m pip install -r requirements-asr-whisper.txt

FunASR / SenseVoice

FunASR 用户应先安装与你电脑 CPU/GPU 匹配的 PyTorch + torchaudio，然后：

双击 install_asr_funasr.bat

或：

python -m pip install -U funasr modelscope

验证 GPU：

python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

第七步：安装 FFmpeg

如果你要：

• 本地 ASR；
• 导入 MP3/M4A/MP4；
• 处理长录音；

建议安装 FFmpeg。

在 CMD 中：

ffmpeg -version

能显示版本即正常。

也可以在后台填写 ffmpeg.exe 的完整路径。

第八步：配置 AI Provider

进入：

后台设置 → AI 与调度

可以连接：

• 本地 Ollama；
• OpenAI-compatible Provider；
• 其他兼容接口。

配置成功后可以使用会议智能、AI 助手、AI 笔记、ASR 语义校验等能力。