在会议、访谈、课程和日常工作中，录音和录像越来越容易，但真正困难的往往不是“记录下来”，而是从大量音视频中找到真正有价值的信息。
AudioNotes 是一个开源的本地音视频转文字与智能笔记工具。它试图解决的并不只是“语音转文字”这一件事，而是建立一条从音视频 → 文字稿 → 结构化笔记 → AI 问答的完整工作流。项目源码公开在 GitHub，核心特点是本地优先，尽量让用户的数据留在自己的电脑上。
一、AudioNotes 到底解决什么问题？
传统录音工具的工作方式很简单：录下来，然后把文件保存起来。
问题是，当录音越来越多之后，“保存”本身就变成了新的负担。
例如，一场两小时的会议结束后，你可能需要回答：
- 谁提出了某个方案？
- 最终决定是什么？
- 某个任务由谁负责？
- 截止时间是什么？
- 某个观点是在什么时候提出的？
- 这次会议最重要的结论有哪些？
如果没有文字稿，寻找这些信息往往意味着重新听一遍录音。
AudioNotes 的思路则是先利用本地语音识别模型将音视频转换为文字，再利用本地大语言模型对内容进行整理，最终形成 Markdown 笔记，并允许用户继续针对这份内容进行提问。项目 README 将其定位为适合会议、访谈、课程以及语音备忘录等场景的工具。
因此，它更像是一个“音视频知识库入口”，而不只是一个录音转写器。
二、完整工作流：从声音到知识
AudioNotes 的使用流程可以概括为四个阶段。
1. 获取音视频
用户可以直接上传音频或视频文件，也可以使用浏览器进行录音。
这意味着它既适合处理已经存在的素材，也适合在会议、访谈等场景中直接使用。
2. 本地语音识别
AudioNotes 默认使用 FunAudioLLM/Fun-ASR-Nano-2512 进行语音识别。
相比单纯把文件上传到在线转录服务，本地识别的最大价值是：音视频内容可以留在用户自己的设备中。
当然，本地识别并不意味着结果一定完美。录音质量、环境噪声、口音以及多人同时说话，都可能影响最终文字稿，因此对于合同、会议决议等重要内容，仍然应该回听原始音频进行确认。
3. AI 自动整理
得到原始文字稿之后，AudioNotes 再利用 Ollama 运行的本地大语言模型进行笔记整理。
项目默认使用 qwen3.5:2b，目标是在普通电脑的资源条件下，在中文处理质量和资源消耗之间取得平衡。
这一步非常关键。
因为一份完整的会议转录可能包含大量口语化表达、重复内容和无关信息，而结构化笔记真正需要的是：
结论、观点、任务、时间、人物以及上下文。
因此，AudioNotes 的价值并不只来自 ASR，而来自 ASR 与 LLM 两个环节的组合。
4. 针对内容继续提问
笔记生成后，用户还可以继续与音视频内容进行对话。
For example:
“这次会议最终确定了哪三个方案？”
Or:
“张三负责哪些任务？”
这让音视频从一个需要人工“阅读”的文件，变成了一个可以直接询问的信息源。
三、本地优先，是 AudioNotes 最值得关注的特点
今天的 AI 产品大量依赖云端 API，但音视频往往包含非常敏感的信息。
一场公司内部会议可能包含商业计划；一次用户访谈可能包含个人信息；课程录音可能涉及尚未公开的研究内容。
AudioNotes 的设计理念恰恰是在这里做了不同的选择。
项目 README 明确说明，音视频、文字稿、笔记和对话历史默认保存在本机；语音识别通过 FunASR 在本地完成，大语言模型通过 Ollama 在本地运行，默认不会把音视频内容发送给第三方 AI 接口。Web 页面默认也只允许本机访问。
从架构角度看，这是一种典型的 local-first AI Thought:
Data enters the computer → complete ASR locally → complete LLM processing locally → save the results locally.
This is particularly attractive in privacy-sensitive scenarios.
Of course,"running locally" does not mean absolute security. The computer's own accounts, disk permissions, backup and network environment still need to be protected. The project also specifically reminds users to use and properly protect local data on a trusted computer.
4. Docker makes deployment relatively simple
For ordinary users, the biggest threshold is often not using AI, but installing AI.
AudioNotes provides Docker methods that encapsulate the main operating environment.
The officially recommended basic environments include Docker Desktop and Olama, and it is recommended that Docker Desktop allocate at least 12GB of memory; if processing longer audio and video, 16GB is recommended, and about 10GB of disk space is reserved.
The basic process is:
ollama pull qwen3.5:2b
mkdir audionotes
cd audionotes
curl -fsSL https://github.com/harry0703/AudioNotes/raw/main/docker-compose.yml -o docker-compose.yml
curl -fsSL https://github.com/harry0703/AudioNotes/raw/main/.env.example -o .env
docker compose up -d --remove-orphans
Access local 15433 Port to enter the Web interface. When performing speech recognition for the first time, you also need to download about 2.15GB of speech models; after the download is completed, the model will be saved on the local machine, and there is no need to download repeatedly in the future.
It's worth noting that the project's Docker configuration does not simply put a Web service up and running. Its Compose configuration also includes settings such as data persistence, health check, automatic restart, process recycling, and container permission shrinkage. For example, by default, only services are bound to 127.0.0.1, and passed cap_drop: ALL and no-new-privileges Reduce the risk of containers being abused.
This shows that in addition to being "operational", the project also considers some security and stability issues in actual deployment.
5. Where is the data stored?
Another key issue with native AI tools is:Where did my data go?
When AudioNotes uses Docker, it maps the data to the created audionotes Folder, which includes uploaded files, model cache, logs, and conversation history. In the source code operation mode, personal data is mainly stored in the project's storage folder, and Ollama's models are managed by Ollama himself.
This brings a very practical advantage: data management is relatively intuitive.
For example, project recommendations:
- Copy entire when backing up
audionotesfolder; - Copy the folder as a whole when migrating a computer;
- When historical data is no longer needed, you can stop the service and delete the folder.
In other words, users do not have to rely on a third-party SaaS platform to manage their historical notes.
6. Who is it suitable for?
AudioNotes is best suited for thoseHave a lot of voice or video data, but want the content to remain localizedpeople.
meeting minutes
This is the most direct application scenario.
After the meeting, you don't need to spend a lot of time manually sorting out the recording. You can let the AI complete the preliminary transcription and structured organization first, and then manually confirm the key conclusions.
Interviews and Interviews
Interviews often involve a lot of natural language and long conversations. After converting the recording to text, it will be much easier to find people, opinions and key answers.
curriculum and learning
For long course recordings, transcripts and AI Q & A can reduce the cost of listening back.
For example, instead of relistening to an hour of class to find a certain concept, you might as well ask directly:
"What examples did the teacher give of this concept?"
Personal voice memo
Inspiration, to-do items, and temporary thoughts can all be saved through recording, and then let AI help organize them into easier to read text.
7. Advantages and limitations of AudioNotes
If evaluated from a product perspective, the advantages of AudioNotes are relatively clear.
First, local priority is given.
This is its core competitiveness that distinguishes it from a large number of cloud AI note-taking tools. Sensitive audio and video do not have to be uploaded to third-party AI services by default.
Second, the workflow is complete.
It does not just solve "speech to text", but further completes note sorting and content Q & A.
Third, open source and self-deployable.
Users can view the source code, deploy it themselves, and adjust the model and configuration as needed.
Fourth, data management is relatively transparent.
Upload files, historical data, and model caches all have clear local storage locations.
But it also has obvious thresholds.
First of all, local AI has requirements for hardware resources. The project clearly recommends allocating at least 12GB of memory for Docker, and 16GB for long audio videos.
Second, the speed and effectiveness of local models will be affected by computer hardware.
Third, although the local model has privacy advantages, it does not automatically guarantee transliteration accuracy. For important content, manual proofreading still cannot be omitted.
最后，安装 Docker、Ollama、模型以及处理模型缓存，对于完全没有技术背景的用户来说，依然存在一定学习成本。
八、它真正有意思的地方，不是“转写”
如果只把 AudioNotes 看成一个“录音转文字软件”，其实低估了它。
它真正值得关注的地方，是把传统的音视频文件转换成了一个可以被 AI 理解和查询的内容载体。
Past:
录音 → 文件 → 人工回听
Now:
录音 → ASR → 文字稿 → AI 摘要 → 对话式查询
这意味着音视频本身开始成为一种“可交互的数据”。
尤其是在本地模型越来越强、消费级电脑算力不断提高的情况下，这种模式可能会成为传统云端 AI 工具之外的一条重要路线。
九、结语
AudioNotes 是一个很有代表性的本地 AI 应用：它没有试图重新发明一个复杂的 AI 平台，而是把几个已经成熟的技术组件组合起来——FunASR 负责听懂声音，Ollama 负责理解内容，Markdown 负责承载笔记，Web UI 负责交互。
它最值得肯定的地方，是把“隐私”和“AI 能力”放在了同一个产品设计里。
对于经常处理会议、访谈、课程或个人录音的人来说，AudioNotes 提供了一种值得尝试的工作方式：
不再把录音当成需要保存的文件，而是把它当成可以被搜索、总结和询问的个人知识来源。
项目源码与完整部署说明可以在 GitHub 上的 AudioNotes 仓库 查看。本文中的功能与部署信息以项目当前 README 和 Docker Compose 配置为依据。