WeChat Put a Meeting Recorder Inside Its Keyboard, and the Output Is a Web Page

The most interesting AI product shipped in China last week was not a model. It was a button on a keyboard.

On October 27, WeChat's official channel announced that the WeChat Input Method — the keyboard a very large number of Chinese phone users already type with — now includes an "AI voice recorder". You record. It transcribes in real time. Then, instead of handing you a wall of text, it hands you a structured web page: typos corrected, filler stripped, the logic of the conversation reorganised, key points pulled out.

And then the part that made me sit up. According to the announcement, it can match charts, diagrams and interactive prototypes to what was said.

Why the keyboard is the interesting part

Every productivity company on earth has shipped a meeting recorder. Most of them require you to install something, grant permissions, pick a workspace, and remember to press record. Adoption stays in the single digits because the friction is at the exact moment you have the least attention to spare.

WeChat's version has no adoption problem to solve. The keyboard is already installed, already has microphone access, and is already the surface where a hundred million people start writing. The entry points reflect that: on mobile you open it from the keyboard toolbar or the input method app. On desktop there is a hotkey, plus the Mac menu bar and the Windows toolbar.

That is a distribution advantage no startup can buy, and it is the reason this is worth writing about at all. The feature itself is not novel. The placement is.

What "structured web page" actually implies

Transcription is a solved-ish problem. Turning speech into something you would voluntarily read is not.

Anyone who has run a meeting through an ASR model knows the output: accurate words, no shape. People repeat themselves, trail off, say "right, so, anyway" fourteen times, and bury the decision in the middle of a digression about parking. The cleanup is where the value is, and it is where most tools quietly give up and hand you the transcript with a summary stapled to the top.

The announced behaviours — correcting typos, filtering meaningless filler, organising logic and highlights using context — describe a pipeline that treats the recording as raw material rather than as the deliverable. The output is a page, which means it has sections, hierarchy, and probably a heading that you did not say out loud.

Chinese input methods have an advantage here that is easy to underestimate. Pinyin input means the keyboard already runs a language model over everything you type, all day, correcting and predicting. Applying that same class of model to speech is a shorter jump than it is for a Western keyboard vendor, because the correction layer is already load-bearing infrastructure rather than a nice-to-have.

The diagrams claim deserves scrutiny

Matching charts, illustrations and interactive prototypes to a recording is a big claim in a small sentence, and it splits into two very different capabilities.

The modest version: you talk about a three-step process, and the tool renders a three-step diagram because it heard "first, then, finally". That is template selection, it is genuinely useful, and it is well within reach.

The ambitious version: you describe a product flow, and the tool produces an interactive prototype you can click. That requires the model to infer state, transitions and layout from unstructured speech, and to be right often enough that the user does not immediately delete it.

My guess, and it is a guess, is that the shipping version is closer to the first. That is not a criticism. A generated diagram that is roughly right is a good first draft; a generated prototype that is roughly right is a debugging session.

The privacy question nobody in China is asking out loud

A keyboard sees everything. Passwords, bank messages, arguments with your family, the search you typed and then deleted. Adding a recorder to that surface means the trust relationship gets denser, not simpler.

The announcement does not go into data handling, which is normal for a product post and not reassuring for a product decision. The reasonable assumptions are that transcription runs on-device or through WeChat's own cloud, that the resulting page is stored in the WeChat ecosystem, and that retention is governed by the same terms as everything else in that ecosystem.

For individual users in China that trade is familiar and mostly acceptable. For a company, "our meeting notes live in WeChat" is a compliance conversation, not a productivity one. The feature will spread through personal use and get blocked in enterprises, which is the standard path for consumer AI in that market.

How it stacks up against a dedicated recorder

On paper, the specialist apps win: better speaker separation, longer sessions, integrations with calendars and CRMs, proper export. In practice, most of them lose on the only metric that matters at the start of a meeting, which is whether you remembered to open them.

The trade-off is that you get less control. A dedicated recorder lets you choose where the audio goes. A keyboard does not, and the announcement does not say. For a personal memo that is a fine trade. For a client call under an NDA it is not a trade you get to make unilaterally.

So the realistic split is: keyboard for the things you would otherwise have written down badly, specialist app for the things that end up in a contract. Both will exist, and the keyboard will handle far more volume.

The strategic read

WeChat has spent a decade being criticised for not shipping AI features fast enough. This is the opposite move: instead of building a chatbot, it is putting small models into the surfaces people already touch a hundred times a day. A keyboard. A voice message. A file.

That is a harder strategy to demo and a much better one to distribute. A chatbot needs you to want to chat. A recorder inside the keyboard is there when you were going to type anyway.

It also explains the output format. A structured page is shareable inside WeChat. A transcript is a file you export and forget. By making the artefact native to the ecosystem, the feature creates content that circulates in the place WeChat already owns.

What to watch

Whether the structured page survives editing. If you can fix a sentence and have the layout hold, it is a document tool. If editing breaks the structure, it is a screenshot.

Whether the diagrams are editable as objects or arrive as flat images. The first is a workflow. The second is decoration.

And whether the recorder gets an API. If it does, the interesting products will not be note-taking apps at all — they will be tools that turn a conversation into a deliverable without anyone opening a laptop.

The feature itself will be copied within a quarter. Baidu, iFlytek and Alibaba all have the pieces. What will not be copied quickly is the placement, because nobody else owns the keyboard that half a country types on.

That is the lesson worth taking from this release, and it has nothing to do with model quality. The best distribution channel for an AI feature is the thing your users already have their hands on.