Voice Input in China Learns to Write for You: From Dictation to Finished Text

A tool that "worked fine for everyone" has become one of the tightest arenas in Chinese AI within a year. Between November 2025 and May 2026, at least three new AI keyboards from major players reached the market, and WeChat added a «隔空傳送» ("transfer through the air") feature to its input. The main front is now voice input, and the bet there is on changing the task: the system listens not to the words but to the user's intent, and immediately returns finished text.

Timeline: four moves in seven months

The clock starts in autumn 2025. Here is what is known about how events unfolded:

• Doubao Input: launched in November 2025. Doubao is ByteDance's AI assistant, and the keyboard became a new entry point for it on the phone.
• Zhipu's desktop AI input: December 2025. Zhipu is known as a developer of its own large models, and it went after the desktop segment rather than mobile.
• Qianwen voice input: May 2026. Qianwen is the consumer AI brand of the Alibaba ecosystem, and this time the emphasis is squarely on voice.
• WeChat input with the «隔空傳送» feature: a tool for quickly passing typed content between devices. It draws on the messenger's huge audience.

The launches come almost back to back, one every one to two months. For a market that used to change over years, that is a very fast pace.

Why the keyboard became a strategic foothold

The keyboard sits at the start of almost every user scenario: a message, an email, a search query, a note, a prompt for a chatbot. Whoever controls this layer sees the user's intent before any app does and can build a model directly into the moment text is created. For a company with its own LLM, this is the shortest path from model to daily habit: there is no need to ask people to open a separate chat.

Voice amplifies the effect. Ordinary dictation has long been the norm, but it requires editing: people speak with pauses, repetitions and slips, and jump between thoughts. The new generation of AI inputs takes over that work. The user states the gist, for example "write to my colleague that the meeting is moving to Thursday, and apologize," and gets a neat message in the right tone. Transcription turns into generation, which means quality now depends not only on speech recognition but also on a language model that understands context and style.

What this changes for the market

This approach has weak points too. Generating instead of transcribing means the system may rephrase what was said in a way the author did not intend. That raises a question of trust: how accurately the final text conveys the thought, and where the line lies between assistance and substitution. The second problem concerns data. The input sees everything a person types and says, so on-device versus cloud processing becomes as important for keyboards as recognition accuracy.

Competition, meanwhile, is not running along a single axis. Doubao and Qianwen are tied to the ecosystems of their AI assistants, Zhipu has chosen the desktop, and WeChat uses a link between devices and its built-in audience. The winner will most likely be whoever puts together the best mix of model quality, response speed and convenience in everyday apps.

What to look for when choosing

If you are trying these inputs, check three things. First: how the system handles speech that mixes languages and professional terms. Second: whether you can quickly see the original transcript next to the finished text, to catch distortions. Third: where the voice is processed and what is stored. For teams and companies building products on top of LLMs, the takeaway is simpler: the spot where a model meets the user at the moment of writing has proved no less valuable than the chat window, and the competition there is only beginning.