AI Voice Input Method — Daiti

Daiti AI voice input method: local, private, secure, with a low-latency typing experience

Daiti AI Voice Input Method: Absolute Privacy, Ultra-Low Latency

Official site: https://daiti.ai/

The AI voice input space suddenly got crowded — Doubao Input, Wispr Flow, Spokenly and more — and there are plenty of approaches too: local models, your own API keys, default model providers. It all comes down to the user’s budget and their appetite for privacy and speed.

Daiti is the work of an individual developer, focused on a small local model plus privacy and security. The developer says it uses Tongyi’s funasr model. Because the model is bundled in, the client is a little bulky.

• Supports Windows/Mac and covers input boxes system-wide; no need to switch input software — it works wherever you are.

• Local AI recognition, text appears in roughly 200–300ms; you see words as you speak, with no progress bar to wait on.

• Smart error correction and punctuation cleanup; you can add your own dictionary so industry jargon and code symbols come out more accurately.

• Smooth across documents, code, chat and AI conversations; data is processed only on your own machine, and it’s free forever.

The current version is iterating quickly. The developer has put their personal WeChat on the official site, so you can get in touch with feedback.

Test:

Chinese input is probably fine, but mixed Chinese-English input is prone to errors. Still, it solves most of the problem.

One other thing I haven’t quite gotten used to: I keep instinctively wanting to reach for the keyboard.