Plenty of people don't speak in one language at a time. A developer says "ζ config ιη timeout ζΉζ 30 η§". A designer says "θΏδΈͺ button η padding εε 4 px". These aren't edge cases β they're how bilingual speakers actually talk, especially around technical terms and brand names that simply don't have a natural Chinese equivalent.
Built-in dictation assumes the opposite. Before you start, you pick a language: English or Chinese. Choose English and your Chinese comes out as garbled phonetics; choose Chinese and "config", "timeout", and "px" get mangled into the nearest-sounding characters. Either way the sentence above is unusable, and you end up fixing half of it by hand β which defeats the point of dictating.
The root cause is architectural. Traditional dictation engines load a per-language model: an acoustic and language model trained and tuned for one language at a time. That single-language assumption reaches all the way down β into how audio is segmented, how sounds are mapped to words, and how the system decides where one word ends and the next begins.
When you code-switch mid-sentence, you violate that assumption. The English model has no Chinese words to choose from, so it forces every syllable into an English-shaped guess (and vice versa). There's no "both" setting because the engine was never designed to hold two languages open at once. The segmentation itself is wrong, so even a perfect transcription of the wrong vocabulary comes out broken.
Voicefall was built around mixed speech from the start. There is no language to pick. You hold the key, speak your sentence exactly as it comes out β Chinese, English, or both β release, and a transcription model tuned for mixed Chinese-English handles the whole utterance together, keeping "config" and "timeout" as English while the surrounding Chinese stays Chinese.
The delivery matters too. Voicefall is push-to-talk and global: there's no window to switch to and no button to click. Whatever app your cursor is in β an editor, a terminal, a chat box, a browser field β the text lands right there, about a second after you let go. So the sentence you'd have spent thirty seconds correcting just appears, correct, the first time.
Want the exact gesture and on-screen cues? See how to use Voicefall.