Every dictation tool I have used starts by asking which language I am about to speak. I do not know. I am going to speak both, and the switch is going to happen in the middle of a noun.
Here is a thing I would actually say to a colleague, pointing at a screen:
"Yeh button dabao toh state reset ho jaata hai, but sirf tab jab modal already open ho."
Two languages, one clause. The technical nouns stay English because they are the names of things in the code. The grammar around them is Hindi because that is how the sentence assembled itself. Nobody who talks like this is doing it for effect — it is just faster.
Set your dictation to Hindi and button, state and modal come out as approximate Devanagari of the English sounds. Set it to English and the Hindi becomes confident nonsense. Both settings are wrong for the same sentence, and you get to pick which half to lose.
Because for most of their history they had to. On-device recognisers — including Apple's, which is genuinely good — load a language model per language. The model is the thing that decides whether a sound is a word, so the language is not a preference, it is a prerequisite. Ask it for English and it will find English in anything, including in Hindi.
Cloud models built on Whisper do not have this constraint in the same way. They were trained on a lot of languages at once and will detect what they are hearing, and — more useful here — they can be asked to translate rather than transcribe. Point a mixed-language sentence at a translation endpoint and what comes back is one coherent English sentence, not two half-broken ones.
The Hinglish sentence above comes back as something close to:
"If you press this button the state resets, but only when the modal is already open."
Which is what you meant, in the language your codebase and your coding agent are both already in. You did not translate it. You said the fast thing and the accurate thing arrived.
This is worth being precise about, because it is a trade and not a free win: you get English out, always. If you wanted the Hindi preserved verbatim — a quote, a message to someone who speaks it, a transcript for a person rather than a machine — this is the wrong tool and you want a straight transcription in Hindi instead.
There is no language setting, because there is no question to answer. You hold the hotkey, point at the thing, and talk however the sentence comes out. The audio goes to whisper-large-v3's translation endpoint and comes back as English, which then goes into the brief alongside the screenshot of whatever your cursor was on.
The audio is the only thing that travels, and it is deleted as soon as the text exists. Screenshots, window titles and the brief itself never leave the Mac. If you would rather nothing at all left, bring your own API key and our servers are out of the path entirely — or fall back to Apple's on-device recogniser and accept the single-language limit above.