If you write in a language that is not your first, dictation is often faster and less tiring than typing, especially when the keyboard layout fights you. But it fails in a specific, predictable way: people leave the recogniser set to one language, speak another, and conclude that dictation cannot handle their accent. Almost always it was handling a different language. This guide covers picking the right language and regional variant, what accents genuinely cost you, how to deal with mixing two languages in one sentence, and the editing pass that turns a rough transcript into writing you can send.
Set the language and the regional variant first
Zahvox: choose the language in the picker above the transcript before you press record.
Mac: System Settings, Keyboard, Dictation — add every language you use under Dictation Languages.
Windows: Settings, Time and language, Speech, then add the speech language pack you need.
iPhone or Android: add the language to the keyboard, then long-press the microphone to switch between them.
Pick the regional variant too — English (India), English (UK), Spanish (Mexico), Arabic (Egypt) — because vocabulary and vowel models differ.
Never leave it on a language you are not currently speaking; that single mismatch produces most of the nonsense output.
What an accent actually costs, honestly
Speech models are trained on far more first-language audio than second-language audio, so a strong accent does reduce accuracy — but far less than most people assume, and unevenly. Ordinary sentences with common words come through well. What breaks is specific: proper nouns, brand names, technical vocabulary, and numbers spoken quickly. The practical consequence is that you should not slow your whole delivery down. Speak at a normal pace and slow down only for names, figures, and unusual terms. Speaking uniformly slowly actually hurts, because these models use the rhythm of surrounding words to decide what a sound was, and an unnaturally stretched sentence removes that context.
The variables that matter more than your accent
Microphone distance: six to twelve inches from your mouth beats any settings change.
Room noise: a fan, a TV, or traffic costs more accuracy than a second-language accent does.
Sentence length: eight to fifteen words gives the model enough context without drifting.
Ending flat: trailing off makes the last two or three words unreliable — finish the sentence with intent.
A cheap wired headset usually outperforms laptop built-in microphones and many wireless earbuds.
Code-switching: mixing two languages in one sentence
This is the hardest case, and it is worth being straight about it. Browser and operating-system dictation recognise one language at a time. When you mix — English technical words inside an Urdu sentence, Spanish inside English, Hinglish — the recogniser forces the foreign words into the sound patterns of the language it is set to, and you get near-miss spellings rather than the word. Two workable strategies. First, set the dominant language, the one carrying most of the grammar, and accept that the inserted words will need a fix in editing — usually a handful per paragraph, quick to correct. Second, split by block: dictate the passage in one language, switch the language setting, then dictate the next passage. Slower to set up, noticeably cleaner output, and worth it for anything a client or examiner will read.
Speaking punctuation is not optional in a second language
Say "comma", "period" or "full stop", and "question mark" out loud — pauses alone are unreliable.
Say "new line" and "new paragraph" so structure survives.
Speak numbers slowly, and say currencies as words when accuracy matters.
Spell out unusual names once by letters, then use them normally afterwards.
Learn the punctuation words in the language you are dictating in — they only work in that language.
Write in your second language, or in your first and then translate?
Both work; the choice depends on the audience. Drafting directly in the target language keeps the phrasing natural and avoids the stiff, over-literal feel of translated text — the right call for client emails, marketplace replies, and anything conversational. Speaking in your first language and translating afterwards is better when the reasoning is complex and you cannot afford to lose nuance while hunting for vocabulary. A middle route works well for many bilingual writers: dictate the structure and argument in your stronger language, then dictate the final version in the target language now that you know exactly what you want to say. The second pass is always faster and cleaner than the first.
The editing pass that makes it sendable
Second-language dictation reliably produces three fixable things: wrong homophones, missing articles, and a first sentence that meanders because you were still deciding what to say. Dictate into the free Zahvox tool at /tool rather than straight into Gmail or a client chat, because the raw pass is not what you want to send. Read the transcript once from the top, delete the meandering opening sentence — the second sentence is usually the real one — fix the proper nouns, then copy it across. That pass takes under two minutes and is the difference between text that reads as dictated and text that reads as written.
A realistic expectation for the first week
Day one: set the languages up properly and dictate a few short notes only.
Days two and three: dictate replies you would have typed anyway, and time the difference.
Day four: try code-switching deliberately so you learn where it breaks for your pair.
Day five onward: dictate first drafts, always edit before sending.
Expect roughly a week before it stops feeling awkward — this is a speaking habit, not a settings change.
