Multilingual Voice Typing on iPhone

Three settings people mix up

Almost every multilingual voice typing problem on iPhone comes from confusing three separate settings. They live in different places and do different jobs.

• Keyboard language: the layout, autocorrect, and suggestions you get while typing. Auri Keyboard supports more than eighty.
• Recognition language: what the speech model listens for. This one causes the nonsense output.
• Translation target: the language you convert finished text into. Auri Keyboard translates into 20+ languages.

Recognition is the setting that produces the confusing failures. Speak Spanish while the model expects English and you get English words that sound vaguely like what you said, which reads like a broken app rather than a mismatch.

Check recognition first whenever output looks like phonetic gibberish. Check keyboard language when autocorrect fights you word by word. Those are two different problems with two different fixes.

Set up the two or three languages you actually use

Keep the list short. Every extra language adds taps to the switch cycle, and a six-language list means hunting for the one you need while someone waits for a reply.

1. Add your languages under Settings > General > Keyboard > Keyboards.
2. Order them so your two most-used sit next to each other in the cycle.
3. In the keyboard app’s settings, confirm the voice input language for each.
4. Dictate two sentences per language into Notes and check the output.
5. Note which language needs a download before it works offline.

Test the switch itself, not only recognition. The failure people hit in real conversations is switching the keyboard language and forgetting the voice input language, then speaking into the wrong model.

If a third language shows up a few times a year, handle it with translation rather than adding it to the cycle. Optimize the setup for the messages you send weekly.

What breaks when you code-switch

Bilingual speech is normal and recognition engines are bad at it. Each take is transcribed as one language, so the model resolves the other language into whatever sounds closest in the one it expects.

The habit that fixes most of this: one language per take. Finish the thought, stop, switch, speak again. Two short takes beat one mixed take you have to rebuild.

Number formats deserve their own check. Decimal commas, thousands separators, and day-month order differ by language, and a transcript that looks right can carry the wrong convention into an invoice or a schedule.

Write directly or write then translate

Both routes are valid, and the choice depends on how fluent you are in speech versus writing in that language. Pick per message rather than settling on one rule.

Translation is strongest on plain, literal sentences. Speak in short declaratives, avoid idioms from your first language, and name the thing instead of gesturing at it. Long sentences with three clauses come back accurate and unreadable.

Read the translated version before sending, even when you are not confident in the language. Names, numbers, and dates are checkable without fluency, and those are the errors that cause real problems.

Words that fail in every language

Some items fail regardless of which language you set, because they were never common words in any training data. Handle them with a note or a word list instead of repeating yourself.

• Your own name and hometown, especially with diacritics.
• Colleague and client names that cross language boundaries.
• Brand and product names spelled unlike how they sound.
• Street addresses and postal codes.
• Field-specific terms borrowed from another language.

Keep a note titled with correct spellings for the twenty terms you use most, and paste rather than fight the engine. Where the app supports custom words, add them there instead.

Accented characters are worth checking every time. A name that loses its accent looks careless to the person reading it, and it is the one error a reader always notices.

A daily routine for mixed-language chats

A short routine keeps mixed-language messaging from becoming a series of corrections. It fits in the seconds before you start talking.

1. Decide the language of the message before you tap the field.
2. Switch keyboard and voice input language together.
3. Speak one take, one language, one point.
4. Scan names, numbers, and accents.
5. Translate only if you drafted in the other language on purpose.

When a chat moves between languages every few messages, batch by language instead of matching each reply. Answer the Spanish thread, then the English thread, and you switch settings twice rather than eight times.

For grammar cleanup after speaking a language you are still learning, one grammar pass and one style pass is enough. More passes drift the meaning while fixing wording nobody would have questioned.

Limits worth knowing

Multilingual voice typing has honest limits, and knowing them prevents an embarrassing message to someone who matters.

• Formality is easy to lose. Machine output picks a register that may be wrong for your reader.
• Idioms translate literally and land strangely.
• Regional variants differ, and a model may produce phrasing from the wrong region.
• Legal, medical, and contractual wording needs a fluent human reader.
• Offline behavior varies. Server-side features stop without a network.

For high-stakes messages in a language you do not write confidently, draft by voice for speed and have a fluent colleague read it before it goes out. That is a two-minute favor that prevents a real misunderstanding.

On a phone, Auri Keyboard defaults to cloud voice typing and offers local speech to text. If a conversation covers something you would rather not send anywhere, that is the setting to change before you start speaking.

FAQ