How to translate a conversation in real time on iPhone, iPad or Mac
Translation apps are good at a sentence you type. A conversation is harder: two people, two languages, nobody wants to stop and type. A translation session turns the assistant into an interpreter that works while people are still speaking.
What a translation session is
Translation is switched on with the Translate button, in any session; the Translation template sets the assistant's role, but the switch is what turns translation on. It changes what the assistant is for: instead of answering questions, it translates. You choose the language to translate into; the language being spoken is detected, or you set it. Each sentence appears translated moments after the speaker pauses, next to the transcript of what was said.
It works in one direction at a time: from the other person's language into yours. When it is your turn to reply, open Translate and exchange From and To, and your words appear in theirs.
It is not a short list of popular languages: a Bulgarian speaker and a Japanese speaker are as possible as English and Spanish.
Setting it up
A minute before the conversation, not during it.
- Tap Translate and switch on “Translate chat and live session”. Set To to the language you want to read.
- Set From to the other person's language, or leave it on Auto Detect if you are not sure which language they will use.
- Put the device between the two of you, screen up, where both can read it. An iPad is easier to share than a phone.
- Start the session, and speak in turns. Short sentences translate better than long ones. To reply, open Translate and exchange From and To. Keep the app open with the screen on: the session ends if the phone locks or you switch apps.
Speak in turns, and in sentences
Live translation is at its best with one person speaking at a time. Two people talking over each other produce a transcript that mixes both, and the translation cannot untangle what the audio did not separate.
Complete sentences help too. A speaker who stops halfway, starts again and changes direction is hard to translate for a human interpreter, and no easier for a model. It is not about speaking slowly — normal speed is fine — it is about finishing the thought.
What it still gets wrong
Honestly: names, jargon and numbers deserve a second look. A surname the model has never heard may come out as a common word. Industry shorthand can be translated literally. A number said quickly can be misheard before it is translated at all.
The fix is cheap. When something matters — a price, a date, an address — say it again, or write it down and hold it up. Live translation is for keeping a conversation moving, not for signing a contract on the strength of it.
Where the words go
In a cloud session the audio is streamed to OpenAI to produce the transcript and the translation. The session itself — transcript, translation and history — is stored on your device, not on the app's servers. The other person should know an AI is translating; most people are relieved rather than bothered.
Which languages can it translate between?
Any language the model understands, auto-detected or chosen, into the language you pick. It translates one way at a time; to reply, open Translate and exchange From and To.
Does the other person need the app?
No. One device, placed between you, is enough. Both of you read the translation from the same screen.
Can I use it on a phone call?
On a Mac, a call's audio can be picked up through system audio. On iPhone the session hears what the microphone hears.