The path from speech to translation
You speak → the call captures audio → a translation worker receives the audio after both participants agree → cloud speech processing recognises and translates the speech → translated text and, where supported, spoken audio return to the call. Text can be displayed as subtitles; speech synthesis makes the spoken output. These outputs are part of one live process, not a requirement to display subtitles before any sound can play.
The same process runs in the opposite direction for the other speaker. Translation is for calls inside AiThao, not ordinary carrier phone calls. An internet connection is needed for the call and for cloud processing.
What you hear
In the normal translation-only mode, the app subscribes to translated audio. Before first-time cloud consent, and when translation cannot start, it plays the original voice. Hearing an original voice is therefore not proof that translation has started.
Translation is selected by default for calls, but both people must agree. If both use the same language, the app does not need to translate. The language pair on the call screen is the first thing to check if the result seems wrong.
What you see
Subtitles show the translated text when enabled. The other person’s latest line is larger; your own line shows how your words came out in their language. Subtitles start off until you enable them and the preference is shared with Settings.
For a language with spoken output, text is optional. For a language supported only through subtitles, enable subtitles to read the result. Subtitles do not move translation onto the phone or remove cloud processing.
Why a live translation still takes time
Speech arrives over the network, the model needs enough context to interpret it, and translated speech must be produced and played. Noise, overlapping voices, language choice and connectivity all affect the result. There is no fixed delay guaranteed for every conversation.
Short turns make it easier to notice a wrong phrase. Confirm names, dates and amounts in writing. A subtitle is a useful view of the result, but it can repeat the same recognition or translation error you heard.
When the cloud can hear the call
Translation starts only after both participants agree. For an encrypted call, the call key is then given to the translation worker; decoded audio reaches cloud translation and the call is not end-to-end encrypted. The encryption indicator disappears. Video is not sent to the translation provider.
Turning translation off stops cloud processing, but it does not take that key back. For a later conversation without cloud translation, begin a new untranslated call and check the encryption indicator on compatible phones. The call-security page describes that boundary.
Questions
Is my call recorded to be translated?
No. Audio is streamed to the translation service while you speak, and AiThao does not record the call.
Why is there a short delay?
Speech has to be recognised, translated and spoken again. The delay depends on the network, the languages and how long each sentence is.