How do I translate a recording into another language?
Translating speech is two steps people collapse into one, and the wrong order is why the output reads badly. The order that works, on iPhone.
On this page
Translate the transcript, not the audio. Speech-to-text first, then text-to-text: transcribe the recording in its original language, check the transcript, then translate. Doing it the other way (asking one tool to listen in Spanish and write English) compounds two error rates and gives you nothing to check when the result reads oddly. AI Transcribe does both steps on the phone: transcription in 32 languages with automatic detection, then translation of the transcript into 46.
Why the order matters
A transcription error and a translation error behave differently. A transcription error is visible, you read the transcript and see that a word is wrong. A translation error built on top of a mis-heard word is invisible: the output is fluent, grammatical, and quietly says something that was never said. When the two steps are collapsed, you lose the checkpoint between them.
Keeping them separate gives you a place to intervene: fix the name the model misheard, then translate.
The steps
- Transcribe in the original language. Automatic detection handles this without you selecting anything; if the recording switches languages mid-way, expect the dominant one to win.
- Read the transcript first. Names, numbers, technical terms, these are where errors cluster and where a translation will amplify them.
- Translate the transcript. The transcript keeps its speaker labels and timestamps, so a translated interview stays navigable.
What this is good for, and what it is not
Good: understanding a recording in a language you do not read; sharing an interview with a colleague who does not speak the original; producing a working draft for a translator to polish.
Not good enough on its own: anything published, legal, or medical. Machine translation of speech carries the errors of both steps, and for consequential text a human pass is not optional. Say plainly which draft is machine-produced when you pass it on.
Not the same thing as subtitles. A translated transcript is text. Burning translated subtitles onto a video is a separate video task.
The honest constraints
Both steps run in the cloud, so a connection is required. Accuracy depends far more on the recording than on the language pair, a clean recording in a less-common language beats a noisy recording in a common one. And a translated transcript inherits every gap in the original: what was inaudible stays missing, it just becomes missing in a second language.
Frequently asked questions
How many languages can it transcribe and translate? Transcription covers 32 languages with automatic detection; transcripts can be translated into 46.
Can I translate a WhatsApp voice message? Yes, import the audio, transcribe it, then translate the transcript. See transcribing WhatsApp voice messages.
Will the translation keep who said what? Speaker labels and timestamps stay with the transcript, so a translated interview is still navigable.
Is it good enough to publish? Treat it as a working draft. For anything published or consequential, have a human check it.