Why the audio never leaves your phone
The Better Translate team ·
There is a version of this app that uploads your audio. It would be easier to
build. We did not build it, and the reason is more practical than principled.
The framework has no cloud path
Speech becomes text on the phone, using Apple's SpeechAnalyzer, introduced in iOS 26. This is
not a setting we switched on. The framework is on-device only and has no server-side path to
switch to. Audio is processed locally, which is why there is no recording to upload, retain, leak,
or hand to anyone.
The accuracy is not a compromise either. In published benchmarks SpeechAnalyzer reaches a word
error rate of 2.12% on the clean portion of LibriSpeech and 4.56% on the noisy portion, ahead of
OpenAI's Whisper Small at 3.74% on the same clean set. On-device used to mean worse. It no longer
does.
What actually leaves
Text, and only in order to be translated:
- The transcribed text of what you said, and the language pair.
- The text found in a menu photograph, and the language pair. It is read on the device too.
On a fast connection the photograph goes with it, resized, because the model reads a page
more accurately from the image than from its text alone.
- Any context note you wrote, so the translation fits the situation.
- Recent turns of the current conversation, so pronouns and follow-ups make sense.
It goes over an encrypted connection to servers we run, which pass it to a translation
provider and return the result. Our relays keep no database of translations; text and photographs are held
only for as long as the request takes.
The honest caveat
On-device transcription is not the same as end-to-end privacy, and we are not going to blur
that line. The words you speak do leave your phone, as text, because that is what translating them
requires. What does not leave is your voice, your conversation history, and any identifier tying
a request to you. A menu photograph leaves only on a fast connection, for the one request. There is no account to attach it to.
The practical consequence: do not type anything into a free-text context note that you would not
send to a translation service. A context note is sent verbatim. It is not the place for a password,
a card number, or a medical detail.
What it costs us to do it this way
Transcription on the device means the app needs a recent iPhone, and it means we cannot improve
the transcriber ourselves. We took that trade. An app that holds a microphone open at a restaurant
table should be able to answer "where does the audio go" with "nowhere", and that answer is only
true if you build it that way from the start.