We make OmniTalk, a live voice and camera translation app for iPhone. This is the technical explanation behind our offline mode. It is not a claim that offline translation is better across the board. It is not. Offline translation is usually worse at translation quality. The reason to use it is that it works at all when there is no signal. If you are comparing an offline translation app for iPhone, the important question is what you give up to get offline mode.
What happens on the phone versus in a data centre
When you speak into a translation app, the phone does not simply "translate" your voice. It runs a sequence. First it captures audio. Then speech recognition converts the audio to text. Then the text is translated. Then the result is either spoken aloud or displayed.
The difference between on-device and cloud is where those middle steps run.
Apple's on-device speech recognition and translation runs speech recognition and translation on the iPhone's processor. The language model is already downloaded. The audio is not sent to a translation server. The phone does all the work locally.
A cloud path works differently. The phone captures the audio, compresses it and sends it to a data centre. A premium cloud speech engine transcribes and translates the speech. The server returns text or audio. The phone then displays or speaks the result.
The latency budget of live conversation
Live conversation is the hardest case. If the translation lags too long, people repeat themselves. Someone talks over the translated voice. The exchange falls apart.
On-device translation removes the network round trip and the server queue. The phone starts processing immediately. That makes the first text appear faster in many cases. It does not mean the local model is faster at the actual translation task. It means there is no network wait.
Cloud translation adds several steps: uplink, server processing, downlink. On good Wi-Fi this can be quick. On airport Wi-Fi, train-station 4G or a hotel ballroom connection, it can slow down. Inside a plane, a basement car park or a rural valley, it is not slow. It is absent. Offline still works.
On an older iPhone, local inference can be slow. The gap narrows. But local slowness is predictable. Network slowness is not.
Why a local model is smaller and less capable
A phone has limited memory, a small battery and a processor that must also run the camera and display. A cloud data centre has racks of accelerators and power to spare.
That forces cuts in the local model. The vocabulary is smaller. The model holds less context. It is worse at long sentences, numbers, names, idioms and background noise. Cloud models can be much larger and are updated without an app update.
Offline translation is not better across the board. It is worse at quality in most conditions. It works best with short, clear speech: directions, check-in questions, simple purchases, a menu item. It starts to break with long subordinate clauses, code-switching, sarcasm, strong accents and noisy rooms.
If quality is your only concern and you have a stable connection, cloud translation is the stronger option.
Privacy: what actually leaves your phone
In offline mode, the translation audio stays on the phone. Camera frames stay on the phone. The speech recognition and translation run locally. OmniTalk BASIC uses Apple's on-device speech recognition and translation frameworks, so the audio is handled by the phone rather than shipped to a translation service.
Switch to PRO and that changes. Audio is sent to a premium cloud speech engine. The speech content is processed by someone else's computer. That is not automatically malicious. It is a different privacy model. The provider may retain audio, store transcripts or use them for service improvement under its own policy. You should assume cloud audio leaves your control unless the provider says otherwise in writing.
OmniTalk saves conversation transcripts, and this is the part worth reading carefully: they are stored in your own iCloud private database, which means they sync across your devices and survive an uninstall. That is your Apple account, not ours — we cannot read them — but it is not the same as "never leaves the phone".
If privacy is the priority, offline is the cleaner answer. If quality is the priority, cloud is usually stronger. These are different jobs.
Language packs, storage and downloading before you fly
OmniTalk launched on 5 August 2026. The current version is 1.0.6 and the app download is 18.9 MiB. That is small because the language models are separate downloads. You must download a language pack before you go offline.
Do it on decent Wi-Fi. Do it before you leave. If you only realise you need offline translation on the plane, you are too late. Airport Wi-Fi is often congested. Hotel Wi-Fi the night before is better.
Language pack sizes vary by language, and the app shows the size before you commit to a download. They can take meaningful iPhone storage. If you download several languages, check the app and delete packs you no longer need. iOS may offload an unused app under storage pressure. If that happens, re-download and check your packs are still there before you fly.
For two-way live translation, make sure you have the pair you actually need. Downloading French does not help in rural Japan. If you are going somewhere with multiple languages, install the right packs before you travel.
OmniTalk works offline once a language pack is downloaded. That includes the speech and camera translation features. You cannot switch on offline mode on the plane without the pack.
If you want a printable backup, OmniTalk has 30 travel phrasebooks at /phrasebook. These are separate from the offline language packs.
What this means in OmniTalk
| Mode | Engine | Works offline | Live conversation cap |
|---|---|---|---|
| BASIC | Apple's on-device speech recognition and translation | Yes, after language pack download | Free tier: 1 hour per month; PRO: unlimited |
| PRO | A premium cloud speech engine | No | Free tier: 1 hour per month; PRO: unlimited |
Camera translation and text translation are not metered in the free tier. The one-hour cap applies to live two-way voice conversation. PRO removes the cap.
BASIC is the on-device mode. If you are going somewhere with no signal, download the pack and use BASIC. PRO will not help you without a connection.
The free tier is not a trial of PRO quality. It is one hour per month of live conversation. If your live conversation needs are heavier, PRO removes the cap and switches you to the premium cloud engine.
When OmniTalk is the wrong tool
OmniTalk is an iPhone app for two people around one phone. It does not plug into Zoom, Teams, WhatsApp or Instagram. If you need translation inside a video call, other tools do that. OmniTalk does not.
If you only need to translate a menu once, the built-in iOS translation tool is simpler and already installed. If you need high-stakes translation for medical or legal situations, a human interpreter is the right call. OmniTalk can help with practical conversation, but it is not professional interpretation.
OmniTalk is iOS only. If you carry Android, this app is not for you.