Travel Routine
Translation apps fail in the places you need them
Machine translation depends on connectivity, clear input and a widely supported language pair, and all three tend to be weakest exactly where a traveller is most stuck.

A phone translator works impressively in a calm room with a good connection. The situations that require one are usually noisy, offline and involve a language with less training data.
The model runs somewhere, and often not on the phone
High-quality translation is computationally heavy and typically runs on a remote server. Without a connection, the app falls back to a smaller local model or refuses entirely.
Offline language packs exist and should be downloaded before departure, since the moment they are needed is generally the moment there is no data to download them with.
Those packs are smaller and noticeably worse, which is acceptable for a menu and unreliable for anything where precision matters.
Speech input is the weakest link
Conversation mode requires the phone to hear clearly, segment the speech and identify the language. A busy street, a station concourse or a market defeats the first step.
Accents and dialects compound this. A speaker using a regional variety may be transcribed incorrectly, after which the translation is fluent and wrong.
Typed input avoids the transcription problem entirely, which is why handing over a phone with text on it works better than speaking into it.
Language pairs are not equally supported
Translation quality depends on how much parallel text exists for a pair. Widely used European pairs are strong; less documented languages are considerably weaker.
Many systems also translate through an intermediate language, so an uncommon pair passes through a third language and accumulates errors at each step.
The output remains grammatical throughout, which is the trap: fluency is produced by the model regardless of whether the meaning survived the intermediate step.
Camera translation solves a different problem
Pointing a camera at a sign or a menu handles printed text well, since the input is clean and the vocabulary is narrow and repetitive.
It struggles with handwriting, stylised type and scripts it must first recognise, which is why a chalkboard menu is harder than a printed one.
It is nonetheless the most reliable mode for a traveller, because the failure is visible: an unreadable output is obvious, whereas a wrong translation of speech is not.
What to prepare rather than translate
The phrases needed under pressure are predictable and few: an address, a dietary restriction, a medical condition, a request for help. Having them written correctly in advance removes the dependency.
A written card carried in a wallet works with no battery, no signal and no explanation, which is the standard any travel tool should be judged against.
Questions readers ask
What do people lose most often when travelling?
Phones, chargers, passports and glasses, overwhelmingly at security, in seat pockets and in accommodation. All are transition points rather than random events.
Is a checklist worth it for regular travellers?
More so, not less. Frequent travellers pack on autopilot, which is exactly the mode in which a single item is quietly omitted.
Also by Nadia Halloran
- Jet lag has a direction, and that changes what worksTravel Routine
- Rail passes are usually the wrong answer, and sometimes the right oneTrains & Ground
- Hostels, guesthouses and the middle nobody booksWhere You Sleep
- Doing laundry on the road is a planning decisionTravel Routine





