Connected translation is genuinely good. Point a camera at a menu and get a readable rendering; type a sentence and get something close to idiomatic back. For most purposes on most days it is the right tool.
The problem is that its failures are not randomly distributed. They cluster precisely in the situations where you needed the language help most — which makes it worth being deliberate about what sits on the device.
Where connectivity actually fails
| Situation | What happens |
|---|---|
| Landing in a new country | No local SIM yet, roaming off, and you need directions immediately |
| Metro systems | Signal drops between stations, exactly where the signage is |
| Rural travel | Coverage thins where guidebooks are least useful |
| Aircraft and long trains | Hours of study time with no connection |
| Data caps abroad | The tool works but you ration it |
| Congested venues | Full bars, no throughput — stadiums, festivals, airports |
The metro case is the sharpest. Underground is where the station names, exit signs and transfer notices are, and it is where a phone reliably has nothing.
Latency is a real cost, not a small one
A round trip to a server takes a second or two on a good connection and longer on a weak one. That sounds negligible until you are reading a page and looking up twelve words. Two seconds each, plus the pause to re-find your place, is a meaningfully broken reading session.
Local lookup returns effectively instantly, and the difference changes behaviour: when a lookup is free, you look up the word you were only half sure about. When it costs a wait, you guess and move on — and the half-known word stays half-known.
Lookup cost shapes how much you learn. The words you skip because checking felt like too much friction are exactly the ones sitting at the edge of your knowledge, which is where learning happens. Reducing lookup cost to near zero is not a convenience feature; it changes what you end up acquiring.
Coverage is uneven in a way that matters
Machine translation quality tracks training data, and training data tracks the size of a language's web presence. Major European languages are served extremely well. Then it drops off, and it drops off exactly where fewer resources exist in the first place.
For languages such as Khmer, Lao, Hmong, Mizo, Ojibwe or Cheyenne, a carefully compiled bilingual dictionary is often more reliable than an automatic system — it was built by people who know the language, rather than inferred from whatever text happened to be online. The gap is widest for the languages with the least material available, which is also where a learner has the fewest alternatives.
A dictionary and a translator are not the same tool
| Machine translation | Bilingual dictionary | |
|---|---|---|
| Input | whole sentences | single words |
| Output | one rendering | all senses, with register |
| Grammar info | none shown | part of speech, forms, gender |
| Good for | getting the gist fast | understanding a word properly |
| Weakness | hides ambiguity | slower for long text |
A translator commits to one reading and shows you nothing about the choice. When a word has four senses and you needed the third, a dictionary tells you that the ambiguity exists. For a learner, seeing the range is often the entire point.
Privacy and battery
Two smaller points that add up. Everything typed into a connected translator leaves the device, which for medical vocabulary, legal documents or work material is a consideration people are increasingly making deliberately. And a local lookup does not wake the radio — on a long travel day with the phone also handling maps and photos, that is real battery.
The reasonable arrangement
Not either-or. Keep the connected tools for camera translation, long passages and languages where they are strong. Keep an offline dictionary installed for the language you are actually working on, so that the underground station, the aircraft cabin and the rural bus stop are covered.
It costs tens of megabytes and one download while you still have signal. The rule of thumb is simple: install it before you leave, not when you need it — by then, by definition, you cannot.
45 dictionaries that work with the radio off
NDT Studio's dictionaries store the full database on the device. No account, no connection, no waiting. Free on Android and iOS.
Open the dictionary appFrequently asked questions
Are offline dictionaries less accurate than online translators?
They are different tools rather than worse ones. A dictionary gives you a word's senses, register and grammatical information; a machine translator gives you a rendering of a whole sentence. For understanding a word properly, the dictionary is usually the better instrument, connection or not.
How much storage does an offline dictionary need?
A comprehensive bilingual dictionary is typically tens of megabytes — small next to a single video or a handful of photos. The trade of a modest amount of storage for guaranteed availability is generally a good one on any modern phone.
Do offline dictionaries still get updates?
Yes, through ordinary app updates when you are connected. The difference is that you choose when to update rather than depending on a connection at the moment of use — the database already sits on the device.