99 languages. 16 we can prove.

How this works in the app

• You choose the language. It stays chosen. Pick it once in Settings, or switch it in two clicks from the tray menu. The language you picked is named in the tray menu and on the tray tooltip, so you can never be dictating into a language you did not intend — and while automatic detection is still deciding, both read “Detecting language”.
• Nothing is mixed. One language is active at a time. InkBeepAI does not try to guess per sentence, because guessing needs several seconds of speech and a dictated phrase does not have them. We measured that: on dictations of 3 to 15 seconds, guessing cost nothing in 13 of 14 languages; on short phrases it made 10 of 14 worse, and Hindi was misread as Urdu often enough to ruin a session.
• Automatic detection is there if you want it, off unless you turn it on, and it works the safe way round: it keeps using the language you picked until one dictation has at least four seconds of actual speech, identifies the language from that one, switches to it, and then stays there. The tray menu shows what it picked.
• It never translates. You speak Spanish, you get Spanish text. Dictation, not interpretation.
• Recognition is still entirely on your PC. Both speech models ship inside the download. Adding 98 languages did not add a single network request — the licence check is the same one it always was.

The 16 we verified

The figure is the error rate: out of every hundred words spoken, how many came back wrong (for Japanese, written without spaces, out of every hundred characters). Lower is better and zero is perfect. It counts a missing word, an extra word and a wrong word alike, and it is measured against the reference transcript that came with the recordings — including their own conventions for numbers and names, which we did not tune for.

Measured on recordings of native speakers

English’s 7.8% is for the small English-only model the app runs by default, because that is what an English user actually gets. The same recordings through the large multilingual model score 5.5%, and you can switch English to it in Settings. The fast one is the default deliberately: the gap in speed is far larger than the gap in accuracy, and dictation lives or dies on the wait.

Nine passed on synthetic speech and failed on real speakers

Every language below cleared our synthetic tests, several of them with almost no errors. On recordings of actual native speakers they missed the bar — some on the average, some because too many individual sentences came back wrong. They are still in the app, because they run and some people will want them; the app tells you they missed when you pick one.

To be verified, a language needs an average under 15% on native speakers and at least 70% of its sentences individually under 15%.

That gap is the entire argument for the human test. Slovak scored 0.0% on synthetic speech and 13.1% on real speakers, with 7 of its 20 sentences outside the bar. It is why we do not simply repeat the model's own language list back to you as a feature.

How it was tested

Four passes, all on one machine, using the same speech engine and the same settings the shipped app uses.

A language has to be good on average and good sentence by sentence, so one lucky sentence cannot carry it. On native speakers that means an average under 15% and at least 70% of sentences under 15%. The synthetic sentences were chosen without digits, because “2026” against “twenty twenty-six” scores as an error without anybody having misheard anything. The native-speaker transcripts are used exactly as they came, so there a number the model writes as digits does count against it.

What this costs you

Said plainly, because you will find out anyway:

• The download is much bigger. Two speech models ship inside it — a small English one and a large multilingual one. It is a one-time download and it is what keeps everything on your own machine.
• Languages other than English want a graphics processor. The graphics built into the processor counts. We have measured speed on one laptop, with Intel Iris Xe graphics; older or weaker graphics will be slower. Without a usable one, the multilingual model runs on the processor alone and is far too slow to dictate with. InkBeepAI times the larger model on your own PC when it loads it, and if it is far too slow it says so once and offers the way out that applies — turning the graphics processor on, or switching back to English. It does not let you find out by waiting.
• English does not need a graphics processor. It keeps the same small English model it has always used. On a PC with a graphics processor it now runs on that; without one it runs on the processor, as it always did.
• 83 languages are not verified. They run, and the app tells you when you pick one. We would rather offer them with that caveat than pretend to a hundred verified languages.

The other 74

Like the nine above, none of these is verified. They run; the app labels them as not verified when you pick one.

Did not clear our synthetic screen (44)

Tested on synthetic speech: most came in outside the bar, some of them far outside, and two cleared the quiet pass and fell over once we added room noise. A few of these results are partly scoring artefacts — Chinese output mixes simplified and traditional characters, and Hebrew spelling varies — so read this as “not proven”, not “broken”.

Afrikaans, Albanian, Amharic, Arabic, Azerbaijani, Bengali, Bosnian, Bulgarian, Burmese, Chinese, Croatian, Galician, Georgian, Greek, Gujarati, Hebrew, Icelandic, Kannada, Kazakh, Khmer, Korean, Lao, Lithuanian, Macedonian, Malay, Malayalam, Maltese, Marathi, Mongolian, Nepali, Pashto, Persian, Serbian, Sinhala, Slovenian, Somali, Sundanese, Swahili, Tamil, Telugu, Thai, Urdu, Uzbek, Welsh.

Not in our tests (30)

The model accepts these, and we have no measurement to give you either way.

Armenian, Assamese, Bashkir, Basque, Belarusian, Breton, Faroese, Haitian Creole, Hausa, Hawaiian, Javanese, Latin, Lingala, Luxembourgish, Malagasy, Māori, Norwegian, Norwegian Nynorsk, Occitan, Punjabi, Sanskrit, Shona, Sindhi, Tagalog, Tajik, Tatar, Tibetan, Turkmen, Yiddish, Yoruba.

Common questions