TL;DR: Romanian uses five diacritic letters — ă, â, î, ș, and ț — that general-purpose Latin-script OCR dictionaries often only partially cover, so extracted subtitles can come out with the wrong letter or the diacritic dropped entirely: "căutat" gets read as "cutat," and "să" gets read as "s." GeekLink recognizes Romanian with a character set built to include all five diacritics, and flags any line it isn't confident about for a quick manual check instead of silently guessing.
Extracting subtitles from a Romanian video? GeekLink runs OCR locally on Mac and Windows — draw a box, run OCR, get a timestamped SRT. Free tier, no account required.
Download FreeWhy does OCR get Romanian diacritics wrong?
Most OCR engines that support dozens of Latin-alphabet languages at once — English, Spanish, Portuguese, French, German, Italian, and many more — share a single character dictionary to keep the recognition model small and fast. That shared dictionary handles accents used across many widely-spoken languages well (é, ñ, ü), but letter-and-diacritic combinations used by fewer languages get thinner coverage.
Romanian's five diacritics — ă, â, î, ș, ț — are used only in Romanian. A dictionary optimized for dozens of languages at once can under-cover them, so recognition falls back to the closest character it does know: the diacritic gets swapped for a similar-looking mark or dropped entirely. The word is still readable, which is exactly what makes the error easy to miss on a quick skim of the subtitle file.
There's a second, older wrinkle specific to ș and ț: Romanian's correct Unicode letters use a comma below the S and T, but for years many fonts and keyboard layouts substituted the visually similar cedilla-below forms (ş, ţ) instead. Software that was tuned around the more common cedilla forms can still stumble on either variant showing up in the same video.
What do the errors actually look like?
In testing against a real subtitle video, two failure patterns showed up repeatedly:
| Correct Romanian | Common OCR miss | What happened |
|---|---|---|
| căutat | cutat | ă dropped entirely |
| să | s | ă dropped entirely |
| preț | pret / preţ | ț misread as a plain t or the cedilla variant ţ |
The first two rows are drawn directly from real-world testing; the third illustrates the same class of ț/ş confusion described above. In each case the output is a real, plausible-looking word — it just isn't the word that was on screen, and a reader skimming a long episode can easily read past it.
Which other languages have the same OCR gap?
The same shared-dictionary gap doesn't stop at Romanian. Checking every Latin-alphabet language GeekLink supports against the recognition dictionary turned up three more languages with real character gaps:
| Language | Missing characters | Impact |
|---|---|---|
| Vietnamese | Tone-combination characters (ạ, ấ, ế, ộ, and similar) | Largest gap — these combinations cover most common Vietnamese words, so accuracy was severely affected |
| Latvian | Macrons (ā, ē, ī) and soft signs (ķ, ļ, ņ) | Long vowels and palatalized consonants read as their plain equivalents |
| Azerbaijani | ə (a high-frequency letter in Azerbaijani) | One of the most common letters in the language misread or dropped |
Romanian's fix was validated against a real subtitle video with known problem lines. Vietnamese, Latvian, and Azerbaijani were checked by rendering real sentences and comparing character-by-character against the previous model, which is a strong signal but not the same as a real captured video — we recommend testing your own footage in each of these languages and telling us if anything looks off.
How does GeekLink handle these languages now?
GeekLink recognizes Romanian, Vietnamese, Latvian, and Azerbaijani with a character set built to cover each language's full diacritic inventory, instead of the general multi-language Latin dictionary. Every other Latin-alphabet language GeekLink supports (Spanish, Portuguese, French, German, Italian, Indonesian, Turkish, Polish, and more) was checked individually against its official character set and already has full coverage in the standard dictionary.
OCR from video is never going to be 100% perfect — font, resolution, compression, and background contrast all play a role. That's why GeekLink's editor flags the lines it's least confident about after extraction, so you can spot-check those instead of proofreading every line of a long episode by hand.
How to get accurate Romanian subtitles from a video
- Draw a box around the subtitle region so OCR only reads that area, not the whole frame.
- Run OCR — GeekLink auto-detects Romanian or you can set it explicitly if the video mixes languages.
- Check the flagged lines in the editor before exporting, especially in videos with small text, low contrast, or heavy compression.
- Export to SRT or ASS with correct timing.
For a large batch — a season of episodes, or a folder of hour-long recordings — the same workflow runs per file, so extraction and the confidence-flagging review scale with the batch instead of requiring a different process for long content.
Frequently Asked Questions
Does OCR support Romanian diacritics (ă, â, î, ș, ț)?
Support varies by tool. Many OCR engines share one dictionary across dozens of Latin-alphabet languages, and that shared dictionary can under-cover diacritics that are specific to a single language, like Romanian's ă, â, î, ș, and ț. GeekLink recognizes Romanian with a dedicated character set that covers all five.
Why does ă sometimes come out as a or ã?
When an OCR model's dictionary doesn't fully cover a diacritic character, recognition falls back to the closest character it does know — usually the base letter or a visually similar accented form. The result still looks like a real word, which is why the error is easy to miss without a proofreading pass.
What other languages have OCR diacritic issues like this?
Vietnamese, Latvian, and Azerbaijani have the same kind of gap in commonly shared Latin OCR dictionaries — Vietnamese tone-combination characters, Latvian macrons and soft signs, and the Azerbaijani letter ə. GeekLink uses an expanded character set for all four languages.
Does OCR recognize Vietnamese tone marks correctly?
Not always. Vietnamese tone-combination characters (ạ, ấ, ế, ộ, and similar) cover most common words, so a shared Latin OCR dictionary that doesn't fully include them can misread the majority of a Vietnamese subtitle line. GeekLink recognizes Vietnamese with an expanded character set built to cover these combinations.
Why does OCR miss Latvian macrons and soft signs (ā, ē, ī, ķ, ļ, ņ)?
The same shared-dictionary gap: Latvian's long vowels (macrons) and palatalized consonants (soft signs) are specific to Latvian, so a general Latin OCR dictionary can under-cover them and return the plain letter instead. GeekLink recognizes Latvian with a character set that includes both.
Does OCR support the Azerbaijani letter ə?
ə is one of the most frequently used letters in Azerbaijani, so a dictionary that doesn't include it affects a large share of words, not an occasional edge case. GeekLink recognizes Azerbaijani with a character set that includes ə.
How do I catch OCR mistakes in a long Romanian video without proofreading every line?
GeekLink's editor flags the lines it's least confident about after extraction, so you can focus review on those instead of reading through an entire hour-long episode line by line.