Whisper was trained on a large amount of Arabic, most of it close to Modern Standard Arabic: news, lectures, formal speech. That is where it performs best. Dialects such as Moroccan Darija, Algerian, Tunisian, Egyptian or Gulf Arabic are recognised less reliably, and the model often rewrites dialect words into their nearest standard form or spells them inconsistently.
Darija mixed with French is the hardest case, because the model has to switch between two languages and two scripts mid-sentence. The smaller models (“Fast” and “Balanced”) also make more spelling mistakes in Arabic than in English. Choosing “Best” helps noticeably; a Pro version with Whisper large-v3 turbo, which handles dialects better, is coming.
Arabic is written right to left, but subtitle files are plain text that many players lay out left to right. Without help, a full stop or question mark at the end of a sentence can jump to the wrong side. filetidy adds an invisible right-to-left mark (U+200F) to each Arabic line when “Right-to-left marks for Arabic” is on, which is the default, so punctuation stays where readers expect it in VLC and on TVs.
SRT files are saved as UTF-8 with a byte-order mark. Older players and some smart TVs use that mark to recognise the encoding; without it they may show Arabic letters as meaningless symbols.
- Best recognised
- Modern Standard Arabic (Fusha)
- Less reliable
- Dialects, especially Darija with French code-switching
- RTL handling
- Invisible U+200F marks, on by default
- SRT encoding
- UTF-8 with byte-order mark
- Translation
- Optional English subtitles from Arabic speech
- Recommended accuracy
- “Best” (Whisper small) where WebGPU is available