The rewrite problem
Ask a general-purpose transcription model to transcribe a Moroccan podcast and something odd happens: the words on the page are Arabic, but they are not the words that were said. "بزاف" becomes "الكثير". "كيفش" becomes "كيف". A French word dropped mid-sentence disappears or is transliterated. The model has been trained mostly on Modern Standard Arabic, the register of news and formal writing, and it quietly normalises speech towards it.
For a written summary that might be acceptable. For captions burned into a clip it is not. Captions are supposed to be what the person said. When a Moroccan viewer hears one thing and reads another, the clip feels dubbed, and the guest sounds like a newsreader.
What "keeping the dialect" looks like in practice
A transcript that respects Darija has three visible properties:
- Dialect vocabulary stays as spoken. بزاف, دابا, واخا, كيفاش appear as the speaker said them, not as their Standard Arabic equivalents.
- Code-switching is preserved. Moroccan speech mixes French and sometimes Spanish or English inside a sentence. Those words should appear in Latin script, in the middle of an Arabic line, in the right visual order.
- The spelling is consistent. Darija has no single standard orthography, so a transcript should at least pick one convention and hold it. A tool that writes the same word two ways in one clip has not decided.
Our Darija subtitles page shows how ClipFlap handles this, with real clips: Arabic recordings are transcribed by a model that writes the dialect as it is spoken and gives every word a timestamp, and each clip is re-timed against the audio before the captions are drawn.
What to review before you publish
Even with a dialect-aware transcript, review the captions. Here is the short list that catches most problems:
- Names. Proper nouns are the most likely words to be misheard in any language, and Darija speakers often pronounce them with French phonology.
- Numbers. Check whether the caption uses Western or Eastern Arabic numerals and whether it matches what the audience expects; Moroccan viewers are used to Western digits.
- Spelling preference. ة versus ه at word endings, and similar choices, are matters of preference rather than errors. Decide once and apply throughout.
- Punctuation. Dialect transcripts often arrive with little or no punctuation. Without it, a five-minute monologue is one sentence and clip boundaries suffer. Check that questions end with ؟ and that pauses are marked.
- Direction and shaping. On the rendered clip, confirm that letters join correctly and that mixed-direction lines read in the right order. Some rendering paths break the joining at word boundaries, which produces detached letters that are unreadable at caption size.
About other Arabic varieties
ClipFlap uses one transcription path for every variety of Arabic. It was measured first on Moroccan Darija; Gulf, Egyptian, Levantine and other varieties go through the same path, but the results have not been validated in the same way and need a fluent review before publishing. We would rather say that plainly than claim a coverage we have not measured.
The honest summary
Keeping the dialect is not a stylistic choice. It is the difference between a caption that matches the audio and one that does not. If your tool rewrites Darija into Standard Arabic, the clip is mis-captioned, however clean it looks. Review names, numbers, spelling consistency and punctuation, and check the rendered clip for joining and direction before it goes out.
For the caption workflow in any language, see how to add captions to podcast clips; for the whole clipping process, the complete guide to podcast clips.