Arabic Transcripts: Colloquial or Standard Spelling?

Decide before you start reviewing whether the transcript should show the words as said, in the speaker's dialect and its spelling, or as you want them read, closer to Modern Standard Arabic. There is no single right answer. A researcher or journalist usually needs what was said; a subtitler for a wide audience may prefer to move toward the standard. Decide once, write the decision on a short style sheet, and apply it to the whole text, so the same word is not spelled three ways.

What does automatic transcription do with a dialect?

It does not ask which dialect you speak. There is one Arabic setting among 33 languages, with no separate Egyptian, Gulf, Levantine or Maghrebi setting, and no switch that says keep the dialect or convert to standard. What comes back can be a mix: some words in spelling close to the standard, some in colloquial spelling, and some words that mean nothing in context.

A small real example: the Arabic sample on our home page is twelve short sentences read by one volunteer, and its text is exactly what the free preview returned, unedited. Even on those read sentences, which are easier than spontaneous talk, the output contains words that mean nothing in context, such as موتع and ثراقك. So the dialect is not the only source of error, and standard speech is no guarantee of a clean draft.

Five purposes, five defaults

Eight spelling decisions for the style sheet

1. Colloquial words with no fixed spelling. A word for "what" is شو in the Levant, وش in Saudi Arabia and إيه in Egypt. Pick one form for each word that recurs, and search for its other forms.
2. Foreign words. English words in Latin letters or in Arabic letters? A recording that mixes the two languages comes out mixed, so decide and apply (see recordings with more than one language).
3. Digits. Arabic-Indic (١٢٣) or Western (123)? Words or figures for amounts and dates? Do not let the text mix them.
4. Diacritics. Do not expect a vowelled text. A stray mark can appear; in our home-page sample جداً came back with a tanween. Decide whether to strip such marks.
5. Hamza, teh marbuta and yeh forms. أ, إ and ا; ة and ه; ي and ى. The variants rarely change the meaning but make a text look irregular. Our normalisation function shows one way to unify them for measurement.
6. Punctuation. The Arabic comma ، and question mark ؟ are different characters from the Latin ones. For players that draw them on the wrong side, see the Arabic guide to the direction fix.
7. Non-speech sounds. Laughter, throat-clearing, hesitations: delete them or mark them in brackets? Research often keeps them; subtitles usually do not.
8. False starts and repetition. Keep them in a verbatim text and remove them in a clean one.

Applying the decisions quickly

• Write the eight decisions on one page before you open the transcript.
• Check names, numbers, dates and amounts first, against the audio, since they cost the most when wrong.
• For each word that recurs in several spellings, search for every spelling in your editor and fix them together.
• If you do not understand a dialect word, do not guess from the draft's spelling: ask a speaker of the dialect or listen to the sentence twice.
• In the app, the Vocabulary hints field (after you create an account) can nudge the model toward the spellings of names and terms. It does not guarantee them.

When a review by yourself is not enough

If the dialect is not yours and the text will be published, quoted or relied on, have a speaker of that dialect check it while listening to the recording. A draft that reads fluently is not proof it is right. See how to transcribe Arabic dialects for the review workflow, and which dialects tools list for what vendors say they support.