Maghrebi Arabic call transcription

What makes Maghrebi acoustically distant from Khaleeji and MSA

The single most consequential Maghrebi feature for ASR is short-vowel deletion. Where MSA writes "kataba" (he wrote), Moroccan Darija reduces this to "ktib" or just "kteb" — collapsing the verb into a consonant cluster that mainstream Arabic acoustic models, trained on vowel-rich MSA and Mashriqi dialects, cannot decode. Algerian and Tunisian Arabic apply similar reductions. The result is dense word-initial and word-medial consonant clusters ("ktbt" for "I wrote", "shrbt" for "I drank") that are phonotactically illegal in Eastern Arabic varieties.

Phoneme inventory diverges further. Moroccan Darija realises MSA ق (qāf) variably as [q], [ɡ], or [ʔ] depending on lexical item and region — a single speaker may use all three within one utterance. The MSA ث (thāʼ) merges with [t] in urban speech and with [s] in some rural varieties; ذ (dhāl) merges with [d] or [z]. Tunisian Arabic preserves [q] more reliably than Moroccan but introduces its own innovations. Libyan dialect, sometimes treated as a transition between Maghrebi and Egyptian, sits acoustically closer to Egyptian than to Moroccan but retains Maghrebi vowel-deletion patterns.

Stress placement is the third axis of divergence. Maghrebi Arabic uses penultimate stress in many environments where Mashriqi varieties stress the antepenult; this reshapes the spectral envelope of multisyllabic words enough that an MSA-tuned model misaligns word boundaries. CallScribe's Maghrebi acoustic adapter is trained on telephony-grade Maghrebi speech rather than on broadcast MSA — see [/dialects/msa](/dialects/msa) for why MSA-tuned models fail on Maghrebi audio that they handle well in formal Eastern dialects.

Lexical strata: Berber/Tamazight, French, Spanish, Italian

Maghrebi Arabic carries a Berber/Tamazight substrate that no other Arabic variety has. Toponyms, plant and animal names, agricultural terminology, and culturally rooted vocabulary in Moroccan and Algerian Arabic come from Tamazight rather than Arabic. A Casablanca speaker discussing a domestic issue may use words like "ttsa" (twin) or place names with Berber roots that an Arabic-only language model cannot resolve. CallScribe's Maghrebi lexicon includes the most frequent Berber-origin terms in colloquial use.

Above the Berber substrate sits a heavy Romance-language superstrate. Moroccan and Algerian Darija borrow extensively from French — not just nouns ("la voiture", "le téléphone") but full functional phrases ("c'est bon", "d'accord") and code-switching at clause boundaries. Tunisian Arabic borrows from French and Italian. Libyan Arabic carries some Italian loans from the colonial period. A typical Maghrebi customer-service call alternates Arabic and French at a rate seen nowhere else in the Arab world — even Lebanese Arabic, with its famous French code-switching, switches less densely than Moroccan Darija. CallScribe handles Arabic-French code-switching at the segment level and preserves French in Latin script in the transcript.

Why mainstream ASR fails on Maghrebi

Maghrebi training data is scarce relative to Eastern Arabic varieties. The Whisper paper (Radford et al., 2022) reports Arabic word-error rate as a single aggregate number, but breakdowns from independent benchmarks consistently show 15-25 percentage points higher WER on Moroccan Darija than on Egyptian or Levantine for the same model. Commodity ASR APIs — Google, AWS, Azure — typically claim "Arabic" support and quietly mean "MSA-with-some-Egyptian-coverage". Operations leaders running calls into the Maghreb on those tools see transcripts that look like word salad and assume the audio is unusable. The audio is fine; the model is wrong for the dialect.

CallScribe's positioning here is direct: we transcribe the dialect family that mainstream Arabic ASR effectively cannot. Our Maghrebi WER on clear telephony audio is 16-22% — substantially worse than our 8-12% Khaleeji number, but substantially better than the 30-45% you typically see on commodity ASR with Maghrebi audio. We are honest about the gap because the alternative is unusable rather than imperfect.

Where Maghrebi shows up in GCC operations

Three patterns. First, Maghrebi expatriates in the GCC — particularly Moroccans and Tunisians in the UAE, Qatar, and Saudi Arabia — use Darija on personal calls but shift toward MSA or Khaleeji on formal calls; mixed-register speech is common. Second, GCC businesses with Maghreb operations (banks with Casablanca branches, telcos with Algerian or Tunisian subsidiaries, hotels with North African clientele) handle outbound and inbound calls in Maghrebi dialect. Third, regional brand campaigns and customer research occasionally surface Maghrebi voice content for sentiment and topic analysis — see our [/use-cases/coaching](/use-cases/coaching) workflow for how we structure dialect-tagged coaching evidence packs.

CallScribe accuracy and limits on Maghrebi

Word-error rate on our internal Maghrebi telephony benchmark is 16-22% for clear single-speaker Moroccan or Algerian audio, 14-20% for Tunisian (which sits closer to Mashriqi norms), and 18-26% for noisy multi-party calls with French code-switching. Libyan dialect, where coverage is sparsest, sits at 22-30%. We treat Maghrebi as a dialect family rather than a single homogenous model: the country tag (MA, DZ, TN, LY) biases the lexicon and applies country-specific French-borrowing patterns. We do not currently support Hassaniya (Mauritanian Arabic) — that is a separate engagement.