Why Arabic-first voice AI

Most voice AI platforms treat Arabic the way US car manufacturers treated right-hand drive in the 1990s: a regional variant, designed somewhere else, shipped late, and never quite right. You can get Arabic out of them, the characters come out, a voice reads the words, but ask anyone from Riyadh, Cairo, Beirut, or Amman to listen for thirty seconds and they'll tell you the voice isn't quite from anywhere.

We built Hakim to flip that default. Arabic isn't a locale we bolt on; it's the first shape the stack took, and the one we optimize for.

The shape of the problem

Arabic isn't one language, it's a spectrum. Modern Standard Arabic (MSA) is the register of newsrooms, contracts, and classroom announcements. The dialect you actually hear on the street in Casablanca, Cairo, Amman, or Riyadh is something else: different consonants, different vowel lengths, different loan words, different rhythms. A model trained on web-scale Arabic that's 90% MSA will speak fluently in exactly one register nobody uses at home.

On top of that, you get the writing system's ambiguity. Arabic script rarely marks short vowels. A sentence like كتب الكتاب could be kataba l-kitāb ("he wrote the book") or kutiba l-kitāb ("the book was written"). Human readers disambiguate from context. A TTS model has to guess, and most models guess wrong often enough to sound non-native.

Finally, there's prosody. Arabic rhythm doesn't map to English rhythm. When you pipe an Arabic sentence through a model that learned its intonation patterns from 10,000 hours of English podcasts, you get something that's technically audible and completely, unmistakably, off.

What "Arabic-first" actually means

We use the term carefully. It doesn't mean "supports Arabic." Every major model supports Arabic. It means:

1. The dialect tree is first-class. Hakim Arab v2 ships with MSA plus fourteen regional dialect varieties, spanning the major families from the Levant to the Maghreb (Levantine/Shami, Egyptian, Gulf, Najdi, Hejazi, Moroccan/Darija, Tunisian, Algerian, Libyan, Iraqi, Sudanese, Yemeni and the rest in the dialect-coverage post), each trained on dedicated corpora, not finetuned on the side. Picking a dialect isn't a prompt-engineering trick; it's an argument to the synthesis call.
2. Diacritics are recovered, not guessed. A preprocessor restores short-vowel marks from context before synthesis, and the model's decoder attends to both the restored diacritics and the bare glyphs. The result: the كتب problem above resolves with the same accuracy as a native reader.
3. Prosody is learned on Arabic speakers reading Arabic. Our training set doesn't share an intonation prior with our English models. We treat pitch accent, phrase-final lengthening, and emphasis separately per register. MSA newsroom cadence is not the same function as Najdi conversational cadence, and the model knows.
4. Latency is measured in Frankfurt, UAE, KSA, and Qatar. The synthesis call lives in the region your user is in. The p95 we publish (113 ms) is the number a user in Riyadh actually sees, not a Northern-Virginia benchmark dressed up as global.

Why we didn't just "add dialects"

We tried. The first version of Hakim Arab, a finetune on top of a multilingual base, could hit a passable Egyptian accent after enough prompting and enough post-processing, and we almost shipped it. We didn't, because the error distribution was wrong.

A multilingual-base approach gets the easy cases right. Common words, common phrases, common sentence structures. Where it falls apart is the tail, the exact place a native listener notices. The emphatic ض in رمضان, the glottal stop in سؤال, the velarization that separates صيف from سيف. These aren't edge cases in the corpus. They're table-stakes phonology.

So we retrained the acoustic model from scratch on Arabic-native audio, with MSA and dialect splits held out as separate conditioning heads. It cost us nine weeks. It was worth it.

What this unlocks

Arabic-first voice AI isn't a rounding error in the roadmap. It's the reason a customer-support center in Riyadh can answer calls with a voice its customers actually want to hear. It's why a fintech in Dubai can localize its IVR without explaining to product why the English version feels crisp and the Arabic version feels "like a robot." It's why educational content built for MENA students doesn't sound like a classroom from Indiana.

We built Hakim because we needed it. We made it generally available because you probably need it too.