Arabic voice · Veeza
She takes visa questions the way Veeza already does on WhatsApp — in Arabic or English, in whatever dialect the caller actually speaks.
Click Talk to us in the corner. About two minutes, no typing, mic only.
Three prompts, ninety seconds. Each one targets a claim in the table below, so you can falsify us rather than take it on trust.
عايز أسافر إيطاليا الصيف الجاي، إيه الورق المطلوب؟
"I want to travel to Italy next summer — what documents do I need?" (Egyptian)
Dialect both ways. Speak Egyptian and she answers in Egyptian — عايز, دلوقتي, إزيك — not the fusha most Arabic agents fall back to.
طيب the appointment — هل ينفع أحجزه من الموبايل؟
"Okay, the appointment — can I book it from my phone?" (mid-sentence switch)
Code-switching. The switch happens inside the sentence and she never comments on it.
كام تكلفة تأشيرة اليابان بالظبط دلوقتي؟
"Exactly how much does a Japan visa cost right now?"
The refusal. If it is not in the knowledge base she says so and offers a human, instead of inventing a number that sounds right.
Veeza asked which Arabic speech models to use. This is the answer, including the parts we have not measured — routing you cannot audit is just a vendor lock by another name.
Arabic-first, dialect-aware. Auto-detects Egyptian, Levantine, Gulf, Iraqi, North African, Yemeni and MSA, and handles Arabic-English mixed speech natively. Best-on-dialect of their models — 3.5% CER on Levantine against 4.8% for the previous one.
What we can't claimWe have not measured Hamsa ourselves. Our public STT board publishes no Arabic streaming number at all, so this is a reasoned pin on vendor evidence, not a benchmark result.
Picked for refusal, not speed. On our LLM board it completes every multi-step tool task with zero fabrication, zero dead air and zero tool silence — and inventing an embassy requirement is the one failure that actually costs a traveller money.
What we can't claimThat board is measured in English. Grounding reliability is known to move by language, so read it as one English-measured pick applied across.
The leader on our Arabic TTS board — 4.15 Internal MOS in a blind A/B listening study, against 3.73 for the runner-up. ElevenLabs, which wins Spanish, German and French, comes third here at 3.33.
What we can't claimSmallest rater panel on the board, only four systems ran, and one FLEURS corpus with no dialect split. Carina is a multilingual voice, not an Egyptian recording — she reads Egyptian wording with an MSA-leaning accent. A dialect-native Arabic voice is a gap nobody has filled.
No single vendor is best at all three legs for Arabic. The transcriber that handles Egyptian and Levantine is not the one with the best Arabic voice, and neither of them is the model least likely to invent an embassy requirement. Speko routes each leg independently, so swapping the voice does not mean re-tuning the prompt or losing the dialect coverage.