For years, Arabic voice input meant speaking slowly in something close to Modern Standard Arabic and hoping for the best. Most people simply didn't bother. That may be about to change. On October 15, MBZUAI and Inception42 released Jais 3 together with Jais Speech, an open Arabic speech-recognition model. According to Gulf News and WAM, it handles MSA and dialects from 19 countries, keeps English and French words in Latin script, detects the dialect and can add vowel marks. TII has also released a speech model focused on Arabic and Emirati speech, as FourWeekMBA reports.
So dialect-aware voice input is suddenly within reach for regional products. But a better model doesn't fix the interface. Once voice works in dialect, designers have a new set of decisions to make. This is a follow-up to my post on designing Arabic-first AI assistants, focused on voice.
Key takeaways
- Open, dialect-aware Arabic speech recognition makes voice input practical for regional products, but the interface still decides whether people trust it.
- Transcripts will mix Arabic and Latin script. Design the live transcript for bidi text from the start.
- Decide whether to show what the user said (dialect) or a cleaned-up version (MSA), and make that choice visible.
- For payments, bookings and other high-stakes actions, always confirm in a structured form, not just in text.
- Tell users where their voice goes, and know when voice is the wrong input altogether.
What changed, in a few sentences
Until now, the strongest speech models were mostly built for English first, and Arabic support often meant MSA plus a handful of dialects of varying quality. Jais Speech is Arabic-native and open, so regional teams can run it themselves rather than depending entirely on a big platform's speech API. That matters for cost, for data residency, and for being able to tune the experience to local users.
For designers, the important part isn't the model architecture. It's that people can now speak the way they actually speak, Egyptian, Gulf, Levantine or Maghrebi, with English brand names mixed in, and get a usable result.
Transcript UI: mixed scripts and live text in RTL
A real Arabic voice transcript will look something like a sentence in Gulf dialect with "Apple Pay" or "Uber" in the middle of it. That's good, because it's accurate. It's also exactly the kind of text that breaks RTL layouts.
Things to design for:
- Bidi ordering. Latin words inside Arabic sentences must stay in the right place as the transcript grows. Test with brand names, numbers and email addresses.
- Streaming text. Live transcripts update word by word and often revise earlier words. In RTL, new words appear on the left side of the line. Make sure revisions don't make the whole line jump around.
- Line breaks. Long transcripts should wrap cleanly, with the cursor or "listening" indicator at the logical end of the text, not stuck on the wrong side.
- Typography. If the model adds vowel marks, your line height needs room for them. My post on Arabic typography for UI covers which fonts handle diacritics well.
Dialect vs. MSA output
This is the most interesting new decision. If someone says a sentence in Egyptian Arabic, should the text field show exactly what they said, or a version converted to MSA?
There's no single right answer. It depends on context:
- Chat and messaging: show the dialect. People write to each other in dialect already, and converting to MSA would feel strange and formal.
- Search: show what they said, but search on both the dialect and a normalised form, so "عايز" and "أريد" find the same results.
- Forms and official records: a cleaned-up version may be better, but show it as a suggestion the user can accept, not a silent change.
- Notes and dictation: let the user choose in settings, and remember the choice.
Whatever you choose, make it visible. A transcript that quietly rewrites someone's words feels like being corrected, and people notice.
Diacritics are a similar choice. They help in learning apps, Quran and religious content, and for names that are easy to misread. In everyday chat they mostly add clutter. Show them where they help, hide them where they don't.
Correction and confirmation for high-stakes actions
Voice errors are cheap in a chat and expensive in a payment. "Send 500 to Ahmed" in colloquial Arabic can go wrong in several ways: the amount, the currency, or which Ahmed.
For anything that moves money, books something or can't be undone:
- Turn the voice request into a structured summary. Show the recipient, amount, currency and date as separate fields, not just the transcript.
- Use clear digits. Show amounts in the digit style the user has chosen in the app, never in a style the model happened to output.
- Make editing easy. Let the user tap any field to fix it, rather than asking them to say the whole thing again.
- Require an explicit confirm step. A button, a biometric check, or both. Never complete a payment from voice alone.
For lower-stakes text, inline correction is enough: tap a word, see alternatives, pick one. If the model detects the dialect, you can even use that to suggest better alternatives.
[Waleed: add a sketch of a voice-to-payment confirmation screen in Arabic, showing structured fields.]
Privacy and trust
Voice feels more personal than typing. People will want to know where their voice goes. Be clear about:
- Whether audio is processed on the device, on your own servers, or by a third party
- Whether recordings are stored, for how long, and whether they're used to improve the model
- How to delete their voice history
Open models make it easier to answer these questions well, because you control where the model runs. Use that as a design advantage and say it plainly in the interface, in Arabic, not only in a privacy policy.
When voice is the wrong answer
Voice input is great when hands are busy, when typing Arabic on a small keyboard is slow, or for users who are more comfortable speaking than writing. It's a poor fit when:
- The user is in a public place and the content is private
- The input is a password, a code or something that must be exact
- The screen is mostly choices, where a tap is faster than a sentence
Offer voice as an option, make it easy to find, and never make it the only way in.
Closing thoughts
Dialect-aware speech recognition removes one of the biggest barriers to Arabic voice input. What's left is design work: transcripts that handle mixed scripts, honest choices about dialect and MSA, careful confirmation for high-stakes actions, and clear answers about privacy. The model lets people speak their own Arabic. The interface decides whether they feel understood.
Sources: UAE launches Jais 3 (Gulf News), MBZUAI, Inception42 launch Jais 3 Arabic AI models (WAM), TII launches Falcon-ASR for Arabic and Emirati speech (FourWeekMBA).
FAQs
Should an Arabic voice transcript show dialect or MSA?
It depends on context. Show dialect in chat and messaging, offer a cleaned-up version as a suggestion in forms, and let users choose for dictation.
How do you handle English words in an Arabic voice transcript?
Keep them in Latin script, as the newer models do, and test bidi ordering carefully so brand names and numbers stay in the right place as the transcript updates.
Can users make payments by voice safely?
Only with a structured confirmation step. Turn the request into clear fields, let users edit them, and require an explicit confirm before anything happens.