Hisham AlhussainI've been playing with voice agents a lot lately, and I think we've crossed a line. A year or two ago, voice AI meant a stitched-together pipeline: speech-to-text, then an LLM, then text-to-speech, with awkward pauses between each step. Talking to it felt like a bad international call.
Realtime speech models changed that. Low latency, natural interruptions, and much more natural-sounding voices. For the first time, I can picture a voice agent handling real customer calls without people hanging up in frustration.
But I work with Arabic-speaking users a lot, so naturally my next question was: does it actually hold up in Arabic?
Where it's good
The progress is real. For clear, fairly formal Arabic, recognition and responses are solid. Voices sound far more natural than they used to. And for simple, structured calls, like confirming an appointment, checking an order status, or answering common questions, I'd be comfortable putting it in front of customers with the right guardrails.
Where it gets hard
- Dialect. This is the big one. People speak in their dialect on the phone, and Gulf dialect in particular is still weaker than MSA in a lot of models. Recognition errors in voice are worse than in text because you can't see what was misheard.
- Code-switching. People switch between Arabic and English mid-sentence all the time, especially for names, numbers, and technical terms. Some models handle this well, others get confused.
- Names, numbers, and addresses. These matter most and are the easiest to get wrong. A misheard digit in a phone number or a booking time is a real problem.
- Latency on top of everything. Every extra check you add for safety adds delay, and in voice, even a second of silence feels long.
How I'd approach it
If I were building an Arabic voice agent today, here's what I'd do:
- Start narrow. One clear use case, like appointment confirmations, not "handle any call."
- Read back the important stuff. Always confirm names, numbers, dates, and times out loud before acting on them. It's slightly slower, but it saves a lot of pain.
- Make handoff easy. If the agent is unsure, or the caller is frustrated, route to a human quickly. A smooth handoff is worth more than a clever agent.
- Test with real callers and real dialects. Recordings from actual users, with background noise, different accents, and people talking fast. Clean studio audio will lie to you.
- Measure what matters. Task completion, handoff rate, and how often a confirmation had to be corrected.
Worth watching closely
I'm genuinely excited about this. A lot of businesses in the region still run on phone calls and voice notes, and a good Arabic voice agent could make a real difference for them. We're not all the way there on dialect yet, but it's moving fast.
If you're building in this space, I'd love to compare notes.