An Arabic AI voice agent is software that answers or places phone calls in Arabic: it transcribes what the caller says, decides what to do, and speaks a reply. The hard part is that callers speak their own dialect (Gulf, Najdi, Hijazi, Egyptian, Levantine, Maghrebi), mix in English, and do it over 8 kHz phone audio, while most speech models are trained mainly on Modern Standard Arabic. So the only reliable way to choose one is to test it per dialect, on recordings of your own calls, before you put it on your main line.
TL;DR
- Dialect is the main risk. In a public benchmark on the Saudi SADA dataset, every model tested did best on MSA. One model's word error rate went from 19% on MSA to 48% on Khaliji.
- Phone audio makes it harder. Phone lines usually carry 8 kHz audio, and most benchmarks use cleaner recordings. Test on real call audio, not studio clips.
- Code-switching, numbers and names break agents more often than greetings do. Put them in your test set on purpose.
- Measure four things per dialect: word error rate on your own calls, task completion, handoff rate and response latency.
- Pilot one dialect at a time, with a human handoff ready, and only widen the rollout when that dialect passes.
- Ask vendors for per-dialect results on telephony audio. "Supports Arabic" tells you very little.
Why Arabic is harder for voice than for chat
A voice agent has to work out the words before it can act on them, and a misheard order number leaves it nothing to recover from. Arabic adds three problems on top of that.
- Training data skews to MSA. Modern Standard Arabic (fusha) is the language of news broadcasts, and broadcast audio is the easiest speech data to collect. Customers on the phone almost never speak it.
- Dialects differ in sounds, not just words. The letter qaf, for example, is usually a "g" in most of the Arabian Peninsula and a glottal stop in much of the Levant and Egypt (Wikipedia, Qoph). A model has to map all of those back to the same word.
- Callers don't say which dialect they speak. A Riyadh line gets calls from Najdi, Hijazi, Egyptian and Levantine speakers, often in the same hour. Speechmatics, a speech recognition vendor, makes the point that forcing a dialect choice at integration time "adds a fragile assumption to every single call" (Speechmatics).
For Arabic support in text, see our guide to Arabic AI chatbots for customer support.
What the research shows about Arabic dialects and speech recognition
The best public evidence comes from the Open Universal Arabic ASR Leaderboard, a benchmark from researchers at ELM in Saudi Arabia (Wang et al., arXiv 2412.13788). It tests open-source speech recognition models on several Arabic datasets, including SADA, a Saudi dataset covering MSA, Egyptian and dialects spoken in Saudi Arabia. The authors write that "all models achieve their best results on MSA, but exhibit a significant decline" on dialects, and they attribute it to MSA making up most of the public training data.
Word error rate (WER) by dialect on the SADA test set, from Table 3 of the paper:
| Model | MSA | Egyptian | Hijazi | Khaliji | Najdi |
|---|---|---|---|---|---|
| NVIDIA Conformer-CTC-large-ar (with language model) | 19.2% | 41.0% | 37.0% | 48.2% | 36.3% |
| OpenAI Whisper large-v3 | 28.0% | 59.3% | 50.0% | 59.9% | 48.6% |
These are open-source models, not the commercial systems most vendors run, and SADA is recorded audio rather than phone calls. What carries over is the pattern: the same model can make more than twice as many errors on Khaliji as on MSA.
A second paper, Arab Voices, maps 31 dialectal Arabic speech datasets covering 14 dialects and finds they vary widely in domain, dialect labelling and recording conditions (Sullivan et al., arXiv 2601.13319). So an "Arabic accuracy" figure means little until you know the dataset behind it.
Commercial vendors publish their own numbers too. Speechmatics reports a 4.5% WER on colloquial Arabic and 6.3% on mixed Arabic-English speech in its own tests, against 5.9% and 9.7% for Google. These are vendor-run results with no published method, so use them as a reason to run your own test, not as a ranking.
The dialects your test set needs to cover
Start from your call logs. Tag a month of calls by dialect (a native speaker can do this quickly) and weight your test set the same way.
| Dialect group | Where callers come from | What to listen for in testing |
|---|---|---|
| MSA | Formal callers, read-out scripts | Your baseline. If the agent fails here, stop. |
| Gulf / Khaleeji | UAE, Kuwait, Qatar, Bahrain, Oman, eastern Saudi | "g" for qaf, Gulf words for common requests, English mixed in |
| Saudi Najdi | Riyadh and central Saudi Arabia | Najdi vocabulary and pronunciation, fast speech |
| Saudi Hijazi | Jeddah, Makkah, Madinah | Hijazi vocabulary, which overlaps with Egyptian in places |
| Egyptian | Egypt, and expat callers across the Gulf | Glottal stop for qaf in Cairo speech, Egyptian vocabulary |
| Levantine | Jordan, Lebanon, Syria, Palestine | Glottal stop for qaf in city speech, Levantine vocabulary |
| Maghrebi | Morocco, Algeria, Tunisia, Libya | Switching into French as well as English, vocabulary far from MSA |
Speech platforms split Arabic by country rather than by dialect. Google Cloud Speech-to-Text lists 18 Arabic country locales, from ar-SA to ar-MA (Google Cloud), and Microsoft Azure lists 18 as well (Microsoft Learn). A locale code tells you what a model was set up for. It doesn't tell you how well it handles a Najdi caller on a Jeddah number.
Code-switching, numbers, dates and names
These carry the data the agent needs to act, so errors here cost the most.
- Code-switching. Callers drop English into Arabic sentences all the time: "ابي اسوي cancel للـ order". The agent has to transcribe both languages in one stream and treat the English words as part of the request.
- Numbers. Order numbers, phone numbers and amounts may be read in Arabic, English or both, digit by digit or in groups. Test each pattern, and check that the agent reads numbers back to confirm them.
- Dates. Customers may give Hijri or Gregorian dates, and "after Eid" is a real answer. Decide what the agent should do with each, then test it.
- Names and places. Arabic names have several English spellings, and your CRM probably stores one of them. Test whether the agent can match a spoken name to the record, and whether it asks for spelling when it can't.
- Product names. Your product names are often English words said with an Arabic accent. Ask how to add custom vocabulary.
Latency and turn-taking on Arabic calls
A voice agent that answers correctly but pauses too long feels broken. Every turn adds up speech recognition, the model's reasoning and speech synthesis. The agent also has to judge when the caller has finished, which is harder when callers pause to find a number or switch language. Test:
- Time from the end of the caller's speech to the first audio of the reply, measured from call recordings. Report the median and the slowest 10% of turns, not the average.
- Whether the agent stops talking when the caller interrupts (barge-in).
- Whether it cuts in while the caller is still reading out a long number.
Choosing the TTS voice
The voice your customers hear shapes trust as much as accuracy does. A formal MSA voice can sound like a government recording to a Gulf caller. A strong dialect voice can feel odd to callers from another country.
A common choice is a lightly regional voice that matches your main market, with plain wording. Major platforms offer country-specific Arabic voices. Azure, for example, lists separate Saudi, Emirati and Egyptian neural voices (Microsoft Learn). Play the candidate voices to a few customers or front-line agents from each market and ask which one they'd trust with a refund.
What 8 kHz phone audio does to Arabic recognition
Many calls still reach a voice agent as narrowband audio. Twilio's media streams, for example, are always 8,000 Hz, mono, mu-law encoded (Twilio docs). At that sample rate, sound above 4 kHz is lost. That removes high-frequency detail that helps tell similar sounds apart, on top of mobile compression, background noise and speakerphone use.
The SADA results above come from recorded audio, not calls, so expect your own calls to score worse. Also check which model the vendor runs on phone calls: some platforms ship separate telephony-tuned models. When we checked Google Cloud's language table, en-US had dedicated telephony models and the Arabic locales did not (Google Cloud). That doesn't mean Arabic fails on calls, only that you need to test it there.
How to test an Arabic voice agent: a step-by-step plan
This plan takes three to four weeks and works for any vendor, Open included.
- Build a test set per dialect. Take 50 to 100 recorded calls per dialect you serve, from your own phone system, with consent and personal data handled under your privacy policy. Include easy calls, angry callers, noisy lines and speakerphone. Have a native speaker write the correct transcript for each.
- Add the hard cases on purpose. For each dialect, script 20 to 30 short turns with order numbers, phone numbers, dates, names, English product names and mid-sentence language switches.
- Measure WER on your own calls. Run the recordings through each vendor's recognition and score word error rate against your transcripts, per dialect. Normalise spelling variants (alef forms, ta marbuta) before scoring, or you'll count spelling as errors.
- Run live test calls per dialect. Have native speakers from each dialect call the agent with your top 10 call reasons. Record whether the task was completed, whether the agent handed off, and whether the handoff was correct.
- Red-team it. Call with heavy dialect, fast speech, long pauses, background noise, interruptions, and requests outside the agent's scope. Try to get it to confirm a wrong number or invent a policy.
- Pilot one dialect at a time. Route a share of real calls from your biggest dialect group to the agent, with warm handoff to your team. Widen it only when that dialect meets your targets, then add the next.
The scorecard
| Metric | How to measure it | What good looks like |
|---|---|---|
| WER per dialect | Your recordings, scored against human transcripts | Set a target per dialect from your MSA baseline, and watch the gap |
| Task completion | Share of live test calls where the caller's goal was met | Close to your human team's first-contact resolution for the same call types |
| Containment and handoff | Share of calls resolved by the agent vs handed off, and whether each handoff was the right call | Handoffs happen early on calls the agent can't handle, with full context |
| Number and name accuracy | Share of spoken numbers and names captured exactly | Every critical value is read back and confirmed |
| Response latency | Median and slowest 10% of turns, from recordings | Consistent across dialects, not just in English |
| Caller rating | Short post-call survey per dialect | No dialect group rates it much lower than the others |
Keep the dialect breakdown in every report: an overall score can hide one group of callers getting a much worse service. For how to hand calls over cleanly, see our guide to AI-to-human support handoff.
What to ask Arabic voice AI vendors
- Which Arabic dialects do you support for recognition, and which for speech output? What evidence do you have for each?
- What WER do you see per dialect on 8 kHz telephony audio, and on which dataset? Will you run our recordings through the same test?
- Do you need to know the caller's dialect in advance, or does one model handle all of them?
- How do you handle Arabic-English code-switching in one sentence?
- What's the median and slowest-10% response latency on Arabic calls, measured end to end?
- Which Arabic voices can we choose from, and can we hear each one reading our numbers and product names?
- Does it connect to our current phone system over SIP, or do we have to move numbers?
- Where are call recordings and transcripts stored, and for how long? For Saudi callers, our guide to Saudi PDPL and AI customer service lists the data questions to add.
- How is it priced: per minute, per call or per resolution, and are telephony costs separate?
If you're still deciding between an AI voice agent and a speech-enabled menu, start with conversational IVR vs AI voice agent. For the general build, see how to set up an AI phone agent.
Where Open fits
Open's AI voice agents answer and place calls, run actions in your tools during the call, and hand off to a human with the transcript and context attached. They run on top of the carrier and phone system you already have: Open connects over SIP, so there's no carrier switch or number port (Open Telephony). The same agent, Agent 5, also answers WhatsApp, chat and email.
Open's voice page lists 100+ languages. It doesn't publish dialect-by-dialect accuracy for Arabic, so put it through the same test set as every other vendor on your list.
Open charges $0.75–$0.90 per resolution, depending on volume. Handoffs to a human aren't billed, and carrier and telephony costs are passed through at cost. The details are on the pricing page. Open's security documentation is at trust.open.cx.
Sources
- Wang, Alhmoud and Alqurishi, "Open Universal Arabic ASR Leaderboard," arXiv 2412.13788, https://arxiv.org/abs/2412.13788 (per-dialect results from https://arxiv.org/html/2412.13788v1, Table 3), checked 21 Nov 2026
- Sullivan, Elmadany, Alcoba Inciarte and Abdul-Mageed, "Arab Voices: Mapping Standard and Dialectal Arabic Speech Technology," arXiv 2601.13319, https://arxiv.org/abs/2601.13319, checked 21 Nov 2026
- Speechmatics, "Your voice agent speaks perfect Arabic. That's the problem," https://www.speechmatics.com/company/articles-and-news/your-voice-agent-speaks-perfect-arabic-thats-the-problem, checked 21 Nov 2026
- Google Cloud, "Speech-to-Text supported languages," https://docs.cloud.google.com/speech-to-text/docs/speech-to-text-supported-languages, checked 21 Nov 2026
- Microsoft Learn, "Language and voice support for the Speech service," https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support, checked 21 Nov 2026
- Wikipedia, "Qoph" (qaf pronunciation by region), https://en.wikipedia.org/wiki/Qoph, checked 21 Nov 2026
- Twilio, "Media Streams WebSocket messages," https://www.twilio.com/docs/voice/media-streams/websocket-messages, checked 21 Nov 2026
- Open, AI voice agents, https://www.open.cx/voice, and Open Telephony, https://www.open.cx/telephony, checked 21 Nov 2026