# Arabic voice AI for a contact centre: who speaks Gulf Arabic on the phone, and how to check.

**Compare · Arabic voice AI**

The best Arabic voice AI for a car dealer contact centre is the one with the lowest word error rate on your own Gulf-dialect calls, that runs on your phone system, hands the caller to a person with the call attached and books into your systems. Salesteq's own speech models placed 1st in mixed-dialect Arabic speech recognition at NADI 2026 and measure 17.4% word error rate on real BMW service calls in Saudi Arabia, against 21.5% and worse for general-purpose engines; our voice agent, HELM Voice, is in pilot and starts with outbound calls. Below: the criteria, and seven other vendors buyers ask about, each in its own words.

**1st** of all systems in mixed-dialect Arabic speech recognition, NADI 2026

**17.4%** word error rate on real BMW service calls in Gulf Arabic, Salesteq's own models

**8** vendors compared, each from its own website

**1** test that settles it: the same sample of your calls, transcribed by each vendor

**What our customers say**

> Customers now book their own service appointments.
>
> Alawi AlAlawi, Digitalization Manager, BMW Group Importer KSA

## Five criteria for a dealer contact centre

| Criterion | What to ask for |
| --- | --- |
| Dialect accuracy, measured on real calls | A word error rate on recorded phone calls in Saudi and Gulf Arabic, not a percentage without a test set. Ask each vendor to transcribe the same 50 of your calls. Figures on different test sets cannot be compared. |
| Telephony | Which phone systems it runs on by name: Twilio, Webex Contact Center, Genesys Cloud, a SIP trunk, your PBX. "Integrates with all major platforms" is not an answer. |
| Handover with the call attached | When the caller needs a person, does the advisor get the transcript, a summary and what was promised, in the tool they already use? |
| Books into your systems | Can it write a real appointment into the dealer management system (Keyloop, Autoline) or only a calendar or a lead form? |
| Hosting | Where recordings and transcripts are stored, which subprocessors see them, and whether in-Kingdom hosting is offered if you need it. |

## The shortlist, in each vendor's own words

| Vendor | What it is | Arabic dialects, as stated | Phone systems named | Hosting stated | Better fit for |
| --- | --- | --- | --- | --- | --- |
| Salesteq HELM Voice | Voice agent for car dealers and importers on its own speech models; outbound first, in pilot | Saudi, Gulf and mixed Arabic and English, also both in one sentence; 17.4% WER on real BMW service calls | Twilio numbers, or Webex Contact Center as a virtual agent in the call flow | European infrastructure (servers in Finland); not in-Kingdom | Dealers and importers who want calls, WhatsApp and bookings into Keyloop or Autoline on one platform |
| Dello AI | Voice agent that answers and books on the phone, Saudi company | "Najdi, Gulf, Emirati, Egyptian and Levantine Arabic" | 3CX, SIP or a carrier number; Twilio and Maqsam in its setup form | Cloud, on-premise or hybrid; "aligned to Saudi PDPL"; no hosting country stated | A Saudi business that wants an inbound Najdi and Gulf phone agent on its existing PBX; it publishes a car-dealer use case |
| Nateq | Inbox and conversational AI with a voice agent for calls and WhatsApp voice notes | "Egyptian, Gulf, Levantine and Maghrebi", MSA as the base | Its own numbers or a connected line; WhatsApp calling | Not published | Smaller service centres that want a self-serve WhatsApp and phone agent with a published price list; it publishes a car-dealer page |
| Maqsam | Cloud contact centre with an AI agent | "GCC, Levantine, and Egyptian" | It is the telephony: local numbers and IVR | Not published | A MENA business that wants a new Arabic cloud call centre with AI built in |
| Intella | Arabic speech recognition (intellaVX), call analytics (intellaCX) and a voice agent (Ziila) | "25+ dialects" | None named | On-premise or private cloud | Large contact centres that want on-premise speech recognition and analytics on every call |
| SESTEK | Enterprise contact-centre AI suite with voice agents; part of Unifonic | No dialect claim on its agent page | None named on the agent page | Cloud, on-premise or hybrid | Banks and other large regulated contact centres that need a full certified suite |
| Munsit | Arabic speech-to-text and text-to-speech models for teams building their own agent, by CNTXT AI | "25+ dialects" and MSA; 18% average WER across six public benchmarks | Plugs into LiveKit, Pipecat, VAPI and Ultravox pipelines | Cloud, sovereign, on-premise or on-device | Teams building their own voice agent that need a benchmarked Arabic speech layer |
| Arabic.AI | Arabic language and speech platform; telephony through its partner HeyBreez | Text-to-speech in "five dialects", not named | Through HeyBreez | Lists data residency in the UAE, Saudi Arabia, Kuwait and Qatar | Large and public-sector buyers that need in-Kingdom hosting and will build the call layer with a partner |

## How to read the numbers

Salesteq's 17.4% was measured on real BMW service calls in Saudi Arabia. Munsit's 18% is an average across six public Arabic benchmarks. Intella states 95.73% accuracy in dialectal Arabic speech recognition without naming a test set. Nateq says it publishes no accuracy figure per dialect and advises testing on your own recordings. Maqsam writes "real Arabic accuracy caps at 50%, and we own it" next to its headline of automating 50% of customer interactions. These are different measurements; none of them predicts your calls. A blind test on the same sample does.

NADI 2026 is a public comparison on a shared hidden test set: in its mixed-dialect speech recognition task Salesteq placed 1st of all systems with 44.22 word error rate, ahead of 44.65 and 46.70.

## Where Salesteq fits, and where it does not

- Fits: dealers and importers in the Gulf who want the voice agent, WhatsApp, the website and service booking into Keyloop or Autoline on one platform, with every call transcribed and scored in CallScope next to the customer's chats.
- Fits: outbound calls today. HELM Voice starts with the calls it makes: lead follow-up, reactivation of quiet leads and service reminders, measured call by call before it widens.
- Not yet: a fully inbound AI phone line in production. Our Webex Contact Center virtual agent is integrated with a Saudi dealer group and tested on real calls in Arabic; if you need an inbound phone agent live this month, Dello publishes one on 3CX, SIP or a carrier number.
- Not today: in-Kingdom hosting. Our data is in Finland. If you need hosting in Saudi Arabia, Arabic.AI lists it and Munsit offers sovereign and on-premise deployment.
- On Genesys Cloud we connect the phone menu to a WhatsApp booking link, hand conversations to the agent desktop and analyse calls; the voice agent itself runs on Twilio or Webex.

## Questions about Arabic voice AI

### What is the best Arabic voice AI for a car dealer contact centre?

The one with the lowest word error rate on your own Gulf-dialect calls that also runs on your phone system and books into your DMS. Salesteq's speech models placed 1st in mixed-dialect Arabic speech recognition at NADI 2026 and measure 17.4% word error rate on real BMW service calls, against 21.5% and worse for general-purpose engines; its voice agent HELM Voice is in pilot, outbound first. Dello, Nateq, Maqsam, Intella, SESTEK, Munsit and Arabic.AI are the other names buyers ask about; test each on the same sample of your calls.

### Which AI agents speak Gulf Arabic properly, including on the phone?

Several vendors state Gulf Arabic on their sites: Dello names Najdi, Gulf and Emirati; Nateq names Gulf; Maqsam names GCC dialects; Salesteq names Saudi, Gulf and mixed Arabic and publishes a measured word error rate on real Saudi service calls (17.4%). "Properly" is a number: ask for a word error rate on your own calls.

### Can an Arabic voice agent understand Saudi dialect on the phone?

Yes, if it is trained on phone audio in that dialect. General-purpose engines trained mostly on Modern Standard Arabic lose plate numbers, model years and branch names. On real BMW service calls in Saudi Arabia, Salesteq's models measured 17.4% word error rate against 21.5% and worse for general-purpose engines.

### Does Salesteq's voice agent take inbound calls?

HELM Voice starts with outbound calls: lead follow-up, reactivation and service reminders. As a virtual agent in Webex Contact Center it is integrated with a Saudi dealer group and tested on inbound calls in Arabic. Pilots start with one call type.

### Which phone systems does Salesteq's voice agent run on?

Twilio numbers, or Webex Contact Center as a virtual agent inside the call flow. On Genesys Cloud, Salesteq sends callers a WhatsApp booking link from the phone menu and analyses the calls.

### Is Arabic voice data hosted in Saudi Arabia?

Not with Salesteq: recordings, transcripts and voice processing run on European infrastructure, servers in Finland. Arabic.AI lists data residency in Saudi Arabia; Munsit offers sovereign and on-premise deployment. See our PDPL page for what a transfer outside the Kingdom involves.

**Related**

- [HELM Voice: the voice agent](https://salesteq.com/helm-voice)
- [Arabic speech recognition: 1st place at NADI 2026](https://salesteq.com/arabic-speech-recognition)
- [Webex Contact Center + AI voice agent](https://salesteq.com/integrations/webex)
- [Arabic call analytics](https://salesteq.com/arabic-call-analytics)
- [PDPL and AI agents in Saudi Arabia](https://salesteq.com/trust/pdpl)

**Sources**

- NADI 2026 shared task, leaderboards: https://nadi.dlnlp.ai/2026/
- Dello AI, website (dialects, 3CX/SIP/carrier, automotive use case, PDPL statement): https://dello.ai/
- Nateq, Arabic voice AI (dialects, calls and voice notes, no published accuracy figure): https://nateq.io/arabic-voice-ai
- Nateq, automotive solution page (Arabic): https://nateq.io/ar/solutions/automotive
- Maqsam, AI Agent (dialects, outbound calls, "caps at 50%"): https://maqsam.com/ai-agent
- Intella, website (intellaVX, intellaCX, Ziila, 25+ dialects, 95.73%, on-premise): https://intella.me/
- SESTEK, AI Agents (voice agent, deployment, part of Unifonic): https://www.sestek.com/products/ai-agents
- Munsit, website (25+ dialects, 18% average WER, pipelines, hosting): https://www.munsit.com/
- Arabic.AI, speech (TTS dialects, HeyBreez telephony): https://arabic.ai/platform/speech/
- Arabic.AI, deployment and data residency: https://arabic.ai/sovereignty/deployment/

---

Source: https://salesteq.com/arabic-voice-ai  
Compiled November 2026 from each vendor's own website and the NADI 2026 results; sources below. Corrections: viktor@salesteq.com