xAIa's voice agents do not transcribe speech themselves. For the speech-to-text Arabic layer that turns a caller's words into text an AI agent can act on, xAIa, the Dubai voice-AI company, says it runs on a third-party vendor called Soniox — and on Soniox's own published pages, the prices and speeds xAIa cites check out.
xAIa, which describes itself on its own site as the AI consultancy and implementation arm of PRO TECHnology Group, set out the relationship in a blog post published 24 September 2026. The company says it has processed more than 1,000,000 production calls through Soniox's speech-to-text engine over the period described, and now runs more than 600,000 calls a month across 63 AI agents on the same vendor.
What xAIa says Soniox gives its speech-to-text Arabic layer
On that same blog post, xAIa describes Soniox's transcription as running at sub-200ms streaming speed, inside an overall call-response latency the company puts at roughly 500ms end to end across its production deployments, and prices Soniox's real-time transcription at $0.12 an hour. xAIa also says the vendor processes audio “in memory and never stored,” and that it moved its own deployment from an earlier Soniox model, stt-rt-v2, to a newer one, stt-rt-v5, over the course of the year described.
Checking those numbers against Soniox's own site
Soniox's own pricing page confirms the figure xAIa cites: real-time, streaming speech-to-text is billed at $0.12 an hour, against $0.10 an hour for transcribing an already-recorded file, with speaker diarisation, language identification and formatting bundled into that rate rather than billed as extras, according to Soniox's published pricing. Soniox's own homepage separately states that its transcription runs with “sub-200ms latency”, and it lists SOC 2 Type 2, ISO/IEC 27001:2022, HIPAA and GDPR as the standards it says it adheres to. On pricing, latency and certifications alike, Soniox's own published claims match what xAIa's blog post describes — though those certifications belong to Soniox, the vendor, and are not a certification xAIa holds in its own name.
On language coverage, Soniox's own documentation lists Arabic as one of more than 60 supported languages, identified by the code “ar,” alongside a published total of 60+ languages through one unified model. That documentation does not break Arabic down by dialect — it is a single entry, not a list of Gulf, Egyptian or Levantine variants — which matters for a Dubai company that elsewhere markets dialect-specific Arabic support: that broader dialect claim is xAIa's own, not something Soniox's own language page documents.
Why real-time Arabic transcription is a harder problem than the language count suggests
Arabic speech-to-text is not one problem but several, because spoken Arabic varies sharply by region while most written Arabic stays close to a single standard form. A 2024 research paper introducing a multidialectal Arabic speech dataset called Casablanca put the underlying issue plainly: “the majority of world languages and dialects remain uncovered” by existing speech technology, a gap the paper's authors wrote “only furthers an already wide technological divide, thereby hindering technological and socioeconomic inclusion.” The dataset the paper built covers eight Arabic varieties — Algerian, Egyptian, Emirati, Jordanian, Mauritanian, Moroccan, Palestinian and Yemeni — precisely because a single Arabic model, trained mostly on the written standard, tends to perform unevenly across them.
That research describes the field in general, not Soniox or xAIa specifically — neither company is named in it — but it explains why a broad language count on a single unified model leaves open a real question for an Arabic-speaking market: which Arabic, spoken by whom, with what accuracy. Soniox's own pages answer the languages-supported question; they do not answer the dialect-accuracy one, and neither does xAIa's blog.
The honest limits
Neither xAIa nor Soniox publishes a dialect-by-dialect accuracy figure for Arabic, so there is no public number to check a "60+ languages" claim against for Gulf, Egyptian or any other Arabic variant specifically. xAIa's own claim that its platform handles specific Arabic dialects describes what the company says it has built on top of Soniox — it is xAIa's description, not a claim made on Soniox's own site.
Frequently asked questions
What is speech-to-text, and why does an AI voice agent need it?
Speech-to-text, sometimes written speech to text, is the software layer that converts spoken audio into written text in real time, which an AI voice agent then reads to decide what to say or do next. xAIa says it buys this layer from Soniox rather than building its own.
Does Soniox support Arabic?
Yes — Soniox's own documentation lists Arabic among the more than 60 languages its speech-to-text API supports, identified by the code "ar," though the page does not publish a separate breakdown by Arabic dialect.
How fast is real-time Arabic transcription supposed to be?
Soniox's own site states its streaming transcription runs with sub-200ms latency. xAIa separately says its overall call-response latency, including the Soniox layer, runs at roughly 500 milliseconds end to end — both figures are the companies' own published claims, not independently measured here.
Who is behind xAIa?
xAIa says on its own site that it is the AI consultancy and implementation arm of PRO TECHnology Group, which it describes as serving MENA enterprises and governments since 1998 across five companies, with xAIa its newest addition, founded in 2025.
Product details and capabilities in this article are as published by xAIa and Soniox at the time of writing.
