Pick any text-to-speech vendor, open the language menu, and you will find "Arabic". What you will rarely find is which Arabic. In practice the word means Modern Standard Arabic read in whatever accent the training data happened to carry, most often Egyptian or Levantine, and the country code next to it describes a billing market, not a dialect. We spent 1 November 2026 reading the documentation and pricing pages of fourteen text-to-speech products, from the big clouds to a handful of Saudi start-ups, and looking for any dated, independent listening test of Arabic voices. We listened to exactly one product ourselves, HUMAIN Voice, in a short session the same evening (see "What we tested ourselves" below), and the dalili.tools publishing pipeline does not use text-to-speech. Everything else is a map of who promises what, in writing.
Who actually writes "Saudi" on a voice
Among the hyperscalers, Microsoft is the only one with a voice pair labelled "Arabic (Saudi Arabia)": Zariyah and Hamed, two of thirty-two voices spread across sixteen Arabic locales from the UAE, Bahrain, Kuwait, Qatar and Oman to Egypt, Morocco and Yemen. The catch is in the same table: every Arabic voice sits in the standard neural tier. No high-definition voice, no multilingual voice and no speaking style exists for any Arabic locale on the language-support page updated in October 2026. Amazon Polly is the one vendor that literally prints "Arabic (Gulf)", on Hala, launched in November 2022, and Zayd, from August 2023, both filed under the UAE code and described by Amazon as bilingual between Gulf Arabic and English; Zeina is an older standard-engine voice for formal Arabic, and Polly's generative and long-form engines still have no Arabic voice as of the October 2026 documentation history.
Google offers no Gulf label at all. Its sole Arabic locale for standard, WaveNet and the new Chirp high-definition voices is "ar-XA", which Google itself glosses as Modern Standard Arabic and lists as "Arabic (Generic)"; the only regional Arabic anywhere in its catalogue is Egyptian, generally available in Gemini-TTS, with a "World" Arabic in preview. ElevenLabs writes "Arabic (Saudi Arabia, UAE)" in the language list of its older multilingual model and plain "Arabic" for its newer ones; its Arabic landing page poses the question "Can the Arabic TTS handle different regional accents?" and shows no answer. What ElevenLabs does have is a Voice Library in which contributors tag their own voices, and third-party mirrors of that library show between seven and fifteen voices tagged "Saudi" and two tagged "Gulf", alongside many more tagged Egyptian and Modern Standard. That is a community label, not a company commitment. OpenAI has no Arabic voice as such: its voices are "optimized for English", Arabic is supported only insofar as the Whisper language list is, and on 1 October 2026 the company announced that its current text-to-speech models will be shut down on 6 January 2027 in favour of the Realtime API.
The Saudi voices come from Saudi Arabia
The picture inverts once you look at regional products. HUMAIN Voice, from the PIF-owned HUMAIN, is an Arabic-first platform with speech recognition, text-to-speech and speech-to-speech APIs and, according to its own FAQ, data residency inside the Kingdom. Its speech engine is called Nebula. The developer documentation says the voice list contains seven multilingual profiles, each carrying a gender and a dialect field, that any Arabic-script letter in the text selects Arabic while numbers and punctuation do not, and that cloning works by sending a reference recording with its transcript instead of a voice ID. The developer documentation does not say which dialect values exist or whether any is Najdi or Hijazi; the free web demo, however, labels its voices plainly, four of seven as "Arabic (Saudi Arabia)" and a further group as "Arabic (Egypt)", as we saw ourselves on the evening of 1 November 2026. Details in the test section below.
Nabrah, in Riyadh, lists ten public voices "for Saudi and Gulf Arabic" with names such as Daniah, Abdullah, Norah and Khalid, and prices only in riyals. Hams.AI, also Riyadh, demos voices it calls "Saudi dialect" and states that it is hosted in Saudi data centres, without public prices. Sarj names five "Saudi voices" and tags their IDs with "ars", the ISO language code for Najdi Arabic, the most precise dialect label we saw anywhere. SILMA offers Modern Standard and a "Saudi dialect" option in its app. Hamsa sorts its voices by dialect code, Saudi, Gulf, Emirati, Qatari, Kuwaiti, Omani and Bahraini next to Levantine, Egyptian, Iraqi and MSA, and announced a sovereign deployment inside Saudi Arabia through OmniOps in July 2026. Munsit, from the UAE, labels voices "Saudi (Riyadh and Jeddah)" and hosts in the Emirates. And from the other direction, Groq now serves a model called, literally, "Orpheus Arabic Saudi", with six Saudi voices, in preview and capped at two hundred characters per request.
Diacritics, numerals and mixed sentences
These are the three places where Arabic voices fail, and few vendors write about them. Microsoft documents number, date and currency reading through the "say-as" tag for every Arabic locale, plus IPA pronunciation hints. Murf's FAQ concedes that an unvocalised word can be read more than one way and offers per-word pronunciation overrides. SILMA and Telnyx document optional diacritics in the input with automatic vocalisation when they are absent. The frankest statement we found is on the model card of an open Saudi-dialect model on Hugging Face, which says plainly that missing diacritics affect pronunciation accuracy and that number normalisation needs work. On code-switching, Polly describes Hala and Zayd as fully bilingual, HUMAIN picks one language for the whole request from the presence of a single Arabic letter, and nobody else documents what happens to an English brand name in the middle of an Arabic sentence. A failure list published by Munsit in July 2026, a vendor rather than a neutral party, names the usual suspects: ambiguous unvocalised words, numbers in the wrong gender, an audible engine switch on English words, and mangled proper nouns.
Voice cloning and the consent rules
ElevenLabs allows instant cloning from under two minutes of audio on most plans and professional cloning from the Creator plan up, with at least thirty minutes of recordings, a voice captcha in which you read lines aloud, and a rule stated without hedging: you may clone only your own voice, and another person's consent does not change that. Microsoft requires a recorded consent statement in the language of the recording for its professional voice and keeps personal voice behind a limited-access form. Google's instant custom voice needs a consent recording read from a fixed per-language script, with its own Arabic text, and access through a sales allow-list. HUMAIN documents the mechanics of cloning but we found no page on its consent rules. Nabrah clones from four to seven seconds with a rights confirmation. OpenAI offers no cloning.
Where your text is processed
For anyone who needs processing inside the Gulf: Azure's UAE North region offers neural text-to-speech, batch synthesis and hosting of custom voices, but not high-definition voices, personal voice or custom-voice training; Qatar Central offers neural synthesis only; there is no Saudi region. Polly is available in Bahrain, but neural voices are not, so Hala and Zayd cannot be served from there, only the standard Zeina, and the UAE is not a Polly region. Google has no text-to-speech endpoint in the Middle East, neither Doha nor Dammam. ElevenLabs offers data residency in the EU, India and Singapore for enterprise customers only, nothing in the region. If the requirement is the Kingdom specifically, the options are HUMAIN and Hams; if the UAE is acceptable, Munsit and Azure.
Prices in November 2026
Per million characters: Google charges $4 for standard and WaveNet voices and $30 for Chirp high-definition voices, with a free monthly allowance. Polly charges $4 for standard and $16 for neural, which is what Hala and Zayd cost. Azure charges $15 for standard neural, the only tier Arabic has, a figure we took from Microsoft's official retail-price API because the pricing page itself rendered no numbers for us. ElevenLabs' API works out to roughly $80 for its full models and $40 for its fast ones, scaled up from the per-thousand-character rate; its subscription plans as displayed today run from a free tier with ten thousand credits through $6, $22 and $99 to $299 and $990 a month. Orpheus Arabic Saudi on Groq costs $40. HUMAIN prices in riyals: five thousand free credits with no card, enough for about fifteen thousand characters, then plans from SAR 49 to SAR 649 a month, and without a plan about a quarter of a riyal per thousand characters, roughly SAR 224 per million. Nabrah runs from SAR 45 to SAR 850 a month. No vendor in this list documents accepting mada, and the only two that price in riyals are the two Saudi ones; our payment-methods guide covers the rest.
What independent tests say
Very little. The only 2026 academic paper we found that scores a commercial system on Saudi speech is Habibi, from Shanghai Jiao Tong University, revised in March 2026: on its Saudi subset the ElevenLabs model scored 3.70 out of 5 for dialect fidelity against 3.58 for the open model, and 3.45 against 3.51 for naturalness, judged by ten to twenty native listeners per dialect. Azure, Google and Polly were not in the test. Navid, a Riyadh company, launched an open Arabic TTS Arena in March 2026 with blind pairwise votes over about fifteen open models; it has published no rankings yet, and its authors call country-level dialect labels "a lossy abstraction". The SawtArabi benchmark that HUMAIN and SDAIA presented at Interspeech 2025 covers Modern Standard, Egyptian, English and Egyptian-English switching, not Saudi. A University of British Columbia survey of Arabic speech data from January 2026 found no Gulf text-to-speech corpus at all. In short, there is no independent number today that says one voice sounds more Saudi than another.
What we tested ourselves: four sentences in HUMAIN Voice
On the evening of 1 November 2026, a few hours after this guide went up, one of us opened HUMAIN Voice's free demo (the Text to Speech tab, no sign-up, Chrome on Ubuntu, in Riyadh). The demo settles a question the developer documentation left open. Its voice picker lists seven voices: four under the heading "Arabic (Saudi Arabia)", named Sara, Nourah, Abdullah and Abdulaziz, followed by an "Arabic (Egypt)" group that starts with Youssef, marked beta; every voice is also tagged English. The demo offers a clone-a-voice button, six delivery presets (call centre, happy, sad, whisper, news, storytelling), a 500-character box and a cap of five requests. We used Nourah, default preset, and downloaded the four clips.
Sentence one was Najdi with numbers: "I want to book an appointment Thursday at 4:30 in the afternoon in Al Malqa, what do you think?" followed by a Saudi mobile number in Western digits. To a Saudi listener the ten-second result sounded Saudi, not formal Arabic with an accent and not generic Gulf; the colloquial "wish rayik" was spoken as written, the time was read as a time and the phone number digit by digit. Sentence two was the diacritics trap, "kataba al-talib al-dars, thumma qara'a kutub al-maktaba kullaha", submitted once fully vocalised and once bare: the engine read the first "ktb" as the verb and the second as the plural noun in both versions, and the two renders were indistinguishable by ear, so it resolves the vowels from context and did not need the marks for this sentence. Sentence three mixed scripts, "I downloaded the HUMAIN Chat app from the App Store yesterday and tried the Voice Agent feature in Arabic": the three English names came out as native English from the same voice, not spelled out and with no audible engine switch, consistent with the documentation's description of the voices as multilingual profiles.
The limits are plain. One listener, one voice out of seven, four short sentences, and no side-by-side with Azure's Zariyah and Hamed, Polly's Hala and Zayd or any "Saudi"-tagged voice in the ElevenLabs library. What the test establishes is narrow and useful: the claim a reader can check in minutes, that HUMAIN labels its voices Saudi and that one of them reads a Najdi sentence with numbers and English words correctly, held up. A ranking of the voices against each other still has no independent number behind it, exactly as it did this morning.
What we could not verify
We listened to one HUMAIN Voice voice and nothing else, so every description of how any other voice sounds is the vendor's. We could not find the dialect values of HUMAIN's API or its cloning consent policy, although the demo interface does label dialects. We could not confirm that ElevenLabs' own library exposes "Saudi" as a filter, nor why its two largest plans display at prices lower than we expected. Murf and Speechify showed no dialect label on the pages we read. PlayHT, whose team moved to Meta in July 2025, still serves a marketing page with no prices and no shutdown notice while third-party reports say the platform closed at the end of 2025; we found no first-party confirmation. We did not verify the names of Apple's Arabic Siri voices, and Apple's own pages list "Arabic (Saudi Arabia)" for dictation and voice control, not for spoken voices. Any "fewer pronunciation errors" figure for Microsoft's Arabic voices reached us only second-hand.
What to choose
If you want a voice the company itself calls Saudi, want to pay in riyals and need your text to stay in the Kingdom, start with HUMAIN Voice on its free credits, where our one Najdi sentence came out Saudi to a Saudi ear in a short test, then Nabrah. If you build inside Azure or AWS, Zariyah and Hamed are the only hyperscaler voices explicitly labelled "Saudi Arabia" and Hala and Zayd the only ones labelled "Gulf", with the understanding that neither comes in a high-definition tier or with processing in Saudi Arabia. If expressiveness, cloning and many languages matter most, ElevenLabs is the richest platform, but its Saudi accent is a community tag rather than a company label, so choose the voice by ear and not by the list. Whatever you pick, write one sentence containing a number, a Hijri date, the name of a neighbourhood in your city and one English word, run it through three voices before paying anything, and trust ten minutes of your own listening over every page of documentation above.