Arabic Text to Speech: Dialects, MSA and Best Practices

Arabic text to speech turns written Arabic into spoken audio, and the first decision is which Arabic to speak. Modern Standard Arabic (MSA) suits news, education and formal announcements. A regional dialect suits conversation and social content. As of October 2026, Azure lists the most Arabic locales, Google offers one MSA locale, and vendors differ, so test first.

What is the best Arabic text to speech voice?

There is no single best voice. The right one depends on the variety your listeners expect. Arabic is diglossic: most speakers learn Modern Standard Arabic, which is used mainly in writing, broadcasting and prepared speech, and they also learn a local colloquial variety first. Speakers often mix the two in one conversation, according to Wikipedia's article on Varieties of Arabic. A voice that reads a news bulletin well may sound stiff in a shop-counter role, and the reverse can also be true.

MSA or a dialect: which Arabic should your voice speak?

Use MSA when the audio is formal, or when a wide audience across many countries must follow it:

News summaries, course narration and public announcements
Product documentation and e-learning modules
Scripts that will be read as much as heard

Use a dialect when the audio should sound like one person talking to another:

Customer-facing phone prompts, where a regional voice exists
Social clips, ads and character voices
Dialogue between two people from the same region

The Varieties of Arabic article groups modern varieties into five geographic groups: Maghrebi, Egypto-Sudanic, Mesopotamian, Levantine and Peninsular, which includes Gulf varieties. It says the groups do not match state borders, so labels such as "Gulf Arabic" or "Egyptian Arabic" describe broad families rather than exact boundaries. It also notes that Egyptian Arabic is one of the most widely understood varieties, largely because of twentieth-century Egyptian media.

A practical test: write two sentences, one in MSA and one in the dialect your audience uses. Generate both with the same voice and listen to them back to back. Keep the version that sounds right to a native listener.

Which vendors cover which Arabic variants?

Coverage varies a lot. The table lists what each vendor's documentation shows as of October 2026.

| Vendor | Arabic coverage | Notes | |---|---|---| | Microsoft Azure AI Speech | 16 locales, 32 voices (two per locale, one female and one male) | Country locales such as ar-EG (Egypt), ar-SA (Saudi Arabia), ar-MA (Morocco) and ar-LB (Lebanon) | | Google Cloud Text-to-Speech | One locale, ar-XA, described as Modern Standard Arabic | 30 Chirp 3: HD voices, plus Standard and WaveNet voices (A to D) | | Amazon Polly | "Arabic" (arb): Zeina, Standard engine only. "Arabic (Gulf)" (ar-AE): Hala and Zayd, both Neural | Three voices in total | | ElevenLabs | Saudi Arabia and UAE variants on Multilingual v2 and Flash v2.5; Arabic also appears in the v3 and v4 language lists | ElevenLabs advises choosing a voice whose accent matches the target region | | Gemini API text-to-speech | "Egyptian Arabic" listed by name; input language detected automatically | 30 prebuilt voices, up to two speakers per request |

Azure offers the widest set of country-level choices. Google's Arabic is a single MSA locale, though it comes with 30 Chirp 3: HD voices. Polly's Gulf pair is a smaller choice. Azure's two Egyptian voices are Salma and Shakir, and its Saudi voices are Zariyah and Hamed.

Do you need tashkeel for Arabic text to speech?

Tashkeel, also called harakat, are the short-vowel marks written above and below letters. Most everyday Arabic is written without them. Full marks appear mainly in the Quran, poetry, children's books, learner materials and dictionaries, according to Wikipedia's article on Arabic diacritics. The one mark common in ordinary spelling is the shaddah, which doubles a consonant. The article's example is دِين (religion) and دَين (debt): the same letters, with different vowel marks and different words.

Requirements differ between systems. The model card for Arabic-TTS-Spark says its text "must include full Arabic diacritization (tashkeel)", and that the model is optimised for MSA rather than dialects. The model card for lahgtna-omnivoice-v2 treats diacritics as a way to improve pronunciation, not as a requirement. Read the documentation for the tool you use before you write the script.

A practical rule: add marks only where a word could be read two ways, then listen to the result. Full tashkeel on every word takes time, and only some systems require it.

How do you write an Arabic script that sounds natural?

Write numbers, dates and units the way you want them spoken. Azure's say-as tags for cardinal numbers, ordinals, dates, telephone numbers and currency are supported for every Arabic locale, according to its SSML documentation.
Spell out abbreviations the first time they appear.
Write dates in an unambiguous form. A date such as 01/02/2023 can be read two ways. ElevenLabs makes the same point in its English-language best-practice guidance, and the risk applies to Arabic too.
Use punctuation for pauses. Full stops, commas and line breaks shape phrasing.
Respell names and loanwords. If a name is misread, change its spelling in the script and listen again.
Test any Arabic and English mix. The sources reviewed for this post do not verify how a specific voice handles English words inside Arabic sentences.

Is dialect coverage improving?

Dialect coverage is an active research area. A January 2026 preprint, Habibi, describes an open-source Arabic TTS model trained on repurposed speech recognition corpora covering more than 12 regional dialects. Its authors report a seven-subset benchmark of over 11,000 utterances and claim zero-shot synthesis without text diacritization. They also describe the model as "highly competitive with ElevenLabs' Eleven v3 (alpha)". These are the authors' own claims in a preprint, not an independent benchmark, so treat them as a direction to watch.

A February 2026 preprint, NileTTS, offers 38 hours of transcribed Egyptian Arabic speech from two speakers. Its authors describe Egyptian Arabic as severely under-resourced for TTS, which helps explain why Egyptian coverage is still uneven across tools.

Arabic text to speech with OndaLabs

OndaLabs has a dedicated Arabic text to speech page. Its 30 prebuilt voices are usable in every supported language, so you can run the same paragraph through several voices and compare. Delivery is directed in plain words and with supported expressive tags in the script. For conversation scripts, the dialogue tool handles two speakers. The verified feature list does not include dialect selection, so the variety you hear comes from your script and the voice you choose.

Plans, as published in October 2026: the Free plan gives 2,000 credits per month, about 2 minutes on Onda Lite at 1 credit per character. Starter at $9 adds Onda Flash, the most expressive model, at 1.5 credits per character, plus commercial use. Creator at $22 includes 3 custom voices, and Pro at $49 includes 10. Custom voices are either cloned from a recording you have permission to use or designed from a text description. You can try the text-to-speech app on the free plan before comparing plans and prices.

For related reading, see the Spanish text to speech guide and the Hindi text to speech guide.

Frequently asked questions

Is there a free Arabic text to speech?

OndaLabs has a Free plan with 2,000 credits per month, about 2 minutes of Onda Lite audio, and five saved clips. Google's Cloud pricing page lists a free tier of 0 to 4 million characters per month for WaveNet and Standard voices. Free tiers change, so check each provider's pricing page before you commit.

Can I use Arabic AI voices commercially?

OndaLabs includes commercial use from the Starter plan at $9 per month and above. For other vendors, check the terms that apply to the voice and plan you use. The sources reviewed for this post do not cover licence terms for Arabic voices.

Do I need diacritics for Arabic text to speech?

It depends on the system. Some model documentation requires full tashkeel, while other models treat diacritics as an aid to pronunciation. Add marks only to words that could be read two ways, then listen to the output. Full diacritics on every word take time, so use them where the output needs them.

Which Arabic should I use for a global audience?

Modern Standard Arabic is the usual choice for content that must be understood across many countries, such as news and formal announcements. Egyptian Arabic is widely understood because of 20th-century Egyptian media, according to Wikipedia, but it is not the same as MSA. Test both on a short sample from your audience first.

Can Arabic text to speech mix Arabic and English?

Some systems detect the input language automatically. The Gemini API TTS page says its models detect the input language automatically, for example. The sources do not confirm how any specific Arabic voice handles English words inside Arabic sentences. Keep English brand names short, test them, and respell any word that comes out wrong.