Arabic Text to Speech: MSA, Dialects, and the Tashkeel Problem

Arabic text to speech converts written Arabic into spoken audio, and the central engineering problem is unique among major languages: everyday written Arabic omits its short vowels. The same three letters كتب can read *kataba* ("he wrote"), *kutub* ("books"), or *kutiba* ("it was written") — the diacritics (tashkeel) that would disambiguate them are simply left out of most real-world text. So a good Arabic text to speech engine has to infer the vowels from context before it can say a single word. This guide covers how engines solve that, the Modern Standard Arabic vs dialect question, which engines are best, and how to handle right-to-left text in your workflow.

Why tashkeel (diacritics) decides pronunciation quality

Arabic script is an abjad: the 28 letters record consonants and long vowels, while short vowels appear only as optional marks above and below the letters — fatha, damma, kasra, sukun, shadda, and the tanwin endings. Fully vocalized text exists mainly in the Qur'an, poetry, and children's books. News articles, websites, business documents, and chat messages are almost entirely unvocalized.

That means the TTS engine must perform automatic diacritization: predicting the vowels (and grammatical case endings) from context, essentially parsing the sentence before speaking it. This is where Arabic engines differ most:

• Modern neural engines diacritize internally. Amazon Polly's Arabic voices were built with an in-built diacritizer that handles unvocalized text, and Microsoft has reported large accuracy gains from fine-tuning diacritic-prediction models — a 2025 Azure update cut word-level pronunciation errors in its Arabic voices by 78% (Microsoft's announcement).
• You can force pronunciations by adding tashkeel yourself. Every major engine respects diacritics present in the input. If a proper noun or ambiguous word keeps coming out wrong, vocalizing just that word in your source text is the reliable fix — no SSML required.
• Case endings are a style choice. Fully pronounced iʿrab (grammatical endings) sounds like formal newscast recitation; most engines target a natural pausal style. Test which your chosen voice produces and whether it fits your content.

Quick quality test: paste an unvocalized sentence containing كتب or بعد in two different senses and listen. Engines that guess vowels from context will read them differently; weak engines pick one reading everywhere.

MSA vs dialects: which Arabic should your voice speak?

Modern Standard Arabic (MSA / فصحى) is the written standard used in news, education, and formal content across all 22 Arab League countries — and essentially nobody's mother tongue. Daily speech happens in dialects (Egyptian, Gulf, Levantine, Maghrebi, Iraqi) that differ from MSA and from each other substantially.

For TTS this cuts a clear line:

Content type	Use
News, e-learning, corporate, documents, accessibility	MSA — universally understood, matches how written Arabic "should" sound
Ads, social video, chatbots, entertainment for one country	Dialect voice — sounds local and warm where MSA sounds stiff

Dialect coverage is the differentiator between platforms. Amazon Polly ships Zeina (MSA) plus Hala and Zayd, Gulf Arabic (ar-AE) voices that also support MSA. Azure covers Arabic locales for 16 countries (ar-SA, ar-EG, ar-AE, and more), giving country-flavored MSA and some dialectal delivery. Specialist Arabic platforms like Hamsa focus on dialects (Egyptian, Gulf, Levantine, Iraqi) with tashkeel support, and ElevenLabs' multilingual models handle MSA with expressive delivery. One caveat: written dialect has no standardized spelling, so dialect TTS quality varies more than MSA.

Best Arabic text to speech engines

Engine	MSA	Dialects	Strengths	Pricing model
Microsoft Azure Speech	Yes, 16 country locales	Country-flavored voices	Best locale coverage, documented diacritization improvements, SSML	Free monthly allowance, then per-million-character pricing
Amazon Polly	Yes (Zeina)	Gulf (Hala, Zayd)	Built-in diacritizer, cheap at scale, AWS integration	Per million characters; 12-month free tier
Google Cloud TTS	Yes (ar-XA)	—	Strong neural quality, generous free tier	~$4–$16 per 1M characters by tier
ElevenLabs	Yes	Some via multilingual models	Most expressive delivery, voice cloning	Free tier; paid from around $5/month
Hamsa (specialist)	Yes	Egyptian, Gulf, Levantine, Iraqi	Dialect-first, tashkeel-aware	API pricing; see vendor

For accessibility and document reading rather than audio production, consumer apps also matter — NaturalReader supports Arabic among its languages and is a simpler on-ramp than a cloud API. For the full multi-language landscape, see our text to speech software guide.

RTL workflow: practical tips for producing Arabic audio

Arabic is written right-to-left, and most TTS dashboards, scripts, and subtitle tools were designed left-to-right. The pronunciation engine doesn't care — text is text — but your workflow will:

1. Beware mixed-direction text. Arabic sentences containing Latin brand names, numbers, or URLs are where copy-paste errors happen — the visual order in your editor may not match the logical character order. Always proof by listening, not by looking.
2. Numbers read naturally either way. Engines normalize both Eastern Arabic numerals (٢٠٢٦) and Western ones (2026). Keep one style per document for consistency.
3. Vocalize sparingly, strategically. Don't diacritize whole documents — add tashkeel only to the words the engine misreads (proper nouns are the usual offenders). Free online diacritizers (e.g., tools built on the Tashkeela corpus) can help vocalize a tricky passage.
4. Store text in UTF-8 and test your pipeline end-to-end. Encoding bugs that silently strip diacritics or reorder characters are far more common in RTL pipelines than engine mispronunciations.
5. Subtitles and players need RTL support too. If the audio accompanies video, verify your captioning tool renders Arabic correctly — right-aligned, correctly shaped letters.

If you also produce audio in other non-Latin-script languages, the input-handling lessons carry over — our Chinese text to speech guide covers the parallel problem of tone and character ambiguity, and the French text to speech guide covers locale selection for another global language. And if the bottleneck is producing the written multilingual content in the first place, that's what AutoSEO automates.

Frequently Asked Questions

Does Arabic text to speech work without tashkeel (diacritics)?

Yes — this is the default case. Modern engines from Amazon, Microsoft, Google, and ElevenLabs automatically predict the missing short vowels from context (automatic diacritization) and read normal unvocalized Arabic correctly most of the time. When an engine misreads a specific word, adding diacritics to just that word in your input forces the correct pronunciation, since all major engines respect explicit tashkeel.

Should I use Modern Standard Arabic or a dialect voice?

Use MSA for formal, informational, and pan-Arab content — news, e-learning, documentation, accessibility. Use a dialect voice (Egyptian, Gulf, Levantine) when you want to sound local and conversational: ads, social content, chatbots for one market. MSA is understood everywhere but sounds formal; dialects sound natural but only travel within their region.

Which engine has the best Arabic voices?

Azure has the broadest coverage, with voices across 16 Arabic country locales and documented improvements to its diacritic prediction. Amazon Polly is a strong, economical choice with MSA (Zeina) and Gulf Arabic (Hala, Zayd) voices and a built-in diacritizer. ElevenLabs offers the most expressive MSA delivery. For deep dialect coverage, specialist Arabic platforms like Hamsa go further than the big clouds. Test each with your own unvocalized text before committing.

Can Arabic TTS read numbers and dates correctly?

Yes. Major engines normalize both Western (123) and Eastern Arabic (١٢٣) numerals into spoken Arabic, including dates and currency in common formats. Grammatical agreement of numbers is genuinely hard in Arabic (numbers inflect for gender and case), and neural engines get everyday cases right — but proof financial or legal figures by listening before publishing.