Arabic is written with less than it is spoken. When an Arabic reader meets an unvowelled sentence, they rebuild the short vowels from context without noticing. A speech model has to do the same thing with statistics. That is where the trouble starts: a single text can allow more than one pronunciation, and each pronunciation changes the meaning. On top of that, Arabic has an enormous range of dialects, and well-organized speech data for them is scarce. In this article we read what recent research papers say about these obstacles, using the numbers their authors published, and then draw out what this means in practice for anyone producing Arabic audio content.
A text that hides its pronunciation: the short-vowel problem
Arabic is usually written without its short vowels (fatha, damma, kasra, sukun and tanween). The word "كتب" is a familiar example: it can be read "kataba" (he wrote), "kutub" (books), or "kutiba" (it was written). A reader settles the question from the whole sentence.
A paper on preparing Arabic speech data by Hala Al Masri and Muhy Eddin Zater describes this problem clearly: one word can carry several meanings when the vowels are missing, and both the voice actor and the AI module benefit from diacritics that reduce ambiguity. The paper proposes partial diacritization, placing marks only in the ambiguous spots, because fully diacritized text is scarce and expensive.
Can we add diacritics automatically? The numbers show progress, and a gap
The logical fix is to add diacritics automatically before the text is spoken. Automatic diacritization has improved a lot, but it is still far from perfect. The CATT paper (a character-based Arabic tashkeel transformer) reports results for its model on two benchmarks:
- On the WikiNews benchmark with case endings, the encoder-decoder model reached a character error rate of about 5.96% and a word error rate of about 20.06%.
- On the harder CATT benchmark, the character error rate was about 8.62% and the word error rate about 34.19%.
The difference between the benchmarks matters. A word error here means the word contains at least one diacritization mistake. Even though the paper's model beats the models it was compared with, it gets about one word in five wrong on the easier benchmark and about one in three on the harder one. Case endings (the grammatical endings of words) are the hardest part of diacritization, because they depend on sentence structure and not on the word alone. That is why the paper's numbers improve when case endings are excluded (about 3.63% character error on WikiNews).
The takeaway: automatic diacritization reduces ambiguity but does not remove it, and any mistake in it carries straight into the pronunciation.
Can data replace diacritics? An experiment from QCRI
There is an alternative research direction: instead of adding diacritics to the text, train the model on a huge amount of speech so that it learns to guess correctly. That is what the paper "More Data, Fewer Diacritics" (March 2026) tested, from researchers at the Qatar Computing Research Institute (QCRI). The team collected about 20,000 hours of audio from social media platforms and, after cleaning and segmenting, kept about 4,000 hours suitable for training. They then trained models of different sizes, with and without diacritics, and measured word error rate (WER) using a speech recognition model.
The main results:
- With 4,000 hours and diacritics: WER of 1.67, the best result.
- With 4,000 hours and no diacritics: WER of 3.18.
- The undiacritized model trained on 1,000 hours (3.12) beat the diacritized models trained on 100 hours (3.48 and 3.62).
So diacritics still give the best result when the amount of data is held constant, but a lot of data makes up for much of their absence. The researchers note important limits: the experiment covered Modern Standard Arabic only, the measurement was objective with no human listening tests, and the same family of recognition model used for scoring is tied to the data-preparation pipeline, which may bias the results.
Standard Arabic and dialects: a written language and a spoken one
The second problem goes deeper than diacritics: which "Arabic" do we mean? The Habibi paper (January 2026) says Arabic has more than 30 spoken varieties, that Modern Standard Arabic is rarely used in everyday speech, and that dialects differ greatly in vocabulary and pronunciation. It adds that most available Arabic speech data was built originally for speech recognition, is often noisy and undiacritized, and that there is no unified benchmark for evaluating multi-dialect speech synthesis.
The "Arab Voices" study (University of British Columbia, January 2026) confirms the picture from the data side. It catalogued 31 Arabic speech datasets and found that dialect labels are not standardized and that regions such as North Africa (Algerian, Hassaniyya, Moroccan and Tunisian) and Iraq are under-represented. In zero-shot speech recognition tests, the error rate of some Whisper configurations exceeded 200 and reached 271.25 for Mesopotamian Arabic, while the error rate for Standard Arabic dropped below 12 in several models. These are speech recognition figures, not speech synthesis, but they reflect the same gap between Standard Arabic and dialects in the data.
One striking result from Habibi is that training a model on Standard Arabic first and then on the dialect gives much better results than training on the dialect alone: on the SADA dataset, WER was about 13.30 with this approach versus 55.63 when training from scratch on the dialect. In other words, Standard Arabic works as a foundation to build on, and dialect data alone is not enough.
Before the model speaks: text normalization and recording
Many of the errors we hear do not come from the speech model but from the text before it arrives. Al Masri and Zater explain that the data pipeline converts numbers, dates, non-Arabic words and abbreviations into spoken Arabic text before diacritization, spell checking and segmentation. A text such as "1994", "5 km" or "iPhone" is not pronounced the way it is written, and unless it is converted explicitly it will be mispronounced or skipped.
The paper also covers recording quality: noise, hiss and plosives in an untreated studio spoil training data, and it advises preventing noise at the source, keeping a steady and natural pace, and placing pauses only at punctuation marks. This is a reminder that the quality of synthetic speech depends on the quality of the human speech it learned from.
How do we know a voice is good? The measurement problem
Most studies rely on automatic metrics such as WER, speaker similarity and predicted naturalness (UTMOS), and all of them have limits. Habibi warns that some of these metrics do not generalize well outside the training data, which is why the researchers added human evaluation with 10 to 20 native speakers per dialect. In its comparison with a commercial model, neither side won on every metric: Habibi did better on speaker similarity and on error rate in several dialects, while the commercial model did better on naturalness. The lesson is that the "best system" depends on what you measure, and that actually listening cannot be replaced.
What does this mean for anyone producing Arabic audio content?
Some practical habits follow from the research above. This is my own analysis, not something the papers state word for word:
- Add diacritics to the ambiguous spots yourself. Partial diacritization on words that allow more than one pronunciation is cheaper and more accurate than relying on guesswork, which is what Al Masri and Zater suggest.
- Write numbers, dates, abbreviations and foreign names the way you want them spoken.
- Choose Standard Arabic or a dialect on purpose, and avoid mixing them in one paragraph unless it is an artistic choice.
- Listen before you publish. No automatic metric replaces a human ear, especially for names, terms and numbers.
- Split long text into short parts and review them part by part, since a single mistake in a full hour is hard to catch.
Moknah today
Moknah handles part of these steps inside the platform: for Arabic it offers three text processors, Basic, Tashkeel and AI Poet, and Tashkeel and AI Poet use more credits than Basic. The platform supports text-to-speech in more than 130 languages. It does not replace human review; it reduces repetitive manual work. You can start free and try the text that matters to you from the pricing page.
Related reading
- Why does the Arabic language pose a major challenge for global AI voice-over systems?
- How do you choose the most suitable voice for your brand using the Moknah platform?
- Text-to-speech technology
Sources
- Musleh, Zhang, Darwish, "More Data, Fewer Diacritics: Scaling Arabic TTS", arXiv:2603.01622, March 2026: https://arxiv.org/html/2603.01622
- "CATT: Character-based Arabic Tashkeel Transformer", arXiv:2407.03236: https://arxiv.org/html/2407.03236
- Al Masri, Zater, "Arabic Text-to-Speech (TTS) Data Preparation", arXiv:2204.03255: https://arxiv.org/pdf/2204.03255v1
- Chen et al., "Habibi: Laying the Open-Source Foundation of Unified-Dialectal Arabic Speech Synthesis", arXiv:2601.13802: https://arxiv.org/html/2601.13802v2
- Sullivan et al., "Arab Voices: Mapping Standard and Dialectal Arabic Speech Technology", arXiv:2601.13319, January 2026: https://arxiv.org/html/2601.13319v2