Money and business in the Middle East.

AI

MBZUAI method cuts Arabic diacritic errors to 3.8%, university says

A new approach to adding Arabic diacritics to speech transcripts recorded a 3.8% diacritic error rate on classical Arabic data, against roughly 9% for a text-and-speech baseline, according to Mohamed bin Zayed University of Artificial Intelligence.

· Originally published by ontime+ · Last verified: 19 Nov 2026

Key Points

  1. MBZUAI researchers built a method that automatically restores diacritics on transcripts of spoken Arabic.
  2. Most speech-to-text systems skip diacritics, leaving Arabic transcripts ambiguous for readers and downstream machines.
  3. Results show the method holds up when training and test dialects differ, where older approaches collapse.

The latest:

A new approach to adding Arabic diacritics to speech transcripts recorded a 3.8% diacritic error rate on classical Arabic data, against roughly 9% for a text-and-speech baseline, according to Mohamed bin Zayed University of Artificial Intelligence. The team presented the findings at the Interspeech conference in Sydney. The method constrains speech recognition so it only inserts diacritics without rewriting the transcript underneath.

Details:

  • The problem: Written Arabic omits short vowels and other pronunciation marks in most texts, which can make words ambiguous, MBZUAI said. Diacritics can record those sounds, but most speech-to-text AI systems do not add them automatically to Arabic transcripts, creating difficulties for both human readers and machines processing the output.
  • The method: The approach builds on a speech recognition technique known as connectionist temporal classification, according to the university. The researchers constrain the search lattice so the system is allowed only to add diacritics, and cannot alter the underlying transcription it has already produced — a task the field calls diacritic restoration.
  • The team: The work was led by research engineer Rufael Marew, postdoctoral researcher Amr Keleg, and Hanan Aldarmaki, assistant professor of natural language processing at MBZUAI. The findings were presented at Interspeech in Sydney, and published on the university’s news page on Nov 16, 2026.
  • Test data: Two datasets were used, MBZUAI said. The first covers classical Arabic read by a single speaker from a classical religious text. The second covers modern standard Arabic news read by multiple speakers, giving the models a range of voices and a different register to work against.
  • Cross-dataset results: Trained on classical Arabic and tested on modern standard Arabic, the constraint-based method posted a diacritic error rate of about 12%, against 19% for the text-and-speech baseline, according to the university’s figures.
  • The gap: Reversing the setup exposed the sharpest difference. Trained on modern standard Arabic and tested on classical Arabic, the baseline’s error rate reached nearly 77%, while the constraint-based method stayed around 12% — roughly a sixfold spread between the two approaches on identical data.
  • Best case: Trained on both datasets together, the constraint-based method recorded 3.8% error on classical Arabic and 8.7% on modern standard Arabic. The baseline scored approximately 9% and 10% on the same two datasets, MBZUAI said.
  • Stated limits: The researchers acknowledge the approach has not been tested on more conversational data. Both datasets consist of carefully read speech, which is cleaner and more predictable than ordinary spoken Arabic, and the university did not announce a timeline for testing on conversational material.

Background:

Diacritics are small marks placed above and below Arabic letters to indicate short vowels and other pronunciation details. They appear in religious texts, poetry and teaching material, but are left out of most everyday writing, where readers infer pronunciation from context.

Between the lines:

The headline number is 3.8%, but the more telling figure is the 77% baseline failure when the training and test registers do not match. A method that holds near 12% across that mismatch is less fragile to the kind of register shift real Arabic audio produces — though the university’s own caveat, that only carefully read speech was tested, limits how far that stability can be assumed.

What’s next

Watch for tests on conversational Arabic audio, the condition the researchers flag as untested, and for whether the constraint-based method is adopted in Arabic speech-to-text and transcription tools used by regional developers.

Read on ontime+ ↗