📄 KSAA-2026 Shared Task on Arabic Speech Dictation with Automatic Diacritization

Summary

The KSAA-2026 shared task, held at OSACT7, targets a long-standing problem in Arabic NLP: restoring full diacritics to text that comes from speech. Most written Arabic omits diacritics, which creates lexical and syntactic ambiguity. Text-only diacritizers also miss phonetic and prosodic cues, and ASR systems seldom produce well-diacritized output. The organizers therefore framed diacritization as a multimodal problem, where systems receive both the audio and its undiacritized transcript.

The task had two tracks:

• Data contribution: volunteers recorded and peer-reviewed diacritized prompts on the VoiceWall platform.
• Diacritization modelling: teams built systems on a roughly 5-hour corpus of MSA and dialectal speech with fully diacritized references.

Systems were scored with DER, WER and SER under settings that include or exclude case endings and undiacritized characters. Four of the five submitting teams beat all three baselines. Both speech input and in-domain fine-tuning clearly helped. Case endings and deciding when a character needs no diacritic remained the hardest parts. The task provides a public benchmark for speech-informed Arabic diacritization, with uses in education, accessibility and speech-driven writing tools.

Key points

• Dataset: about 5 hours of speech from male and female speakers of nine nationalities (Saudi, Egyptian, Syrian, Algerian and others). Clips last at most 9 seconds and cover politics, economy, sports, religion and news. The train split has 2,327 files (~4.5 h), dev is ~30 min, and the test set (~40 min, hidden) was kept for evaluation.
• Data contribution track: VoiceWall gathered 2,160 recordings, with 360 prompts assigned to each participant. Quality control combined automatic checks with cross-team review.
• Participation: 29 teams registered for the diacritization subtask, and 5 submitted valid runs (Thaka, TantaArabNLP, Abjad AI, Eraserhead, Fine-Tashkeel). Their approaches included:
  • Abjad AI: grouped speech conditioning;
  • Eraserhead: ASR consistency filtering combined with speaker-adaptive post-processing.
• Baselines: three Transformer systems.
  • Text-only, trained on Tashkeela.
  • Text + ASR, trained on Tashkeela and CartTTS following Shatnawi et al. (2024).
  • A fine-tuned version of the text + ASR model, which was the strongest baseline: WER 31.84, DER 9.91.
• Results (primary metric): the primary metric is WER with case endings, including characters without diacritics. Thaka ranked first with WER 23.26 and DER 6.87. The text-only baseline reached a WER of 49.85, showing a large gain from adding speech information.