Reproduce “Dialectal Coverage And Generalization in Arabic Speech Recognition” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Amirbek Djanibekov, Hawau Olamide Toyin, Raghad Alshalan, Abdullah Alitr, Hanan Aldarmaki

Researchers from Mohamed bin Zayed University of Artificial Intelligence and STC developed and evaluated Arabic ASR models that integrate diverse dialects and multilingual data during pre-training and fine-tuning. Their approach improved performance on various Arabic dialects, including low-resource ones, and significantly enhanced recognition for code-switched utterances, demonstrating reductions in Word Error Rate by over 10% on code-switched test sets.

Key insights

  • Incorporating broad dialectal data during pre-training improves downstream dialectal ASR performance without degrading MSA performance.
  • Joint multi-dialectal fine-tuning benefits low-resource dialects and zero-shot transfer but can lead to performance degradation for high-resource dialects compared to dialect-specific fine-tuning.
  • Multilingual pre-training is crucial for recognizing code-switched Arabic, but it introduces a trade-off with monolingual Arabic dialectal performance due to language interference.
  • Allowing the model to infer the dialect ID during inference generally yields better performance for joint multi-dialectal models than explicitly forcing a dialect ID.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.