GigaSpeech 3 Scales Arabic Speech Recognition via Weak Labels
November 2, 2026
GigaSpeech 3 provides 39,000 hours of multi-dialect Arabic speech data, including Algerian, Egyptian, Iraqi, Saudi, and Syrian dialects. The corpus was built using speech-recognition-biased self-supervised learning and noisy-student refinement.
HOW THIS AFFECTS YOU
●
builderYou can improve speech recognition performance for diverse Arabic dialects using this large-scale weakly labeled corpus.