TL;DR of the paper, generated automatically.

GigaSpeech 3 is a 39,000-hour weakly labeled Arabic speech corpus for five dialects.

• Data: about 10 million segments from Algerian, Egyptian, Iraqi, Saudi and Syrian Arabic.
• Method: started from 51 to 63 hours of manual transcription per dialect, then auto-labeled web audio.
• Results: dialect models hit 23.86% macro-average word error rate on GigaSpeechBench, per the paper.