A dialect-balanced multi-dialect Arabic speech-recognition corpus: 54,600 clips /
267.3 hours across 13 Arabic dialects, 16 kHz mono. Each dialect is evenly
represented — 4,000 train + 200 test clips per dialect — so models and
evaluations aren't skewed toward high-resource dialects (e.g. Egyptian/Gulf).
Used to train the oddadmix v2 dialectal-ASR model family.
