Arabic speech recognition has far more web audio than manually transcribed dialect data. This release tackles that imbalance by turning a small human-labeled seed—about 51–63 hours per dialect—into 39,028 hours of weakly labeled speech. The key tradeoff is scale for label certainty: the corpus makes dialect-specific training practical, but its transcripts are not ground truth.
Key Value
- Covers five dialects spanning Maghrebi, Egyptian, Mesopotamian, Gulf, and Levantine varieties, with about 10 million segments. This supports comparisons across dialects rather than treating Arabic as a single homogeneous speech domain.
- Provides about 11,461 hours for Algerian, 8,533 for Egyptian, 10,162 for Iraqi, 4,512 for Saudi, and 4,361 for Syrian Arabic. The uneven distribution matters when balancing training data or interpreting cross-dialect results.
- The accompanying study reports a 23.86% macro-average word error rate for dialect-specific Zipformer models on GigaSpeechBench. This indicates the data can support useful systems, while the benchmark remains separate and human-transcribed.
Who Should Use It
A good fit for researchers training or evaluating Arabic dialect ASR, weak supervision, and cross-domain transfer. Look elsewhere for commercially licensed audio, gold-standard transcripts, or fine-grained regional and social dialect labels: the release is research-only, labels are machine-generated, dialect codes are broad, and web-source biases and code-switching may affect results.
