Quranic ASR Benchmark: 600 Leakage-Free Audio Clips

600 verified audio clips of Quranic recitation in the Hafs riwayah, held out from training data to test model generalization. The benchmark provides a standardized evaluation set with a live leaderboard for comparing Arabic automatic speech recognition models. It was created by Muno459 and last updated in June 2026.

Use Cases

• Benchmarking Arabic ASR model generalization based on leakage-free, held-out test clips.
• Comparing model performance on Quranic recitation using a standardized scoring framework.
• Training ASR models on classical Arabic pronunciation with a verified evaluation set.

Strengths

• 600 audio clips, each verified absent from the creator's training data.
• Clips are standardized (same clips, same scoring) for fair model comparison.
• Includes a live leaderboard for tracking model performance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the specific Quranic recitation sources used.

Provenance

Source

huggingface

Collection Method

Clips sourced from multiple recitation collections (e.g., everyayah_heldout).

Freshness

Last updated 2026-06-08 19:08:48; freshness should be verified.