Automatic Quranic recitation assessment requires accurate detection of phonetic and Tajweed-related pronunciation errors while providing low-latency feedback. However, most speech recognition models are designed for offline processing and are too large for deployment on consumer mobile devices.

We present a real-time Quranic recitation assessment system based on a multi-head Wav2Vec2-BERT model trained to jointly predict phonemes and Tajweed Sifat attributes using Connectionist Temporal Classification. To enable streaming inference without retraining, we introduce an Adaptive Eager-Emission framework that combines heuristic output commitment, dynamic context expansion, and overlap recovery, reducing perceived latency while preserving long-range temporal features.

For efficient on-device deployment, we propose a structured compression pipeline based on encoder pruning, recovery fine-tuning, and INT8 quantization with ONNX optimization, reducing a 600M-parameter, 2.4 GB model to a 24.5M-parameter, 36 MB model while maintaining recognition performance.

The streaming framework consistently outperforms static sliding-window approaches across the accuracy–latency trade-off, achieving responsiveness as low as 1.0 s at roughly 2.1% phoneme error rate (PER), or near-offline performance below 1% PER at 6.0 s average chunk latency, while the production streaming configuration maintains 99.45% average Sifat attribute accuracy. On a mid-range Android CPU, the compressed model runs at a real-time factor of 0.225 with a peak memory of about 600 MB.