DA-ICL: Distribution-Aware In-Context Learning for Arabic Automatic Speech Recognition Error Correction
TL;DR
A novel two-stage framework for accurate and efficient Arabic ASR enhancement, combining HuBERT-based acoustic modeling with LLM-based DA-ICL correction and LoRA-efficient adaptation yields a robust, accurate, and scalable solution for Arabic ASR, effectively bridging the gap between acoustic signal and linguistic knowledge.
Abstract
Arabic Speech Recognition (ASR) faces compounded challenges due to rich dialectal variation, morphological complexity, and data scarcity. While self-supervised speech models such as HuBERT excel in acoustic representation, they lack the deep linguistic reasoning needed to resolve ambiguities unique to Arabic. Large Language Models (LLMs) offer complementary grammatical and semantic knowledge, yet their role in systematic, real-time Arabic ASR error correction remains underexplored. In this work, we propose a novel two-stage framework for accurate and efficient Arabic ASR enhancement. First, we fine-tune a HuBERT model on the Common Voice Arabic corpus, establishing a baseline word error rate (WER) of 19.3%. Second, we introduce Distribution-Aware In-Context Learning (DA-ICL), a prompting strategy that supplies the Arabic LLM Aya-23-8B with a curated set of few-shot examples derived from a systematic taxonomy of ASR error types, including phonetic confusions and morpho-orthographic errors. DA-ICL enables precise, structurally faithful corrections, reducing WER to 9.6% without undesirable sentence rephrasing. To address domain shift and catastrophic forgetting, we further apply Low-Rank Adaptation (LoRA) to adapt a pre-trained HuBERT model to new domains parameter-efficiently. This approach reduces out-of-domain WER from 67% to 24% while preserving in-domain performance, demonstrating improved generalization without full fine-tuning. Our results confirm that combining HuBERT-based acoustic modeling with LLM-based DA-ICL correction and LoRA-efficient adaptation yields a robust, accurate, and scalable solution for Arabic ASR, effectively bridging the gap between acoustic signal and linguistic knowledge. Our framework achieves a WER of 9.6% on Common Voice Arabic, significantly outperforming Whisper-large (47.49% zero-shot, 37.89% with LoRA fine-tuning) and demonstrating the effectiveness of our linguistically-aware approach for Arabic speech recognition.