Error-Aware TF-IDF Retrieval-Augmented Generation for ASR Error Correction

Short summary

Researchers introduce an error-aware TF-IDF retrieval framework to correct automatic speech recognition hallucinations, particularly for rare entities and domain-specific terms. Using a sparse penalty matrix based on historical phonetic errors, the method improves Persian speech recognition accuracy from 23.06% to 18.83% word error rate. This lexical approach offers a practical alternative to heavyweight neural models for low-resource language ASR systems with minimal inference latency.

• •Error-aware TF-IDF algorithm corrects ASR hallucinations without expensive neural models
• •Achieves significant accuracy improvement: 23.06% → 18.83% word error rate on Persian speech
• •Practical for low-resource languages with near-zero latency inference overhead

Generated with AI, which can make mistakes.