Abstract
Abstract To address the challenge of balancing real-time processing and accuracy in spoken English translation, this paper proposes a real-time error correction model based on Mistral and self-supervised learning. This model integrates parallel decoding and hierarchical noise simulation strategies, which significantly mitigates the challenges of data scarcity and inference latency. Experiments show that the model reduces the inference delay from 180ms to 50ms, and improves the accuracy for technical term correction from 54.2% to 88.6%. In a low-resource scenario with only 1% labeled data, the model achieved a BLEU score of 25, far exceeding the baseline of 12, confirming its excellent performance and robustness in complex spoken scenarios.