基于多模态一致性指导的无参考数据选择用于自动语音识别口音适应

Automatic speech recognition (ASR) systems often perform poorly on accented speech due to acoustic-phonetic and prosodic shifts that create mismatches with training data, making labeled accent adaptation costly. Common pseudo-label selection heuristics are primarily text-centric, such as perplexity (PPL) filtering, which may favor fluent yet acoustically mismatched hypotheses, leading to error amplification during fine-tuning. To address this issue, a multimodal consistency-guided, reference-free data selection pipeline is proposed for ASR accent adaptation under a transductive, label-free protocol. The pipeline begins with a target-aware preselection step based on submodular mutual information to enhance query relevance and reduce downstream computation. This approach significantly improves the efficiency and accuracy of accent adaptation.