disfluency

该论文评估了 ASR 转录文本归一化（如去除填充停顿、标点）对对话行为分类的影响，发现重度归一化会在多个模型与语料库上一致降低 Macro-F1，主张应采用任务相关的选择性归一化而非一刀切处理。

This paper proposes a continual learning approach to integrate disfluency tokens into pretrained ASR models, addressing catastrophic forgetting and improving recognition of disfluent speech.