Dual-Memory Multi-Modal Learning for Continual Spoken Keyword Spotting with Confidence Selection and Diversity Enhancement.