In this paper we investigate the word extraction task in on-line recognition of cursively handwritten text lines. For the segmentation we propose a method which is based on the assumption that the size of gaps between consecutive words may considerably vary, but humans usually leave more whitespace between two consecutive words than between two connected components that belong to the same word. We use several metrics known from off-line word segmentation for measuring the distances between two adjacent components. Then we apply different procedures to get the initial threshold for segmentation. Using these techniques we could significantly increase the segmentation rate compared to methods which are usually applied in on-line text recognition systems.