Speech segmentation by statistical learning depends on attention