Pseudo-supervised training improves unsupervised melody segmentation