Synthetic moving-speaker dataset