Speech parsing by downward request search based on the divide and conquer method