Do speech foundation models really learn words?