Better speech synthesis through scaling