Gravar-mail: How learning to abstract shapes neural sound representations