Speech parametrization and recognition using block and recursive linear prediction with data compression