User-customized password speaker verification using multiple reference and background models