The paper analyzes vowel data using reduced-rank linear discriminant analysis (RRLDA), reduced-rank quadratic discriminant analysis (RRQDA) and principal component analysis plus linear discriminant analysis (PCA+LDA). Then it drew some curves of false classification about the three model. A curved surface of the best classification has drawn for RRLDA and PCA+LDA after reduced rank to two dimensions. From the result, it can be conclude that RRLDA is good than PCA+LDA. The false classification of RRQDA is considerably big, because PCA ignores the information of classification about data and only disperses data during reducing rank. Simultaneously these curves prompts RRLDA owning the best generalizing ability when its dimension is 2 in subspace, and PCA+LDA owning the best generalizing ability when its dimension is 4 in subspace, and RRQDA owning the best verify error rate in tenth dimension.