Correlation between observations of different states is an important apriori information reflecting speech characteristics, which is a key factor improving speech recognition system robustness. Since speech and noise are statistically independent, correlation information can be used to reduce noise effect on speech recognition performance in noisy environment. This paper proposed a new speaker adaptation method using Projection to Latent Structure (PLS) algorithm. PLS can extract a group of basis vectors reflecting the characteristics of codebook and adaptation data, so as residual, the noise can be subtracted from speech. Different from Eigenvoice method, PLS extracts basis vectors using adaptation data. Furthermore, it needs very small storage space. Experimental results on large vocabulary continuous speech database show that this method is superior to baseline, MAP, EV and MLLR, etc. Keywords: Projection to Latent Structure (PLS), Speaker Adaptation, Correlation.