The main ovjective of this dessertation is the development of a speaker adaptive speech recognition system which can yield the acceptable recognition rate even for speakers who have not provided enough speech to train the recognition system. This system consists of the baseline system and the speaker adaptation system which is made up of two stages:codebook adaptation and HMM parameter adaptation. First, we presented a speaker-dependent system based on HMM. This system has been used the baseline system for speaker adaptation. Second, wo proposed a modified Viterbi scoring algorithm to imorove the discriminability of phonetically similar words. The proposed algorithm weights the Viterbi scores of state which are considered to be perceptually important. When the candidate words were so similar that the phonetical difference between the top 1 and top 2 candidates was one phoneme, the modified Viterbi algorthm reduced the recognition error rate by about 19\% as compared to the conventional method. Third, we proposed a codebook adaptation scheme using a neurallyinspired LVQ whith highly descriminat ability. By the proposed scheme, the codebook was generated to have the descriminant feature rather than the minimum distortion for adaptation speech. From the adaptation speech. From the adaptation experiment, we found that the adaptation with LVQ codebook resulted in higher destortion error than that with conventional codebook but the recognition rate was better, and that LVQ2 codebook, in which K-means each codebook was used to initialize, yielded the best recognition rate. Fourth, we presented a modified corrective training algorithm as a method to improve the performance of HMM parameter adaptation. The observation probability parameters of HMM are re-estimated by this algorithm after performing the spectral mapping algorithm. From the experiment, we found that the performance of the speaker adaptation system was improved after adopting the modified CT algorithm, and...