介绍了一种基于词网的最大似然线性回归无监督自适应算法，并进行了改进。根据解码得到的词网估计变换参数，词网的潜在误识率远小于识别结果，因此可以使参数估计更为准确。传统的一个很大缺点是计算量极大，较难实用，对此本文提出了两个改进技术：1利用后验概率压缩词网；2利用单词的时间信息限制状态统计量的计算范围。实验测定,误识率比传统相对下降了。-Introduced the term network based maximum likelihood linear regression unsupervised adaptive algorithm, and an improved. According to decode the received word net estimated transformation parameters, the word error rate of net potential is far less than the recognition results, it can make parameter estimation more accurate. A major drawback is that the traditional calculation enormously difficult practical, this paper presents two improved technology: 1 compression using word posterior probability network　2 time information using the word limit state statistic calculation. Experimental determination of the relative error rate than traditional down.