使用无监督网络MLLR自适应改进算法的语音识别

介绍了一种基于词网的最大似然线性回归(Lattice-MLLR)无监督自适应算法,并进行了改进。Lattice-MLLR是根据解码得到的词网估计MLLR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。Lattice-MLLR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Lattice-MLLR的误识率比传统MLLR相对下降了3.5%,改进技术使Lattice-MLLR计算量下降幅度超过了87.9%。

Unsupervised Lattice-MLLR in Speech Recognition

The lattice-based maximum likelihood linear regression(MLLR) for the unsupervised adaptation algorithm is discussed.Assuming that the 1-best hypothesis is accurate,Lattice-MLLR passes through the word-lattice to accumulate statistics for the MLLR transform estimation procedure.Since the oracle word error rate of word-lattice is much lower than that of the 1-best hypothesis,word-lattice is more likely to provide correct models to estimate the transform.The disadvantage of Lattice-MLLR is that it costs too much computation and storage,so,this paper proposes the two improved measures:(1) compressing the word-lattice;(2) utilizing the information of word time to limit the statistics accumulation.Recognition rate on the national speech recognition evaluation database indicate a relative 3.5% reduction in the word error rate of Lattice-MLLR over traditional MLLR,and the proposed measures can decrease more than 87.9% in computation of Lattice-MLLR.