关键词: 潜在语义分析, N元文法, k均值聚类, 连续语音识别

Abstract: The theory of Latent Semantic Analysis（LSA） for speech recognition is described，and the related techniques for implementing LSA-based language modeling in speech recognition systems are presented.An LSA-based semantic model is constructed on the WSJ0 text corpus.This paper uses the interpolation method to combine this semantic model with conventional 3-gram to form a hybrid language model（i.e.，LSA+3-gram）.To optimize the performance of the hybrid model，it applies k-means algorithm to perform vector clustering in the LSA vector space while the density function is used to initialize the centroid.The constructed hybrid language model outperforms the corresponding 3-gram baseline：Continuous speech recognition experiments conducted on the WSJ0 test corpus show a relative reduction in word error rate of about 13.3%.

Key words: latent semantic analysis, N-gram, k-means clustering, continuous speech recognition