Abstract: Integrating high level semantic knowledge in speech recognition has been a hot topic in the field. Latent semantic analysis (LSA) technology can model long span correlation of words in the language efficiently. How to utilize LSA in language modeling and guide searching procedure more accurately is an urgent task. We analyzed major problems on the modeling and decoding procedure of LSA language model, realized modeling of LSA language model, and proposed a method to use LSA language model in the first-pass decoding of continuous speech recognition system. Experimental result shows that the proposed method can improve the performance of large vocabulary continuous speech recognition of Chinese language significantly.

Key words: language model, vector space model, latent semantic an alysis, speech recognition