Research On Speech Recognition Based On Word Activation Force

The technology of Speech recognition mainly focuses on the following four aspects:the preprocessing of speech signal, acoustic modeling, language modeling, and post processing. Cause the Word Activation Force has two main characters, one is that the Word Activation force has rich semantic information, the other is that the Word Activation force has long-range dependency information,The main research revolves around word activation force based language modeling and the word activation force based lattice rescoring according to the characteristics of word activation force. The main work of this paper is shown as follows:1.This paper researches how to use word activation information to smooth and model language modelBy using the semantic information of word activation force, this paper constructs a high-level word pair and proposes a method of language model smoothing to solve the problem of data sparse. Compared with the traditional this method performs better. This paper uses the word activation information to model long distance language model, Judging from the results, the word activation force based language model gives a better performance than baseline system.2. This paper proposes a word activation force based lattice rescoring algorithm.This paper uses the close semantic information and long-range dependency information to rescore lattice. The results show the using of the information does a lot good the post-processing, it can help the system to eliminate the bad candidates. This paper proposes three specific ways to work out the rescoring algorithm.