Speech Key Words Recognition Technology Research

Key words spotting and recognition (KWSR) is a kind of speech recognition whose principal task is to find and recognize one or more given words in a continual speech flow. Due to its wide practical applicability, KWSR is a research hot spot in this field. Two acoustic models are discussed and applied in the paper to implement an experimental system of KWSR—Flight Information Query System.Firstly it introduces two different ways of the speech features extraction, presents features extraction used in the paper and then expounds the advantages of the used ways by using experiments.Secondly, the paper gives an account of the two various acoustic models by which corresponding key words recognition systems are constructed. The two acoustic models are based on the improved Gaussian Mixture Model (GMM) and Conditional Random Field Model (CRFM) respectively. The GMM is such a model that has been approved to bring about good effects in the key words recognition, while the improved GMM is brought in under the circumstance that it can improve the recognition of the isolated words. The CRFM owns the better classification ability than the HMM, hence it is introduced into the key words recognition model.Thirdly, in the key words recognition model, filler model occupies a very important position. However, the three traditional filler model cannot meet the present needs. So according to the analyses of the particularity of Chinese and Pinyin classification polymerization, this paper establishes an effective filler model designed for Chinese. The experiments show that the filler model constructed in this paper has better performance than that of the traditional one.Fourthly, in the final stage of key words recognition, the speech confidence of confirmation is used to improve accuracy and reduce error. Based on the Posterior probability and Likelihood Ratio, the paper introduces two traditional confidence calculation methods. But a new confidence calculation method, on the basis of mutual information, is applied in this paper. And it is proved to improve the confidence degree of confirmation.The last but not least, based on the improved GMM and CRFM within key words recognition system respectively, a certain number of the training sample library and test sample library are recorded and set up. These libraries are also simulated under the circumstance of matlab software. Through system simulation, properties are discussed under the two situations. Besides, the special situations in which the two libraries can be applied are discussed as well.