Speech Keyword Spotting And Confidence Measure

As the Information age coming, the computer becomes an important tool in human life. In order to achieve a more friendly user interface, and let the computer can understand human language, speech recognition technology have being develeped. Speech keyword spotting is an important research direction of speech recognition. Its purpose is to recognize a given set of words from the continuous and unrestricted flow of natural speech. Keyword spotting technology can be applied to a number of occasions with specific requirements. Besides, it has features of high speed , high detection rate and great practical ect . So, keyword spotting has wide application prospect and great research value. Meanwhile,in order to improve the overall performance of recognition system, the confirmation of the recognition results is also very important. Therefore, the research of confidence measure in speech is also particularly necessary. This paper focuses on speech keyword spotting technology and confidence measure program , the main research contents are described in details as follows:1.The process and theory of the keyword spotting is introduced in this paper. A new search algorithm, i.e., Lattice-T algorithm is also introduced whichi is based on the Lattice structure. The detection of keyword needs to store with the Trie tree, we can search the node of lattice structure from the Trie tree . It just needs traverse the Lattice once to get the result. Compared to the traditional method based on the N-Best structure, the Lattice-T method increased 12.22% detection rate, the time efficiency is improved by11.58%. The new method achieves good results.2. Compared to the method of Arithmetic average,confidence measure obtained by support vector mechine (SVM) can improve overall system performance. The relevant knowledge about the SVM is introduced in this paper, It also introduces the new method to get the confidence measure (CM) which uses SVM. Some issues need to be addressed: SVM probabilistic output problem, CM combination problems.Compare with the method of arithmetic average,confidence measure obtained by SVM can get better result. In the same Rejection rate, the recognition rate can get 7% improvement,identify the overall performance enhancements.This paper build a keyword spotting system using the 114 center of China Telecom telephone speech. The system includes keyword search module, the search results view module, presented results module, the parallel retrieval module, report generation module ect. Through the introduction and implementation of these modules , we can be more intuitive to learn the practical application of keyword spotting technology.