Design And Implementation Of Keyword Spotting System

Keyword spotting which contains signal processing,pattern recognition and several other technologies is an important branch of speech recognition and has broad application space.In this article,the development status and the main technologies of keyword spotting are researched,and two practical system are well designed and implemented based on two different technologies.And the two systems' performance are tested and analyzed respectively.An innovative improvement is introduced on the basis of the second system,improving the system's recognition efficiency.In this article,the research focuses on the design and implementation of keyword spotting system,the analysis and implementation of each module's main technologies,the algorithm's design and implementation of aiming at some specific goal,the analysis of the improved algorithm's effectiveness and availability,etc.The main works are as follows:1.The technologies of keyword spotting based on filler model and the implementation of practical systemIn this paper,we have researched the technologies of keyword spotting based on filler model and implemented a practical system.In the process of system implementation,a two-phase training method is used to train the acoustic models reaching the goal of reducing the manual annotation of the training speech effectively.And the LDA is well researched and exploited to improve the system's performance.2.The technologies of keyword spotting based on filler model and the implementation of practical systemConsidering that training the models requires a lot of training data,we research the method of template matching based on DTW which requires only several keyword utterance examples and implement it to construct another practical keyword spotting system,completing the extraction of Gaussian posteriorgram and template matching.3.A fast approach to keyword spotting based on prosodic dynamic features and the implementation of practical systemThe model-based method and the conditional DTW-based method both are based on frame by frame to complete recognition and template matching which leads to huge computing and somewhat limited speed.So in this article we propose an innovative approach by introducing the prosodic features to fast locate the hypothesis regions of the keywords based on segment owning to the prosodic features' characteristic that one prosodic feature vector can represent continuous several frames speech and then in the next stage we can only implement template matching within the hypothesis regions frame by frame,so as to achieve the purpose of reducing the amount of calculations and ensure the performance being not affected.In this part,we mainly focus on the research of the prosodic dynamic features' specific application in keyword spotting.And we complete the design of the algorithm,the analysis of the algorithm's effectiveness and availability and the compromise between system performance and efficiency.Finally,the experiments of system efficiency and performance and comparative analysis are carried out to prove the efficiency of our proposed approach which can improve the speed of the process of the whole keyword spotting system and achieve a similar recognition accuracy.