The Research And Application For Smart Home Of Speech KWS Based On HMM

There are characteristics of convenience and facilitation by using the speech recognition control technology in smart home.It can better get rid of the bad experience of the manual control and meet the much more intelligent needs of the public.In practice,smart home system requires higher recognition rate and recognition efficiency for speech recognition algorithms.Neural network algorithm relies on Internet resources and advanced server platform so it is still unable to satisfy practical requirements for lower-configured hardware and lightweight arithmetics.At present the main speech recognition system includes the isolated word recognition system based on DTW(Dynamic Time Warping),the isolated word recognition system based on HMM(Hidden Markov Model)and the keyword recognition system base on the Filler and HMM.These algorithms have the problems of low recognition rate or large computational complexity or low recognition efficiency or unsuitable for lower-configured hardware or poor experience of isolated word recognition technology.Therefore,improving algorithm for promising both high speech recognition rate and recognition efficiency,in order to satisfy the users' smart home control experience,has important significance to the further development of smart home.Firstly,the speech recognition system is analyzed and the basic principles of speech recognition technology are introduced and analyzed in detail.It includes preprocessing,feature extraction,model building and pattern matching.Secondly,it introduced the commonly used lightweight speech recognition algorithms in smart home,and analyzed the advantages and disadvantages of the algorithms.The two algorithms corresponding to the isolated word recognition system and keyword recognition system are introduced and analyzed in detail.And then an improved algorithm for keyword recognition is given.Thirdly,the improved algorithm is simulated by MATLAB,and the feasibility of the algorithm,the recognition rate and efficiency are tested.Finally,designed a speech recognition system based on web platform using apache server.The main task of the program is in the preprocessing stage.It improved the traditional low-pass denoise,using wavelet denoise for speech enhancement in suppressing ambient noise.Endpoint detection uses syllable segmentation instead of traditional method.Slide matching all syllables in recognition stage and extracted the segment that score the highest.If it passes the score threshold,the segment is confirmed as the keywords.The simulation results showed a good effect on the improvement of keyword recognition ratio and recognition efficiency.To some extent,it improved experience of speech control.For the existing problems,the future direction can be improved is offered.