Speech Recognition For Finite Passwords In Smart Home

With the improvement of living standards and the development of speech technology,the use of speech control home appliances has become one of the direction of the development of smart home.In recent years,famous companies such as Apple,Google,i FLYTEK,has launched their own voice smart home products respectively and these products bring a new experience to customers.Speech recognition mainly includes the process of voice activity detection,feature extraction,template training and matching.The basic of speech recognition for smart home is an isolated word,small vocabulary,speaker-independent speech recognition.And it must solve the following problems.Firstly,the voice activity detection algorithm must be robust to adapt to the family environment.Secondly,the recognition algorithm must has relatively high recognition rate.Thirdly,on the basis of high recognition rate,the recognition speed should be high and the real-time.Aiming at the problems existing in the voice activity detection,this paper proposes a two-threshold voice activity detection algorithm based on a priori signal-to-noise ratio.The algorithm has quickly detection speed and the detection threshold is automatically adjusted with the priori signal-to-noise ratio so the detection rate is highly increased.In order to improve the robust and recognition rate,the multi-band spectral subtraction is used to do digital filter.The 12 dimension MFCC and its first order difference total 24 dimensional features are selected as recognition features.For the training and matching part of speech model,this paper makes a comparative study of DTW algorithm,GMM algorithm and HMM algorithm.,And a lot of research work has done especially for DTW algorithm.This article designed 14 passwords for home appliances which are commonly used in the smart home.There are a total of 6272 samples of data and each password has 448 samples.All speech samples are collected by 112 people,including 72 boys and 40 girls.Three recognition algorithms is experimented with these data.In this paper,an improved DTW algorithm is proposed to limit the search path to two parallel lines which are parallel to the main diagonal because the DTW algorithm has a large search range.In order to improve the recognition rate of the system,multiple templates for each password is used in the recognition system.As for password recognition based on the DTW algorithm,five kinds of template libraries are established.In the template library 3 to 5,the pitch of each training date will be extracted as the feature.K-means clustering and DTW algorithm will be done to use these pitch feature to cluster the data into the defined parts.And each parts will be established into a template by DTW regularization.In the password recognition experiment based on GMM and HMM algorithm,two template libraries are established in this paper.Each password in template library 1 has only one template and each password in template library 2 has two template which are established by female or male training data respectively.When using multi-template recognition,the recognition rate is improved while the recognition time is also increased.To solve this problem,this paper uses a multi-CPU based on parallel computing to accelerate,without reducing the recognition rate,the recognition speed is greatly increased.In this paper,the following conclusions are obtained: The average recognition time of a single password can be reached 59 ms when the GMM algorithm and parallel computing are adopted.The recognition speed of the algorithm is the fastest,but the average recognition rate of all the passwords is 92.14%,less than the recognition rate of the continuous HMM algorithm.The average recognition time of DTW algorithm is 963 ms when using parallel computing technology,and it is too long when compared with GMM or HMM algorithm.The recognition rate is 92.05%,which is lower than GMM algorithm.Therefore,we can see that continuous HMM algorithm is relatively best algorithm for limited passwords recognition when combining with multi-template and parallel computing method.It can better meet the high recognition rate at the same time with a faster recognition speed.The work done in this paper has certain practical value for the realization of speech recognition system for smart home,but there are still many problems that need to be solved due to the restrict of time and experiment conditions,such as small sample library,single template training method and so on.Though the subject still has a very broad application prospects and research value and it is worth further study.