Study Of Speech Recognition In Intelligent Assistive Bath Appliances

With the development of society, people’s living standards are getting higher andhigher,in the meanwhile, the problem of population aging is becoming more and moreserious, which makes assistive appliances for disabled people in more demand. With thisbackground, we have developed a speech control system for the Auxiliary bathingapparatus for disabled people.The speech recognition system is the important part of speech control system. Theprocess of speech recognition always includes the feature extraction from the voice signals,the building of template training, and the matching process of the speech character. Inorder to achieve better recognition accuracies, the exploration began with the building oftemplate training and the matching process of the speech character. According to our study,a speech recognition system of isolated word in small vocabulary is concerned, so theDynamic Time Warping (DTW) based methods are chosen for the voice recognition in thispaper. From the aspect of the building of template training, we compared three kinds ofthe template training methods. The results of the experiments of template training methodswere also compared. And the k-means clustering method was found best. In addition, weproposed two kinds of incremental feedback learning algorithm to train referencetemplates and improve speech recognizing rate for DTW-based voice recognition, whichhas better acclimatization by updating the reference templates with the featurevectors of the testing speech, including the variable gain coefficient (VGC) based methodand the time warping average (TWA) based method. And A series of experiments weredone to analyse the performance of the methods. From the aspect of the matchingprocess of the speech character, we proposed the time-frequency mixed recognitionmodel and the single-double word mixed recognition model,which were proved to beeffective by experiments.With respect to the specific noise in the bathroom, we introduce a modified spectralsubtraction algorithm which can eliminate the musical noise.It was programmed and usedin the experiments.The results of the experiments were also compared and analyzed. Then, we developed a speech recognition software by Microsoft Visual C++6.0integrated development environment under windows operating system. The speechrecognition system applied the k-means clustering method as the method of templatetraining, the DTW-based method as the method of speech recognition, and MFCC as thefeature extraction, including incremental feedback learning methods. At last, weintroduced the hardware and software integration environment and the method of thesystem application. The experiments were done to test the software.