Isolated Word Recognition System Based On Neural Network Implemented On Android

With the development of Machine Learning and Artificial Intelligence, Speech Recognition begins to play a role in daily life. Especially, the isolated word recognition is adopted in many fields because it is easy to detect its endpoint and the system is quite simple. In addition, the smart cellular phone has been blooming recently, which offers a new platform for the isolated word recognition to be used in much more circumstances. For example, when human beings are riding or taking exercises and so forth, they cannot handle the cellular phone by hands directly. So it can improve people’s life quality and make the cellular phone smarter and more friendly by developing a system that can recognize the isolated word on the cellular phone.The paper consists of four parts. The first part introduces the history of speech recognition, presents its basic model and recognition process, discusses about several algorithms in detail and make a comparison between them. Then it picks up the BP neural network as the recognition algorithm. The second part presents the basic theory of neural network, introduces the artificial neural network model and its learning methods and so on.Then it figures out how the BP network works and its running process. At last it make an analysis about BP network’s drawback and presents some improved BP networks. The third pare presents Android’s structure and feature, shows how to install the Android SDK,and introduces the components and tools involved in the paper, such as JNI and NDK. The fourth part shows the implementation of the isolated word recognition system on the cellular phone and its result. Then it discusses about its drawbacks. The main work of the paper as follow:First, build the sample library of isolated words and get the characteristic parameters from the library. The library is divided into two parts, one part is used for specific recognition, the other is used for Non specific recognition. The paper got the characteristic parameters by MATLAB through the process of pre-treatment and feature extraction and restored them into files.Second, make a BP network model by Visual Studio, put the characteristic parameters gotten into the model. Change the model’s parameters, compare the results,get the best network, and restore the network model.Thirdly, develop the speech program on the Android. The abundant hardware resources on Android make it easy to develop the speech’s collection and encapsulation program. In order to make the speech recognition in time real time and the program strong, the paper develops the program by C++to finish the process of pre-treatment and feature extraction. And it works.Fourthly, implant the neural network on the Android, read the files of network, reconstruct the network, input the characteristic parameters into the network, and get the recognition result at last. This part is also programmed by C++.Fifthly, compile and package the C++Program by NDK, make that a shared library. Add the library to the project in the Eclipse. Then develop different applications according to the results. That make the cellular phone respond to the voice and the application becomes more useful. At last, package all files into.apk,and download it on the Android. Then run it and make a test.In the end, the paper summarizes the work and achievement, make an analysis about the system’s performance and present some method to improve the system. The performance shows that the system can be useful and can satisfy the more requirements with some extension.