Design Of Complete Speech Control Mobile Phone For The Blind Based On Speaker-independent Speech Recognition

The development of speech recognition technology in recent years has provided and promoted a new human-machine interaction method, i.e., the approach of natural voice communication between user and devices. Such a kind of way has been found in some special circumstances and devices, and is now being welcomed for its natural and convenient feature. For the blind, Defect of vision always keeps them from benefiting from the most advanced technologies. Now this can be compensated by the human-machine speech interaction method, which is deeply studied in the thesis.So far the research on speech recognition has yielded many valuable accomplishments after several decades’promotion of researchers in this field. Many successful theories and principles have promoted the research greatly. However, most of the studies and researches have been carried out on the large computers or the personal computers, and ideas seem to concentrate on new recognition theories or methods for better recognition accuracies. For the speech recognition on embedded systems, which is one of the step for the research result to become widely used technology, has not yet gained enough support from both research and real application. Now most of the speech recognition methods implemented on embedded systems adopt the speaker dependent isolated word recognition with simple algorithms. Hence it is really a challengeable job to make the complicated algorithms implemented on the embedded systems to meet the goal of high-accuracy embedded speech recognition.Research in this paper has concentrated on the continuous hidden Markov model and its application in speaker independent isolated word recognition systems, beginning with study on the theories and principles. Specifically, problems of initial parameters, training methods and data underflow are addressed under the multi sample series. Further a complete speech control mobile phone system is designed based on these researches.The system adopts the Baum-Welch algorithm to train the models of 25 speech command and 10 numbers (0~9). These models are later transmitted to the embedded system for further recognition. 25 speech commands refer to different operations. When wireless communication function is needed, a composition of AT commands as well as data package is then transmitted to the GPRS module. Referring to the attribute of mobile phone menu item system, a method of graded command recognition is proposed to achieve better recognition accuracies. To enhance the convenience for the blind to use this system, functions of voice remind and message speak out are also designed through storing the 3817 Chinese characters and numbers’voice data into one mass-storage NandFlash. And an OOV (out of vocabulary) word detection method is used to meet the real application need. Experiments are also carried out in different environments and improving methods are also discussed.