Research On Yangzhou Dialect Speech Recognition Based On Isolated Words

In recent years, along with the rapid development of speech recognition technology, a variety of developed speech recognition products have been continuously invented. To a large extent, this has extremely changed the inconvenient interface between person and computer, made it easier to operate those complex electronic devices in the past, and greatly promoted the development of electronic information, computer and other related fields. However, in the practical applications, due to the factors and constraints of the complicated external environment and users themselves, such as the background noise and the dialect accent, the accuracy of the speech recognition and the robustness of the system are also faced with severe challenges. It still remains an unresolved problem all across the world that how to improve the adaptability of the speech recognition system on the dialect accent in a variety of complicated environments. The development of the speech recognition technology in China starts late, and the foundation is relatively weak, but it has developed very quickly. With the support of the "863" national program, there appears many kinds of scientific research achievements. However, there are many different types of dialects in Chinese, which contains eight major dialect languages, including a total of more than80kinds of regional dialects that they can not communicate with each other, thus increasing the difficulty of the speech recognition of Chinese from the objective perspective. Yangzhou dialect belongs to the official language family of Jianghuai areas, and has extensive mass basis in the middle part of Jiangsu. It has certain representativeness, and can be used as an example of a higher value to solve the problem of dialect speech recognition.This paper was supported by National Language Application "11th Five-Year Plan" Research Project-Jiangsu Province Language Resource Audio Database Construction Subproject. It focuses on the research of the speech signal pretreatment, feature parameters extraction and the recognition algorithm. Combined with audio database dialect in Yangzhou, we have constructed the speech recognition system of Yangzhou dialect on the basis of the isolated words, and the main contents include: (1) Described the background and significance of this research, the history and current situation of the development of speech recognition technology, and the composition, classification and evaluation criteria of the speech recognition system.(2) Described the features and generation principle of the speech signal, the pretreatment of the speech signal. This paper pays great attention to analyze the steps and methods of endpoint detection. Introduced the calculation and extraction method of the feather parameters of LPCC and MFCC. On this basis, we have further introduced the differential mixing coefficient of the MFCC, and make the experiential simulation.(3) Described an overview of the development of the worldwide speech corpora database, outlined the general steps to construct a speech corpus database, and further introduced the construction process of the audio database of Yangzhou dialect.(4) Introduced the two classic speech recognition algorithm of DTW and HMM. Based on the description of the basic principles, the algorithm realization, and the parameter estimation, we have made an analysis of the advantages and disadvantages of the DTW algorithm, and put forward the improvement of the HMM algorithm to solve the specific problems such as the data transfer bandwidth in the calculation process. On the period of experimental simulation, we have taken the isolated words"1-10" in Yangzhou dialect as an example. At first, we made a speaker-dependent dialect speech recognition experiment using the DTW algorithm, then we made a speak-independent recognition experiment through the HMM algorithm, and selected the differential mixing coefficient of MFCC as the feather parameters to obtain a higher recognition rate.(5) Introduced the Yangzhou dialect speech recognition system on the basis of HMM algorithm, the fractional lower order statistics and the related theories of the non-Gaussian signal processing. Selected the LMP algorithm to optimize the system in the non-Gaussian noise environment. On the period of experimental simulation, the number "1-10" in Yangzhou dialect and50typical dialect words were tested to prove that the system has a higher recognition rate and a certain degree of adaptability under the non-Gaussian noise environment.