Research On Wuhan Dialect Speech Recognition

Since the 1950 s, the speech recognition has been one of the hot. Compared with other countries, China began to study speech recognition technology later. Since the 1980 s, setting up the special agency for the speech recognition technology, our country has made a lot of breakthroughs. However, there are still many technical problems in speech recognition has not been resolved. The speech recognition is an extremely complex task. Besides, the complex of Chinese dialect also has increased the difficulty of the speech recognition research. Nowadays, some researchers have begun to study speech recognition of local dialects. Due to immature technology and developing late for the speech recognition in our country, many key issues need to be studied.The completely speech recognition system is mainly consisted by four parts: preprocessing, feature extraction, acoustic model and language model. The pre-processing module is used to process the input speech signal, including noise filtering and enhanced voice. The feature extraction module is used to extract the speech signal characteristic. The parameter model is constructed by training speech database in the acoustic model, used to get parameters stream in the recognition phase. The speech model is used to predict the probability of a sentence. And then, the paper mainly studies three parts, complete the following main tasks.(1) The small group of language depot which is used for Wuhan dialect recognition iscreated by analyzing the characteristics of Wuhan dialect pronunciation.(2) Dealing with background noise smoothly and reconstructing speech signal bywavelet, which can improve the accuracy in the condition of low SNR.(3) Voice activity detection is proposed based on multifractal, which is firstly using thefractal characteristics of the speech signal to calculate Fractal dimension of the frame,and then determining the endpoint of voice signal by its relevance.(4) Building Wuhan dialect speech recognition system based on the HTK, we doexperiments by the different characteristic parameters.