Research On Uighur Connected Digit Speech Recognition System Based On HTK

Presently, speech recognition goes into the large vocabulary, continuous speech recognition stage. But there still is system's accuracy low. A high rate of speech recognition is used statistical model technology. This paper designed speech recognition acoustic model of the statistical model. Uighur connected digit speech recognition established on HMM technology, including speech corpus, acoustic model and language model. Using HTK (HMM toolkit) tool, Uighnr continuous speech recognition system realized, in this paper, work as follows:(1) A suitable small corpus is established of Uighur connected digit speech recognition, Thus we established the speech and text database file according with the standard of corpus.(2) Uighur connected digit speech recognition's acoustic model established HTK. This model was improved and optimization. To solve the Uighur coarticulative problems the context of triphones model is introduced. By using the middle tree, establishing audible dictionaries and triphones-binding, mending mute, adding Gaussian mixture components of algorithm and adjusting system template parameters, the HMM model parameters could be quite well optimized.(3) Language model is established on rule. Syntax, semantic knowledge and voice recognition effected integration in this system. It used hreadth-first algorithm search words node network. If recognition results is similar, language model may exclude illegal sentence. It improved recognition rates and reduced the search scope. This is a major innovation in this paper.(4) Used Visual C++ for the second development, by adding thread multithreading, increasing the memory management, the procedure can run beyond CMD.(5) Finally, after procedure has been compiled, we did the three experiments: used different identification element, increased the number of Gaussian mixture components and different language model to compare the rate of system. Through the analysis of experiment results the system framework is certificated. The phrase's recognition rate of system can approach 80.00％, and that of the word can reach 91.19％.