A Research On Uyghur Continuous Speech Recognition Based On Julius

In recent years, test system performance of Uyghur speech recognition has dropped dramatically due to the lack of data of the minority and the inconsistency between of environment of data collecting and training and practical environment. How to overcome these difficulties, seek to the balance between training environment and practical environment, improve the performance of speech recognition systems, which has become a hot research filed of computer human-machine interface and communication technology.This paper takes Uyghur as research object, using spoken data under different backgrounds to research process of Uyghur speech recognition and the mathematical model, create recognition under reading speech and research the problem about how to improve the robustness of it under the environment of telephone voice.This paper's main studies as follows:1. Analyzed the characteristics of acoustic and linguistic about Uyghur, and then discussed Uyghur acoustic model by HTK tool under Linux system.2. Analyzed the Uyghur text corpus in the different areas by using Palmkit tool (CMU-Cambridge) and generated N-gram model, and then evaluated the perplexity of language model to discuss the optimal model.3. Researched the transmission characteristics of telephone voice signal and discussed the VAD method based pitch fo to reduce the influence of noise and long silent interval among voice interval.4. Designed and created a Uyghur continuous speech recognition system based Julius to improve recognition accurate and speed of system.5. The whole system is carried out under embedded Linux by writing Perl script and shell script, combined with HTK and Julius tools.